Early Neurodevelopmental Disturbances During Sensitive Periods of Stem Cell Differentiation

Early neurodevelopmental disturbances during sensitive periods of stem cell differentiation

Dissertation zur Erlangung des akademischen Grades eines Doktors der Naturwissenssenschaften (Dr.rer.nat.)

vorgelegt von

Dreser, Nadine

an der

Mathematisch-Naturwissenschaftliche Sektion

Fachbereich Biologie

Tag der mündlichen Prüfung: 17.12.2020

Referent: Prof. Marcel Leist Referent: Prof. Alexander Bürkle Referent: Prof. Daniel Dietrich

Konstanzer Online-Publikations-System (KOPS) URL: http://nbn-resolving.de/urn:nbn:de:bsz:352-2-1vdk3v1zzwe7j7

List of publications

Publication integrated in this thesis

Results chapter 1: Waldmann, T., Grinberg, M., König, A., Rempel, E., Schildknecht, S., Henry, M., Holzer, A.-K., Dreser, N., Shinde, V., Sachinidis, A., et al. (2017). Stem Cell Transcriptome Responses and Corresponding Biomarkers That Indicate the Transition from Adaptive Responses to Cytotox- icity. Chemical Research in Toxicology 30, 905-922.

Results chapter 2: Dreser N., Madjar K., Holzer A.-K., Kapitza M., Scholz C., Kranaster P., Gutbier S., Klima S., Kolb D., Dietz C., Trefzer T., Meisig J., van Thriel C., Henry M., Berthold M.R., Blüthgen N., Sachinidis A., Rahnenführer J., Hengstler J.G., Waldmann T. and Leist M.(2019) Impaired formation of neural rosettes from stem cells as teratological correlate of toxicant-induced transcriptome disturbances. Arch Toxicol 94(1):151-171

Results chapter 3: Dreser N.*, Meisig J.*, Waldmann T., Hengstler J., Sachinidis A., Rahnenführer J., Blüthgen N. and Leist M. Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed neuroectodermal differentiation. Submitted for publication to Nucleic Acid Research *shared first authorship

Publications not integrated in this thesis

Zimmer, B., Pallocca, G., Dreser, N., Foerster, S., Waldmann, T., Westerhout, J., Julien, S., Krause, K.H., van Thriel, C., Hengstler, J.G., et al. (2014). Profiling of drugs and environmental chemicals for functional impairment of neural crest migration in a novel stem cell-based test battery. Arch Toxicol 88, 1109-1126.

Dreser, N., Zimmer, B., Dietz, C., Sugis, E., Pallocca, G., Nyffeler, J., Meisig, J., Bluthgen, N., Berthold, M.R., Waldmann, T., et al. (2015). Grouping of histone deacetylase inhibitors and other toxicants disturbing neural crest migration by transcriptional profiling. Neurotoxicology 50, 56-70.

Table of Contents

1. Zusammenfassung ...... 9

2. Abstract ...... 11

3. General Introduction ...... 13 3.1. Embryonic development in vivo ...... 13 3.1.1. From fertilization to germ layer formation ...... 13 3.1.2. Neurulation in brief ...... 14 3.1.3. Mechanisms and signaling of neural tube folding ...... 15 3.1.3 Neural border formation and delamination of the Neural crest ...... 18 3.2. Embryonic development modeled in vitro ...... 19 3.2.1. Pluripotent stem cells ...... 19 3.2.2. Rosettes - the in vitro counterpart of neural tube formation ...... 20 3.2.3. Tissue specific markers to determine regional specification ...... 22 3.3. Neural tube defects: ...... 23 3.3.1. NTD induced by genetic defects ...... 25 3.3.2. NTD induced by drugs (HDACi) ...... 25 3.4. Developmental toxicity and developmental neurotoxicty (DNT) ...... 27 3.4.1. From Animal models to in vitro models- towards more predictive DNT testing ... 28 3.4.2. Concepts of a new DNT testing strategy: AOPs, TEPs, PoTs ...... 29 3.5. DNT Test assays ...... 32 3.5.1. Transcritomics as endpoint ...... 33

3.5.2. The UKN1/ STOPtoxUKN assay ...... 35

4. Aims ...... 37

5. Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity ...... 38 5.1. Abstract ...... 39 5.2. Introduction ...... 40 5.3. Materials and Methods ...... 44 5.4. Results ...... 49 5.5. Discussion ...... 67 5.6. Funding ...... 70 5.7. Acknowledgments ...... 70 Abstract

Supporting information available ...... 71 5.8. Supplementary information ...... 72

6. Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances .... 77 6.1. Summary ...... 78 6.2. Introduction ...... 79 6.3. Materials and Methods ...... 82 6.4. Results ...... 88 6.5. Conclusions and overall outlook ...... 105 6.6. Acknowledgements ...... 106 6.7. Conflict of interest ...... 106 6.8. Supplementary information ...... 107

7. Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed neuroectodermal differentiation ...... 119 7.1. Abstract ...... 120 7.3. Introduction ...... 121 7.4. Material and methods ...... 124 7.5. Results ...... 130 7.6. Discussion ...... 142 7.7. Availability ...... 144 7.8. Accession numbers ...... 144 7.9. Acknowledgement ...... 144 7.10. Supplementary information ...... 145

8. Concluding discussion and perspectives ...... 147 8.1. History of the UKN1 test method ...... 148 8.2. Relevance and uniqueness of the test system ...... 152 8.3. Readiness for validation ...... 154 8.3.1. Readiness criteria: How far did we get? ...... 155 8.4. Biological mode of action of HDAC inhibitors ...... 157 8.4.1. NEPs differentiate towards neural crest cells ...... 157 8.4.2. Development of an AOP ...... 159 8.5. Limitations and outlook ...... 161

9. Abbreviations ...... 164

Abstract

10. Record of contribution ...... 166

11. Danksagung ...... 168

12. References ...... 0

Zusammenfassung

1. Zusammenfassung

In den letzten Jahren hat sich das Interesse an neuen Alternativ-Methoden für die Neuro- Entwicklungstoxizitäts(NET)-Testung stark gesteigert und immer mehr Tests wurden entwickelt. Zudem wurde dieser Bereich auch vermehrt durch die Regierung bzw. durch europäische Mittel gefördert. Die UKN1 Test-Methode wurde bereits vor 10 Jahren entwickelt und 2013 erstmals als solche beschrieben. Durch die Differenzierung von humanen embryonalen Stammzellen bzw. induzierten pluripotenten Stammzellen zu neuroektodermalen Vorläuferzellen (NEV) wird die frühe Neuroentwicklung in vitro nachgestellt. Der Entwicklungsstatus entspricht der Neuralplatte bzw dem frühen Neuralrohr in vivo. Als Endpunkt wird die Veränderung der Genexpression analysiert. Hierzu werden entweder Expressionsunterschiede im ganzen Transkriptom (basierend auf Mikroarrays) gemessen oder es werden nur wenige Markergene (PAX6, OTX2, NANOG, OCT4) durch eine qPCR Messung erfasst. Die UKN1 Testmethode diente auch dazu, Mechanismen der Entwicklung zu erforschen wie z.B. epigentische Prozesse. Außerdem trug sie dazu bei, Rahmenbedingungen für NET-Testungen zu untersuchen und zu definieren wie z.B. zur Konzentrationsfindung und um plausible Behandlungsmuster und Zeitfenster festzulegen. Zusätzliche wurde die Testmethode verwendet um Pilotstudien bezüglich der Verwendung von Transkriptomdaten als NEP-Messung durchzuführen. Diese Arbeit beschäftigt sich mit der Verbesserungen und Weiterentwicklung der UKN1 Test Methode selbst. Das erste Teilprojekt dieser Arbeit untersucht die Auswirkungen von akut toxischen Substanzkonzentrationen auf die Genexpression, um die Konzentrations- abhängigkeit von Transkriptomregulation zu identifizieren und um zwischen einem zytotoxischen Effekt und einem entwicklungstoxischen Effekt zu unterscheiden. So wurde eine wichtige Basis geschaffen, um NET-Substanzen eindeutig zu definieren. Durch die Erfassung von Genmarkern als Signatur für Zytotoxizität konnten Regeln definiert werden, um zu hohe Substanzkonzentrationen automatisch auszusortieren. Im zweiten Teil des Projektes wurden die NEV zu einem Stadium weiter differenziert, in dem die Zellen beginnen, sich zu neuralen Rosetten zu formieren, welche das in vitro Äquivalent zum Neuralrohr darstellen. Rosettenformation wurde als neuer Endpunkt etabliert, welcher nun ermöglicht die Genexpression an einen Endophänotyp zu koppeln und die Relevanz der Genveränderungen zu beurteilen. Darauf basierend wurde ein zuverlässiges Vorhersagemodell erstellt. Außerdem konnte so das genaue Zeitfenster bestimmt werden, in dem ein Histondeacetylase-Inhibitor (Trichostatin A) entwicklungstoxisch auf das System wirkt, während er zu anderen Zeitpunkten keinen Einfluss auf die differenzierenden Zellen ausübt. Im dritten Teilprojekt wurde ein Computermodell generiert, das die Auswirkung von

Zusammenfassung verschiedenen Valproinsäurekonzentrationen auf die Expression einzelner Gene zu verschiedenen Zeitpunkten der Differenzierung vorhersagt. Hierfür wurden Gen- expressiondaten einer ungestörten Differenzierung über die Zeit und Daten von Zellen, die mit verschiedenen Valproinsäure(VPA)-Konzentrationen behandelt wurden, verwendet. Das Computermodell lieferte Informationen über Signalwege, die durch VPA an nicht gemessenen, intermediären Zeitpunkten aktiviert wurden. In einem Interventionsexperiment wurde gezeigt, dass durch das Blockieren eines solchen Signalweges (WNT) die Rosettenbildung wieder stattfinden konnte, die zuvor von VPA gestört wurde. Die drei Teilprojekte dieser Arbeit bilden eine neue wissenschaftliche Basis für den UKN1 assay, die in Zukunft vorraussichtlich eine breitere Anwendung im industriellen wie auch im regulatorischen Bereich ermöglicht.

Abstract

2. Abstract

The high need for appropriate in vitro tests for regulatory and industrial purposes as well as the focused governmental and EU support leads to a constant increase in assays addressing developmental neurotoxicity (DNT). The UKN1 assay is such a test method that has been set up already 10 years ago. By using human embryonic stem cells/induced pluripotent stem cells differentiating towards neuroectodermal precursor cells (NEP), the early neurodevelopment is modeled in vitro. The NEP state corresponds to the neural plate or the early neural tube in vivo. The readout for disturbances in the differentiation is based on the measurement of gene expression of several marker genes (PAX6, OTX2, NANOG, OCT4), it has also been assessed by transcriptome analysis using oligonucleotide microarrays. The UKN1 test has served to study processes of neurodevelopment such as epigenetic mechanisms. Moreover, the test method was used to develop framework conditions for setting up DNT methods in general, for example, strategies have been developed for compound concentration range finding and for quantifying transcriptomics datasets as a DNT endpoint. Especially, for the setup of transcriptome conditions for DNT testing, the UKN1-based studies have pioneered the field. In this thesis project the UKN1 assay was further developed. Firstly, the effect of acutely toxic compound concentrations on the transcriptome have been investigated. The aim was to understand the mechanisms of transcriptomic regulation in a developing system in unperturbed and perturbed development. Additionally, it aimed to distinguish between an acute and a developmental effect. Since the test should only identify specific DNT compounds, this represents an important basis. By using genetic markers as a toxicological fingerprint for acute cytotoxicity, rules were established to exclude compounds tested in the cytotoxic range. In the second part of the project, the NEPs were further differentiated until the cells formed neural rosettes. These represent the in vitro equivalent of the in vivo neural tube. The rosette formation was established as a new endpoint for the UKN1 assay. The new phenotypic readout provides a robust tool to anchor transcriptomics changes to an endophenotype and to investigate the relevance of transcriptome changes at early neurodevelopmental stages. A reliable prediction model was setup based on the acquired information. Moreover, this approach served to define a narrow time window of toxicity of a HDAC inhibitor (trichostatin A). HDAC inhibitors are used as anticonvulsants and can induce congenital malformations such as spina bifida in the developing embryo. In the third part of the project, a computational model was developed to predict the influence of different valproic acid (VPA) concentrations on the expression of single genes at every time point during differentiation. The model was based on experimental datasets over time and on data from cells exposed to various VPA concentrations. The computational model delivered

Abstract information (predicted) about the pathways that were activated by VPA at intermediate (not measured) time points. The VPA-induced disturbance of rosette formation (endophenotype) was successfully rescued by blocking the wnt pathway. The three parts of this work have together provided a new scientific basis for the UKN1 assay and they are expected to lead to a broader application for regulatory and for industrial purposes.

General introduction

3. General Introduction

The embryonic development and especially neurodevelopment involves highly complex processes that are regulated time and region dependent. It is obvious that developmental processes are extremely vulnerable to any interfering factors from the environment. In animal models, it is possible, in principle, to study late and very pronounced effects of such interfering compounds. However, pathways that lead to an adverse outcome remain not fully characterized which makes it impossible to allow a proper risk calculation. This and many other argu- ments underline the high need for rigorously validated, reliable DNT in vitro tests. This thesis will deal with the promises but also the shortcomings of such a test.

3.1. Embryonic development in vivo

3.1.1. From fertilization to germ layer formation

Development (embryonic day 0) starts with the fertilization of the oocyte by the sperm. The fertilized oocyte develops into the zygote, which makes its way through the oviduct. During this time the first cell division (cleavage) takes place. The zygote divides symmetrically twice, until the four-cell stage, in which the cells are still totipotent and could give rise to an embryo indi- vidually (Gilbert et al. 2006; Wolpert 2007). Then the first asymmetrical cell cleavage takes place, which is the beginning of the specialization of cells. After the 4th cleavage, in 16 cell- stage, the morula starts to form, the cells become smaller and are tightly connected by gap junctions. The cells organize to a first tissue; a layer of external cells gain polarity and surround the inner apolar cells, now called the inner cell mass. The outer cell layer represents the throphectoderm, which will give rise to the embryonic part of the placenta, the chorion. The inner cell mass will give rise to the embryo and the yolk sac, allantois, and amnion. Beginning from the 64-cell stage, the outer trophoblast cells secrete a fluid that leads to the formation of a cavity and the re-organization of the inner cell mass. This results in the formation of the blastocyst at the end of the first embryonic week. At this stage, the embryo reaches the uterus and begins to adhere to the uterus wall. During this time the inner cell mass begins to further differentiate into hypoblast and epiblast (Gilbert et al. 2006; Wolpert 2007). The gastrulation- the folding and invagination of the blastula- starts in the third week after fertilization (embryonic day 12-14). This process results in the formation of endoderm, mesoderm, and ectoderm, which are the three germ layers of the embryo. These three tissues will give rise to all the cells of the body. From the mesoderm mainly muscles, bones, and blood cells will originate. The endoderm will give rise to many organs such as the respiratory tract,

General introduction digestive tract, glands, but also the notochord. The ectoderm will develop into the surface ectoderm (epidermis), the nervous system, and the neural crest (Gilbert et al. 2006; Wolpert 2007). At the end of embryonic development, one fertilized oocyte will have given rise to more than 200 different cell types organized in many tissues and organs. To orchestrate this highly complex and fast diversification, the development of each cell has to be tightly controlled and guided specifically. Therefore, a highly complex spatio-temporal regulation and organization is required (Groves and LaBonne 2014).

3.1.2. Neurulation in brief

Neurodevelopment describes a highly complex process, which summarizes all critical events that have to take place to result in a fully developed central nervous system (brain and neural tube), and peripheral nervous system. This includes proliferation, differentiation, migration, pathfinding, patterning, apoptosis, synaptogenesis, neurite outgrowth, and myelination. Immediately after generation of the three germ layers, the first step in neurodevelopment- the neurulation- follows in the third week after gestation (Groves and LaBonne 2014). The neural tube formation looks rather simple and straight forward from the outside. However, numerous pathways and processes guide the single steps. The ectoderm germ layer is organized as an epithelium. It will give rise to the surface ectoderm, the neural crest that differentiates to numerous different cell types and the whole central nervous system. As a first step, the early neuroectoderm (neural plate), an area in the ectoderm, begins to thicken. The thickening is induced by factors secreted by the notochord, which lays ventrally to the neuroectoderm. The neural plate bends into the neural fold until the dorsolateral hinge points (DLHP) meet and fuse to finally close the neural tube (Figure 1) (Eom et al. 2012). The closure is discontinuous and the initiation occurs at two regions in the developing human embryo: the hindbrain/cervical boundary and the rostral end of the forebrain (O'Rahilly and Muller 2002). From these regions, the tube closes zipper-like along the spine in both directions rostral and caudal (Greene and Copp 2009). Initiation starts on day 18 after fertilization and the spinal cord closure is com- pleted at around day 35 postfertilization. Neurulation strongly varies among species. For example in rodents, the neurulation is initiated at three different regions, in Xenopus closure happens almost simultaneously along the body axis. Moreover, the key processes, important for neural tube formation exhibit clear inter-species differences (Copp and Greene 2010).

General introduction

Figure 1 Folding of the neural tube (Adapted from (Gammill and Bronner-Fraser 2003)). (A) Neural plate begins to thicken, before convergent extension is induced. (B) The neural groove forms by bending at the median hinge points (the middle area of the neural plate) until the (C) dorsolateral hinge points at the neural plate border region meet and begin to fuse. (D) Neural crest cell delamination starts already during fusion but also at later time points.

3.1.3. Mechanisms and signaling of neural tube folding

Signaling: Induction of all key signaling cascades relies on external morphogen gradients, which are provided by the surrounding tissues (Groves and LaBonne 2014). The key signaling molecules involved in organization of neurulation are Sonic hedgehog (Shh), Fibroblast growth factors (FGF), Wnts (wingless/Int1), bone morphogenetic proteins (BMPs) and BMP inhibitors (eg noggin), released by the organizer regions. At the neural plate stage, a wnt gradient (an- tero-porsterior) and BMP gradient (mediolateral) exist already. Dorsal-ventral patterning is achieved on one side by BMP gradient, which is secreted by the non-neural ectoderm. On the other side, the Shh gradient is present that is released by the notochord ventrally of the neural plate. Additionally, the notochord secretes BMP antagonists (noggin, follistatin, chordin). At the boundary between mid and hindbrain FGF8 is expressed. The signaling molecules control the expression of all the downstream transcription factors and thereby drive the cell differentiation, regulate the key processes, all the mechanisms required for neural tube formation, and determine the cell fate. The differentiation and diversification of cells are determined by the compo- sitions and the concentration gradients of the various morphogens (Groves and LaBonne 2014; Nikolopoulou et al. 2017). 15

General introduction

Convergent extension: Before the neuroectoderm develops into the neural tube and the hinge points are enabled to meet and fuse, several tightly regulated processes have to take place to enable proper neural tube formation. The first key step is the convergent extension: the tissue narrows along its mediolateral axis and elongates along its anteroposterior/ros- trocaudal axis. The convergent extension takes place in the neuroectoderm as well as in the underlying mesoderm tissue. The whole process is driven by the planar cell polarity signaling (PCP), which is dependent on the non-canonical wnt signaling. This pathway shows the highest correlations in terms of neural tube defects since many different genes (or rather gene alterations, e.g. in the PCP complex) have been proven to be associated with failure in neural tube closure. For details see chapter 3.2.1. (Ciruna et al. 2006; Nikolopoulou et al. 2017; Tawk et al. 2007).

Bending: Sonic hedgehog that is secreted by the notochord, which lays on the ventral side of the neuroectoderm, induces the formation of the median hinge point (MHP) on the neuroepithelial (Smith and Schoenwolf 1989). On the other hand it inhibits the DLHP (Figure 2. 5). TGFβ/BMP signaling is repressed by noggin at the MHP, which leads to the stabilization of apical junctions (Amarnath and Agarwala 2017; Eom et al. 2012).The MHP defines the neural plate center and, in this area, the tissue begins to bend to bring closer the future DLHP. Among other mechanisms, PCP signaling plays still a major role during this process and links the convergent extension and the bending. A PCP-regulating cadherin (Celsr1), is concentrated in adherents junctions located on the apical part of the cells. Celsr1 cooperates with a protein complex and thereby upregulates Rho kinase. This causes an actomyosin-dependent contraction in a planar-polarized manner (Figure 2. 7) (Escuin et al. 2015). The resulting apical con- striction leads to the bending of the neural plate and formation of the neural groove (Nishimura et al. 2012). Another important mechanism of convergence is the interkinetic nuclear migration leading to a wedge-shaping of cells at the MHP. The cells of this region are mainly in S-phase and that determines the nuclei position to the basal part of the cell (Figure 2. 4) (McShane et al. 2015). During the process of bending, already the first wave of neural crest cells, the sacral neural crest cells delaminate from the neuroectoderm. This process is also thought to have an influence on the bending of the epithelium (Copp 2005). Besides the aforementioned morphogen gradients, many factors of the extracellular matrix such as proteoglycans, laminin, fibronectin as well as proteases have been shown to contribute to neural tube closure (Figure 2. 8) (Nikolopoulou et al. 2017).

Fusion and zipping: After the neural plate invagination, the dorsolateral hinge points (DLHP) need to meet and to fuse and the non-neuronal ectoderm (NNE) tissue needs to separate 16

General introduction

(Copp et al. 2003). The DLHP formation is regulated by antagonistic interactions between BMP, Shh, and noggin signaling (Figure 2. 4). Since the cells of the DLHPs are already local- ized next to each other, caused by the processes of bending, the cells have to recognize each other and to make cell-cell contacts. This process is facilitated by cellular protrusions, which are dynamic actin-rich filopodia and lamellipodia (Figure 2. 2). The protrusions are formed by the cells of the NNE as well as by the NE. Which cells make the first contacts differs among species and is even dependent on the axial level (Geelen and Langman 1977; Geelen and Langman 1979). Finally, at the boundary apoptosis occurs to ensure the separation of the two newly formed tissues (Yamaguchi et al. 2011).

Secondary neurulation: The process of fusion and zipping describes the first neurulation. However, another process- the secondary neurulation- results in neural tube formation in the lower sacral region. Unlike the first neurulation, the second neurulation leads to epithelializa- tion and neural tube formation without the formation of the neural fold. Tail-bud cells start to condense into a rosette-like appearing structure, a primitive neural tube that strongly reminds of in vitro neural rosettes. This structure represents the medullary cord, which further develops into an epithelium surrounding a lumen, the secondary neural tube (Copp et al. 2015).

General introduction

Figure 2: Key processes guiding neural tube formation (adapted from (Nikolopoulou et al. 2017)). (1) Key transcription factors regulate cell-cell contacts. (2) Protrusion make the first cell-cell contacts during fusion mediated by the non-neuronal ectoderm as well as the neural plate border. (3) membrane-bound proteases in the non-neuronal ectoderm. (4) Interkinetic nuclear migration (IKNM) lead to wedge shaped cells actively contributing to the bending process. (5) Dorsolateral hinge point (DLHP) formation is regulated by BMP, noggin, Shh signaling. (6) Noggin inhibits BMP signaling (smad inhibition). (7) Myosin contraction leads to bending at the mediolateral hinge point (MLHP). (8) ROCK dependent actomyosin turnover is required to keep dynamics.

3.1.3 Neural border formation and delamination of the Neural crest

During neurulation and later in development neural crest cells (NCC) develop from the neural plate border, the region where the tube fuses (see Figure 1, D). NCC delaminate in distinct waves during neural tube formation as well as at later time points. The delamination underlies tightly regulated signaling pathways to ensure cell survival after losing cell-cell contacts as well 18

General introduction as correct pathfinding. Neural crest cells can differentiate into most subtypes of peripheral neurons and all types of glial cells of the peripheral nervous system. Furthermore, they can give rise to various cell types: pigment cells, major parts of the craniofacial cartilage and bone, endocrine cells, cardiac tissues, smooth muscle cells and tendon (Dupin et al. 2006; Grenier et al. 2009; Hall 2008; Kirby and Hutson 2010; Le Douarin and Kalcheim 1999; Minoux and Rijli 2010). To delaminate from the neuro-epithelial cells have to undergo a partial or complete epithelium-to-mesenchymal transition (EMT) (Ahlstrom and Erickson 2009). This process is triggered by the surrounding tissue secreting factors such as bone morphogenic protein (BMPs), fibroblast growth factor (FGFs), and Wnts (Sauka-Spengler and Bronner-Fraser 2008). Induction of the neural border transcription factors leads to the further expression of neural crest transcription factors like SNAIL/SLUG, TFAP2, FOXD3, TWIST, CMYC, and SOX9/10 (Sauka-Spengler and Bronner-Fraser 2008). Many intracellular changes are required such as reorganization and the rearrangement of the cytoskeleton, the membrane compart- ments, and the extracellular matrix (ECM). Furthermore, and most importantly, the cells have to lose cell-contacts, detach from the epithelial and survive (Le Douarin and Kalcheim 1999). Cadherins are one of the key players during EMT and migration: Type I Cadherins which are characteristic for epithelial cells are down-regulated while Type II Cadherins are up-regulated. This change allows the cells to migrate (Chu et al. 2006). Another main feature of epithelial cells is the apico-basal orientation in the cell layer, which is also lost in the EMT by downregulation of occludin and claudin, the two major components of tight junctions (Newgreen and Gibbins 1982; Newgreen et al. 1982). Moreover, in order to delaminate and migrate, also the extra-cellular matrix (ECM) has to be removed (Perris and Perissinotto 2000). After neurulation and delamination, NCCs begin to sort into migratory streams, which involve highly dynamic processes (Le Douarin and Kalcheim 1999; Roffers-Agarwal and Gammill 2009). Migratory cells then possess a flattened, stretched morphology with filopodia and lamellipodia (Newgreen and Gibbins 1982; Savagner 2001).

3.2. Embryonic development modeled in vitro

3.2.1. Pluripotent stem cells

Pluripotent stem cells (PSC) have become an important tool in research since they bear an enormous potential to model and understand the early steps of embryonic development. More- over, especially in regenerative medicine, stem cell-based therapy proves to be a promising approach, e.g. (Fattahi et al. 2016; Zhang et al. 2001). Embryonic stem cells (ESC) are derived

General introduction from the inner cell mass of the blastocyst of a developing embryo (Thomson et al. 1998). Hu- man embryonic stem cells (hESC) are obtained for example from blastocysts that were fertilized but not used for in vitro fertilization. The cells bear an indefinite proliferation potential in vitro and once they are cultured, they provide an unlimited source for embryonic stem cells. Moreover, they keep their in vivo potential to (theoretically) differentiate into any cell type of the body (Amit et al. 2000). However, as described in chapter 3.1 “early embryonic development” they cannot give rise to a whole embryo since they are not totipotent as the zygote. They lack the capacity to give rise to extraembryonic tissues which are derived by the outer cell mass (De Los Angeles et al. 2015).

A major change in this young field was made by the discovery that somatic cells can be repro- grammed to induced pluripotent stem cells (iPS), showing basically the same features as ESC. This technique is based on the overexpression of few crucial transcription factors (OCT4, SOX2, KLF4 and c-myc) (Takahashi and Yamanaka 2006). The crucial transcription factor set could even be reduced to three factors: OCT4, SOX2 and NANOG. This technology opened up new possibilities such as the generation of patient-specific iPSCs and their differentiation towards clinically relevant cell types (e.g. cardiomyocytes (Chow et al. 2013), hematopoietic cells (Kaufman et al. 2001) or neurons (Fattahi et al. 2016; Lee et al. 2007). In the field of developmental toxicity testing, iPSCs could be used to replace hESC since ethical regulations prohibited the use of hESC in toxicity testing. However, the research field was also skeptic about the reprogramming process and asked whether this reprogramming step might proceed incompletely especially regarding the epigenetic landscape. Epige- netics patterns have already been shown to be inherited to future generations besides the DNA code itself (Yu et al. 2018). These concerns were addressed by numerous studies comparing hESC and iPSC lines, which lead to the result that no significant difference between the group of hESC and iPSC existed (Choi et al. 2015). Still, since every PS cell line behaves specifically and often shows a preferential differentiation track, stem cell culture maintenance and differentiation protocols may have to be adjusted to each distinct cell line respectively.

3.2.2. Rosettes - the in vitro counterpart of neural tube formation

Pluripotent stem cells (hESC as well as iPSC) have been used to generate neuroectodermal progenitor (NEP) cells that are prone to further develop into functional neurons such as dopaminergic neurons. This differentiation is guided by the small molecules noggin, SB431524 (Chambers et al. 2009) (and dorsomorphin) (Balmer et al. 2012) inhibiting BMP and TGFbeta signaling.

General introduction

BMP (inhibitors are noggin, chordin, and receptor-dorsomorphin) and TGFβ (receptor inhibitor is blocked by SB431524) signaling lead to the induction of smad by phosphorylation of smad1/5/8. P-smad and smad4 form complexes and are translocated into the nucleus where they induce the transcription of target genes (also see Figure 2. 6). Inhibition of this pathway prevents endoderm and non-neuronal ectoderm formation and cells differentiate towards the neuroectoderm (Figure 3, early NEPs). If NEPs are allowed to further differentiate and organize in cell complexes, the cells change morphology and begin to form flower-like structures the so- called neural rosettes (Chambers et al. 2011). Other differentiation protocols make use of embryoid body cultures in which cells are pre differentiated and, at the later stage, they are plated to form neural rosette structures (Elkabetz et al. 2008) (Figure 3, R-NSC). Morphology of neural rosettes strongly recapitulates early in vitro neural tube and the rosettes share its key features. Depending on the age of the neural rosettes, they consist of one or more cell layers and the cells surrounding an inner lumen are tightly connected via tight junctions mainly composed of zonula occludens I (ZO1) and N-cadherin. This suggests an apicobasal polarity of the cells with the apical part facing the central lumen (Conti and Cattaneo 2010; Elkabetz et al. 2008). The nuclei of the cells migrate period- ically during the cell cycle from apical to basal (and back) by interkinetic nuclear migration facilitating the cells to divide asymmetrically at the outer side of the rosette. (Nasu et al. 2012). Neural rosettes forming NEPs can be maintained in high cell densities (cell to cell notch signaling) with Shh and FGF8 conditioned medium (Fasano et al. 2009) or can further differentiate into regional specific neurons of the nervous system. For example, application of ventralizing factors such as Shh and FGF8 lead to dopaminergic neurons and motor neuron formation (Colleoni et al. 2011; Lee et al. 2007), while application of growth factors such as FGF2 leads to the release of neural crest cells (Mica et al. 2013). Neural crest cells can then be further differentiated into e.g. sensory neurons (Hoelting et al. 2016), melanocytes or cartilage. Mark- ers that have been shown to be specific for neural rosettes or neural rosette-forming neural stem cells/NEPs are FORSE1, ZIC1, DACH1, PLAGL1, PLZF, SOX2, NESTIN, PAX6 (Elkabetz et al. 2008). Depending on the axial level, also MSX1 and EMX2, LIX1 and LMO1 serve as marker genes for Rosette forming cells (Delli Carri et al. 2013; Koch et al. 2009).

General introduction

Figure 3: Generation of different cell types of neuro development in vitro and their in vivo counterpart (adopted from (Conti and Cattaneo 2010): Schematic overview of the cell types for which distinct differentiation protocols are available and which stages of in vivo development they mimic. For this thesis a differentiation protocol based on the NEP differentiation developed (Chambers et al. 2009) was adapted and the cells were further differentiated towards neural rosettes similar to the R-NSCs as published by (Elkabetz et al. 2008).

3.2.3. Tissue specific markers to determine regional specification

Marker genes are important genetic tools to identify certain tissues or cell types. However, presumably no gene is exclusively expressed in only one distinct cell line or tissue. For example, NANOG is one of the most important stem cell markers; however, it is known to also play a role in cardiac development (Kuegler et al. 2010). Therefore, it is important to use several markers to identify and characterize cells. By using gene expression profiles, a unique fingerprint or transcriptome patterns to determine cell type/state for categorization and spatiasliza- tion, can be obtained Kuegler et al. 2010 provide in their review a hand-selected marker set (for murine cells) that allows to characterize cells according to their cell type (stem cells, NEP, astrocytes, non-neuronal germ layer) as well as to gain information about their localization in the developing brain/spinal region (Figure 4).

General introduction

Figure 4: Summary of marker genes of regional specialization in the developing brain and spine (adopted by (Kuegler et al. 2010).

3.3. Neural tube defects:

During the last century, infant mortality decreased dramatically because of better medical supply for prenatal and neonatal care. Hence, neural tube defects (NTD) among others have therefore become a significant percentage for life-threatening medical conditions in newborns (Copp and Greene 2010). Neural tube defects describe the condition where neural tube closure fails and yields in neural tissue that is exposed to the environment (the amniotic fluid). The incomplete closure and the resulting environmental exposure can lead to prenatal neurodegeneration. The severity of the defect mainly depends on the level of the body axis that is affected. Since the closure is occurring discontinuous from day 18 to day 35, the effect is highly dependent on the timing, and especially during the 2-3 weeks of neurulation, the developing nervous system is extremely vulnerable. The consequences of NTD range from lethal defects in the brain, as a result of incomplete closure in the cranial region (e.g. Microcephaly, Exencephaly or Anencephaly) to spina bifida in various severity grades (Copp and Greene 2010) (Figure 5). In children with spina bifida parts of the spinal cord remain completely open or are only covered by skin. In the tail bud region often the mesodermal region does not separate from the spine 23

General introduction completely. This results in lesions with attached lipomas (Finn and Walker 2007). The latter case might not be lethal, however, such cases often require postnatal or even in utero surgery and usually lead to a lack in sensation below the lesion and other functional impairments (e.g. incontinence or inability to walk) (Copp and Greene 2010; Johnson et al. 2006; Sutton et al. 2001). Furthermore, the lesion may block the flow of cerebrospinal fluid, which results in an enhanced pressure on the brain and as a final consequence in the formation of hydrocephalus (Malcoe et al. 1999). Pregnant women are strongly advised to take in folate, especially before pregnancy and during the first weeks of pregnancy since several studies clearly showed that folate deficiency leads to an enhanced (reoccurrence) risk for NTD (Hibbard 1964; Laurence 1981). The exact mechanism is not fully understood, however, studies have linked folate one-carbon metabolism for which folate serves as a mediator of one carbon transfer for several biosynthetic processes to its mode of action (Beaudin and Stover 2009). However, in mice low folate supply does not lead to NTD unless combined with another risk factor and also folate nutrition succeeds to prevent only a proportion of NTD (Copp and Greene 2010).

Figure 5: Spatial timing of neural tube closure and their disturbance leading to NTD adopted from (Greene and Copp 2014). Based on the time when the neural tube does not close completely, different congenital malformation manifest in the developing embryo.

General introduction

3.3.1. NTD induced by genetic defects

Many studies have shown that neural tube defects in humans are multifactorial implying that one single factor is not sufficient to induce the disease. However, combinations of certain risk factors such as genetic causes, bad nutrition (such as lack in folate, vitamin B12 or inositol supply), environmental factors and maternal diseases such as diabetes mellitus, maternal obe- sity (glycemic dysregulation) but also hyperthermia can lead to failure of neural tube closure (Harris and Juriloff 2007). Recurrence risk for siblings is 2-5%, which means that it is about 50 fold higher than in the general population (Wyszynski 2006). Epidemiologic cohort studies suggested about 80 candidate genes that potentially increase the risk for NTD (Pangilinan et al. 2012). The most significant (1.8% increased risk) candidates were two polymorphisms in the homocysteine re-methylation gene MTHFR (Boyles et al. 2005). Experimental data have been generated in mouse models which resulted in more than 300 genes found to be associated with NTD (Harris and Juriloff 2007; Wilde et al. 2014). VANGL2 (in homozygotes), one of the key players of the non-canonical wnt pathway was identified to cause craniorachischisis. This is a severe consequence of failure of in the first closure site (Figure 5, a) that results in open midbrain, hindbrain, and the entire spinal cord. (Kibar et al. 2011; Murdoch and Copp 2010). Moreover, other key players of the PCP have been shown to play a role in the manifestation of NTD (e.g. CELSR1, FZD3 (Harris and Juriloff 2007)). But also knock out of other proteins e.g. related to the extra-cellular matrix (such as laminins, proteoglycans, fibronectin, proteases) or to transcription factors such as in PAX3 have been shown to result in NTD in combination with folate deficiency (Burren et al. 2008). Loss of function of any key player in neural tube formation leads to a failure in the closure or at least to an increased risk of NTD. Thus knock out experiments of distinct proteins (in mice, chicken, Xenopus and rats) lead to the investigation and confirmation of the pathways that are involved in neural tube formation (Greene and Copp 2009). Genetic related neural tube defects involve components controlling/contributing to cytoskeleton, cell proliferation, differentiation, transcriptional regulation, signaling (Shh, smad, BMP, PCP, RA, inositol) (Copp and Greene 2010). However, the majority of mutations only result in a phenotype if other factors such as folate deficiency co-occurs.

3.3.2. NTD induced by drugs (HDACi)

As reviewed by Grandjean and Landrigan only few environmental toxicants are known to act as DNT compounds (Grandjean and Landrigan 2006; Grandjean and Landrigan 2014a). These compounds have been confirmed by epidemiological cohort studies. However, such epidemiologic studies clearly giving evidence for causality are not easy to obtain, since the exposure 25

General introduction metrics are very difficult (Sim 2002) or as Smirnova et al stated: “it took epidemiology some 50 years to show that smoking induces cancer” (Smirnova et al. 2014). However, histone deacetylase inhibitors (HDACi) are known to cause severe neural tube defects in the developing embryo. They are used as antiepileptic drugs also in pregnant women. The risk for malformation or stillbirth of the embryo caused by epileptic seizures even exceeds the risk of HDAC inhibitors preventing the seizures. Due to this, many clinical data in human as well as studies in mice for comparison are available for this compound class. While for others hardly any data is available. Therefore, HDACi serve as highly promising tool compounds. The most well-known HDACi is valproic acid (Blotiere et al. 2019), but also man other compounds such as trichostatin A have been used. Recently the effect of 10 different HDAC inhibitors has been reviewed: the HDACi effect links to 23 different congenital malformations that include spina bifida, cleft palate, microcephaly, Ebstein anomaly, atrial septal defect and many more (Blotiere et al. 2019). HDACi lead to a massive reorganization of the epigenetic landscape, which results in deregulation of gene expression and consequently to a complex secondary effect. The primary DNT causing mechanism is not fully understood yet. Many different genes have been linked to the HDACi “disease mechanism” already and some have even been confirmed by knockout experiments in rodents. For example, MARCKSL 1 has been shown to be associated with the valproic acid effect (Chanda et al. 2019), but also other factors such as alteration of NMDA receptor signaling induced by VPA has been shown to cause NTD (Sequerra et al. 2018). HDACi inhibit the deacetylation of histones, that regulates DNA accessibility and thereby (among other epigenetic mechanisms e.g. DNA methylation) the transcription of genes. His- tones may be modulated by different marks, that either increase or decrease DNA accessibility. While methylation at different histone sites cannot be clearly linked to act as an active or repressive mark, histone ubiquitination acts as a repressive mark, mainly in hetero-chromatin regions. Histone acetylation has been shown to be mainly present in eu-chromatin, the active, accessible Chromatin (Dulac 2010). The usual dosage of VPA (as anti-epileptogen) is 1000-3000 mg/day. They result in blood concentrations of about 50-100 µg/mL and higher, which corresponds to 350-700 µmol/L (Ben- tué-Ferrer et al. 2010; Omtzigt et al. 1992; Turnbull et al. 1983; Vajda et al. 2004). At the concentrations (therapeutic range), valproic acid does induce birth defects such as spina bifida, exencephaly or other congenital malformations (Omtzigt et al. 1992; Vajda et al. 2004). It has been shown in the clinic that by lowering the VPA dosage from 1600 to 1100 mg (Omtzigt et al. 1992; Vajda et al. 2004), which corresponds to about 60µg/ml (Turnbull et al. 1983) the risk for birth defects is already significantly reduced.

General introduction

3.4. Developmental toxicity and developmental neurotoxicty (DNT)

Severe congenital (neuro) developmental malformations or disorders such as NTD are a huge concern. Additionally, worrying high percentage (about 10-15%) of newborns are affected by disorders of neurobehavior development such as autism spectrum disorders, learning deficits or other developmental delays (Grandjean and Landrigan 2014a). Since the rate of affected children has increased intensively, it is very likely that also the increasing amount of environmental toxicants, food additives etc. contribute to this trend and many compounds may act as DNT compounds (Grandjean and Landrigan 2014a). The number of chemicals produced in high volumes (more than 100 t/year) is steadily, and the chemical universe of compounds produced in medium and small volumes (more than 1 t/year) already exerts 30000. This leads to a continuous influx of chemicals in our environment. Until now, only few compounds, present in the environment in high concentration, have been tested on their potential toxic risk. Thou- sands of industrially produced compounds remain to be tested as set out in the council regulation (EG) No 1907/2006 (REACH) (Smirnova et al. 2014). The guideline strategy includes testing for neurotoxicity but not for developmental toxicity. Only if a compound generates an alert during testing it will be considered for further testing. This poses an enormous underesti- mated risk since the developing nervous system is way more vulnerable than the adult ones due to (i) immature blood-brain barrier, (ii) increased absorption versus low body weight and (iii) low ability to detoxify (Smirnova et al. 2014). Additionally, in the developing brain many complex processes are taking place that represent a potential target for toxicants, respectively. Moreover, while toxicant effects in adults are often only transient, in developing embryos/children, the effect becomes persistent or might even be enhanced when occurring early. The Barker hypothesis suggests that an effect that is not apparent first might lead to a delayed manifestation of a disease in later life (diabetes (Hales and Barker 1992)). DNT effects might be linked to an (earlier) outbreak of Alzheimer, Parkinson or similar age-related diseases. However, not only for each individual dealing with DNT impact is huge, DNT affects the whole society. Lowering the IQ of only 5 points results in a shift of the bell curve and leading to a significantly higher percentage of the population not able to care for themselves as well as to a decrease in the highly gifted people. Additionally, the costs arising from DNT are extremely high. For example, the annual costs of childhood lead poisoning in the US were estimated to exceed 50 billion US dollars (Grandjean and Landrigan 2014a).

General introduction

3.4.1. From Animal models to in vitro models- towards more predictive DNT testing

Safety assessment of chemicals and chemical mixtures still relies mainly on animal experimentation, which is mainly due to the lack of alternatives - especially in the field of developmental neurotoxicity (DNT). The OECD guideline advises in the TG246 how to test for DNT in animals. About 1000 animals are needed to test one compound, the time that is required exceeds three months and the costs are about $ 1.4 million (Smirnova et al. 2014). It took several decades to realize and to consider that animals apparently feel pain and suffer from such experiments. The 3R concept (refine, reduce, replace) lead at least to a change in thinking and resulted in seeking for better, more humane conditions for the laboratory animals (Russell and Burch 1992). Regarding 2Rs (Reduce and Refine), for example, animal suffering was reduced by handling improvement and fewer animals were sacrificed and wasted by spending more effort in planning experiments. In terms of replacement, there is still a long way to go. Some animal experiments used for toxicity testing may be easily replaceable and also the test method replacement development is already at an advanced stage: for example for eye corrosion, skin corrosion, skin sensitization robust in vitro methods have been validated by the ECVAM already (Liebsch et al. 2000). Developmental toxicology still represents a major discipline of toxicology where extensive work is required. The working model (as set out in TG426) remains a “black box” approach, since the mechanisms that lead to a toxic effect completely remain in the dark. This black box problem cannot be addressed by a complex animal experiment that does not allow for further research on the mechanistic part. Moreover, the predictivity is still very low because of their limited scope to assess specific outcomes (e.g. neuropsychiatric disorders) (Hartung 2009; Leist and Hartung 2013) and inter-species differences, which can be dramatic if missed (e.g. Contergan or Cytokine storm (Attarwala 2010; Stebbings et al. 2007)). To allow the development and evolution of this young field of toxicology, a new “way of thinking” has to be adopted. A new guideline was firstly proposed in the 21st century toxicity testing (Hartung 2009; Leist et al. 2008) and was complemented by many new strategies and sugges- tions to meet upcoming problems (e.g.(Bottini and Hartung 2009; Hartung 2008; Hartung et al. 2009; Hartung and Leist 2008; Hartung and McBride 2011; Leist et al. 2010)). Development of alternatives is pressurized by the vast amount of untested chemicals and a new regulation that prohibit the usage of animal experiments for cosmetic products (the 7th Amendment Directive of the Cosmetics Regulation). This hopefully leads to an accelerated recognition process of new methods in the OECD and in the EU by ECVAM.

General introduction

3.4.2. Concepts of a new DNT testing strategy: AOPs, TEPs, PoTs

To meet the needs and problems of DNT, new in vitro testing strategies have been designed. A new approach of testing focuses on a more optimal and purpose-oriented manner of testing. Decades of “black-box” testing (as mentioned in the last chapter) should be replaced by assays that consider the single steps of development and investigate not only the outcome without knowing the in-between processes. Certainly, unlike complex animal models- no single in vitro assay will be able to model and capture all processes that take place during embryonic development. However, this is also not the aim of a new testing strategy. New testing approaches suggest a combination of new approach methods (NAM) in a well-defined test battery that covers all key processes of (neuro) development. However, not only the key processes of development should be focused on, but rather the mode of action (MOA) of the test compounds should be addressed and summarized in pathways of toxicity (PoT) (Hartung et al. 2013b; Leist et al. 2008). PoT describe the network of cellular regulations that lead to the eventual cell fate (Balmer and Leist 2014; Hartung and McBride 2011). A model has been proposed by integrating a concept from a medical discipline: the toxicological endophenotype (TEP) (Balmer and Leist 2014; Kadereit et al. 2012a). In the field of psy- chology (Gottesman and Gould 2003) the endophenotype describes behavioral symptoms rep- resenting the phenotype that is linked to a genetic component (John and Lewis 1966). The toxicological endophenotype links an effect of a compound to a disturbed biological process (which can be modeled in vitro) that is further linked to an in vivo disease pattern (exophenotype). A disease such as autism or schizophrenia represents the in vivo exophenotype that is caused or at least linked to changes in the development such as migration, neurite outgrowth, differentiation or synaptogenesis, which would represent the toxicological endophenotype. (Figure 6 A, B). Such distinct mechanistic, biological processes can be modeled in in vitro systems. Unlike the highly complex systems biology approach of unraveling the molecularly defined PoT, a relatively easy, linear concept to reduce complexity for pragmatic decision-making, was proposed. The concept of an “adverse outcome pathway” (AOP) links a molecular initiating event (MIE) induced by a compound (for example blocking of channel x) to a final adverse outcome (AO) such as a disease (e.g. Schizophrenia) in a linear way (Bal-Price et al. 2015; Leist et al. 2017; Smirnova et al. 2014). MIE finally results in an AO, linked via several key events, which describe the disturbance of biological processes on any functional level. Key events range from molecular events to disturbance on whole organ level (Figure 6A, B). Since the completeness of key events is not of main interest, this concept is more or less easy to follow. This approach allows to leave gaps, e.g. the key event relationship (depicted as arrows) do not have to be all necessarily captured. Even if not all of the information is provided by an 29

General introduction

AOP, several AOPs can be combined and common key events (cKE) may be identified in this procedure. Finally, the aim is to gain more and more information about the in-between key events and to fill the gaps in an AOP. Data has been collected and several AOPs have been generated and combined based on assays and test compound knowledge. More than 300 AOPs have already been summarized in the AOPwiki database and many more are yet to come since many huge projects such as the horizon 2020 program EU-Toxrisk use this approach of generating AOPs. It has to be considered, that this approach also bears shortcomings and the concept is criticized by many toxicologists. Reducing the complexity has drastic consequences since compensatory mechanisms are missed, feedback loops are not considered and the key events are not seen in the context of other regulatory units (Leist et al. 2017). Furthermore, one compound often has more than one target. Often the AOPs are not defined properly and contain only poor information. Apparently, the basic concept of linking the key events to an adverse outcome (in vivo) is often poorly reflected and understood, since evidence for the KE-AO relation is missing. Combinations of AOPs or KE cannot be easily facilitated. Ultimately, it does only predict hazard, but not risk since treatment duration or concentration of a compound is not considered (Leist et al. 2017). The concept of an integrated approach of testing and assessment (IATA) has been evolved in the last 50 years by the OECD member states and describes a strategy that uses diverse lines of evidence for hazard prediction of chemicals. The IATA concept follows an iterative approach to assess the hazard and risk of a given compound by taking into account the acceptable level of uncertainty associated with the decision context. The IATA strategy is based on adverse outcome pathways (AOPs) and utilizes all available information in a weight of evidence assessment relying on animal data, in vitro data, in silico data, read-across. Furthermore, at the same time, it aims to generate new information (Bal-Price et al. 2015; Bal-Price et al. 2018b; Fritsche et al. 2017).

General introduction

Figure 6: The concept of adverse outcome pathways and toxicity endophenotypes. (B) Linear concept of adverse outcome pathways. (C) Test methods investigating AOPs.

General introduction

3.5. DNT Test assays

A toxicological test assay describes a method to obtain data that monitors the effect of a given compound on a specific biological process (Figure 6 A, C). The assay or test method is always structured in the same way and consists mainly (i) of a (stable) biological system (eg differentiation cells), (ii) a logic treatment scenario that exposes the biological system to a defined compound concentration in a defined time window, (iii) one or more defined endpoint(s) and a corresponding readout (e.g. viability, gene expression) and finally, (iv) a prediction model that assesses the result and defines a compound as positive or negative. Based on the limitations of the test method, acceptance criteria need to be defined (Aschner et al. 2017). The test assay is based on a distinct AOP and may also serve for data and information generation regarding the mode of action of a compound and for investigation of underlying biological processes (Figure 6 C). Building up a robust, reliable and relevant test assay requires extensive characterization. The system has to be challenged by positive controls, negative controls, and pathway-specific controls, and acceptance criteria have to be defined. The performance of a test regarding specificity (performance of identifying positives correctly; number of false positives), sensitivity (performance of defining negatives correctly, number of false negatives) and robustness needs to be assessed (Leist et al. 2012b). For some testing purposes, it may be preferable to have a test that is highly specific, while sensitivity is not of major interest. This might be the case for example for pharma pretesting. While for other purposes (such as for testing of industrial chemicals used in the environment) the vice versa situation might rather an advantage. The requirements for each test system differs depending on the output of the test. E.g. for some test assays a borderline range, which defines a compound as not clearly negative, might be useful. The borderline compounds might be prioritized for further testing in another assay/in more concentrations (Leontaridou et al. 2019; Leontaridou et al. 2017). In terms of DNT assays, no matter if in vivo or in vitro model, it has to be considered that the system itself is continuously changing, which may pose problems that are not apparent from first sight. Usage of e.g. differentiating cells means, that the cells from day 0 will not be the same on day 1 or 2 or n. This means in consequence that the cells depending on their state of differentiation might not share the same characteristics (Smirnova et al. 2014). Therefore the vulnerability to distinct compounds may not only vary (i) depending on the compound concentration and the (ii) exposure time but also (iii) the time window has to be taken into consideration for proper risk assessment. Experimental planning to define the distinct aims and expec- tations are inevitable. Depending on the desired endpoint that is measured, the baseline has to be adjusted and the toxicant effect has to be calculated relative to controls. E.g. if gene expression is used as an endpoint, gene expression has to be expressed as gene expression

General introduction change depicted relative to (i) time control (start of diff vs end of diff) and (ii) relative to an untreated control(Aschner et al. 2017; Bal-Price et al. 2015) DNT test system are often highly complex requiring many validation steps. However, designing and building up such a test is not the only difficulty that has to be addressed requiring a lot of attention, time and money. There are certain issues that further pose problems to the development of DNT tests:

- Identification of positive controls: Only few environmental compounds are known to cause DNT (Grandjean and Landrigan 2014a). However, for the evaluation process of a test assay, positive controls are needed (Aschner et al. 2017). - Gold standard problem. It was shown that numerous compounds have effects on only several species but not at all on others (e.g contergan, cytokine storm (Attarwala 2010; Stebbings et al. 2007)). Interspecies-extrapolation does not work in many cases and animal data do not hold true to predict human hazard reliably. However, the present model still remains the “gold standard”. Even if many good alternative models are available (such as for eye irritation), if the data are not absolutely clear, it is advised to test in animals as a final approach (IATA) (Casati 2018). - Regulatory pressure: Straight forward regulations are needed, to put more pressure on the development and validation of alternatives such as the 7th Amendment Directive of the Cosmetics Regulation that prohibits animal experimentation for cosmetic products (Smirnova et al. 2014). - Validation: The validation process as outlined by ECVAM orientates at a 1:1 replacement of animal experiments. There is no guideline for validating test batteries or assays that cover one step in a test battery (following chapter). - Irrational decisions: for example, an “ethical” regulation prohibits using human embryonic stem cells for testing purposes. Certainly, a lot of money and time was spent to adjust the tests to new induced pluripotent stem cells.

3.5.1. Transcritomics as endpoint

Among the most important tools to gain insight into cellular pathways and processes are the so-called ‘omics’-technologies. “Omics” describes the whole (micro) environment in a cell and includes proteomics, fluxomics, lipidomics, genomics, metabolomics, and many more. We are now able to obtain vast amounts of data relatively fast and easy by using transcriptomics as a toxicological endpoint. Unlike other endpoints, transcriptome data do not only deliver a toxic/nontoxic decision, but the huge amount of data can also be used for mechanistic investigation. Transcriptome data may be used in two different ways: (i) data can be used for toxicity

General introduction testing to define markers or patterns e.g. to define regional identity as in (Kuegler et al. 2010) or marker signatures that are specific for compounds. These approaches are often unbiased, relying on statistical/mathematical methods (Rempel et al. 2015) while the biological background is not considered at all. (ii) In the second approach, transcriptome data is used to gain information about underlying biological processes, which can be facilitated by data classification tools such as KEGG pathway or gene ontology (GO) databases, which are available online. The following examples for transcriptomics data interpretation (Figure 7) are split into the mentioned categories: Category i: A promising concept in disease, as well as toxicology, is the concept of using biomarkers as identifiers. With regards to toxicology, transcription factors (or other genes) may serve as markers for (un)disturbed development or compound-specific patterns may be obtained which would allow the generation of fingerprints to facilitate grouping and referencing (Hartung and Leist 2008). This might be a useful tool for compound evaluation and prioritiza- tion. However, how to define markers does not follow one principle, but can be achieved by many projections (e.g. toxicity biomarkers, endophenotype biomarkers, biomarkers for compound groups). An example of an advanced and complex approach is the analysis of combinations or patterns of markers like it was used by (Rempel et al. 2015) for the calculation of a classifier. Tran- scriptomic data was also used to calculate a toxicity index, as described in (Shinde et al. 2016a), which allows the ranking of compounds by their potential to disturb development just by quantifying the amount of DNT genes affected by a compound. Category ii: Interesting, new ways of interpreting data are to track the development of the DNT system and its disturbance caused by compounds in the context of “the whole body” or within the “whole development”. This was firstly set up as a small scale “gene correlation network” (Hoelting et al. 2016), where the development of sensory neurons was visualized in frame of other human sample cell types by PCA (Figure 7C). The data for other cell lines and tissues were obtained from databases available online. However, by continuously evaluating and adding data, this approach can be still improved. A more simple variant of this approach is the “cell lineage tracking” that was published recently (de Leeuw et al. 2020). In this method, marker genes for certain cells and tissues that arise during development (e.g. germ layers, neural crest) as well as markers for certain adult cell types (e.g. sensory neurons, astrocytes, cardiac cells), were chosen and used for endpoint-assessment in a DNT test assay. Using markers for germ layer formation was already used as an endpoint before (Hessel et al. 2018; Krug et al. 2013b; Shinde et al. 2015).

General introduction

Figure 7 How to use and interpret transcriptome data: Examples from different approaches. (A) The table summarizes approaches that have been used to capture transcriptomic changes. (B) Graphical output from the classification of two compound groups based on a PCA (adapted from Rempel et al., 2015). (C) Graphical output from “gene correlation network” adapted from Hoelting et al, 2016)

3.5.2. The UKN1/ STOPtoxUKN assay

The STOP-tox(UKN) assay/UKN1 uses human embryonic stem cells (hESC) or induced pluripotent stem cells (iPSC) as a test system to model neural induction (Chambers et al. 2009) and neural tube formation. By dual-smad inhibition (addition of the molecules noggin, SB and dorsomorphin) BMP and TGFbeta signaling is blocked and cells differentiate towards neuroepithelial precursor cells (NEP) within 6 days (Figure 8). However, in a second step, the cells are allowed to further differentiate. By reseeding the cells at DoD11 they gain polarity and self-

General introduction organize to form neural rosettes at DoD 15. These neural rosettes represent the in vitro counterpart of the in vivo neural tube (See Introduction chapter 3.1.3). In the standard setup, the developing cells are treated with compounds during neural induction between 0-6 DoD (Krug et al. 2013b) to identify disruptors of early neurodevelopment. As endpoint 1 at DoD6, cell viability (resazurin assay) as well as gene expression (transcriptome) is analyzed via microarray or qPCR. The major markers for successful NEP differentiation are PAX6 and OTX2. Fur- thermore, the downregulation of stem cell markers OCT4 and NANOG are monitored. Disturb- ance of the development leads in consequence to deregulation of these markers (less up- regulation of differentiation markers and less downregulation of stem cell makers). As a second endpoint, viability (cell count) as well as neural rosette formation are assessed by automated counting of the neural rosettes (manuscript 2).

Figure 8: UKN1 test method. Pluripo- tent stem cells are differentiated towards neuroectodermal precursor cells within 6 days. Cells are reseeded on d11 and the cells form neural rosettes until DoD15.

Established neural tube toxicants that lead to spina bifida in human embryos in vivo (such as valproic acid or trichostatin A) also lead to an altered gene expression profile in the UKN1 assay and rosette formation is disturbed already at low concentration. In frame of the Horizon 2020 EU-Toxrisk project, an AOP based on the HDAC effect (MIE) leading to “neural tube defects” as an adverse outcome was developed for the UKN1 test assay (Figure 9). In terms of the endophenotype concept, rosette formation assessed by the assay would represent the toxicity endophenotype linked to neural tube defects as an exophenotype.

Figure 9: Adverse outcome pathway induced by HDACi (AOP wiki database No. 275). Inhibition of Histone deacetylase (KE1) leads to activation of several key events (KE2-KE4) that manifest in vivo as neural tube defects.

Aims

4. Aims

The UKN1 system is a model system that has been used and improved for almost a decade. It has firstly been used as a test to identify DNT compounds; however, a lot of effort has been invested to use this system to understand early neurodevelopment and the underlying processes such as gene expression, epigenetics etc. The focus of this thesis was to further develop the test method by implementing a new phenotypic endpoint to the UKN1 test system, as well as to use the new assay to study early neurulation in vitro.

The major aims of this thesis were:

- to investigate the effect of acute toxic and non-cytotoxic concentrations on the transcriptome of a DNT-cell model. This should help to understand the mechanisms of transcriptomic regulation in a developing system in unperturbed and perturbed development as well as to distinguish between an acute and a developmental effect.

- to establish a differentiation protocol for neural rosettes as a new assay endpoint for

the UKN1/STOP-tox(UKN). The novel endpoint provides a robust tool to anchor transcriptomics to an endophenotype and to investigate the relevance of transcriptome changes at early neurodevelopmental stages (DoD6)

- to set up a model (together with Johannes Meisig and Nils Blüthgen) that is able to predict the influence of different valproic acid dosage (time x concentration) on expression of single genes by using a very limited dataset. To prove the information delivered by the computational model on affected pathways by rescuing the VPA-induced disturbance of rosette formation (endophenotype).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5. Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Tanja Waldmann*1, Marianna Grinberg*2, André König2, Eugen Rempel2, Stefan Schildknecht1, Margit Henry3, Anna-Katharina Holzer1, Nadine Dreser1, Vaibhav Shinde3, Agapios Sachinidis3, Jörg Rahnenführer2, Jan G. Hengstler4, Marcel Leist#1

*shared first authorship

1 In Vitro Toxicology and Biomedicine, Dept inaugurated by the Doerenkamp-Zbinden Chair foundation, University of Konstanz, 78457 Konstanz, Germany, 2 Department of Statistics, TU Dortmund, D-44221 Dortmund, 3Center of Physiology and Pathophysiology, Institute of Neu- rophysiology, University of Cologne (UKK), D-50931 Cologne, 4 Leibniz Research Centre for Working Environment and Human Factors (IfADo), Technical University of Dortmund, D-44139 Dortmund

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.1. Abstract

Analysis of transcriptome changes has become an established method to characterize the reaction of cells to toxicants. Such experiments are mostly performed at compound concentrations close to the cytotoxicity threshold. At present, little information is available on concentration-dependent features of transcriptome changes, in particular at the transition from non- cytotoxic concentrations to conditions that are associated with cell death. Thus, it is unclear in how far cell death confounds the results of transcriptome studies. To explore this gap of knowledge, we treated pluripotent stem cells differentiating to human neuroepithelial cells (UKN1 assay) for short periods (48 h) with increasing concentrations of valproic acid (VPA) and methylmercury (MeHg), two compounds with vastly different mode-of-action. We developed various visualization tools to describe cellular responses, and the overall response was classified as ‘tolerance’ (minor transcriptome changes), ‘functional adaptation’ (moderate/strong transcriptome responses, but no cytotoxicity) and ‘degeneration’. The latter two conditions were compared, using various statistical approaches. We identified: (i) genes regulated at cytotoxic, but not at non-cytotoxic concentrations; (ii) KEGG pathways, gene ontology term groups and superordinate biological processes that were only regulated at cytotoxic concentrations. The consensus markers and processes found after 48 h treatment were then overlaid with those found after prolonged (6-day) treatment. The study highlights the importance of careful concentration selection, and of controlling viability for transcriptome studies. Moreover, it allowed identification of 39 candidate ‘biomarkers of cytotoxicity’. These could serve to provide alerts that data sets of interest may have been affected by cell death in the model system studied.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.2. Introduction

Transcriptome analysis is a powerful tool to describe changes of cells in a comprehensive way, but few studies have addressed the question of concentration-dependent qualitative and quantitative changes in gene expression at a genome-wide level. Information on concentration- dependent changes is one of the approaches that may help to distinguish transcript regulations related to a compound’s hazard from those changes that rather indicate a cell’s adaptive response (Blaauboer et al. 2012). Such data will become important for systems toxicological approaches of hazard assessment (Heise et al. 2012; Jennings 2013; Jennings et al. 2013; Sauer et al. 2015; Wilmes et al. 2013). Moreover, information on concentration-dependent changes is relevant for classical applications of transcriptomics in toxicology, such as grouping of compounds based on similar transcriptome patterns, for the identification of pathways-of- toxicity (PoT) or modes of action (MoA) and for ranking of compound potency within a set of experiments.Transcriptome changes triggered by a given toxicant obviously depend on the exposure time (short, long, repeated). This is taken into account in some large studies (e.g. TG-GATES), where data are obtained after different exposure periods.(Uehara et al. 2010) However, most published studies use fixed time exposure conditions to evaluate toxicant effects. Besides timing, also the test compound concentration plays a major role.(Hermsen et al. 2013; Schulpen et al. 2015a; Schulpen et al. 2015b; Theunissen et al. 2012) For non-transcriptome endpoints, this factor has been widely recognized, so that e.g. quantitative high throughput (qHTS) screens have been developed in the Tox21 and ToxCast programs (Judson et al. 2016; Richard et al. 2016; Sand et al. 2016) to cover large concentration-ranges. Concentra- tion-dependent data also form the basis for extensive modeling and toxicity prediction.(Kleinstreuer et al. 2014; Shah et al. 2015; Thomas et al. 2013; Zhang et al. 2014) None- theless, the majority of transcriptomics studies have used a single fixed concentration, even if complex hazard predictions (carcinogenicity) were the objective.(Herwig et al. 2015; Schaap et al. 2015) This also applies to cases where the assay is well characterized for prediction of a regulatory endpoint of high relevance to the industry (e.g. GARD assay for skin sensitization), and with a need for potency information (Forreryd et al. 2015; Johansson et al. 2013). Only few studies have used at least 2-3 concentrations, (Schulpen et al. 2015b; Wilmes et al. 2013) and also the TG-GATES project did not explore the concentration issue more deeply (Igarashi et al. 2015). Although so few concentrations are not sufficient for robust concentration-response modelling, such studies have clearly indicated the importance of dose variation for the characterization of cellular responses (Krug et al. 2013b) Dose-response dependency on gene expression endpoints has been demonstrated for in vivo (Ahlborn et al. 2008; Andersen et al. 2008; Thomas et al. 2011; Thomas et al. 2013) and for in vitro (Gao et al. 2015; Li et al. 2015;

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Romero et al. 2015; Waldmann et al. 2014) toxicological test systems. If too low test concentrations are chosen, many responses will be missed, whereas a too high concentration may induce cytotoxicity with its associated signaling processes. This latter process can distort the transcriptome response and jeopardize the elucidation of the real MoA of a toxicant. In the field of developmental toxicity, exposure timing is particularly complex, as the system itself changes over time. (Balmer and Leist 2014; Smirnova et al. 2015) This means that the effect of a 48 h exposure will differ from that of a 144 h exposure, but also the effect of a 48 h exposure from day 0 to day 2 will most likely be different from the effect of a 48 h exposure from day 4 to day 6. The reason for this is that toxicants interfere with different waves of developmental gene expression, (Zimmer et al. 2011b) and for instance in neurally-differentiating pluripotent stem cells, different windows of exposure have led to distinct transcriptome responses.(Balmer et al. 2014; Balmer et al. 2012) Data from such experiments suggest that the transcriptome response triggered by toxicants present during the entire developmental period mostly reflects the altered differentiation state of the cells, while it informs only to a minor extent on the MoA of the toxicant.(Balmer et al. 2014; Smirnova et al. 2015) Short exposures to toxicants are required to obtain information on the signaling pathways triggered directly by toxicants. This applies to test systems of developmental toxicity or direct organ toxicity, such as neurotoxicity. If a study is mainly designed to elucidate mechanisms or biomarkers relevant to neurotoxicity, then (i) concentrations need to be chosen that are beyond the neurotoxic threshold, (ii) the exposure duration should be short enough to prevent too many secondary reactions from occurring (≤ 48 h), and if the cells chosen as target are differentiating, (iii) a time window needs to be selected in which little cellular change occurs. Systems toxicology (Sturla et al. 2014) attempts to assign the regulated transcripts to biological processes. In many standard approaches, the transcript data are overlaid with background biological information to filter signal from noise. Different databases contain such guiding biological background information and algorithms to such databases have been developed. The two most used databases are the gene ontology (GO) database (http://geneontology.org) (Ashburner et al. 2000) and the KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway database (http://www.genome.jp/kegg/pathway.html).(Kanehisa and Goto 2000; Ogata et al. 1999) The GO database is hierarchically structured and annotates every known gene to several GO terms (e.g. neuronal development or cell death). In contrast the KEGG pathway database consists of manually drawn pathways assembled by experts. These pathway databases can be used to bring biological sense into large sets of genes by overrepresentation analysis. Both databases have been used to characterize the gene expression response to the treatment of several in vitro test systems (Balmer et al. 2014; Pallocca et al. 2016; Rempel et al. 2015; Schulpen et al. 2015b; Waldmann et al. 2014)

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

The most common use of the above databases is to analyze whether certain GO groups or pathways are overrepresented amongst the regulated genes. At present it is unclear, how such information can eventually be used in toxicology, in particular for regulatory applications. There is a large need to further synthesize and categorize such information and to convert it into graphical displays that are intuitive and that confer at least semi-quantitative information on relevant concentrations and effect thresholds. Examples of such approaches are the establishment of a developmental toxicity index, (Waldmann et al. 2014) spider diagrams indicating the finger print of biological responses triggered by a toxicant, (Pallocca et al. 2016) the mapping of responses to an overall human transcription factor network(Rempel et al. 2015) and the establishment of quantitative indices like the ‘developmental potency’ (Shinde et al. 2016a). One in vitro test that has been explored extensively for such toxicogenomics studies is the UKN1 assay. (Balmer et al. 2014; Krug et al. 2013b; Waldmann et al. 2014) It recapitulates early stages of human neuroectodermal tissue development from the pluripotent initial stage to the formation of early neural ectodermal progenitor cells (NEP). (Balmer et al. 2012; Chambers et al. 2009) The differential responses of MeHg (or several other mercurial compounds) and VPA (or several other HDACi) in this system is well-established. (Krug et al. 2013b; Rempel et al. 2015) Most compounds have only been tested at a single concentration, but there is extensive concentration-response information on VPA available. (Waldmann et al. 2014) Moreover, different exposure windows have been explored for VPA, and TSA (another HDACi). It is known from such studies that exposure during the last 48 h of the 6-day differentiation does not affect the cell differentiation. (Balmer and Leist 2014; Balmer et al. 2012) It has also been examined how gene related to the developmental program of the test system (D- genes) overlap in different conditions with genes altered by toxicants (T-genes), and knowledge on these interactions has been used to define toxicity indices. The above background information was used to select the exposure time and the test compounds for the present study. VPA and MeHg were chosen, based on our extensive experience with these toxicants, and as they are associated with two largely different MoA. In order to avoid large overlap of D-genes with the cytotoxicity-related T genes that we aimed to identify in our present study, we exposed the differentiating cells only during the last 48 h of the 6-day differentiation period. We chose this condition, as very few D-genes change during this time, so that the genes identified are mostly T-genes and thus can give clear data on cytotoxicity related biomarker candidates after exposure to VPA or MeHg (Shinde et al. 2016a). In order to get insight into transcriptome changes associated with cytotoxic drug concentrations, we challenged the UKN1 test system with compound concentrations ranging from non-cytotoxic to cytotoxic. This allowed filtering of genes regulated only at cytotoxic concentrations, but not below. We found,

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity that both compounds can affect gene expression in concentration ranges that are not yet cytotoxic. These differentially expressed genes were mainly enriched in gene ontology terms dealing with neurodevelopment, indicating that our test system indeed alerts for developmental toxicity when potential toxicants are used at sub-cytotoxic concentrations. More importantly, by analyzing the response at high toxicant concentrations, we identified genes that may be used as candidate biomarkers of cytotoxicity and may help in the future to better interpret transcriptomics data in developmental toxicology.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.3. Materials and Methods

Materials

Gelatine, putrescine, sodium selenite, progesterone, apotransferin, glucose, insulin, methyl mercury and valproic acid were obtained from Sigma (Steinheim, Germany). Accutase was from PAA (Pasching, Austria). FGF-2 (basic fibroblast growth factor) and noggin and were obtained from R&D Systems (Minneapolis, MN, USA). Y-27632, SB-43154 and dorsomorphine dihydrochloride were from Tocris Bioscience (Bristol, UK). MatrigelTM was from BD Biosci- ences (Massachusetts, USA). All cell culture reagents were from Gibco/Invitrogen (Darmstadt, Germany) unless otherwise specified.

Neuroepithelial and rosette differentiation

Human embryonic stem cells (hESC) (H9 from WiCells, Madison, Wi, USA) were differentiated according to the protocol published by Chambers and colleagues (Chambers et al. 2009) with modifications established in (Balmer et al. 2012; Shinde et al. 2015; Waldmann et al. 2014). Instead of using 500 µM noggin we used the combination of 35 µM noggin and 600 nM dorsomorphine together with 10 µM SB- 431642 for dual SMAD inhibition as described earlier. This was used to prevent BMP and TGF-β signaling, and thus to achieve a highly selective neuroectodermal lineage commitment. For handling details, see supplemental material as decribed by Balmer et al., 2012 (Balmer et al. 2012). All differentiations were performed in 6-well plates containing 2 ml of medium per well. For some experiments, cells were further differentiated until they formed neural rosettes. In detail, cells were differentiated until DoD10 as described above or in in (Chambers et al. 2009). On DoD11, cells were detached using Accutase, and seeded at a density of 150,000 cells/cm2 onto Matrigel-coated 48 well plates. Cells were grown in N2S medium supplemented with 20 ng/ml FGF2, 100 ng/ml FGF8, 20 ng/ml sonic hedgehog and 20 µM ascorbic acid until rosettes formed at DoD15.

Experimental exposure and viability measurement

Treatment with valproic acid (VPA) was done with the indicated concentrations from 0.6 mM to 5 mM and with methyl mercury (MeHg) with concentrations from 1.5 µM to 40 µM. For "neurotoxicity" data generation, cells were differentiated without treatment until day of differentiation 4 (DoD4) as described above. On (DoD4) medium was removed and exchanged for new medium supplemented with the indicated concentrations of VPA or MeHg. All concentrations were prepared from a 1 M stock solution (in Aq. dest.) for VPA and from a 10 mM stock solution (in 44

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Aq. dest.) for MeHg. In order to determine cytotoxicity, a resazurin assay was performed on DoD6, exactly as described previously (Krug et al. 2013b; Stiegler et al. 2011)

Isocitrate dehydrogenase (ICDH) activity assay

ICDH (porcine, Sigma, I-2002) (10 µg/200 µl) in a tris(hydroxymethyl)-aminomethan (Tris)- buffer (20 mM) containing MnSO4 (2 mM), pH 7.4, was incubated with methyl mercury or valproic acid at 37°C for 20 min. ICDH activity was determined by the addition of isocitrate (4 mM) and NADP+ (0.1 mM). The enzymatic reduction of NADP+ to NADPH was photospectro- metrically followed at 340 nm over the course of 15 min at 1 min intervals at 37°C. The enzymatic activity was determined from the slope of the absorbance increase over time. All data were normalized to the activity of untreated enzyme (free of toxicant).

Glutathione reductase (GSR) activity assay

GSR (human, Sigma G-9297) (10 µg/ 200 µl) was incubated in sodium phosphate buffer (100 mM), pH 7.5, containing ethylenediaminetetraacetic acid (EDTA; 1 mM) and methyl mercury or valproic acid for 20 min at 37°C. To assess GSR activity, oxidized glutathione (GSSG) (5 µM), NADPH (0.4 mM), 5,5’-dithiobis(2-nitrobenzoic acid (DTNB) (all from Sigma) were added, the reaction was followed by measurement of the absorbance of 405 nm (37°C) in 1 min intervals over the course of 15 min. The enzymatic activity was determined from the slope of the absorbance increase over time. All data were normalized to the activity of untreated enzyme (free of toxicant).

Affymetrix DNA microarray analysis

After the measurement of viability (resazurin assay) medium was removed, and cells were lysed in RNA protect solution (Quiagen). Affymetrix chip-based DNA microarray analysis (Hu- man Genome U133 plus 2.0 arrays) was performed exactly as described earlier (Krug et al. 2013b; Rempel et al. 2015). The original data used for the comparisons with longer (chronic) VPA treatment (DoD0 to DoD6) were obtained in the context of an earlier study. These data and different form of analyses have been published by Waldmann et al, 2014 (Waldmann et al. 2014).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Microarray statistics

The following analyses were performed using the statistical programming language ‘R-version 3.1.1: for the normalization of the entire set of 36 Affymetrix DNA microarrays, the Extrapola- tion Strategy (RMA+) algorithm (Harbron et al. 2007) was used that applies background correction, log2 transformation, quantile normalization, and a linear model fit to the normalized data in order to obtain an absolute expression value for each probe set (PS) on each array, ranging from about 2 – 14 on a log2 scale. For comparability reasons the normalization parameters obtained in earlier analyses (Krug et al. 2013b) were used (frozen RMA procedure). After normalization, the difference between gene expression (at each toxicant concentration) and corresponding controls was calculated (paired design). Replicates of controls were averaged before subtraction from corresponding exposed samples. Differential expression was calculated using the R package limma (Smyth 2005). Here, the combined information of the complete set of genes is used by an empirical Bayes adjustment of the variance estimates of single genes. This form of a moderated t-test is termed here as ‘Limma t test’. The resulting p-values were multiplicity-adjusted to control for the false discovery rate (FDR) by the Benjamini– Hochberg (BH) procedure (Benjamini and Hochberg 1995). As a result, for each concentration, a gene list was obtained, with corresponding estimates for log fold change and p-values of the Limma t test (unadjusted and FDR-adjusted).

COMBAT batch correction

Batch effects, non-biological experimental variation, are commonly observed phenomena in the area of microarray studies. They occur due to experiments that cannot be conducted all at once, for various reasons. It is important to remove batch effects, as otherwise relevant mode of actions might remain undetected due to strong batch effects. We apply the ComBat algorithm (Johnson et al. 2007) that allows to adjust for batch effects in datasets where the batch covariate is known. It uses a non-parametric empirical Bayes approach for the estimation of an additive and multiplicative batch effect. First the data is standardized with respect to mean gene expression and treatment effect. Then the batch effects are estimated and eliminated from the standardized data by subtracting the additive effect and dividing by the multiplicative effect. Finally the data is back transformed, i.e. mean gene expression and treatment effect are added.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Data display algorithms

Heatmaps were used to visualize hierarchically clustered matrices of gene expression values. The colors encode the magnitude of the values, ranging from blue (low) to yellow (high). Prin- cipal component analysis (PCA) plots were used to visualize sets of expression data in two dimensions. The percentages of the variances covered on each axis are indicated in the figures. The calculation and display of cytotoxicity curves was done using GraphPad Prism 5.0 (Graphpad Software, La Jolla, USA). Ring diagrams were generated with Microsoft Excel (2011). One semi circle represents the amount of all significant oGO (cut offs: p-value elim < 0.01;|FC| > 1.8) determined for each concentration and is set to 100 %. The size of each segment represents the amount of different superordinate biological processes in % normalized to all oGOs. The spider diagrams were done in Microsoft Excel (2011). The axes represent the number of KEGG pathways belonging to the given KEGG categories.

Gene ontology

The gene ontology (GO) group enrichment analysis was performed using R-version 3.1.1 with the topGO package (Alexa et al. 2006) using Fisher’s exact test, and only results from the ontology domain "biological process" ontology. The resulting p-values for overrepresentation of GO terms amongst gene sets were corrected for multiple testing by the method of Benja- mini−Hochberg procedure (Benjamini and Hochberg 1995). In order to avoid redundancy the GO group enrichment analysis was performed by using the the topGO elim algorithm. The GO database is organized hierarchically, i.e. it forms a directed acyclic graph (DAG) (Fig. S5.2A). Every GO term is represented by a node in this graph. Higher-level nodes represent GO terms from more general biological processes and lower-level GO terms from more specific processes. The complex structure of the GO introduces strong dependencies among the GO terms but lower-level nodes are of higher biological relevance since their associated definitions are more specific than the definition of their ancestors. Therefore we use the elim algorithm that iteratively bottom-up removes the genes mapped to significant GO terms from all more general (higher level) GO terms (Fig. S5.2B), unlike the classic algorithm (Fig. S5.2A), which neglects the local dependencies between GO terms in the significance calculations.

GO activation profiles

A new biostatistical method (Waldmann et al. 2014) for measuring activation profiles of GO groups across all concentrations (up- or down-regulation) was applied. The method uses a segmentation test, which uses the log FC values as segement definition (Minguez et al. 2006)

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity to identify significant enrichment with up or down regulated genes (restricted by FDR < 0.05), comparing each concentration with the controls. The large number of enrichment tests is FDR- adjusted according to the two-step procedure of Benjamini-Hochberg (Benjamini et al. 2006) to limit the FDR to 1%. Each GO group obtains a corresponding activation profile. In each profile, the order is from left to right: lowest dose to highest dose and always in relation to the untreated control. The profile ‘0++++++’ means that the corresponding gene set is significantly enriched with up-regulated genes at the second lowest concentration and for all higher concentrations. Analogously a gene set with profile ‘00000--’ is enriched with down-regulated genes only at the two highest concentrations. The full list of GO activation profiles is given in Table S5.6. Further, for each GO activation profile we assigned a key biological process as described earlier (Waldmann et al. 2014) and selected 1-2 examples for each key biological process and calculated a quantitative GO activation score for visualization of the profiles. To this end the GO activation score was calculated for each concentration by multiplying ‘the percentage of genes within the GO that was found to be significantly regulated’ (p-value < 0.05) with ‘the average fold change of their regulations’. Then theses scores were plotted against concentrations of the compounds.

KEGG pathway enrichment analysis

KEGG pathway analysis was performed using the R package hgu133plus2.db (Carlson). Probesets are mapped to the identifiers used by KEGG for pathways in which the genes are involved. The enrichment was then performed analogously to the gene ontology (GO) group enrichment analysis using Fisher’s exact test. To visualize the regulations of the single genes that belong to the pathways we calculated the mean of all significant PS corresponding to each gene and colored it redish for upregulation and blueish for downregulation.

Superordinate biological processes oGO terms obtained by enrichment analysis and GO activation profiles were classified into superordinate biological processes by expert knowledge (Waldmann et al. 2014). The classification for overrepresented KEGG pathways corresponds to the 7 groups given by KEGG pathway: metabolism, genetic & environmental information processing, cellular process, organismal systems, human diseases and drug development (Kanehisa and Goto 2000; Ogata et al. 1999). These superordinate biological processes were visualized by spider and ring diagrams (Microsoft Excel 2011).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.4. Results

Determination of cytotoxic concentrations

We used the UKN1 test system that is based on differentiation of pluripotent cells to neuroepithelial cells within 6 days, and which has been extensively characterized earlier (Balmer et al. 2014; Balmer et al. 2012; Krug et al. 2013b; Shinde et al. 2016a; Shinde et al. 2015; Waldmann et al. 2014). As toxicant exposure period we chose the last 48 h of the overall 144 h differentiation (Fig. 5.1A). Under such conditions, the test system has been shown to be relatively re- sistant to toxicant-induced developmental disturbances, and the transcriptome effects of substances present during this time period reflect mostly direct cellular responses (adaptations, signaling pathways and possibly also cytotoxicity) (Balmer et al. 2014; Balmer et al. 2012). Initially, we performed a pilot study to determine the approximate cytotoxicity threshold for the test compounds (VPA, MeHg) under these test conditions. Guided by the pilot experiments, we selected a range of concentrations in a way that some were expected to be non-cytotoxic and some would be beyond the toxicity threshold. Cells from three different experiments (cell lots) were treated with the indicated concentrations (0.6 - 5 mM VPA and 1.5 - 40 µM for MeHg), and at the end of the incubation period, both the viability was determined, and the RNA was harvested for microarray analysis from the same samples. Concentrations of ≤ 1 mM VPA and of ≤ 15 µM MeHg were found to be non-cytotoxic (> 90% viability). At concentrations of 3- 5 mM, VPA significantly decreased the viability (by 20 – 30%) (Fig. 5.1B); for MeHg we detected a massive (by 40 – 60%) decrease of viability at 20 - 40 µM (Fig. 5.1C). Thus, at least two cytotoxic and two non-cytotoxic condition conditions were available for each compound for subsequent analysis of the respective transcriptome changes. For the quantification of cell viability, we used here as standard approach a resazurin reduction assay, i.e. an analytical method that has worked reliably on stem cell cultures in our previous experiments. The data were fully confirmed by cell counting of H-33342-stained cultures. Nevertheless, one can never be entirely sure that viability tests capture e.g. a subtle functional damage that may have delayed effects on cell survival. For this reason, we used an additional functional test of the developmental potential of treated cultures. For this purpose, treated and non-treated neuroepithelial cells (= DoD6 cultures) were further differentiated towards neural rosettes. Such in vitro structures have been considered as cell culture equivalent of the neural tube stage in vivo(Chambers et al. 2009), and are established toxicological endpoints (Colleoni et al. 2011) Under control conditions, well-structured rosettes were formed. Upon 48 h treatment with VPA or MeHg, we found that the highest non-cytotoxic concentration for each compound was still

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity compatible with subsequent rosette formation (Fig. 5.1D). In contrast to this, all cytotoxic concentrations were incompatible with further differentiation (no cell attachment, no rosette formation). This semi-quantitative assay confirmed that the cytotoxicity threshold identified by resazurin testing clearly correlated with functional capacities of the cells. Such toxicity thresholds depend on the exact experimental conditions. For instance, in previous studies, cells had been exposed to the compounds for the entire 144 h (instead of 48 h). Under these conditions, the toxicity threshold (EC10) for VPA had been at 0.6 mM.(Balmer et al. 2014; Balmer et al. 2012; Krug et al. 2013b) We performed a rosette-formation assay also for these long-term exposure conditions. Also under these conditions, cytotoxic concentrations (as determined in the resazurin assay) were incompatible with the assay conditions, as the damaged cells did not re-attach under the assay conditions. Under non-cytotoxic conditions (0.6 mM VPA), the cells did attach for the rosette assay, but their differentiation was disturbed. Instead of forming neuroepithelial rosette structures, VPA-treated (144 h) cultures formed an epithelial-like tissue structures as indicated by ZO1-positive tight junctions between adjacent cells. This is fully in line with the transcriptome changes and might be a first phenotypic anchoring under these conditions that are indicative of malformations related to the fetal valproate syndrome.(Balmer et al. 2014; Balmer et al. 2012) In summary, these cytotoxicity tests indicate that the associated transcriptome data may refer to three conditions: (i) clearly non-cytotoxic (allowing cell survival, and rosette formation), (ii) clearly cytotoxic ones (corresponding to neurotoxic effects in vivo), and (iii) conditions that are not cytotoxic (high viability), but disturb neurodevelopment (failure to form rosettes).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 1: Transcriptome changes of neurally-differentiating stem cells induced at different toxicant concentrations. (A) Differentiation scheme. Embryonic stem cells (hESC) were differentiated within six days to a pure population of neuroepithelial precursors (NEP). The cells were treated for the last two days of differentiation (DoD), i.e. from DoD4 to DoD6, with the indicated concentrations of VPA (B) and MeHg (C). On DoD6, resazurin-reduction was measured, as measure of viability. Data are means ± SD of three different experiments. The dashed red line indicates a viability of 90%. Values below this were considered here to reflect cytotoxicity. (D) Cells were treated as indicated. After the standard 6 day differentiation, cells were further differentiated until day 11 in the absence of any added toxicants. Then, cells were detached, dissociated and re-plated to test for their capacity to form neural rosettes. After four additional days, they were immunostained for ZO1 (red), GM130 (green) and DNA (H-33342 dye, blue). The micrographs are representative images taken from 3 differentiations. (E) Cells as treated in A and B were used for whole transcriptome analysis. In addition, data from a previous publication (Waldmann et al. 2014) with VPA treatment from DoD0 – DoD6 were included.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

The expression changes of the 100 PS with the largest variance are displayed in form of a two-dimensional principal component analysis (PCA) diagram. Each point represents one experiment (= data from one microarray), and the color coding indicates the concentrations as given in the legend. The form of the data points indicates the compound and treatment duration (circle: VPA DoD4 – DoD6, square: VPA DoD0 – DoD6, triangle: MeHg DoD4 – DoD6) used in the experiment. The axis labeling indicates the percentage of the overall data variance covered by the respective axis. Controls (data from untreated cells) have been subtracted from the data displayed, so that the PCA displays "deviation from control".

Overview of transcriptome changes under different test conditions

The transcriptome changes triggered by 48 h exposure (DoD4 – DoD6) were analyzed by DNA microarray analysis (Affymetrix), and for an initial overview the data were displayed together with those from the 144 h exposure to VPA (legacy data from a previous study) in a 2D principal component analysis plot. The three most conspicuous findings were that (i) data of three fully independent experiments per test concentration clustered closely together; (ii) a clear concentration-dependent progression was seen under all conditions; (iii) the three types of treatment showed three clearly distinct tracks (Fig. 5.1E). The only exception was found for 40 µM MeHg. Under this condition, the high level of cytotoxicity (Fig. 5.1C) had made it difficult (for one sample even impossible) to isolate an appropriate amount of intact RNA. Therefore, this concentration was excluded from some of the further analyses in this study. For a further overview of the data structure, we generated heatmaps and corresponding dendrograms (Fig. S5.1) calculated by hierarchical clustering. This data display confirmed the impression from the PCA (Fig. 5.1E) that the different experimental replicates cluster closely together for all experimental conditions. Moreover, it was confirmed that cells exposed to 1.5 µM MeHg hardly differed from untreated controls (Fig. S5.1B). For a rough quantification of changes, differentially expressed probe sets (PS) were determined. For VPA the number of up- or down-regulated PS gradually increased from the lowest to the highest concentration. The same was observed for MeHg. Only treatment with 1.5 µM MeHg induced no significant change in gene expression (after false discovery rate (FDR) correction) (Fig. 5.2A, Tab. S5.1). The apparent number of PS altered by 40 µM MeHg appears to be lower than for 20 µM MeHg, but this is due to the different statistical basis (only two samples could be analyzed), and these data are therefore excluded from further quantitative analysis (Fig. 5.1E).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 2: Concentration-dependent changes of regulated probesets and biological processes.

(A) The left panel displays the changes induced by VPA and the right panel the changes triggered by MeHg. The numbers of differentially expressed PS compared to untreated control (p<0.05; |FC| >1.8) are displayed. (B) Signif- icantly up- and down-regulated PS were scrutinized for overrepresentation of GO terms (oGO) among the PS. The numbers of oGOs (p < 0.05; GO elim algorithm) among the respective PS are displayed. (C) Differentially expressed up- and downregulated PS were pooled for each condition and analysed for overrepresented KEGG pathways (p < 0.05). (D, E) Neurotoxicity display. To visualize the relationship between individual PS and biological features overrepresented amongst them, the number of PS was plotted on the x-axis and the number of oGO terms (D) or KEGG pathways (E) was blotted on the y-axis. Each of the plots contains information for both, MeHg and VPA. Three separate areas are distinguished within the scatter blots by dotted lines: tolerance to toxicant effects; functional toxicity (altered transcriptome, but no cell death); cytotoxicity.

. 53

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Overview of altered biological processes as consequence of transcriptome changes

To get an overview on how toxicant concentrations affected the number of coordinately regulated genes (belonging to a GO term or a KEGG pathway) we analyzed the significantly (p- value < 0.05 after FDR correction, |FC| > 1.8) regulated PS for statistical over-representation in GO terms and KEGG pathways. First we investigated overrepresented gene ontology terms (oGO). For a quantitative comparison of oGO amongst different experimental conditions, it is important to understand the structure of the GO database. It is organized hierarchically, which implies that genes belonging to a low level group in the hierarchy are automatically also part of all higher-level groups. To face this problem we used the recently developed TopGO elim algorithm(Alexa et al. 2006) that starts with the processing the nodes from the most specific level, and then iteratively moves to nodes on a more general (higher) level. A graphical demonstration of this procedure and its effects on the significance values of GO overrepresentation can be found in supplemental material (Fig. S5.2). The enrichment analysis, based on the TopGO elim algorithm, showed that the number of oGO increased with increasing VPA concentrations, in parallel with the increase of regulated PS (Fig. 5.2B, left panel Tab S5.2). For increasing concentrations of MeHg the number of oGOs increased steeply from 10 µM to 15 µM (both non-cytotoxic), and no further rise (rather some decrease) was found in the cytotoxic range (Fig. 5.2B, right panel, Tab. S5.3). Analysis for overrepresented KEGG pathways was done on the basis of a pool of up- and down-regulated PS, since both may contribute equally to one given pathway. The results showed that already the lowest concentration of VPA triggered > 25 pathways, and this number increased only moderately with higher concentrations (Fig. 5.2C, left panel, Tab. S5.4). Far less KEGG pathway were overrepresented among the MeHg regulated PS. There was no pathway at 10 µM, and only 7-11 pathways at 15-20 µM (Fig. 5.2C, right panel, Tab. S5.4).

Graphical representation of concentration dependent changes in key biological processes

To provide a basis for toxicological judgement, it is necessary to find measures of how far a cellular transcriptome moves away from the normal ground state. In previous work, we defined a two-dimensional teratogenicity index (Waldmann et al. 2014) and captured transcriptome deviations in developmental models by the two measures of developmental potency and developmental index (Shinde et al. 2016a). For the more acute neurotoxic exposure used here, we designed an analogous 2D neurotoxicity display. The underlying principle was to plot the various experimental conditions in a diagram with the number of altered PS as x-axis, and the oGO or overrepresented KEGG pathways on the y-axis. The overall distance from the origin 54

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity describes the extent of cell disturbance, with the direction of the deviation informing on the degree to which coordinated processes are being disturbed (instead of relatively random disturbances of non-related genes). The acute neurotoxic exposure conditions used in the present study showed a progressive activation of oGO/KEGG with increasing overall transcriptome changes (PS), while the developmental neurotoxicity protocol used earlier (Waldmann et al. 2014) had shown a very distinctive step-wise change. As alternative approach to provide a quick display of toxicant-induced changes, we used a display format that informs on the types of key biological processes changed under different experimental condition. For this purpose, the identified oGOs were further assigned to fourteen superordinate cell biological processes, such as “non-neuronal differentiation” or “apoptosis” (Fig. 5.3A). This classification of oGOs into superordinate cell biological process was done as described earlier(Waldmann et al. 2014) to combine the hundreds of possible oGO to a small number of larger categories that could be displayed graphically. The assignment to these categories is documented in the supplementary material (Tab. S5.2, Tab. S5.3). To exemplify the procedure, we chose four conditions. For each test compound we picked the highest cytotoxic concentration and we chose for comparison a concentration triggering a biological response in the non-cytotoxic range (0.6 mM VPA, 10 µM MeHg for upregulated PS; 3 mM VPA, 15 µM MeHg for downregulated PS): On first sight, this display format indicates clearly that different types of processes are affected by up- and down-regulated genes (Fig. 5.3A). These is a strong indication that gene regulation follows an overall biological pattern and is not just a random event. Among the oGOs identified among the upregulated PS for both compounds and independent of the concentration were biological processes such as "non-neuronal differentiation", "signaling pathways". The overview also showed clearly that "cellular stress response" was more prominent for MeHg than for VPA, and "cell death/apoptosis" was only overrepresented for MeHg. Among the downregulated PS, the two main topics were “nervous system associated” and “transcription”. This agrees well with disturbed differentiation to neuronal cell types, as described earlier for VPA.(Waldmann et al. 2014) The analyses up to this point indicated that fewer biological processes are affected on the transcriptional level by MeHg (than by VPA). This is consistent with previous results (Krug et al. 2013b; Rempel et al. 2015) and agrees with the assumed mode of action of VPA, i.e. a direct modification of the chromatin structure. The narrow window between transcriptional changes and induction of cell death that we observed for MeHg might be due to a MoA of MeHg that directly affects vital proteins. MeHg is known to interact with cellular thiols, and it may thereby kill cells without large coordinated changes on the transcriptome. For experimental verification of such an assumed interaction of MeHg with pivotal enzyme activities we chose two protein candidates known to depend on free thiol groups for their function.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 3: Superordinate biological processes and enzymatic activities affected by toxicants

(A) The GO elim algorithm was used to identify oGOs (p < 0.01) amongst up- and downregulated PS. They were then assigned to 15 (14 as specified in the legend plus "others") superordinate biological processes. The number of oGO in each of the 14 defined processes is displayed as ring diagram. The segment sizes (in radiants) correspond to the percentage of oGO in a superordinate biological process, relative to all classified oGO, with 180° corresponding to 100%). Each ring diagram shows four experimental conditions. Two concentraions of VPA (left) and two of MeHg (right). The full information of all concentrations is given in supplementary tables S2 and S3. The toxicant concentrations chosen for display correspond to the highest concentration for each compound (cytotoxic; inner circle) and to the respective lowest concentration (outer circle) triggering a significant biological effect (= regulating GOs). (B) Isolated enzymes were used for activity assays in the presence of the indicated toxicant concentrations. The measured activities were normalized to enzyme activity in the absence of toxicant and expressed in %. Enzyme inhibition was calculated as "100 - % activity". Data are means ± SD from triplicate determinations. *: p < 0.05.

The activity of glutathione reductase (GSR) and isocitrate dehydrogenase (ICDH) was measured under various experimental conditions (Fig. 5.3B). We found that MeHg directly inhibited the two enzymes in a concentration range relevant to our studies. In contrast, no inhibition was observed by VPA, even in the high mM range. These findings may be important for the interpretation of transcriptome studies in developmental toxicology: for some compounds, like VPA,

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity pronounced changes of the mRNA pattern can be observed, and these may inform to some extent on the MoA of the compound. On the other hand, there are also toxicants, like MeHg, that mainly act via direct enzyme inhibition, and the window between cellular responses triggered by such inhibition, and immediately ensuing cell death can be very narrow. Under such conditions we might not be able to clarify the mode of action (MoA) of a drug (with respect to developmental disturbances under non-cytotoxic conditions) by transcriptome analysis. How- ever, we may still identify a common signature of gene expression alterations for cytotoxicity. This approach was further followed in the present study.

Figure 5. 4: Concentration dependent regulation of biological processes.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

(A) Quantitative GO activation scores were calculated for each concentration by multiplying ‘the percentage of genes within the GO that was found to be significantly regulated’ with ‘the average fold change of these regulations’. The analysis was performed for all GO terms that were overrepresented at least at one toxicant concentration. For the exemplary GO terms indicated in the figures, the names are indicated either directly on the lines or in the legend (for those where space constraints precluded direct labelling). Examples are representatives of different profiles of GO activation scores (see Tab. S5 for a complete list). (B) Two major "topics" identified amongst the GO activation profiles in (A), i.e. "cell death" and the "development of other (non-neuronal) lineages" are displayed for direct comparison of curve shapes. To allow this, the concentration data were re-classified as to their relative position to the cytotoxicity threshold. For transparency reasons, the real concentrations for each data point are indicated within the graphs.

Concentration dependence of the activation of biological processes for different compounds

We followed up on the question, in how far different compounds trigger a coordinated transcriptional response already in the non-cytotoxic range or only at concentrations associated with cell death. For this purpose, we decided to follow the regulation of entire GO groups, instead of individual genes, by examining gene ontology activation profiles. To obtain these, we applied a relatively new bio statistical method (Waldmann et al. 2014) that analyzes the activity (up- or down-regulation) of genes annotated to a GO group across all concentrations. For VPA, most of the monotonic activation profiles referred to GO terms from the superordinate biological processes “stress response” and “non-neuronal lineages”. They included "response to inorganic substance", "response to hypoxia" or "circulatory system development". Activation profiles that started only at the highest VPA concentrations referred mostly to “apoptotic processes”, but also an oGO relating to “circulatory system development” emerged at this high concentration (Fig. 5.4A left panel). Thus, several coordinated responses were triggered by VPA in the non-cytotoxic range. This set of responses may inform on the MoA of VPA as developmental toxicant. On the other hand, the responses triggered in the cytotoxic range may be less important for information on the MoA as developmental toxicant. For MeHg, the situation looked different: few GO were activated at non-cytotoxic concentrations, and these included “regulation of apoptosis”. These data indicate that for MeHg, at the conditions used here, there was only a very small concentration span that triggered transcriptional changes independent of cytotoxicity (Fig. 5.4A right panel). The full overview of all GO activation profiles is given in supplementary material (Tab. S5). For a better comparison of VPA and MeHg, concerning the two main topics among the GO activation profiles, “cell death” and the “induction of non-neuronal lineages”, we constructed a direct graphical comparison (Fig. 5.4B): to align information on two compounds with different concentration ranges, we used an x-axis that was

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity scaled according to the study concentrations steps below (negative numbers) and above (positive numbers) the cytotoxicity threshold. By this transformation, the compound concentration ranges were aligned for their cytotoxicity threshold. When GO activation scores were plotted in this way, it got particularly evident that the regulation of cell death pathways was already initiated at low (relative to the cytotoxicity threshold) concentrations of MeHg, but only at high (cytotoxic) concentrations of VPA (Fig. 5.4B, left panel). On the other hand, GO groups that indicate a wrong differentiation track were induced already at low concentrations of VPA, but only above the cytotoxicity threshold of MeHg (Fig. 5.4B, right panel). In summary, by using and comparing two largely different compounds, we found that there are some common biological processes triggered at cytotoxic concentrations, but large differences are observed for processes triggered at non-cytotoxic concentrations. While the latter ones may inform on the MoA of developmental toxicants, the former ones may inform on direct neurotoxicity or other acute effects of chemicals. We proceeded to look for more descriptors of the cytotoxicity indicators.

Figure 5. 5: Grouping of overrepresented KEGG pathways in superordinate biological processes

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

(A)The KEGG pathways determined in Figure 2C were assigned to the pathway groups given by the KEGG data base itself. The number of identified pathways in each group is displayed as spider diagram. The axes correspond to the number of significant KEGG pathways belonging to the groups. For this analysis, the highest cytotoxic (in orange) and the highest non-cytotoxic (green) concentration for each toxicant (VPA: left panel; MeHg: right panel) were used. The pathways that were "activated" (= overrepresented amongst regulated genes) by cytotoxic concentration, but not by non-cytotoxic concentrations were determined. These pathways were then compared for MeHg and VPA, and those that were activated by both toxicants are indicated. (B) Genes were selected based on their role in common KEGG pathways. The main criterion was regulation into the same direction by both compounds (VPA and MeHg). In a second step, we only retained the genes that were regulated under all cytotoxic conditions, and that showed a ≥ 4-fold change for at least one condition. For each gene, the pathway from which it was selected is indicated.

Identification of KEGG pathways over-represented amongst toxicant-altered genes

The KEGG pathway database provides an alternative resource to map genes against potential signaling pathways relevant for toxicity. There are significantly fewer KEGG pathways than GO terms. The pathways are manually assembled, and the curation level is significantly higher than for the GO database. Moreover, the pathways are already assigned to seven superordinate groups (Kanehisa and Goto 2000; Ogata et al. 1999) in the database. We explored here, whether KEGG pathways were over-represented amongst the genes found to be regulated in our study. We pooled the up- and downregulated PS for this analysis (as pathways obviously contain positive and negative regulators). The number of KEGG pathways overrepresented with either the highest non-cytotoxic or the highest cytotoxic concentration were identified. These pathways all fell into five of the seven KEGG pathway groups, i.e. human disease, cellular process, organismal systems, metabolism and genetic & environmental information processing (here named “signal transduction”). Both compounds clearly regulate more pathways related to cellular metabolism, when the concentration was beyond the cytotoxicity threshold. A presentation of the data in the form of spider diagrams also showed that the information extracted from the KEGG data base clearly differed from the one of GO overrepresentation analysis. For instance, mercury affected a relatively high number of pathways, already at non- cytotoxic concentrations (consistent with direct enzyme inhibition), and over ten pathways relevant for human disease were found, while there was only one for VPA (Fig. 5.5A). We used the KEGG analysis to select pathways that were only regulated at cytotoxic concentrations (but not at lower drug exposure) by MeHg and VPA. The full list of overrepresented KEGG pathways is given in supplementary material (Table S5.4). Two metabolism pathways fulfilled the above condition: “glutathione metabolism” and “nitrogen metabolism”. For closer examination, we focused on glutathione metabolism, as there is ample evidence for its involvement in cell death regulation.(Franco and Cidlowski 2012) A large fraction of all genes involved in core

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity glutathione metabolism was regulated, and none of the genes was regulated in opposite directions by VPA and MeHg. However, there was no easily intuitive pattern of regulation that would indicate whether the overall pathway was up- or downregulated. The data mainly indicate a general disturbance and imbalance of this metabolic network (Fig. 5.6A; for full names of the indicated genes see Fig. S5.3). Besides the metabolism pathways, the KEGG p53 network was regulated under four cytotoxic (4/5 mM VPA; 20/40 µM MeHg) and the highest non- cytotoxic concentration of MeHg. Only two out of 24 toxicant-affected genes of this pathway were regulated into different directions (MDM2 and Cyclin D) by VPA and MeHg (Fig. 5.6B).

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 6: Schematic representation of the two main pathways affected by cytotoxic concentrations of VPA and MeHg (A) Schematic representation of the glutathione metabolism. Metabolites are printed in black; enzymes are indicated in red and positioned at the respective metabolic nodes. The genes that were regulated and belong to this pathway are color coded as described in (A). GCLC and GCLM are subunits of γ-glutamyl-cysteine ligase; GSTx are different isoforms of glutathione-S-transferase. (B) Schematic representation of the p53 pathway. The regulation of the genes belonging to the pathway is color-coded with red for upregulation and blue for downregulation. Light colours indicate the regulation by MeHg and full red / blue the regulation by VPA. Non-color coded genes were not regulated in this study but they are displayed to show the relationship between the genes. The pathway is adopted from the KEGG pathway database.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Identification of consensus cytotoxicity genes

A first approach to arrive at potential candidate biomarkers of cytotoxicity was to select the genes from the KEGG pathways that were overrepresented amongst the toxicant-regulated genes. These were then further selected, based on the extent of their regulation (Fig. 5.5B). An alternative approach to identify marker genes of cytotoxicity is to determine individual genes that are regulated only at cytotoxic concentrations, by different compounds and under various experimental conditions. Data sets that would allow such an analysis are still relatively rare for stem cell based toxicity test systems. Therefore, we used the data generated in the current study, and we combined them with data from our recently published concentration study with VPA (Fig. 5.1D).(Waldmann et al. 2014) As the data sets contained more than one cytotoxic concentration, we combined the differentially expressed PS from all cytotoxic concentrations for one given compound (e.g. 3 mM, 4 mM and 5 mM VPA for the two days (d4 – d6) treatment protocol) and the duplicate PS were removed. The same procedure was followed for the non- cytotoxic concentrations; here we pooled e.g. 0.6 mM and 1 mM for VPA (d4 – d6) or 10 µM and 15 µM MeHg. The resulting genes were defined as the "non-cytotoxic pool" of genes. To identify consensus genes, we determined transcripts that were up- (Fig. 5.7A) or downregulated (Fig. 5.7B) exclusively by cytotoxic concentrations. The overlap between the two VPA treatment scenarios (DoD4 – 6 and DoD0 – 6) resulted in 396 upregulated (Fig. 5.7A) and 405 downregulated (Fig. 5.7B) cytotoxicity-associated genes. The overlap between VPA and MeHg (d4 – d6) included 39 up- and 170 downregulated genes (boxes A and B in Fig. 5.7A,B). Finally, the overlaps of all three conditions revealed 12 up-regulated and 48 down-regulated common genes regulated under cytotoxic conditions. The latter group (Fig. S5.4, overall consensus) may be particularly interesting for developmental toxicity studies, as these genes are regulated, both in the standard UKN1 developmental toxicity test and in the more acute neurotoxicity scenarios after short (48 h) exposure. In order to identify potential biomarkers of toxicity, we further filtered the above pool of genes from the present study (Fig. 5.7C). A first filter step eliminated genes that were not regulated by all cytotoxic concentrations, e.g. a gene that was regulated by 4 mM VPA, but not by 3 mM VPA, was sorted out. The next requirement for a gene to be selected for the list of potential biomarkers was that it had to be regulated by more than 2-fold in at least one cytotoxic condition. From this, we obtained a list of Potential Candi- date Biomarkers (Fig. S5.4). In a subsequent filter step, only genes were retained, if they were more than 4-fold regulated in at least one cytotoxic condition. This final pool of candidate biomarkers comprised 24 genes (Fig. 5.8A). Several of the candidate biomarkers were also identified upon 6-day exposure to VPA as cytotoxicity markers, and they may therefore be used directly within the UKN1 system. For others, more exploratory work with a larger set of chemicals will be required. 63

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 7 : Overlap of genes upregulated by cytotoxic concentrations

The differentially up- (A) and down- (B) regulated PS from the cytotoxic and non-cytotoxic concentrations were pooled and the duplicates were removed from the resulting list of genes. The cytotoxic "only" pool was determined by overlapping with the non-cytotoxic pool followed by a subtraction of the common genes. The number of genes overlapping between various "cytotoxic only " conditions (= cytotoxicity consensus genes) is given in the orange boxes.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

The following concentrations were used to define the cytotoxic pool: VPA (DoD4 – DoD6) 3 mM, 4 mM and 5 mM; VPA (DoD0 – DoD6) 800 µM, 1000 µM; MeHg (DoD4 – DoD6) 20 µM. For the non-cytotoxic pool, the following concentrations were used: VPA (DoD4 – DoD6) 0.6 mM, 1 mM; VPA (DoD0 – DoD6) 350 µM, 450 µM, 550 µM; MeHg (DoD4 – DoD6) 10 µM, 15 µM. (C) Selection scheme for the candidate biomarker list. The algorithm used as input the consensus genes between neurotoxic conditions (Box A and B). Then, two further filters were applied to obtaine a list of potential biomarker candidates: the gene should be regulated at all cytotoxic concentrations, and regulations should be more than 2-fold with at least one of the cytotoxic concentrations. A final list was obtained by retaining only the genes that were regulated at least 4-fold in at least one condition. The results are given in in Fig. 5.8A.

Further, we aimed to investigate the separation power of the 39 candidate biomarkers. There- fore we performed a principal component analysis (PCA) on the basis of these 39 genes (Fig. 8B). The PCA showed a clear separation of the cytotoxic and non-cytotoxic concentration range. This could be the basis for the development of new prediction models, however more substances need to be tested in the future to further validate these candidates. To exemplify the performance of the candidate biomarker set, two genes were selected (on the basis of their selective up-regulation). Their expression levels were determined under 5 non-cytotoxic and 4 cytotoxic experimental conditions. The results show that the conditions could be clearly distinguished on the basis of transcript markers (Fig. 5.S5). We therefore believe that it would be worthwhile to further explore the set of marker genes suggested here, and to validate their performance in multiple experimental situations.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure 5. 8 : Candidate biomarkers for cytotoxicity s determined by the overlap approach. (A) All genes that fulfilled the selection criteria from figure 7C are listed. Up-regulated genes are highlighted in red and down-regulated genes are marked with blue boxes. The consensus groups from which the genes were derived are indicated: “neurotox” means that the gene is regulated by both VPA and MeHg after 48 h exposure (DoD4 – 6); “overall” means that the gene is regulated by all neurotoxic conditions AND by VPA DoD0 – 6. Such genes were printed in bold font. (B) The expression changes of the 39 potential biomarker genes are displayed in form of a two- dimensional principal component analysis (PCA) diagram. Each point represents one experiment (= data from one microarray), and the color coding indicates the concentrations as given in the legend. The form of the data points indicates the compound and treatment duration (circle: VPA DoD4 – DoD6, triangle: MeHg DoD4 – DoD6) used in the experiment. The axis labeling indicates the percentage of the overall data variance covered by the respective axis. Controls (data from untreated cells) have been subtracted from the data displayed, so that the PCA displays "deviation from control". 66

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.5. Discussion

The occurrence of cytotoxicity can substantially influence and distort the outcome of a gene expression study. Therefore, we aimed to identify genes that are exclusively changed by cytotoxic concentrations. Availability of such information could provide alerts for transcriptome data sets, and inform on ongoing cell death in a test system as confounder of the transcriptome data. As a first approach to this issue, we have used in our study two largely different toxicants, each at several non-cytotoxic and cytotoxic concentrations; then the associated transcriptome changes were measured and analyzed. We believe that the MoA of the test compounds chosen represent extremes of the potential spectrum relevant to developmental toxicants: (i) on the one hand, the HDAC inhibitor VPA affects gene expression directly by altering the cells’ epigenetic state. Thus, large transcriptome alterations occur at non-cytotoxic conditions, and eventually cytotoxicity results from an excessive disturbance of the cellular gene regulation; (ii) on the other hand, MeHg has no direct effect on transcription, but rather inhibits a large set of enzymes and other proteins with vital functions. This can directly lead to cell death, and changes of transcription are secondary to the failure of cellular signaling and regulations. Therefore, only few genes are altered under non-cytotoxic conditions by MeHg. These differences of the toxicants were reflected by the induction of different KEGG pathways as well as the overrepresentation of different GO groups amongst regulated genes. Despite these large differences in MoA and ensuing transcriptional changes, we identified a small set of 34 commonly regulated genes changed selectively under cytotoxic conditions. Such transcripts could serve as indicators for the onset of cytotoxicity, and should be further profiled as biomarkers for such a process in follow-up work. In our first pathway-based approach, we elucidated which pathways are overrepresented amongst the genes that are regulated in the presence of cytotoxicity but not under non-cytotoxic conditions. Besides glutathione and nitrogen metabolism pathways, the “p53 pathway” was significantly regulated under four cytotoxic conditions. Alt- hough it was also regulated with the highest non-cytotoxic concentration of MeHg, we still selected this apoptosis-related pathway for more detailed analysis: we found both up- and downregulated genes within the pathway. Data on the effector genes of the pathway suggest that there was an overall up-regulation, as p21, 14-3-3-σ, and reprimo (involved in p53 mediated G1 arrest) as well as the apoptosis genes CASP8 and NOXA were upregulated. However, the situation (based on transcript information) is not entirely unambiguous, as the master regulator of the pathway, p53 itself, was downregulated by VPA, and one of the inhibitors of the pathway, p14ARF, was upregulated. We believe that some of the genes that are consistently regulated by both compounds, and that belong to pathways important for the regulation of cell death, may be considered candidate biomarkers of cytotoxicity for gene expression studies, but p53

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity itself may not be such a well-suited candidate due to its pronounced posttranscriptional regulation(Kruse and Gu 2009) and its involvement in too many different regulations, and counter- regulations. In a second approach we mined the transcriptome response data for cytotoxic concentrations of VPA and MeHg to identify consensus genes. This revealed 24 candidate biomarkers. Eleven of these were also regulated upon 144 h exposure to VPA. Some of these latter genes, such as DLL1, NR2E1 or SOX1, are involved in neurodevelopment.(Islam and Zhang 2015; Kiefer 2007; Teodorczyk and Schmidt 2014) It is unlikely that they will be useful as cytotoxicity marker as these genes may be down regulated rather as a secondary response induced by cellular death, and they may rather indicate a disturbed development. In general, we feel that down-regulated genes are difficult to interpret as cytotoxicity markers. We suggest a preferential use of up-regulated markers, which do not have a direct functional role in neurodevelopment. Three candidates found here are: FLT1, RBM20 or SDK2. FLT1 (fms-like tyrosine kinase 1) encodes a member of the vascular endothelial growth factor receptor (VEGFR) family and plays a role in angiogenesis and cell survival.(Orlandi et al. 2010; Shibuya et al. 1990) RBM20 (RNA binding motiv protein 20) codes for a protein that binds RNA and is involved in splicing.(Maatz et al. 2014) SDK2 (sidekick cell adhesion molecule 2) encodes an adhesion molecule that promotes synaptic connectivity in the retina(Krishnaswamy et al. 2015) and is regulated by p53.(Brynczka et al. 2007) Therefore these genes might be the most promising candidates to indicate cytotoxicity in transcriptomics data sets of the UKN1 test system or other neurodevelopmental assays. Further experimental studies are required to test the usefulness of the genes as biomarkers, and they could then be investigated with faster and cheaper technologies (PCR) in the future. Gene expression data are more and more combined with classical readacross methods for risk assessment and several toxicological databases arose over the last years for this purpose.(Lock et al. 2012; Low et al. 2011; Uehara et al. 2010) One reason for this may be that gene expression changes may be a more sensitive and earlier endpoint because they usually precede the occurrence of apical signs (e.g. histopatho- logical changes, organ toxicity) of adverse effects.(Daston 2008; Jennings et al. 2013) For instance, data from the TG_GATES database (Uehara et al. 2010) indicate that gene expression patterns obtained at very early experimental phases could predict later hepatotoxicity to some extent.(Low et al. 2013; Low et al. 2011) However, toxicogenomics approaches still have many shortcomings and often resulted in poor predictions. One reason for this may be that time and concentration-dependencies need to be taken into account. A simple example for this is that too low concentrations have the risk of false negative results, whereas too high concentrations of a toxicant induce cell death and the interpretation of the results may become hard to impossible. Such effects are particularly important for developmental toxicity test methods.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

In this domain, non-cytotoxic concentrations are usually the most important ones, and triggering of cell death will in many cases (except compounds like methylazomethanol or 5-fluoroura- cil(Kadereit et al. 2012b; Smirnova et al. 2015)) lead to distorted transcriptome patterns. While the objective of studies of organ toxicity (e.g. neurotoxicity) is to find transcriptome changes that correlate with organ toxicity (which can involve death of cells in that organ), the situation is different for developmental toxicity. Here, there is a major interest in finding out what is altered at non-cytotoxic concentrations in order to predict risks of e.g. malformations. The in- tentional or unintentional use of higher concentrations may lead to artifacts and misinterpreta- tions. Therefore there is a need for biomarkers of cytotoxicity for such studies. However, such biomarkers are difficult to obtain from the developmental toxicity test systems itself, as alterations of gene expression due to cytotoxicity are hard to distinguish from changes that are due to altered differentiation. In fact, most of the transcriptome change triggered in the UKN1 test system triggered by toxicants is due to altered development states of the cells,(Balmer and Leist 2014; Rempel et al. 2015) and even when cytotoxic concentrations are used, only minor changes in signaling pathways are observed.(Waldmann et al. 2014) A closer examination of the various differentiation phases of the assay showed that most developmental changes occur during the first 96 h,(Balmer et al. 2014; Balmer et al. 2012) while very little change is observed during the last 48 h. In fact, exposure to the toxicant only during the last 48 h, or from day 4-8 (instead of 0-4) did not change the differentiation track at all.(Balmer et al. 2012) For this reason (direct identification of cytotoxicity genes), cells were exposed here to toxicants during the last 48 h of the 6-day time course of UKN1 differentiation. Sometimes it can be hard or even impossible to derive clear mechanistic information from the differentially expressed genes identified by a transcriptomics approach. For instance, it may be hard to distinguish the regulated transcripts relevant for the MoA of the toxicant from other transcripts regulated because of cellular adaptations, counter-regulation or differentiation. Nevertheless, regulated genes may be helpful as biomarkers of response. This is for instance important in cancer studies, or other clinical approaches, in which healthy tissue is compared to diseased tissue. In such studies, an important aim is to identify biomarkers that can eventually be detected by cheaper and easier methods.(Janvilisri 2015) Such biomarkers are unbiased and independent of any expert knowledge, however often they do not give any mechanistic insight into the toxicants’ MoA or AOP. An approach to avoid such issues would be to identify pathways of toxicity (PoT). Even with elaborate data mining approaches the problem remains that activation of a given a pathway can have protective and pro-apoptotic consequences. For example, the p53 pathway, which we identified in this study to be induced by cytotoxic concentrations, is activated by several upstream events, such as DNA damage or telomere shortening.(Jennings et al. 2013) This serves in the first place as an adaption / repair mechanism by inducing cell cycle arrest

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity via the cyclin-dependent kinase inhibitor 1A (CDKN1A or p21) and the growth arrest and DNA damage inducible gene 45A (GADD45A).(Moskalev et al. 2012) But the p53 pathway activation can also lead to the expression of pro-apoptotic genes, such as PUMA or NOXA, eventually leading to cell death.(Kruse and Gu 2009) The type of p53 response that predominates is determined by post-translational modifications of the protein. For instance, acetylation of p53 tends to be associated with apoptosis, whereas the "repair"-function is activated via phosphorylation of p53.(Kruse and Gu 2009) For this reason, we feel that the mechanistic information provided by individual transcripts is limited. However, this does not mean that it is useless. First, we have shown here that there is a coordinated regulation of transcripts across entire pathways or biological domains, and we have higher confidence in the predictivity of such changes of biological patterns instead of individual markers. Second, there may be a value of individual genes as alerts of an ongoing response, even though the mechanistic link may still be unclear. In summary, we have explored the emergence of such markers, using different approaches, either starting from single genes, or from various biological patterns. A marker set, derived from extensive analysis of broad concentration-response information has been assembled here. It will be interesting to design further studies to test whether transcriptomics alone is sufficient to conclude on a PoT. An alternative approach would be to use multi-omics approaches,(Carreras Puigvert et al. 2013; Krug et al. 2014; Wilmes et al. 2013) in order to distinguish if a potential PoT is activated and relevant. Further experimental studies will also be required to assess the usefulness of potential biomarkers identified here by bioinformatics approaches. On the basis of such studies, it will then be necessary to define quantitative prediction models for the use of such biomarkers in defining cytotoxicity.

5.6. Funding

This work was supported by the BMBF, EU-ToxRisk and the PhD schools RGT1331 and KoRS-CB at the University of Konstanz, Germany.

5.7. Acknowledgments

We thank M. Kapitza and Tamara Rotshteyn for excellent technical support.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Supporting information available

Supporting information is available online and contains the names and regulations of the probe sets, gene ontologies and KEGG pathways that are shown in the diagrams of the figures.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

5.8. Supplementary information

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure S5. 1: Heatmap analysis for VPA and MeHg

Heatmap of the transcriptome data of (A) VPA and (B) MeHg (after exclusion of the concentration 40 µm). The absolute gene expression values (log2-scaling) of the top-100 probesets with the highest variance were color-coded for display. Each row represents one gene and each column one experiment. The color-coding legend at the top of the heatmap indicates the concentrations and the independent experiments. All samples (columns) were clustered (Euclidean distance) on the basis of gene expression values for this set.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure S5. 2: Comparison between the classic and the elim TopGO algorithms

Exemplary illustration of the subgraphs related to the GO term “Regulation of transcription“ (GO:0006355) using two different algorithms. (A) The graph represents the TopGO classic. (B) The graph represents the TopGO elim algorithm. While the ‘elim’ algorithm iteratively removes the genes mapped to significant GO terms from more general (higher level) GO terms, the ‘classic’ algorithm does not consider dependencies between GO terms. The subgraphs were generated by using PS downregulated by 3 mM VPA (p-value < 0.05 after FDR correction, |FC| > 1.8). Each node represents a GO term and each edge a familial link between the two respective GO terms. The colors of the nodes encode the magnitude of the p-values, ranging from dark red (most significant) to white (least significant). The p-values indicate the relative significance of the terms with respect to overrepresentation of their members (A) or specific members (B). (C) Table for the three most significant GO terms by the elim algorithm.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure S5. 3: Differentially expressed genes involved in p53 and glutathione metabolism pathways

The differentially expressed genes that were determined to be involved in the p53 pathway (A) and the glutathione metabolism pathway (B) are listed in the table. The official gene name and the full name of the genes are given.

Results- manuscript 1: Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity

Figure S5. 4: Complete list of potential biomarkers

Potential biomarkers for cytotoxicity are chosen according to the scheme in Figure 5.7C. The numbers represent the fold change compared to untreated control. The source of overlap is given according to Figure 5.7. Genes marked in blue are members of the priority list shown in Fig. 5.7.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6. Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Nadine Dreser1, Katrin Madjar2, Anna-Katharina Holzer1, Marion Kapitza1, Christopher Scholz1, Petra Kranaster1,7, Simon Gutbier,1,8, Stefanie Klima1, David Kolb3,9, Christian Dietz3,9, Timo Trefzer1,#, Johannes Meisig4, Christoph van Thriel5, Margit Henry6, Michael R. Berthold3, Nils Blüthgen4 Agapios Sachinidis6, Jörg Rahnenführer2, Jan G. Hengstler5 Tanja Waldmann1,* and Marcel Leist1,* *shared senior authorship 1 In Vitro Toxicology and Biomedicine, Dept inaugurated by the Doerenkamp-Zbinden Chair foundation, University of Konstanz, 78457 Konstanz, Germany, 2 Department of Statistics, TU Dortmund, 44221 Dortmund, Germany 3 Department of Computer and Information Science, University of Konstanz, 78457 Kon- stanz, Germany 4 Institute of Pathology, Charité-Universitätsmedizin, 10117 Berlin, Germany 5 Leibniz Research Centre for Working Environment and Human Factors (IfADo), Technical University of Dortmund, D-44139 Dortmund, Germany 6 Center of Physiology and Pathophysiology, Institute of Neurophysiology, University of Co- logne (UKK), D-50931 Cologne, Germany 7 Konstanz Research School Chemical Biology (KoRS-CB), University of Konstanz, 78457 Konstanz, Germany 8 current address: Roche Pharma Development, Grenzacherstrasse, 4070 Basel, Switzerland 9 current address: KNIME GmbH, 78467 Konstanz, Germany # current address: Digital Health Center, Berlin Institute of Health (BIH), Charité-Universi- tätsmedizin, 10117 Berlin, Germany

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.1. Summary

The first in vitro tests for developmental toxicity made use of rodent cells. Newer teratology tests, e.g. developed during the ESNATS project, use human cells and measure mechanistic endpoints (such as transcriptome changes). However, the toxicological implications of mechanistic parameters are hard to judge, without functional/morphological endpoints. To address this issue, we developed a new version of the human stem cell-based test STOP-tox(UKN). For this purpose, the capacity of the cells to self-organize to neural rosettes was assessed as functional endpoint: pluripotent stem cells were allowed to differentiate to neuroepithelial cells for six days in the presence or absence of toxicants. Then, both transcriptome changes were measured (standard STOP-tox(UKN)), and cells were allowed to form rosettes. After optimization of staining methods, an imaging algorithm for rosette quantification was implemented and used for an automated rosette formation assay (RoFA). Neural tube toxicants (like valproic acid), which are known to disturb human development at stages when rosette-forming cells are present, were used as positive controls. Established toxicants led to distinctly different tissue organization and differentiation stages. RoFA outcome and transcript changes largely correlated concerning (i) the concentration-dependence, (ii) the time-dependence, and (iii) the set of positive hits identified amongst 24 potential toxicants. Using such comparative data, a prediction model for the RoFA was developed. The comparative analysis was also used to identify gene dysregulations that are particularly predictive for disturbed rosette formation. This ‘RoFA pre- dictor gene set’ may be used for a simplified and less costly setup of the STOP-tox(UKN) assay.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.2. Introduction

The assessment of compounds that potentially trigger developmental neurotoxicity (DNT) is a major challenge in the field of toxicology. The standard test, according to OECD test guideline TG426, has only been applied to few chemicals as DNT testing is not usually mandatory (Fritsche et al. 2018b; Smirnova et al. 2014; van Thriel et al. 2012). For instance in the context of REACH, there is only a requirement for DNT data, if there is a neurotoxic alert from standard guideline studies (Terron and Bennekou 2018). However, more and more evidence arises that chemical and other stressors acting during prenatal periods or early childhood are correlated to later neurodevelopmental abnormalities. These include learning deficits, susceptibility to neurodegenerative diseases and to neuropsychiatric conditions, as well as congenital malformations such as neural tube defects (e.g. spina bifida, exencephaly) (Fritsche et al. 2018b; Grandjean and Landrigan 2006; Grandjean and Landrigan 2014b; London et al. 2012; Smirnova et al. 2014). Current DNT risk assessment is based on animal data only. This strategy makes it hard to measure human-specific health deficits, such as language impairments or effects on complex cognitive and social behavior. Even more importantly, it is associated with problems of species extrapolation.The use of batteries of human cell-based new approach methods (NAM) has been suggested as alternative testing approach for different regulatory areas concerned with pesticides, drugs and industrial chemicals (Bal-Price and Fritsche 2018; Fritsche et al. 2017; Harrill et al. 2018; Terron and Bennekou 2018). The concept behind these NAM is that brain development requires the activation of so-called key neurodevelopmental processes (KNDP) (Aschner et al. 2017; Bal-Price et al. 2015), which can be studied one by one with appropriate in vitro systems. These KDNPs comprise the generation and migration of neural stem / precursor cells, the differentiation of neuroepithelial cells into specific cell types, and neuronal network building (Aschner et al. 2017; Bal-Price et al. 2018a; van Thriel et al. 2012). It is thought that DNT can occur when a compound interferes with at least one of these processes. On this basis, test methods have been established that model such KNDP (Fritsche et al. 2018a; Harrill et al. 2018; Schmidt et al. 2017). They detect toxicant effects on e.g. cell migration (Nyffeler et al. 2017a; Nyffeler et al. 2017b), neurite growth (Krug et al. 2013a; Radio et al. 2008; Stiegler et al. 2011) or formation of electrical networks (Frank et al. 2017; Frank et al. 2018). An added value of the use of the NAM strategy is that a better mechanistic information about the tested chemicals’ actions can be obtained, and the concept of adverse outcome pathways (AOP) (Leist et al. 2017) can be used on the basis of such data (Baker et al.

2018). STOP-tox(UKN) (also referred to as UKN1) is a recently developed (Krug et al. 2013a) NAM that models the very early steps of neurodevelopment, such as neural induction and differentiation. The test method is based on the differentiation of pluripotent stem cells (PSC)

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances into neuroepithelial precursor (NEP) cells that correspond roughly to the cells that build up the neural plate / neural groove (Balmer et al. 2014; Balmer et al. 2012). STOP-tox(UKN) was developed to follow the KNDP “early neural differentiation”, and gene expression analysis was used as the primary endpoint (Balmer et al. 2014; Krug et al. 2013a; Shinde et al. 2015; Waldmann et al. 2017; Waldmann et al. 2014). Transcriptomics has proven to be a powerful tool for assessing disturbances of differentiation, also on a fully quantitative level (Shinde et al. 2016a; Waldmann et al. 2017; Waldmann et al. 2014). Extensive characterization has been performed on how to use transcriptomics data in toxicological hit definition (Dreser et al. 2015; Pallocca et al. 2016; Shinde et al. 2016b; Weng et al. 2012). However, chemicals may also trigger transcriptome changes that are not related to altered differentiation patterns (Balmer et al. 2014; Grinberg et al. 2014; Waldmann et al. 2017). To conclude on the toxicological significance of a disturbed gene expression pattern, there is a need for a functional phenotypic anchoring of the gene expression changes to an unambiguously adverse outcome in the culture dish. This can be used to answer the question whether the gene expression changes have a (toxicologically-relevant) functional consequence during subsequent development or in later life. Such a “phenotypic anchor” could be a structural feature or a specific protein expression that can be measured in the cell culture dish. For instance, neurite outgrowth measures of peripheral (Hoelting et al. 2016) and central neurons (Krug et al. 2013a) have been used for toxicant screening purposes (Delp et al. 2018). Alternatively, quantification of oligodendrocyte surface markers (Baumann et al. 2014; Baumann et al. 2016) or synaptic marker proteins (Mundy et al. 2008) have also been used for toxicological approaches to connect gene expression changes to an adverse outcome. Another strategy for a phenotypic anchoring focuses on cellular functions, rather than on static features. A prime example is the mouse embryonic stem cell test (mEST) that measures the beating capacity of cardiomyocytes in vitro (Seiler and Spielmann 2011). Another functional endpoint for phenotypic anchoring is represented by the electrical network properties of mature neurons (Frank et al. 2017; Frank et al. 2018). One functional feature particularly useful for DNT studies could be self-organization of cells to form tissues and organoids (Beccari et al. 2018; Lancaster et al. 2017). The self-organization process relevant for neural tube formation is reflected in cell culture dishes by neural rosette formation (Chambers et al. 2009; Colleoni et al. 2011; Conti and Cattaneo 2010). Neural rosettes are nestin-positive structures formed by neuroepithelial cells in a way that ZO1 and cadherin- N are expressed only at the inside, while proliferation only occurs on the outside. They are characterized by high expression of rosette signature genes, such as PLAGL1, DACH1, and PLZF (Elkabetz et al. 2008). We studied here, whether rosette formation could be used as a robust and quantifiable endpoint for the early neurodevelopmental toxicity test STOP-tox(UKN). We set out to develop improved staining and quantification procedures. The main questions

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances then were, whether rosette formation correlates with (i) the transcriptome changes for different toxicant concentrations, (ii) with incubation times and (iii) with effects of diverse toxicants. Us- ing this information, statistical classifiers were developed to identify DNT compounds. Moreo- ver, transcript changes that were particularly predictive for adverse outcomes, were identified.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.3. Materials and Methods

Materials

Gelatine, putrescine, sodium selenite, progesterone, apotransferrin, glucose, insulin and ascorbic acid were obtained from Sigma (Steinheim, Germany). Accutase was from PAA (Pasching, Austria). FGF-2 (basic fibroblast growth factor), FGF-8b, sonic hedgehog and noggin were obtained from R&D Systems (Minneapolis, MN, USA). Y-27632, SB-43154 and dorsomorphine dihydrochloride were from Tocris Bioscience (Bristol, UK). MatrigelTM was from BD Biosciences (Massachusetts, USA). All cell culture reagents were from Gibco/Invitrogen (Darmstadt, Germany) unless otherwise specified. The compounds used for treatment including the suppliers are listed in the supplement (Fig. S6.1).

Neuroepithelial differentiation and rosette formation

The differentiation protocol was followed as described before (Chambers et al. 2009), with minor changes (noggin concentration was only 35 ng/ml, dorsomorphin was added additionally), as operated in (Balmer et al. 2012; Chambers et al. 2011). Single cells (hESC) were seeded on matrigel-coated plates (in a density of 18000 cells/cm²) and cultured until they reached 75% confluency. Cells were supplied with fresh medium every day (previously conditioned for 24 h on mitomycin C-inactivated mouse embryonic fibroblasts; supplemented with 10 ng/ml FGF2 and 10 µM ROCK inhibitor Y-27632). On DoD0, differentiation was initiated by adding knockout serum replacement medium (KSR) (Knockout DMEM with 15% knockout serum replacement, 2 mM Glutamax, 0.1 mM MEM non-essential amino acids and 50 μM beta- mercaptoethanol) containing 35 ng/ml noggin, 600 nM dorsomorphin and 10 µM SB-431642. On DoD4, KSR medium was started to be replaced gradually by N2S medium (DMEM/F12 medium, 1% Glutamax, 0.1 mg/ml apotransferrin, 1.55 mg/ml glucose, 25 μg/ml insulin, 100 μM putrescine, 30 nM selenium and 20 nM progesterone), also supplemented with equal amounts of noggin, dorsomorphin and SB-431642 as used for KSR medium. On DoD11 cells were detached by incubation with accutase for 20 minutes. Cells were washed with 10 ml DMEM/F12, counted and reseeded in a density of 150000 cells (cm²) in N2S supplemented with 20 ng/ml FGF2, 100 ng/ml FGF8, 20 ng/ml sonic hedgehog, 20 µM ascorbic acid and 10 µM ROCK inhibitor Y-27632. For rosette formation assay matrigel coated 96 well plates, for staining 24 well plates and for mRNA 12 well plates were used. On DoD13 medium was changed to supplemented N2S without ROCK inhibitor. On DoD15 rosettes formed and were analyzed.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Experimental exposure

Standard treatment with toxicants was either performed in differentiating cells from DoD0- DoD6 or as indicated in the respective figures (Fig. 6.7; time window treatment). The compound list including the concentrations is shown in the supplemental material (Fig. S6.1). Cells were either harvested on DoD6 for microarray/RT-qPCR analysis or further differentiated until DoD15. In order to determine cytotoxicity, a resazurin assay was performed on DoD6, exactly as described previously (Krug et al. 2013b; Stiegler et al. 2011) and the EC10 concentration was determined for the test compounds. Usually, the EC10 concentration was used for the analysis of mRNA transcript levels at DoD6 and the rosettes formation at DoD15.

Immunostaining

For immunostaining, cells were either grown on plastic (96 well plates) or on cover slips. At DoD6 or on DoD15, cells were ﬁxed in 4% paraformaldehyde and permeabilized in 0.3% Triton X-100 in PBS. After blocking in PBS (containing 5% bovine serum albumin and 0.1% Triton) for 1 h, primary antibodies were added (antibody concentration is indicated in Fig. S6.2) and incubated for 1 h at room temperature (RT). After a washing step, secondary antibodies were applied for 45 min at RT. DNA was stained with Hoechst H-33342, and cover slips were mounted in FluorSaveTM reagent (Calbiochem, Merck).

Visualizing sialic acids using metabolic glycoengineering

In order to visualize cellular membrane and cellular arrangement into rosette structures, we took advantage of the metabolic glycoengineering method, where cells metabolize a mannosamine variant containing an azide group and incorporate it into their membrane glycolipids/gly- coproteins as sialic acids (Campbell et al. 2007). The azide-modified sialic acids can then be visualized by coupling the azide group to biotin, and its subsequent fluorescent detection by streptavidin-conjugated fluorescent dye (Sletten and Bertozzi 2009; Spate et al. 2014). Cells were differentiated as described. On DoD11 cells were seeded in 300 μl medium/well in matrigel-coated 8-well glass bottom μ-slides (iBidi, Munich, Germany). On DoD13, medium was changed to N2S medium containing FGF2, FGF8, Shh and ascorbic acid. On DoD14, medium was replaced for N2S containing 70 μM of tetraacetylated mannosamine with an azide group

(Ac4ManNAz) (Jena, Germany). 24 h later, on DoD15, the sugar precursor was washed off and sialic acids on the cellular membrane were stained by treatment with 100 μM dibenzocy- clooctyne (DBCO)-biotin (Jena Bioscience, Germany) for 20 min, followed by 15 min incubation with 8 μg/ml streptavidin-AlexaFluor488 conjugate (Life Technologies, US) and 1 μg/ml

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Hoechst-33342. Afterwards, cells were fixed with 2% PFA containing 4% sucrose for additional 10 min and wash twice with PBS. Imaging was performed using a laser scanning confocal microscope LSM 880 (Zeiss, Germany) equipped with gallium arsenide phosphide (GaAsP) detector, using a 40x/NA 1.40 oil objective. For image processing, the Fiji program was used (https://imagej.net/Fiji).

Quantification of rosette formation and viability assays

Cells were differentiated to neural rosette forming cells until DoD15 in 96 well plates. Im- munostaining of tight junction protein zona occludens 1 (ZO1), golgi matrix protein 130 (GM130) as well as nuclear staining with Hoechst H-33342 was performed as described above. For each treatment, at least three replicates and for untreated control, six replicates were stained and about 50 images of the whole well were taken by Cellomics automated microscope (Thermo Fisher Scientific) for each channel. Rosettes per well and the nuclei area per well was determined by Konstanz information miner software (KNIME) (Berthold et al. 2007). Data was displayed as rosettes/nuclei area relative to the untreated control. All rosette data shown in the paper were obtained at non-cytotoxic concentration (no change (of > 10%) in number of cells per well). Cytotoxicity was also controlled at the pre-rosette stage by measuring the viability of the neuroepithelial precursor cells on DoD6 by a resazurin assay (Balmer et al. 2014; Balmer et al. 2012; Krug et al. 2013b; Rempel et al. 2015). The cytotoxicity was determined by benchmark concentration fitting (Delp et al. 2018; Krebs et al. 2018; Krug et al. 2013a) and the BMC10 (concentration leading to a viability drop of 10% on a modelled curve) was use as highest non-cytotoxic concentration.

KNIME analysis

To remove noise and enhance the signal, background preprocessing was applied to all images. The total nuclear area, which is used for normalization, was determined: Hoechst H- 33342 staining was used to create a segmentation of the image using the Otsu Global Thresh- olding Algorithm (Otsu 1979). The number of segmented pixels was counted. Since tight junctions cluster in the rosette center as a ring-like structure, ZO1 staining was used to locate the rosettes in a first step. Images that did not contain rosettes were filtered out. ZO1 spots were detected by applying Otsu thresholding and segmentation of the image. The software discrim- inated between rosettes and other structures (e.g. epithel, neurons) by using the geometry of detected ZO1 segments as a filter. Since the cells and the organelles of a rosette orientate towards the center, the “halo like” shape of a Golgi staining (GM130) was used to identify rosettes: Voronoi Segmentation was applied to the GM130 channel by using the previously 84

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances discovered (ZO1) locations as seed points. This led to a segmentation of positive rosettes. The result was improved by applying a machine learning algorithm which allowed to filter out wrong segments resulting from noise or artefacts. The classification was done using a random forest (Breiman 2001), which was learned on a training set by manual annotation. The features of the training set comprise measures of segment geometry. After classification, the positive segments were used to calculate the results. The number of rosettes (segments) per image was counted. For each well, the number of rosettes/image was summed up and normalized to the total nuclear area. Data is always displayed relative to an untreated control.

Affymetrix DNA microarray analysis and quantitative PCR

After the measurement of viability (resazurin assay), medium was removed and cells were lysed in TriFastTM (Peqlab, Germany). mRNA was isolated as described in the manufacturers protocol and reverse transcribed (iScript, Biorad). Quantitative PCR (qPCR) was performed using EVAGreen SsoFastTM mix on a BioRad Light Cycler (Biorad, Germany). For quantiﬁcation, qPCR threshold cycles were normalized to reference genes [tatabox binding protein (TBP) and ribosomal protein L13 (RPL13A)]. If not stated otherwise, the data of cells treated with compounds were then expressed relative to transcript levels of untreated control cells, which have been grown and differentiated for the same amount of time. For this normalization, the 2^(-delta delta C(T)) method was used (Livak and Schmittgen 2001). The primer sequences are listed in the supplement (Fig. S6.3). Affymetrix chip-based DNA microarray analysis (Human Genome U133 plus 2.0 arrays) was performed exactly as described earlier (Krug et al. 2013b; Rempel et al. 2015). The original data sets of HDAC inhibitors and mercu- rials were obtained in the context of an earlier study. These data and different forms of analyses have been published in (Rempel et al. 2015).

Preprocessing of gene expression data

Three datasets (Balmer et al. 2014; Krug et al. 2013b; Rempel et al. 2015; Shinde et al. 2016a; Waldmann et al. 2017; Waldmann et al. 2014) with RMA normalized gene expression (Harbron et al. 2007) data (log2 scale) measured on the Affymetrix HG-U133 Plus 2.0 array were combined to one gene expression dataset including 54675 different probe sets. For each biological replicate, the control value was subtracted from the corresponding treatment values (when multiple technical controls were available for one biological control replicate, the expression values of the technical controls were averaged to one mean control value per probe set). For the comparison with rosette data, one mean expression value was calculated for each compound and each probe set. Correlation analysis was performed using Pearson’s correlation 85

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances coefficient and the corresponding correlation test for the null hypothesis of zero correlation. The resulting p-values were multiplicity-adjusted to control for the false discovery rate (FDR) by the Benjamini–Hochberg (BH) procedure (Benjamini and Hochberg 1995).

Preprocessing of rosette formation data for the prediction models

For each biological replicate one control value was defined (in case of multiple technical controls per biological control replicate, the rosette formation values of the technical controls were averaged to one mean control value). Next, for each biological replicate separately, the rosette formation values of the treatments were divided by the corresponding (mean) control value. Finally, for the comparison with gene expression data, one mean rosette formation value was calculated for each compound. Mean rosette inhibition values were defined by 1-mean rosette formation values (resulting negative values set to zero, so that the rosette inhibition ranges from 0 to 1) and used as response in the prediction models.

Prediction models of rosette inhibition and associated gene sets

For 24 compounds (belinostat, BIO, BPA, CHIR, CsA/FK506, DMSO, entinostat, estradiol, galnon, geldanamycin, gleevec, HgBr2, HgCl2, IFNbeta, LiCl, MeHg, panobinostat, PCMB, PMA, RA, SAHA, thimerosal, TSA, VPA) rosette formation as well as gene expression values were measured. They were used for correlation analyses and prediction models. Different gene sets were combined with different statistical algorithms to build and validate classification or regression models for the prediction of rosette inhibition based on leave-one-out cross-validation. In the case of regression models, the rosette inhibition was considered as continuous response using the original values. In the case of classification models, the rosette inhibition values were dichotomized at a threshold value of 75% (binary response: "0" = less than 75% inhibition, "1" = at least 75% inhibition). Leave-one-out cross-validation was carried out based on 24 compounds. In each cross-validation step one compound was used as test set to predict the response and the remaining 23 compounds were used as training set to build the model. This procedure was repeated until each compound was considered once for validation. As a result of the leave-one-out cross-validation, we obtained a predicted value of rosette inhibition for each of the 24 compounds that we compared with the true rosette inhibition values to evaluate the prediction performance of the respective model. Prediction accuracy was assessed using Pearson’s correlation coefficient for continuous response, and the area under the curve (AUC) of the sensitivity-specificity plot (ROC-curve) for binary response. Based on the training set in each cross-validation step, two different gene sets were selected: (i) the top-x probe sets with the largest variance in expression values (“top-variance genes”), and (ii) the top-x probe sets 86

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances with the best univariate (positive or negative) correlation between expression values and response. For binary response the AUC was considered as correlation measure (“top-AUC genes”) and for continuous response the Pearson’s correlation (“top-corr genes”). The number of top-x genes was varied with x=100, 200, 1000. In addition, all 54675 probe sets and the 21 neurodevelopmental distance (NDD) measure genes (Supplement Fig. S6.7a) (i.e. the 46 different probe sets corresponding to the 21 genes) were used to construct prediction models. In total, we compared eight different gene sets in combination with the following three statistical algorithms: regularized logistic regression with lasso penalty, regularized logistic regression with ridge penalty, and random forest. Random forest with top-200-variance genes and continuous response was chosen as final prediction model, and consequently fitted and evaluated once based on all 24 compounds. For this final model the top-200-variance genes were determined once based on all 24 compounds (Supplement Fig. S6.5). The R package ‘ranger’ (version 0.10.1) was used for computing the random forest with the following default parameter settings: The random forest is an ensemble of 500 regression trees and its prediction is an average of the predicted response values returned by each tree. An individual tree is constructed based on bootstrap samples of the observations and by recursive binary splitting of the covariate space. At each node the best splitting variable (gene) out of a randomly chosen subset of 14 genes and the best split cut point are determined. The estimated variance of the responses is used as splitting rule. A minimum node size of five observations is chosen.

Calculation of threshold and borderline range for the RoFA prediction model

The threshold T for the definition of positive, rosette inhibiting compounds was set as T= Mn-

(2x SDn) with Mn: mean of negatives and SDn: standard deviation of negatives (Fig. 6.8d). The borderline range (BR) was defined as BR= T ± U(T) (U(T): uncertainty of T) and U(T)= SDpooled (Fig. 6.8e) as defined by (Leontaridou et al. 2019; Leontaridou et al. 2017).

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.4. Results

Characterization of NEP and their differentiation towards neural rosettes

Two problems are still preventing the application of the standard STOP-tox(UKN) test method for predictive toxicology. First, establishment of a general prediction model based on transcriptome changes has proven difficult; second, the functional relevance of gene expression changes alone is hard to prove. Anchoring genotypic changes induced by a toxicant to a distinct phenotype may provide a solution for these issues. Thus, the test protocol, normally end- ing with day of differentiation (DoD) 6 (Fig. 6.1a) (Krug et al. 2013b) was extended by a second phase in which cells were allowed to self-organize to form rosettes (Fig. 6.1a). Transcriptome data from the standard assay highlight the difficulty with building a prediction model (Fig. 6.1b): In such a data set, one can observe strong changes in gene expression over time, already in untreated cells. A k-means clustering (K=9) of the gene expression time courses (Fig. 6.1c) showed, that the expression changes did not follow a single monotonous pattern such as up- regulation/ down-regulation over time (cluster 1 and 2). Some clusters rather showed a wave- like pattern (e.g. cluster 4 and 5) (Fig. 6.1c). This makes it difficult to define marker genes that are reliable for every time point during differentiation as well as for each toxicant applied to the system. In the past, the genes PAX6 and OTX2 (cluster 2a) as well as POU5F1 and NANOG (cluster 1a) have been used as test endpoints (Fig. 6.1d). Wave-like regulated genes, have not been considered at all for the prediction model, and it is impossible to judge the relative importance of all regulation clusters for proper development. Therefore, we sought a more integrated bifunctional endpoint. The self-organisation capacity of genuine neuroepithelial cells, as opposed to other cell types (Chambers et al. 2009) may provide such an endpoint (Colleoni et al. 2011). The self-organization process relevant for neural tube formation is reflected in cell culture dishes by neural rosette formation. To study rosette formation, cell culture was continued, cells were replated and allowed to form rosettes for 4 days (Fig. 6.1a). In a first step, we analysed the expression of relevant neural rosettes marker genes. DACH1, LIX1, LMO3, MSX1 were upregulated on gene expression level on DoD11 and DoD15 (Fig. 6.1e). A preliminary morphological analysis showed that the cells on DoD6 expressed the neuroepithelial marker PAX6 (Fig. 6.1f) and the general neural stem cell marker nestin (data not shown) while they were mostly negative for the neural crest marker TFAP2. The expression of NESTIN and PAX6 was maintained until DoD15 (with continued absence of TFAP2 (not shown)). Mor- phology changes from unorganized clumps on DoD6 to organized rosettes on DoD15 (Fig. 6.1g).

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 1: Extension of the STOP-tox(UKN) protocol to generate rosette stage cells

(a) Exposure and differentiation scheme: Pluripotent stem cells were seeded at a low density three days before the start of differentiation (DoD-3). On the first day of differentiation (DoD0), the neural fate was induced; on DoD6, the standard STOP-tox(UKN) protocol endpoint (gene expression) was measured. For rosette formation, cells were further differentiated, with a re-plating step at DoD11. Neural rosettes were assessed at DoD15 for the rosette phenotype (new endpoint) or gene expression. Toxicant exposure occurred always between DoD0 and DoD6. (b) RNA was harvested from cells differentiated for 6 h, 4 d or 6 d, and gene expression was quantified by microarray analysis. The heat map (one lane per microarray) shows absolute expression values (log2-scaled) of the 1000 genes with the largest variance, after hierarchical 2D clustering. (c) Waves of gene regulation: The time course profiles of gene expression were clustered in nine groups (k-means with k=9), and average expression levels of the genes per groups are indicated for the four time points as row-wise z-scores. For instance genes in cluster-5 are first up- regulated, then downregulated; genes in cluster-1b start high and are then continuously downregulated. (d) Exam- ples of four signature genes: the gene expression levels are displayed relative to the pluripotent state (DoD0). Data are averages from four independent experiments. (e) The cells were cultured as in A, and mRNA was prepared on DoD11 (brown) and DoD15 (blue) for PCR analysis of four rosette marker genes. Their relative regulation (fold change (FC) vs DoD0) is displayed for three independent experiments; black horizontal lines indicate the average FC. (f) DoD6 cells were stained for PAX6 (neuroepithelial marker) and TFAP2 (neural crest marker). (g) The cell cultures were stained on DoD15 for the neural stem cell markers PAX6 (nuclear localization) and nestin (cytoskeletal protein), and representative images are displayed. Two typical rosettes are shown within white circles. 89

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Characterization of neural rosettes

Neural rosettes have been described as the in vitro counterpart to the neural tube, as they look like cross sections of the in vivo developing neural tube (Dhara et al. 2008; Dhara and Stice 2008; Elkabetz et al. 2008; Zhang et al. 2010). To find an appropriate imaging technology to exploit the rosette formation as an assay endpoint, DoD15 neural rosettes were stained in different ways. ZO1 (tight junction marker) staining indicated that rosettes possessed a core structure that looked like a net of cell membranes possibly with an open (cell-free) space in its center (Fig. 6.2a, 2b). Rosettes staining for endothelin beta indicated that this marker is distributed over the whole membrane of the rosette forming cells (Fig. 6.2b) (Elkabetz et al. 2008). Metabolic glycoengineering cell feeding with labelled N-acetylmannosamine was used to label surface sialic acids as typically found on poly-sialyted neural cell adhesion molecule (NCAM), a known neural stem cell marker. This revealed a cellular polarisation, with the sialic acid res- idues rather located in the center of the rosettes (Fig. 6.2c). Also, all organelles showed a clear orientation within rosette cells. The Golgi apparatus was always positioned towards the middle, while cell nuclei pointed towards the outside (Fig. 6.2d). DoD15 rosettes had a diameter of approximately 50-100 µm and they were equally distributed over the plastic or glass surface of the cell culture dish. This allowed to count the formed rosette structures (Fig. 6.2e). The Z- dimension (3-dimensionality) of the rosette was very small and they resembled flat domes with the central core appearing to be covered by a single cell layer (Fig. 6.2b, 2f).

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 2: Immunocytochemical characterization of rosettes

Cells were differentiated for 15 days according to the scheme in Fig. 6.1a. Then they were fixed and stained. Nuclei were visualized with the H-33342 dye. (a) Rosette stage cultures were stained for ZO1 (tight junction protein) and nestin (neural stem cell marker). (b) 3D-image of rosettes stained for endothelin B and ZO1. On the right, two exemplary z-planes of the rosette are depicted with focus close to the cell culture dish (low) or rather in the middle of the rosette height (center). Confocal images were obtained from z-planes spaced 0.3 µm apart. From these data, a 3D-display was obtained. (c) Metabolic glycoengineering was used to label surface sialic acids on rosette cells. At 8 h before fixation, cells were fed the sialic acid precursor N-acetylmannosamine with a tag for immunocytochemical labelling. At 30 min before fixation, surface sialic acids were live-cell labelled for the sugar tag (Sia). Then cells were fixed and additionally stained for ZO1 and DNA (H-33342). (d-f) Rosette stage cells were stained for ZO1 and GM130 (cis-Golgi marker). (e) Individual antigens of rosettes across a cell culture well are shown at a low magnification view. (f) Images were recorded for 20 z-planes (0.35 µm spacing) and rendered for a 3D-display. Three selected z-planes are shown. Scale bars are included in all images for size calibration. 91

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Disturbance of rosette formation by toxicant treatment

If rosette differentiation is to be considered as endpoint of a test method, then DNT toxicants should affect this process. We reason that exposure to chemicals may lead to a change of the neuroepithelial differentiation. This may lead to different outcomes, such as (i) a decreased number of rosettes, (ii) an accelerated differentiation towards neurons, (iii) a conversion of NEP to any other neural cell type, (iv) or even a switch of differentiation track towards other cell lineages (Fig. 6.3a). The HDAC inhibitors trichostatin A (TSA) and valproic acid (VPA) have been used as tool compounds in the STOP-tox(UKN) test system, and they led reproducibly to extensive changes in the gene expression of developing cells (Balmer et al. 2014; Krug et al. 2013b; Shinde et al. 2016a; Waldmann et al. 2014). Therefore, we examined the effect of these two compounds on rosette formation. Differentiating cells were treated with 400 µM and 600 µM VPA and 10 nM TSA from DoD0 to DoD6. On DoD6, the compounds were removed and the cells were further differentiated until DoD15 (rosette stage). Low concentrations of VPA (400 µM) led to a disturbance of rosette formation, and the appearance of few neurons (beta III tubulin (TUJ1) positive). Treatment with higher concentrations of VPA (600 µM) or with TSA (10 nM) completely abolished any rosette formation. The cells formed an epithelial-like structure (Fig. 6.3b) and patches of neuron-like cells were detected (Fig. 6.3b). Further immunostaining experiments showed that reduced formation of rosettes was associated with a variety of alternative differentiations. For instance, treatment with TSA did not affect the neural stem cell marker nestin, but it abolished the expression of the neuroepithelial marker PAX6 almost completely. It also led to a stable expression of the pluripotency marker POU5F1 in some cells, while normally all rosette stage cells are negative for POU5F1. Furthermore, SOX10 (precursors of oligodendrocytes and neural crest cells) was up-regulated and peripherin-positive cells (peripheral neurons) emerged (Fig. 6.3c). This is consistent with findings that VPA/TSA lead to a change of the differentiation track from NEP to a more neural crest- (like) cell fate in chicken (Murko et al. 2013). Therefore, we hypothesized that one of the developmental disturbances that can be triggered in our test is a switch from CNS neuroepithelial precursors to neural crest cells and their derivatives, such as peripheral neurons (Fig. 6.3c). In summary, the reduction of rosette formation appeared to be the most universal feature of the developmental toxicants tested; in contrast to this, the generation of alternative cell types was highly dependent on concentration and type of toxicant, and it also appeared to depend more strongly on the time of endpoint assessment. Therefore, we decided to use overall rosette formation as endpoint.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 3: Qualitative features of disturbed rosette formation.

(a) Overview of potential toxicant effects during early differentiation: Under normal control conditions cells differentiate towards a neural fate and form rosettes on DoD15. Treatment with teratogens during differentiation may lead to an altered phenotype on DoD15: (i) reduced numbers of rosettes, (ii) emergence of differentiated neurons or of epithelial-like structures, or (iii) generation of other, non-neural cell populations. (b) Cells were differentiated for 15 days according to the scheme in Fig. 6.1A. Treatment (400 µM and 600 µM VPA; 10 nM TSA) was performed from DoD0-DoD6. Cultures were fixed and stained for ZO1 and βIII-tubulin (TUJ1 antibody, neuronal marker). Nuclei were visualized with the H-33342 dye. (c) Characterization of cells after TSA treatment: Cells were fixed on DoD15 and stained for nestin, PAX6, POU5F1, peripherin (Peri) and SOX10. Nuclei were visualized with the H-33342 dye. Exemplary images of untreated and TSA-treated cells are depicted.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 4: Quantification of rosette formation.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Neural rosettes were fixed, stained and imaged with an automated microscope. Then a KNIME protocol was used for automated image recognition and quantification. (a) Working hypothesis for the rosette quantification: It is assumed that treatment of cells from DoD0-6 with neurodevelopmental toxicants leads to a concentration-dependent reduction in rosette number, and it may also lead to the presence of other cell types. (b) Control rosettes were stained as in Fig. 6.2D for ZO1 (red) and GM130 (green), and nuclei (blue). This exemplary micrograph is used to illustrate the quantification algorithm. (c) Quantification algorithm: The three imaging channels underlying the image in b are shown separately. A KNIME protocol is used to count the pixels of the nuclei and from this the nuclear area is determined. ZO1-spots are recognized by their size and roundness. ZO1-spots that are surrounded by GM130 coronas (=pre-rosettes) are classified as rosettes. These “true rosettes” are counted and normalized to the nuclear area. (d) Cells were treated with TSA (10 nM) or solvent from DoD0-6. They were fixed and stained on DoD15 and used for rosette quantification as in c. The graph on the left shows the rosette counts, the graph in the middle shows the rosettes/nuclei ratio and the graph on the right shows the nuclear area. Each dot represents data from one well (technical replicate). Data are from three independent experiments with altogether n ≥ 25 technical replicates.

Quantification of neural rosettes and establishment of the rosette formation assay (RoFA)

We hypothesized that a neurodevelopmental toxicant inhibits rosette formation in a concentration dependent manner, so that a fully quantitative method to evaluate this endpoint was de- sirable (Fig. 6.4a). Therefore, a counting algorithm was developed, based on the production of images from the immunostained rosettes. We decided on labelling the inner tight junction marker ZO1 in combination with staining of the Golgi apparatus by antibodies against the Golgi matrix protein GM130, as this approach yielded the clearest representation of rosettes (Fig. 6.4b). The program we developed uses the combined information from the ZO1 spot in the middle surrounded by a halo of GM130. This was combined with a filter for size and shape (roundness) to identify rosettes. In the next step, the number of rosettes is normalized to the number of cells in an image field (based on nuclear counts) (Fig. 6.4c). Testing this algorithm, we found that the number of rosettes that were counted, was stable for biological replicates as well as for technical replicates. The tool compound TSA (10 nM) inhibited the rosette formation, but did not affect the cell count (nuclear area) (Fig. 6.4d). After the establishment of the rosette formation assay (RoFA) and the development of the quantification software, we started comparing the RoFA performance to the classical gene expression readout of the STOP-tox(UKN) assay. As first approach, we quantified the concentration dependencies for VPA and TSA. For gene expression analysis, cells were harvested on DoD6, and the gene expression changes of marker genes were investigated. MSX1, POU5F1 and NANOG were up-regulated after TSA/VPA treatment, while EMX2, OTX2, PAX6 were down regulated. Gene expression changes required TSA concentrations ≥ 2-8 nM and VPA concentrations of ≥100-300 µM (Fig. 6.5a). Rosette formation was inhibited in a similar

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances concentration range. The BMC25 [benchmark concentration leading to a 25% decrease compared to control (Krebs et al. 2018)] for TSA was ≈ 4 nM and for VPA ≈ 170 µM (Fig. 6.5b). Rosette formation was inhibited by ˃ 50% at 3 nM TSA and 220 µM VPA. In summary, this first RoFA evaluation experiment showed a sensitivity of the new endpoint in the same range as the hitherto used gene expression. Moreover, an important advantage became obvious. Easily interpretable, quantitative inhibition data obtained for the new endpoint. This was not the case for the gene expression data (the genes all behaved slightly different). For gene expression endpoints, a prediction model, taking into account the different weight of multiple expression markers (known to be affected differently by various classes of toxicants (Balmer et al. 2012; Rempel et al. 2015; Waldmann et al. 2014)), would be very difficult and resource- intensive to establish. Notably, such an approach has been used for genome-wide expression data (however, without giving weights to individual genes) (Shinde et al. 2016a). Thus, gene expression endpoints and RoFA can be used in a complementary fashion. Where cost plays a major role, the RoFA may be preferable, at least as an initial filter.

Figure 6. 5: Concentration-dependent changes in gene expression and rosette formation. Cells were differentiated towards neural rosettes as in Fig. 6.1a. During differentiation, cells were exposed to VPA or TSA from DoD0-6. (a) On DoD6, mRNA was isolated and gene expression of marker genes was assessed by RT-qPCR analysis. The relative expression of the marker genes PAX6, OTX2, NANOG, POU5F1, EMX2 and MSX1 (normalized to untreated controls) is shown for TSA- (left) and VPA- treated (right) cells. The cell viability was not affected at the concentrations used (data not shown).

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

(b) Cells (as in A) were continued to be cultured without toxicants until DoD15. Then, rosette formation was quantified as in Fig. 6.4B/c. Rosette count data were normalized to untreated controls. For all experiments, the means from 2-6 independent experiments are shown. For the rosette formation assay, eight out of the 15 concentrations were only included in a single experiment. The error bars ±indicate the standard error of the means.

Correlation of rosette formation with sensitive time windows

Neuroepithelial development is particularly sensitive to toxicants during certain time periods, and not at all sensitive during others (Balmer et al. 2014; Balmer et al. 2012). We used this fact to further investigate the correlation of rosette formation and gene expression changes. Differentiating cells were treated with TSA (10 nM) during different time windows (e.g. DoD2- DoD3), and rosette formation was assessed for all conditions on DoD15. Cells treated with TSA from DoD0 to DoD6 (positive control) showed no rosette formation, while the short treatment during the first three hours of differentiation did not result in any inhibition of rosette formation (negative control). Six hours of treatment was sufficient to inhibit rosette formation, and all the other treatment schedules that exposed cells during the first three days of differentiation also inhibited rosette formation. In contrast, treatment from DoD3-4 did not decrease the number of rosettes significantly and compound exposure that started later than day 3 (DoD3-6, 4- 9, 12-15) did not lead to any effect, even if cells were exposed for up to 4 days (Fig. 6.6).

Figure 6. 6: Dependence of rosette formation on the time window of exposure. Cells were differentiated towards neural rosettes according to Fig. 6.1a. (a) TSA (10 nM) was added to the cell cultures for different time periods, as depicted by the horizontal bars. At DoD15, cells were fixed, stained and rosettes were quantified. A qualitative overview of the outcome is given by color coding of the bars. Red bars indicate that rosette formation was disturbed; green bars indicate that no inhibition of rosette formation was observed for the respective exposure condition.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

(b) Quantitative results for the rosette formation assay after exposure to TSA as specified in A. Rosettes were quantified as in Fig. 6.4b/c and normalized to untreated controls (untr). The grey bars indicate the means from 2-9 independent experiments (n=13 for the untreated control), the error bars indicate the standard error of the means.

Due to the re-seeding step on DoD11, treatment was never performed during this time (to avoid technical artefacts). These data suggest, that the window of sensitivity of developing NEPs comprises the first three days of differentiation. To clarify whether rosette formation on DoD15 was correlated with gene expression on DoD6, samples from the timed exposure to TSA were analysed by RT-qPCR for marker genes. The neuroepithelial genes PAX6, OTX2 and EMX2 were downregulated by TSA. At least 2 days of drug exposure were required for a (≥ 2 fold) reduction in gene expression, and this exposure window had to include the second day of differentiation. The pluripotency transcription factors POU5F1 and NANOG are known to be strongly downregulated during neuroepithelial differentiation (Balmer et al. 2012; Chambers et al. 2009). Here, disturbance by TSA led to higher residual levels of these stem cell markers, and also of the early neural crest specifying gene MSX1 (Fig. 6.7a, b). For such effects a drug exposure of six (MSX1, NANOG) to 24h (POU5F1) only was sufficient, if occurring early during differentiation (first day). Otherwise, the windows of sensitivity for dysregulation of POU5F1/NANOG/MSX1 were similar to those for the neuroepithelial markers (Fig. 6.7a, c). Comparison of the gene expression and the rosette formation endpoints revealed a good overall agreement. The rosette formation assay was slightly more sensitive, if our prediction model (described in Fig. 6.7c) for combinations of marker genes was applied. Notably, alternative prediction models could have been used, and they may have shown identical sensitivity. This comparison again demonstrates the practical difficulty of establishing a robust prediction model based on genetic markers. Prediction based on such markers would require a vast amount of test compounds to find out whether different markers should be considered as equally important, whether they should be assigned to weight factors and whether the average, median or e.g. most sensitive set of changes is to be considered. Altogether this set of experiments showed a high concordance of RoFA and gene expression (STOP-tox(UKN)) endpoints (Fig. 6.7b, c), and we planned for further exploration of the predictivity of inhibited rosette formation.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 7: Sensitivity comparison of gene expression analysis and rosette formation assay

Cells were differentiated for 15 days according to the scheme depicted in Fig. 1a. Treatment with TSA (10 nM) was performed for different time periods (specified as DoD on the x-axes). At DoD6, RNA was extracted and RT-qPCR was performed for PAX6, OTX2, EMX2, MSX1, POU5F1 and NANOG. (a) Relative gene expression (compared to untreated controls) is depicted for the indicated genes. The dotted red line depicts a FC ≥ 2. Regulations beyond this threshold were classified as positive. Data are means from 2-5 independent experiments. The error bars represent the standard error of the means. (b) Classification strategy: Different exposure scenarios were investigated. Gene expressions were classified as toxicant-regulated, if the relative expression was altered by ≥ 2-fold. (c) Com- parison of gene expression and rosette formation. Summary data for gene expression changes were produced by the following rule: If ≥ three out of the six marker genes were regulated, gene regulation was considered to be disturbed (red). Rosette stage cells at DoD15 were fixed, stained and analyzed with the rosette formation assay (Fig. 6.4). Disturbance of rosette formation was classified as specified in Fig. 6.6. No gene expression data were available for exposure of cells only during the first 3 h of the assay.

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Correlation of RoFA and STOP-tox(UKN) for a test set of toxicants

As our initial experiments convinced us that the RoFA shows high sensitivity towards the tool compounds TSA and VPA, we initiated a broader evaluation of altogether 24 test chemicals (see Fig. S6.1 for compounds and their concentrations used here). Cells were exposed to these from DoD0-6, and rosette formation was assessed on DoD15. Whole transcriptome data for DoD6 were obtained from the published literature (our legacy data on the same compound concentrations from (Balmer et al. 2014; Krug et al. 2013b; Rempel et al. 2015; Shinde et al. 2016a; Waldmann et al. 2017; Waldmann et al. 2014)). Based on these results, we evaluated the reciprocal relationship of RoFA data and transcriptional changes by using various correlation and classification strategies. In a first approach, we used an index of toxicity related to gene-expression, the developmental potency (D(P)), that we had developed and evaluated earlier (Shinde et al. 2016a). This correlated with rosette formation, but only to a moderate extent (rho = 0.61) (Fig. S6.4). Correlations in a similar range were found when other relatively crude measures of transcriptome alterations were used (e.g. number of genes deregulated). Therefore, more complex classification methods were explored. Several prediction models based on linear correlations and on random forest, were tested, and various sets of genes (e.g. top100 probe sets, top200, top1000, most variable etc.) were used as input. Good data were obtained by using the top 200 most variant probe sets from the microarray results (Fig. S6.5). The correlation for RoFA vs transcriptome changes (based on these 200 probe sets) was R²=0.97, when a random forest-based prediction model was employed (Fig. 6.8a). The correlation plot furthermore indicates that predictions of RoFA outcomes are quite reliable, if they are in the range of 0-33% inhibition or 66-100% inhibition. For such compounds, transcriptome data of the 200 probe sets (Fig. S6.5) appear to be sufficient to predict phenotypic deficits in self-organization of the differentiating neuroepithelial cells, i.e. for a hazard classification of test compounds. If a reduction of rosette formation of ˃33% and ˂66% is predicted, we would suggest to initiate further testing in the RoFA (Fig. 6.8b). The setup of initial/preliminary prediction models during assay development, using a limited number of compounds, necessarily leads to overfitting. Ideally, further validation with larger numbers of new reference chemicals would be performed, but this requires large resources and time. Alternatively, an internal validation can be performed to obtain at least some measure of reliability of the PM. We used this strategy and chose the leave-one-out cross-validation (LOO) approach: 23 compounds were used to predict the remaining compound, and this process was repeated for each of the compounds.

100

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure 6. 8: Relationship between toxicant-induced transcriptome changes and rosette formation Cells were differentiated for 15 days according to the scheme depicted in Fig. 6.1a. During differentiation, cells were treated with the indicated compounds from DoD0-6. At DoD15, rosette formation was quantified and adjusted for controls. In parallel cultures, RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls). (a) A random-forest prediction model was built based on all 24 compounds and used to predict rosette inhibition from DoD6 transcriptome data (for gene list see Fig. S6.5). The predicted rosette responses for each compound are plotted against the % inhibition of rosette formation actually measured by RoFA. For faster orientation, the areas for which low (< 33%) or high (> 66%) inhibition of rosette formation was predicted, were shaded in green. 101

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

(b) Based on the findings in a, a test procedure is suggested, in which DoD6 gene expression data are used to initially predict inhibition of rosette formation, and thus to asses developmental neurotoxicity (DNT) hazard. For the data range of 33-66%, at which the prediction model has high uncertainties, actual testing in the RoFA is necessary to determine functional effects of chemicals predictive of DNT. (c) To establish a preliminary prediction model for the experimental RoFA, the gene expression data on DoD6 were used to categorize DNT-negative and DNT- positive chemicals. For this, a panel of 21 hand selected biomarker genes (46 probe sets, for details see Fig. S6.7A) was employed: The expression changes by the test chemicals of the corresponding 46 probe sets were used to calculate a neurodevelopmental staging defect distance (NDD) measure (i.e. a deviation from normal development, as detailed in Fig. S6.7b). Twelve test compounds (shown in green) had a NDD score of <1, and were therefore considered negative, those with a score > 1 were considered positive (black). (d) Cultured cells treated as in A were differentiated until DoD15 (RoFA as in Fig. 6.4). Each symbol shows a data point of an independent experiment. Data from compounds with an NDD < 1 are shown in green, with an NDD > 1 in grey. Data on the statistical average and spread of the data of positives (black) and negatives (green) are indicated on the right-hand side. A value separating positive and negative compounds was defined based on the variability of the negatives (=noise band); threshold T= Mn-(2x SDn). (e) Rosette formation was quantified for each compound as in d. Each bar represents the mean percentage of rosette formation (compared to untreated controls) from two to four independent experiments per compound. Error bars represent the SEM. The data from d were used for a preliminary prediction model of the RoFA, so that compounds that affected rosette formation for more that the noise band (T: 2 x SD of the negatives) were considered positive (purple), while the other compounds were considered negative (grey). Com- pounds in the borderline range (grey band) are labelled blue; U(T): uncertainty of T. The borderline range was calculated based on the average variation of all values as described by Leontaridou et al.

The LOO data highly correlated (R2 = 0.82) with the full model (n = 24). Moreover, the LOO prediction data showed again very clearly, that there was particularly high uncertainty about compounds predicted to be 33-66% rosette inhibiting (Fig. S6.6). This confirmed that such compounds would require experimental testing, while compounds with very high or low prediction values agreed to a high degree (Fig. 6.8b). Altogether, our comparison of the transcriptome endpoints and the RoFA suggested that both tests agree well in predicting a potential developmental hazard. On this basis, we undertook some further efforts to set up a binary prediction model for the RoFA, which works without requiring transcriptome data. For this purpose, positive and negative controls needed to be defined. This could theoretically be done by using legacy data from a gold standard test (e.g. human clinical data or animal experimental data). However, this approach has weaknesses related to large data uncertainties, data gaps and problems of species extrapolation. An additional problem is that the definition of positives and negatives refers to specific concentrations reached in vivo. Such data are in many cases not available in standardized form. As alternative approach, we decided to use an internal biological measure, as suggested for strategies of so-called mechanistic validation (Aschner et al. 2017; Hartung et al. 2013a; Leist et al. 2014; Leist et al. 2012b). A list of 21 marker genes (corresponding to 46 probe sets) was selected, based on literature data indicating the link of these genes to different developmental stages expected in our test, or known to be reached 102

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances upon disturbance of the system (Fig. S6.7a). We used mean relative expression levels for these probe sets (ratio of Affymetrix array readouts in the absence or presence of test compound) to calculate a neurodevelopmental distance measure (NDD) (Fig. S6.7b), i.e. a score defining how far the system deviated from its normal state in the presence of a test compound. Compounds with a score of ≥ 1 were classified as positive hits (= positive controls). Com- pounds with a score < 1 were regarded as negative controls. This approach resulted in 12 hits (belinostat, entinostat, retinoic acid, PMA, SAHA, CsA/FK506, panobinostat, LiCl, CHIR, VPA,

BIO and TSA) and 12 negative compounds (DMSO, BPA, estradiol, HgCl2, IFN beta, MeHg, geldanamycin, thimerosal, gleevec, PCMB, galnon and HgBr2) according to their gene expression signature (Fig. 6.8c). This set of reference compounds was then used to define a threshold for hit classification in the RoFA. For this purpose, the average for rosette inhibition of negatives was determined (101 % rosette formation). We then determined the standard deviation (SD) for all negatives (26.8%), and used the value of 2 x SD (53.5%) as proxy for the noise band of negative compounds (Fig. 6.8d). Thus, rosette formation of < 47.6% (101.1% - 53.5%) was considered as definitely disturbed rosette formation in this final prediction model (Fig. 6.8e). The mean value of positive hit compounds was 10.6 ± 17.8% and thus separated generally very well from the negatives. Based on this hit definition, the RoFA prediction model identified 12 compounds as positives and 12 compounds as negatives (Fig. 6.8e). All compounds that were positive in the NDD distance measure were also classified as hits in the RoFA. In the consideration of the prediction model we also propose a further refinement: it has been suggested that binary models may work better if a borderline range (BR) is defined (Leontaridou et al. 2019; Leontaridou et al. 2017). We adopted this suggestion and added the BR around the threshold T (BR= T ± 24.5%) and found that three compounds fall into this range (Fig. 6.8e). Such borderline compounds may easily be classified one or the other way, due to their unclear test results. For practical purposes, such compounds would be mild alerts (compared to the full hits as serious alerts), and they may be prioritized for further testing in complementary assays (Leontaridou et al. 2019; Leontaridou et al. 2017). Altogether, the NDD and RoFA endpoints correlated very well (correlation of R²= 0.84, if continuous measures were taken), and the hit classification agreed in all cases (Fig. S6.8).

Outlook on the definition of better transcriptional markers for the prediction of developmental toxicity and of RoFA outcomes

Despite the encouraging data presented above, further improvements of the RoFA prediction model will be required. The next steps should for instance include some form of validation against human hazard data. In case the validity of the RoFA is confirmed as predictive assay

103

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances for human DNT (or general developmental defects), it may become a gold standard itself. In this function, it would help to define and calibrate other prediction models. For instance, data from the RoFA could be used to define better transcriptional biomarkers for the traditional

STOP-tox(UKN) test (Fig. S6.9). The advantage of this approach would be that one could perform a 6-day assay (STOP-tox(UKN)) instead of a 15 day assay (RoFA), but obtain similar predictive capacity. We used this approach here to identify such a set of transcriptional biomarkers. In follow-up work, they may be refined further, and be validated against new sets of control compounds. The following approach was taken here:For all 23 compounds and all probe sets measured with gene symbol annotation available (> 40,000), the mean differential (control-adjusted) expression levels were correlated with the mean control-adjusted rosette formation results (using Pearson’s correlation coefficient). Genes that were most positively correlated with rosette formation were identified, and they included CHAMP1, CDH2, MYLK, FGF9 and CDK5RAP2. Also genes, correlating negatively with rosette formation were identified and they comprised DDIT4, MEOX1, TFAP2A, GADD45B and DACT1 (Fig. S6.9a, S10). The performance of a prediction model based on these genes will be tested in the future, when data on new (different) compounds become available. A second, complementary strategy to identify biomarkers was also followed (Fig. S6.9b). The starting point for this was that both HDAC inhibitors and Wnt activators were identified as hits in the RoFA, while several mercuri- als were identified as negatives (not affecting the transcriptome). We used this situation to identify a subpool of genes that were specifically regulated by rosette formation inhibiting compound groups. First, transcripts differing between HDACi (VPA, TSA, panobinostat) vs. three negatives (HgBr2, MeHg, thimerosal) were identified (Fig. S6.11, S6.12). Then, transcripts differing between WNT activators (LiCl, CHIR, BIO) vs. negatives (HgBr2, MeHg, Thimerosal) were compiled (Fig. S6.13, S6.14). Finally, the overlap of the negatively-correlated genes was determined. They included EDNRA, DACT1, SLIT2, BMP5, ANXA2, SEMA3C and NPY (Fig. S6.15, S6.9b). These transcripts were dysregulated by six different compounds (falling into at least two classes of largely differing toxicity mechanisms), and they may thus take a more general role in indicating disturbed differentiation. Similarly, an overlap of positively correlated genes was identified and they included DUSP4, LHX2, SIX3, CAPN6 and FZD5 (Fig. S6.15, S6.9b). The two pools of biomarkers will be a valuable starting point to reduce the costs and efforts associated with the STOP-tox(UKN), as they may be determined by a focused approach (e.g PCR), independent of expensive whole gene transcriptomics approaches. Moreover, further exploration of these markers may give hints on biological mechanisms underlying early neurodevelopmental defects.

104

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.5. Conclusions and overall outlook

Gene expression profiling was initially chosen as endpoint for the STOP-tox(UKN) assay (Krug et al. 2013b) as it was considered a proxy for a comprehensive description of the differentiation pattern and may in addition offer the possibility to derive mechanistic information. During the practical use of the test, it became clear that too few well-characterized control compounds exist (Aschner et al. 2017) to validate the assay by a correlation approach. For mechanistic validation (Hartung et al. 2013a; Leist et al. 2012a), the knowledge on transcriptional dysregulation is too limited. Therefore, a definite prediction model has not yet been established, and the assay has been mainly used to develop procedures for generating classifiers and toxicity indices. These worked very well under defined assay conditions within a narrow chemical space applicability domain (Shinde et al. 2016a). In the present work, we developed a different strategy to further refine the STOP-tox(UKN) Test. For this purpose we developed a “phenotypic anchor”, i.e. a morphological phenotypic endpoint that could be correlated on the one hand to the gene expression profile in the in vitro test, and that would on the other hand plausibly link to in vivo morphological changes, i.e. teratogenicity. This approach avoids the need for in vivo (human) altered gene expression data for mechanistic or correlative evaluation of assay performance. Notably, such a strategy has been successfully applied in other toxicological areas. For instance, the GARD test for skin sensitization (Johansson et al. 2013; Johansson et al. 2011; Li et al. 2019; van Vliet et al. 2018) uses a complex transcriptional profile as readout, while phenotypic data (observations in animals and man on an allergic response) were used to define appropriate genes and their combination. Another example are the studies of the Piersma group on zebrafish embryo transcriptome changes related to developmental toxicity (Hermsen et al. 2013). These altered transcript patterns were first anchored to morphological malformations. Then they were used for quantifications or mechanistic conclusion (Tonk et al. 2015; Weigt et al. 2010). Also, assay development based on recursive optimization of mechanistic and phenotypic endpoints is well-established in toxicology. For instance in genotoxicol- ogy, the primary cell endpoint (various forms of DNA damage and mutation) is hard to directly correlate with genetic damage in vivo. Typical phenotypic endpoints used as bridges are micronucleus formation (Fenech 2000; Fenech and Morley 1985), the formation of bacterial col- onies in the Ames test (Leist et al. 1992) or the fluorescence of reporters, e.g. in the GADD45 assay (Walmsley 2008). Also in other fields, complex patterns of tissue stress response activation or of nuclear receptor activation networks have been anchored to simple reporter responses (Legler et al. 1999; Piersma et al. 2013; van der Burg et al. 2013; van der Linden et al. 2014) to provide robust assays with highly quantitative readouts. In developmental toxicology research, such development often started from phenotypic endpoints. E.g. the EST (Seiler

105

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances and Spielmann 2011) or the whole embryo culture (WEC (New 1978; Piersma et al. 2004; Tonk et al. 2013)) test have mainly relied on their phenotypic endpoint, and a full characterisation of associated transcriptome changes is still ongoing (Corvi et al. 2016). With the combination of the new RoFA readout and the established transcriptome patterns, we have taken an important step in the development of a human cell based in vitro assay to predict developmental toxicity. Further research will be performed to put the prediction model on a broader basis. In this context, it is also important to note that different cell lines may perform differently. We observed that various hESC and iPSC lines required extensive adaptation for the standard

STOP-tox(UKN) assay and some cells were not usable at all. The same may be true for the RoFA. Our own experience showed that various iPSC lines did form rosettes, and e.g. the Sigma iPSC line (IPSC0028) showed assay performances similar to the H9 hESC line used for the assay development. This may be an important consideration for setup of the RoFA for commercial testing. However, such issues apply to many assays based on cell lines. E.g. the cell transformation assay to predict carcinogenicity only works with few cells, sometimes even only specific clones of such cells. Another important future development will be the incorporation of the test into a more comprehensive battery, e.g. also including a neurite outgrowth assay (Delp et al. 2018; Stiegler et al. 2011), a precursor cell migration assay (Barenys et al. 2017; Nyffeler et al. 2017b), assays addressing non-neuronal development (Jagtap et al. 2011; Krug et al. 2013b)) and additional tests using model organisms (Bal-Price et al. 2018b; Hunt et al. 2018; Scholz et al. 2008).

6.6. Acknowledgements

This work was supported by the Land BW, the Doerenkamp-Zbinden foundation, the DFG (RTG1331, KoRS-CB) and the Projects from the European Union's Horizon 2020 research and innovation programme EU-ToxRisk (grant agreement No 681002) and ENDpoiNTs (grant agreement No 825759). We are grateful to H. Leisner and D. Fischer and the staff of the bi- oimaging center (BIC) for invaluable experimental support.

6.7. Conflict of interest

The authors declare no conflict of interest.

106

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

6.8. Supplementary information

Figure S6. 1: Overview of the compounds and the respective concentrations that were used for the study. Cells were differentiated for 15 days according to the scheme in Fig. 1a. During differentiation, cells were treated with the indicated compounds from DoD0-6. Cells were either harvested on DoD6 for microarray analysis or further differentiated until DoD15. In order to determine cytotoxicity, a resazurin assay was performed on DoD6, and the EC10 concentration was determined for the test compounds. Usually, the EC10 concentration was used for the analysis of mRNA transcript levels at DoD6 and the rosettes formation at DoD15.

Figure S6. 2: Primary antibodies that were used for the study

Figure S6. 3: Primers that were used for the study

107

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 4: Correlation of gene regulation (D(P)) and rosette inhibition.

(a)The definition of developmental potency as gene expression index is given, as originally defined in Shinde et al, 2016. (b) Cells were differentiated for 15 days according to the scheme depicted in figure 6.1A. During differentiation, cells were treated with the indicated compounds from DoD0-6. At DoD15, rosette formation was quantified. In parallel cultures, RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls). Developmental potency was calculated for each compound treatment based on the gene expression on DoD6 (fold change ≥ 2). The toxicity-related gene expression (quantified as developmental potency) is an index of developmental toxicity, based on altered transcriptome expression profile, as determined by Affymetrix microarrays. Developmental potency was correlated with rosette inhibition. All data are based on averages of at least three independet experiments.

108

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 5: Two hundred probe sets with the highest variance across 24 test compounds.

Cells were differentiated for 6 days according to the scheme depicted in Fig. 6.1A. During differentiation, cells were treated with 24 different compounds (belinostat, BIO, BPA, CHIR, CsA/FK506, DMSO, entinostat, estradiol, galnon, geldanamycin, Gleevec, HgBr2, HgCl2, IFNbeta, LiCl, MeHg, panobinostat, PCMB, PMA, RA, SAHA, thimerosal, TSA, VPA) from DoD0-6.RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls) by microarray analysis. The table shows the 200 probe sets with the highest variance across all compounds.

109

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 6 Evaluation of the random forest statistical model (Fig. 8a) by leave one out cross validation.

Cells were differentiated for 15 days according to the scheme depicted in Fig. 6.1A. During differentiation, cells were treated with the indicated compounds from DoD0-6. At DoD15, rosette formation was quantified. In parallel cultures, RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls). The top 200 variance probe sets were used for a random forest classification model (see Fig. 6.8A). This model was then evaluated by a leave-one-out cross-validation. In each cross-validation step, 23 compounds were used to define the top-200 probe sets and to construct a statistical model according to the rules used for the model in Fig. 6.8A.This was then used to predict the response for the remaining, left out compound. This procedure was repeated until each compound was considered once. The 24 predictions are displayed on the y-axis. The x- axis shows the predicted response from the full n=24 (random forest statistical model).

110

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 7: Set of hand-selected genes to capture neurodevelopmental dysregulation by diverse toxicants.

(a) Based on the literature (ours and others past work) and databases, taking into account good signal-noise ratios, 21 genes were selected that have been described to indicate key events and their disturbance during neurodevelopment. For this list of genes a neurodevelopmental distance (NDD) was calculated for each potential toxicant. The NDD was designed as a measure of dysregulation relative to untreated controls. This measure was used in Fig. 6.8c and S6.8. (b) The NDD was calculated as the sum of squared mean differential regulation of the 46 probe sets corresponding 21 genes for a given toxicant vs. solvent control.

111

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 8: Neurodevelopmental distance measure correlated to rosette inhibition.

Cells were differentiated for 15 days according to the scheme depicted in Fig. 6.1a. During differentiation, cells were treated with the indicated compounds from DoD0- 6. At DoD15, rosette formation was quantified (compared to untreated controls). In parallel cultures, RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls). A statistical distance measure based on the expression of the genes from Fig. S6.7 was calculated for each compound treatment. The NDD was then correlated with rosette inhibition. The red lines indicate the thresholds for the binary classifications (dashed for NDD, solid for the RoFA). The divergent compound is shown in a blue circle.

Figure S6. 9: Definition of biomarkers to predict rosette formation at early timepoints (DoD6).

Cells were differentiated for 15 days according to the scheme depicted in Fig. 6.1a. During differentiation, cells were treated with compounds from DoD0-6. At DoD15, rosette formation was quantified (compared to untreated control). In parallel cultures, RNA was extracted on DoD6, and differential gene expression was measured (compared to untreated controls). (a) The DoD6 transcriptome values of 41941 probe sets with available gene symbol annotation from 24 compounds (belinostat, BIO, BPA, CHIR, CsA/FK506, DMSO, entinostat, estradiol, galnon, geldanamycin,

Gleevec, HgBr2, HgCl2, IFNbeta, LiCl, MeHg, panobinostat, PCMB, PMA, RA, SAHA, thimerosal, TSA, VPA) were correlated with the respective rosette formation results from Fig.6. 8e using Pearson's correlation coefficient. The top 20 correlated genes for positive (i.e. higher expression in rosette-forming samples) and negative (i.e. vice versa) correlation were identified (S10). Ten of these potential genetic biomarkers of rosette formation with the best signal/noise ratio across compounds are displayed. A focus was given to probe sets represented ≥ twice in the top 20 list (underlined). (b) As a second approach to identify biomarkers, genes were identified that were jointly regulated by rosette-disrupting compounds of very different mode of action. 112

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

HgCl2, MeHg, thimerosal were chosen as negatives (not disrupting rosette formation). Pearsons correlation was performed for each probe set based on these negative compounds and three positive HDACi (VPA, TSA, panobinostat). Similarly, Pearson correlation was performed based on negatives and three rosette-disrupting WNT activators (LiCl, CHIR, BIO). The top 100 correlated probe sets (positively and negatively correlated) are shown in S11,12,13 and 14. The overlap by gene symbol of the top 100 positively and top 100 negatively correlated genes of the HDACi correlation and the top 100 lists of the WNT activator correlation was identified (S15). Twelve of these potential genetic biomarkers of rosette formation with best signal/noise ratio across compounds are displayed. Genes that showed. higher expression after treatment with rosette disrupting compounds are colored in orange, those with lower expression are shown in blue.

Figure S6. 10: Identification of biomarkers that correlate with rosette formation.

The DoD6 transcriptome values of 41941 probe sets with available gene symbol from 24 compounds (belinostat,

BIO, BPA, CHIR, CsA/FK506, DMSO, entinostat, estradiol, galnon, geldanamycin, Gleevec, HgBr2, HgCl2, IFNbeta, LiCl, MeHg, panobinostat, PCMB, PMA, RA, SAHA, thimerosal, TSA, VPA) were correlated with the respective rosette formation results from Fig. 6.8e in a continuous Pearson’s regression. The top 20 correlated genes for positive (i.e. higher expression in rosette-forming samples) and negative (i.e. vice versa) correlation were identified. FDR adjusted p-values of the correlation test based on Pearson’s correlation coefficient (null hypothesis of 0 correlation) were calculated across all 41941 probe sets."

113

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 11: Negative HDAC inhibitor biomarkers.

Top 100 negatively correlated probe sets that showed the strongest negative correlation with rosette formation for the negative mercurial compounds (HgCl, thimerosal and MeHg) and the positive HDACi compounds (panobinostat, TSA, VPA). The analysis was restricted to probe sets with (i) available gene symbol annotation, and (ii) mean differential transcriptome value > 2 or < ½ in the negative mercurial compounds and/or in the positive HDACi compounds. FDR adjusted p-values of the correlation test based on Pearson’s correlation coefficient were calculated across all probe sets fulfilling the inclusion criteria (i) and (ii). Fold changes (FC) were calculated between the mean differential transcriptome values in the negative mercurial compounds and in the positive HDACi compounds.

114

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 12: Positive HDAC inhibitor biomarkers.

Top 100 positively correlated probe sets that showed the highest correlation with rosette formation for the negative mercurial compounds (HgCl, thimerosal and MeHg) and the positive HDACi compounds (panobinostat, TSA, VPA). The analysis was restricted to probe sets with (i) available gene symbol annotation, and (ii) mean differential transcriptome value > 2 or < ½ in the negative mercurial compounds and/or in the positive HDACi compounds. FDR adjusted p-values of the correlation test based on Pearson’s correlation coefficient were calculated across all probe sets fulfilling the inclusion criteria (i) and (ii). Fold changes (FC) were calculated between the mean differential transcriptome values in the negative mercurial compounds and in the positive HDACi compounds.

115

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 13: Negative correlated Wnt activator biomarkers.

Top100 the negatively correlated probe sets that showed the strongest negative correlation with rosette formation for the negative mercurial compounds (HgCl,thimerosal and MeHg) and the positive Wnta compounds (BIO, CHIR, LiCl). The analysis was restricted to probe sets with (i) available gene symbol annotation, and (ii) mean differential transcriptome value > 2 or < ½ in the negative mercurial compounds and/or in the positive Wnta compounds. FDR adjusted p-values of the correlation test based on Pearson’s correlation coefficient were calculated across all probe sets fulfilling the inclusion criteria (i) and (ii). Fold changes (FC) were calculated between the mean differential transcriptome values in the negative mercurial compounds and in the positive Wnta compounds.

116

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 14: Positive correlated Wnt activator biomarkers.

Top100 positively correlated probe sets that showed the highest correlation with rosette formation for the negative mercurial compounds (HgCl,thimerosal and MeHg) and the positive Wnta compounds (BIO, CHIR, LiCl). The analysis was restricted to probe sets with (i) available gene symbol annotation, and (ii) mean differential transcriptome value > 2 or < ½ in the negative mercurial compounds and/or in the positive Wnta compounds. FDR adjusted p- values of the correlation test based on Pearson’s correlation coefficient were calculated across all probe sets fulfilling the inclusion criteria (i) and (ii). Fold changes (FC) were calculated between the mean differential transcriptome values in the negative mercurial compounds and in the positive Wnta compounds.

117

Results- manuscript 2: Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances

Figure S6. 15: Overlap of Wnt activator and HDAC inhibitor biomarkers.

Overlap of the Top 100 HDACi and top 100 Wnta gene lists that showed the highest correlation with rosette formation (Negative correlated: S11 and S13/ Positive correlated: S12 and S14).

118

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7. Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed neuroectodermal differentiation

Johannes Meisig1*, Nadine Dreser2,*, Marion Kapitza2, Margit Henry3, Jörg Rahnenführer4, Jan G. Hengstler5, Agapios Sachinidis3 Tanja Waldmann2,, Marcel Leist#,2 and Nils Blüthgen #,1 *,# shared authorship 1 Institute of Pathology, Charité-Universitätsmedizin, 10117 Berlin, Germany 2 In Vitro Toxicology and Biomedicine, Dept inaugurated by the Doerenkamp-Zbinden Chair foundation, University of Konstanz, 78457 Konstanz, Germany 3 Center of Physiology and Pathophysiology, Institute of Neurophysiology, University of Co- logne (UKK), D-50931 Cologne, Germany 4 Department of Statistics, TU Dortmund, 44221 Dortmund, Germany 5 Leibniz Research Centre for Working Environment and Human Factors (IfADo), Technical University of Dortmund, 44139 Dortmund, Germany

119

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.1. Abstract

Thousands of transcriptome data sets are available, but approaches for their use in dynamic cell response modelling are sparse, especially for processes affected simultaneously by two orthogonal influence variables. We approached this problem for neuroepithelial development of human pluripotent stem cell (differentiation variable), in the presence or absence of valproic acid (signaling variable). Using few basic assumptions (sequential differentiation states of cells; discrete on/off states for individual genes in these states), and time-resolved transcriptome data, a comprehensive model of spontaneous and perturbed gene expression dynamics was developed. The model made reliable predictions (>90% correlation with measured data) on 82% of gene expression trajectories. Even regulations predicted to be non-monotonic were successfully validated by PCR in new sets of experiments. Transient patterns of gene regulation were identified from model predictions. They pointed towards an activation of wnt signaling as candidate pathway leading to a re-direction of differentiation away from neuroepithelial cells towards neural crest. Experimental intervention experiments, using a wnt/beta-catenin antag- onist, led to a phenotypic rescue of this disturbed differentiation. Thus, our broadly applicable model allows the analysis of transcriptome changes in complex time/perturbation matrices.

120

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.3. Introduction

Early cell fate commitment towards the neural lineage is accompanied by highly dynamic and large-scale transcriptome changes (Abranches et al. 2009; Gaspard et al. 2008; Gaspard and Vanderhaeghen 2010; Suzuki and Vanderhaeghen 2015; van de Leemput et al. 2014; Zimmer et al. 2011a). Disturbances during this process can alter developmental trajectories perma- nently. Accordingly, key transcription factors as well as large transcriptome modules, such as gene ontologies and constituents of regulatory circuits, are strongly affected (Balmer et al. 2014; Balmer et al. 2012; Boles et al. 2014; Cho et al. 2014; Rempel et al. 2015; Yoo et al. 2011). Pluripotent stem cells (PSC) are good models to mimic the in vivo processes during neurodevelopment. Human embryonic stem cells (hESC) or induced pluripotent stem cells (iPSC) are of particular interest, as this allows experimental access to development of neuroectodermal and later nervous system structures in man. Such neurodevelopmental processes have many species specific features (López-Tobón et al. 2019; Ou et al. 2018; Walker et al. 2019; Zhang et al. 2010) and they could hardly be studied in a systematic way prior to the advent of stem cell technology. Examples for human specific aspects of brain development range from transposon activity, over relative importance of key transcription factors to largely differing time scales and relative steps of wave-like gene activation patterns (Bhinge et al. 2014; Dolmetsch and Geschwind 2011; Dragunow 2008; Marchetto et al. 2019; Marchetto et al. 2013; Zhang et al. 2010). Prior to the classical neurodevelopmental stages, leading from neural stem cells to brain tissue formation and regional specialization (Chandran et al. 2016; Walker et al. 2019) early differentiation processes lead from initially pluripotent cells to neuroepithelial progenitor cells (NEP) and various types of neural stem cells (Conti and Cattaneo 2010). Such steps occur in fetuses during the formation and initial folding/closure of the neural plate (Chambers et al. 2009; Chambers et al. 2011). In vitro, the NEP cells form neural rosettes, a correlate of the neural tube (Dreser et al. 2020). A highly standardized method

(STOP-tox(UKN) assay) has been developed to track early neural development, and its disturbances, via analysis of transcriptome changes, epigenome alterations and the capacity to self- organize in rosettes (Balmer et al. 2014; Balmer et al. 2012; Conti and Cattaneo 2010; Rempel et al. 2015; Shinde et al. 2016a; Waldmann et al. 2017; Waldmann et al. 2014). Disturbances of the regulatory networks that control the spatio-temporal expression of genes can lead to severe congenital malformations (e.g. spina bifida, anencephaly) (Nau 1994; Ornoy 2009), and they may contribute to neuropsychiatric disorders such as attention deficit/hyperactivity disor- der (ADHD), autism spectrum disorders (ASD) or schizophrenia. Such disturbances can be induced by genetics (Kondo et al. 2009), stress experiences during fetal and early childhood development (Herzog and Schmahl 2018) and environmental influences (drugs, environmental

121

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation and industrial chemicals, malnutrition) (Heyer and Meredith 2017; Kondo et al. 2009). Little is known about the biochemical mechanisms that underlie these neurodevelopmental disorders, and how and when transient perturbations, such as by drugs, alter differentiation in a pathological way. Valproic acid (VPA) may be used to establish bona fide models of clinically-relevant neurodevelopmental disorders. It is an anticonvulsant drug that is administered in epilepsy and mood disorders and it is well known to increase the risk for congenital malformation such as neural tube closure defects and craniofacial malformations (Alsdorf and Wyszynski 2005; Chateauvieux et al. 2010; Nau 1994; Ornoy 2009; Tomson et al. 2016). Extensive clinical data exist on VPA´s developmental neurotoxicity effects (Koren et al. 2018), and the relevant plasma concentrations in fetal and maternal blood are known (Sztajnkrycer 2002; Wiltse 2005). The impact of developmental disturbances depends strongly on the time of onset and duration of the disturbance. In this sense, thalidomide (best known under the brand name Contergan) gained notoriety as there were very small time windows of in utero exposure identified that determined the kind of malformation (Ito et al. 2011). Therefore, in order to investigate the molecular mechanisms of such disturbances by using differentiating human PSC, it is important to get a better understanding of how these perturbations propagate to dysregulated gene activities and cellular phenotypes. Mathematical models have been very instrumental to disentangle the complexity of regulatory networks in general, and of gene regulation in the context of stem cell differentiation more specifically. Models have been applied to analyze networks on multiple layers. For stem cells, most models focused on the network of key regulators and the long-term dynamics of cell fate decisions (Herberg and Roeder 2015; Spector and Grayson 2017) . Those models were able to disentangle the network topology of small networks of key pluripotency factors and have helped to define the logic that governs the stability and exit from pluripotency or other cell identity transitions (Bessonnard et al. 2014; Chickarmane and Peterson 2008; Dunn et al. 2014). Such detailed mechanistic models build on extensive mechanistic knowledge and often integrate detailed data from different sources and studies. In contrast, a number of data-driven modelling approaches have been used to model transcriptome changes on a genome-wide scale. Broadly, these models were either used to infer quantitative information about mRNA processing and turnover to describe the response of the transcriptome to acute stimulations, or they have been used to infer hidden variables like transcription factor activity from transcriptome time series data (Barenco et al. 2009; Honkela et al. 2010; Honkela et al. 2015; Miller et al. 2011; Uhlitz et al. 2017). So far, these models were only applicable to describe the transcriptome dynamics to immediate stimuli or the effect of individual transcription factor on gene expression. Neither detailed mechanistic approaches nor the current transcriptome-wide modelling approaches are optimally suited to disentangle the transcriptome dynamics of a differentiation program, such as the

122

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation transition from pluripotency to neural progenitor cells. As these transitions occur in a cascade of steps and lead to waves of gene expression, models are needed that consider the dynamic changes under control conditions to evaluate modulations by stressors/toxicants. Moreover, they need to be simple enough to be parameterized from sparse data, such as transcriptome time series. We therefore developed a coherent modelling framework that allows us to describe and model transcript dynamics during differentiation and to disentangle the effects of perturbation on the transcriptome dynamics. Our model allows to describe time-dependent gene expression in differentiating cells, and it is sufficiently coarse-grained to be applied transcriptome-wide. Using this model, we analyze differences between untreated and disturbed differentiation, where we exemplarily study the influence of VPA on transcript dynamics. We show that our kinetic model correctly predicts gene expression kinetics at concentrations of VPA not used for training. Additionally, the model provides a mechanistic basis to explain experimen- tally observed non-monotonic concentration-response relations that were non-intuitive at first. Finally, the model allows us to identify perturbed regulatory modules with high resolution in the time/compound concentration space. With this approach, we find that treatment with VPA leads to a transient upregulation of Wnt signaling which in turn causes the upregulation of neural crest genes. Our work shows that genome-wide models of transcriptome dynamics during stem cell differentiation may not only help to understand organogenesis/ embryogenesis, but will also shed new light on developmental disturbances due to genetic and environmental stressors.

123

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.4. Material and methods

Materials

Gelatine, putrescine, sodium selenite, progesterone, apotransferrin, glucose, insulin, ascorbic acid, valproic acid and ICRT3 were obtained from Sigma (Steinheim, Germany). Accutase was from PAA (Pasching, Austria). FGF-2 (basic fibroblast growth factor), FGF-8b, Sonic hedgehog and noggin and were obtained from R&D Systems (Minneapolis, MN, USA). Y-27632, SB- 43154, CHIR99021 and dorsomorphine dihydrochloride were from Tocris Bioscience (Bristol, UK). MatrigelTM was from BD Biosciences (Massachusetts, USA). All cell culture reagents were from Gibco/Invitrogen (Darmstadt, Germany) unless otherwise specified.

Neuroepithelial and rosette differentiation

Human embryonic stem cells (hESC) (H9 from WiCells, Madison, Wi, USA) were differentiated according to the protocol published by Chambers and colleagues (Chambers et al. 2009) with modifications established in (Balmer et al. 2012; Chambers et al. 2011). Instead of using 500 µM noggin we used the combination of 35 µM noggin and 600 nM dorsomorphine together with 10 µM SB- 431642 for dual SMAD inhibition as described earlier. This was used to prevent BMP and TGF-β signaling, and thus to achieve a highly selective neuroectodermal lineage commitment. For handling details, see supplemental material as decribed by Balmer et al., 2012 (Balmer, Weng et al. 2012). Beginning on DoD4, KSR medium was gradually replaced by N2 medium (DMEM/F12 medium with 1% Glutamax, 0.1 mg/ml apotransferin, 1.55 mg/ml glucose, 25 μg/ml insulin, 100 μM putrescine, 30 nM selenium and 20 nM progesterone), supplemented with the same amounts of noggin, dorsomorphin and SB-431642 as KSR. All differentiations were performed in 6-well plates containing 2 ml of medium per well. For some experiments, cells were further differentiated until they formed neural rosettes. In detail, cells were differentiated until DoD10 as described above or in in (Chambers, Fasano et al. 2009). On DoD11, cells were detached using Accutase, and seeded at a density of 150,000 cells/cm2 onto Matrigel-coated 96 well plates or glass cover slips. Cells were grown in N2S medium supplemented with 20 ng/ml FGF2, 100 ng/ml FGF8, 20 ng/ml sonic hedgehog, 20 µM ascorbic acid and 10 µM ROCK inhibitor Y-27632. On DoD13 ROCK inhibitor was removed. Cells were fixed for immune staining at DoD15 after rosettes had formed (Dreser et al. 2020).

124

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Experimental exposure

Developing cells were treated from DoD0-DoD6 with different concentrations of valproic acid (0.025, 0.15, 0.35, 0.45, 0.55, 0.65, 0.8 and 1mM) or the Wnt activator CHIR99021 (2µM). For rescue experiments with the β-catenin Inhibitor iCRT3 (Sigma) cells were (co)treated from DoD2-DoD6.

Immunostaining

For immuno staining, cells were ﬁxed on DoD6 or DoD15 in 4% paraformaldehyde and 2% sucrose prior to permeabilization in 0.3% Triton X-100 in PBS. After blocking in PBS containing 5% bovine serum albumin and 0.1% Tween-20 for 1 h, primary antibodies (Supplementary Material, Fig. S7.1) were incubated for 1 h, at room temperature (RT). After washing, secondary antibodies were applied for 45 min at RT. DNA was stained with Hoechst H-33342, and cover slips were mounted in FluorSaveTM reagent (Calbiochem, Merck).

Rosette formation assay

Rosette formation assay was performed in 96 well plated. Number of rosettes was assessed in six technical replicates for control and three technical replicates for treatments. The immune cytochemistry staining of golgi matrix protein (GM130), zona occludens 1 (ZO1) and the DNA with Hoechst H-33342 was used to identify rosettes. ZO1 is located in the centre of a rosette, surrounded by GM130. Images of the whole well were taken by Cellomcis (Thermo Fisher Scientific) automated microscope. Using Konstanz information miner software (KNIME) (Berthold, Cebron et al. 2007), rosettes were detected and counted. The number of rosettes per well was normalized to the total nuclear area. Data is always displayed relative to an untreated control (Described in detail in (Dreser et al. 2020))

Sample preparation for microarray analysis and quantitative PCR

Cells were lysed in Trizol reagent (Quiagen) at the indicated time points. mRNA was isolated as advised by the manufacturers protocol and reverse transcribed (iScript, Biorad). Quantita- tive PCR was performed using EVAGreen SsoFastTM mix on a BioRad Light Cycler (Biorad, Germany). For quantiﬁcation, qPCR threshold cycles were normalized to reference genes [tatabox binding protein (TBP) and ribosomal protein L13 (RPL13A)] (Primer list: Fig. S2). Data was either displayed relative to expression in stem cells or cells treated with chemicals were

125

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation then expressed relative to transcript levels of untreated control cells (which had been grown and differentiated for the same amount of time). For this normalization, the 2^(-Delta Delta C(T)) method was used (Livak and Schmittgen 2001).

Affymetrix chip-based DNA microarray analysis (Human Genome U133 plus 2.0 arrays) was performed exactly as described earlier (Krug et al. 2013b; Rempel et al. 2015). Expression kinetics for unperturbed differentiation was measured in quadruplicates at the following 12 time points: 0h, 6h, 10h, 16h, 24h, 36h, 48h, 60h, 72h, 96h, 120h, 144h. Expression kinetics at 0.6 mM VPA was obtained in quadruplicates at the time points 72h and 144h and in duplicates at 24h, 48h and 96h. Concentration dependent transcriptomes were obtained at 144h in tripli- cates at 8 VPA concentrations (0.025, 0.15, 0.35, 0.45, 0.55, 0.65, 0.8 and 1mM) and was published earlier in (Waldmann et al. 2017).

Data preparation

Gene expression data was normalised using bioconductor and rma as described in (Rempel et al. 2015). For each condition, we computed mean expression of each gene. Robust estimates of the standard deviation were derived using an error model that was parameterised as follows. The empirical standard deviation for each gene and time point was computed using replicates. Data was then divided into 15 equally sized bins based on mean expression, and average empirical standard deviation was computed within each of the 15 bins. For the three bins with the lowest mean, the standard deviation was set to the maximal value across all bins (corresponding to expression < 6). The resulting values were then interpolated using the R function loess with a span of 0.75.

Batch effect correction

As the untreated kinetics and the kinetics obtained from cells treated with 0.6 mM VPA were obtained in different batches (first and second batch), we performed batch effect corrections. For this, we used untreated samples at the time points 0 h, 72 h and 144 h, where we computed the difference in expression between batches for each gene at the three time points. These differences were linearly interpolated to obtain b(t), the time dependent batch correction. Fi- nally, for each gene and time point, expression in the VPA-treated kinetic samples was transformed according to expr'(t)=expr(t)+b(t);

126

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation where expr(t) denotes the measured gene expression at time point t and expr'(t) is the batch corrected gene expression.

Selection of regulated genes

First, genes present on the microarray were restricted to protein coding genes using the En- sembl database (Ensembl 84 from March 2016) through the R package biomaRt. To identify regulated genes, the variation of the replicates of single time points was compared to the variation with respect to time using the Hotelling T2 statistic obtained with the mb.long function from the R package timecourse (www.bioconductor.org). The four replicates for each of the 12 time points of the untreated kinetic served as input to mb.long. The 1500 genes with the top T2 statistic were labeled as regulated.

Model and cost function

The dynamic model consists of a linear chain of n states with transition rates gon and goff. The rate gon is the on-rate of each state and the off-rate of its preceding state. Only the final state n has a different off-rate, goff. This model can be analytically solved (see supplemental information for details) to obtain the quantity Act(t,n,gon,goff), the fraction of cells in the active state. To model gene expression kinetics, we also introduce a time shift dt and the parameters A and

B as the expression in states 0…n−1 and n, respectively. Then the gene expression Ep is given by Ep(t,A,B, gon, goff,n,dt)=A+B Act(t+dt,n, gon,goff). To prevent non-defined values when computing the logarithm, 1+Ep is used instead of Ep in the cost function. This amounts to a redefini- tion of the parameter A.

To model the dependency of the rates gon (VPA) and goff(VPA) on VPA, we use Hill functions of the form g(VPA,k,n)=g+(g’-g)*VPAn /(VPAn+kn) for both rates. The full model thus contains the additional parameters gon', goff', k1, k2, n1 and n2.

2 2 The cost function used for fitting the parameters is obtained by summing (Em,i-Ep,i) /Si over all data points from the untreated and treated kinetic indexed by i where Em,i is the measured expression and Ep,i is the predicted expression and Si is the moderated standard deviation. For

2 2 the concentration-response data at t=144 h the sum (Fm,i-Fp,i) /Si over all data points indexed by i is computed. Here, Fm,i is the measured fold change with respect to the time point 0 h in the untreated kinetic and Fp,i is the predicted fold change with respect to the predicted expression for time 0 h, untreated. The overall cost function is the sum of both contributions

127

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Parameter estimation

We estimated the parameters by optimizing the difference between data and model simulations weighted by the standard-deviation (see above). Optimization is performed in two iterations using first only untreated data and then all data. In the first iteration, we used an optimizer based on simulated annealing (R package GenSA with option max.time=1000) to determine the parameters A, B, n, gon, goff, dt from the untreated data. Initial values were A=min(e),

B=max(e), gon =0.1, goff=0.01, n=1, dt=0, where e is the vector of expression values in linear units for the respective gene. In the second iteration, random starting parameters are generated in the interval (1/5⋅par,5⋅par), where par is the optimized parameter from the first iteration using the latin hypercube sampling implemented in the function randomLHS from the package lhs. The parameters gon', goff ', k1, k2, n1 and n2 that are not determined in the first iteration, are either randomly drawn from the entire allowed interval (k1, k2, n1, n2) or drawn from the interval determined by gon and goff, respectively (gon', goff '). The fit is performed with a deterministic gradient descent optimizer (L-BFGS-B from the R package optim). Starting parameters are randomly picked 50000 times from the interval described above. For each gene, the best fit is determined from all repetitions.

−5 For both fit iterations, optimization parameters are restricted to the intervals 10

−5 −6 −6 4 5 10

−5 −5 −10 10

GSEA

GSEA was performed using the HTSAnalyzeR package for R. As reference gene sets for the GSEA we used gene sets from the MSigDB v6.0 (gsea-msigdb.org), containing manually curated gene sets from pathway databases as well as published experiments (all curated sets, c2.all.v6.0.entrez.gmt and KEGG sets, c2.cp.kegg.v6.0.entrez.gmt). For all curated sets and KEGG sets, we used a minimum gene set overlap of 20 genes and 15 genes for computing significance, respectively. The FDR for each gene set was computed from 10000 permuta- tions. The GSEA exponent was set to 1.

Pathway gene sets

Gene sets for Figs. .7.4C, D, E, F were obtained as follows. Neural crest signature genes were taken from the MSigDB (see above) gene set named

128

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

“LEE_NEURAL_CREST_STEM_CELL_UP”. TGFbeta pathway genes were taken from the MSigDB gene set named “KEGG_TGF_BETA_SIGNALING_PATHWAY”.

Wnt pathway target genes were obtained from (Mica et al. 2013) through the GEO database (GSE45223). Gene expression data was processed as described above in the section data preparation with the exception that for bins with expression < 10 the standard deviation was set to the maximal value. Wnt induced genes were obtained using the fold change between samples "WA09 embryonic stem cells, NC, day 3" and"WA09 embryonic stem cells, DSi, day 3” with a z-value threshold of 3.5. Wnt repressed genes were obtained using the fold change between samples "WA09 embryonic stem cells, DSi, day 3" and"WA09 embryonic stem cells, NC, day 3" with a z-value threshold of 3.5. The z-value was defined as mean expression divided by the moderated standard deviation.

Gene ordering in heatmaps

In order to sort genes by the time of peak expression, gene expression values were interpolated to a resolution of 0.1 h using the R function loess with a span of 0.8. Subsequently, the time point of the maximum value of the interpolated expression was determined for each gene.

UMAP

Dimensional reduction of the transcriptome data for the 1500 top regulated genes was performed using the umap R package with parameters n_neighbors=4, n_epochs=4000 and random_state=798395..

129

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.5. Results

Heterogeneity of gene regulation during early neural differentiation

We used an in-vitro system where human pluripotent stem cells (hPSC) are differentiated according to the double SMAD inhibition neural induction protocol to analyze gene expression dynamics. We first obtained transcriptome profiles at twelve time points during the first six days of differentiation. To trigger a defined developmental disturbance (Balmer et al. 2014; Balmer et al. 2012; Chanda et al. 2019; Rempel et al. 2015), we treated cells with 0.6 mM VPA and obtained the transcriptome samples after 1, 2, 3, 4 and 6 days (Fig. 7.1A). A two-dimensional Uniform Manifold Approximation and Projection (UMAP) representation was chosen to visualize transcriptome changes over time. The progressive gene expression changes over time were clearly visualized by this approach. Compared to the trajectory of unperturbed cells, VPA- treated cells were clearly offset (Fig. 7.1B). As first overview of gene expression dynamics on the individual gene level the 1500 top regulated transcripts were ordered by the time of their peak expression. The gene expression heatmap clearly showed that many genes had their highest expression at the start or end of differentiation, but that also a sizeable number peaked at intermediate times (n=812) (Fig.7.1C). The sequential peaking of groups of genes is typical for a highly coordinated sequence of neurodevelopmental steps. It may be explained by waves of gene activation programs. The observed pattern is also consistent with the sequential occurrence of cell states that express particular genes. We observed that the timing of peak expression remained broadly unaltered in VPA-treated cells, although a detailed comparison of the patterns showed individual genes that shifted their time course of expression (Fig.7.1C). To follow up on the latter finding, we examined whether peak times of VPA-affected genes generally occurred later or earlier than in a normal differentiation. However, transcripts most strongly affected by the drug (log fold changes > 2) showed a peak time distribution that did not differ from little-affected genes (Fig. 7.1D). This suggests that VPA does not selectively affect genes induced at a specific time point in development. The above analysis of the overall expression pattern showed that VPA-affected genes are not peaking at a particular point in time and that the number of genes peaking at a given time is not affected by the drug. However, for individual genes we noticed that VPA treatment can cause dramatic changes in the expression concerning the timing as well as the extent of regulations (Fig. 7.1E). The complexity of the transcriptome modulation by VPA treatment was our incentive to develop a mathematical model of gene expression dynamics that would reflect the observed expression changes over time.

130

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure 7. 1: The effect of HDAC inhibition on gene expression kinetics in spontaneously differentiating hPSC.

(A) Time-resolved transcriptome data was obtained during the differentiation of hPSC from the pluripotent state to the neuroepithelial stage (DoD6), using an established double SMAD inhibition protocol to direct cell fate. Data were obtained for 12 time points (between 0 and 144 h) during undisturbed differentiation (VPA = 0 mM, n = 4 independent differentiations), and for 5 time points in the presence of 0.6 mM VPA (n=2-4). (B) Presentation of dimensionality reduced (UMAP method) transcriptome data (time series for the 1500 most strongly regulated genes during undisturbed differentiation) to probe differences between normal and VPA-affected differentiation. Numbers indicate the time (in h) after initiation of differentiation. (C) Heatmap of gene expression (row-wise z-normalized) over time for unperturbed differentiation (left) and cells treated with 0.6 mM VPA (right). Expression is shown for the 1500 most strongly regulated genes, sorted by the time of peak expression. (D) The distribution of the time point of peaking is shown for a subset of the 1500 genes from C that have their peak expression at a timepoint after day 0 or before day 6. Genes were grouped as strongly affected by VPA (red, n=64) or little affected by VPA (blue, n=748). (E) Examples of expression kinetics for genes with different behavior in the case of undisturbed differentiation and differentiation in the presence of VPA. Data are means ± moderated standard deviation. (F) Schematic of the structure of the kinetic model, based on the assumption that differentiation proceeds via a sequence of discrete cellular states (e.g. stages 1-4), and that the speed of transition can be described by the transition rates gon and goff. A gene is assumed to be expressed only in a single cellular state (here in state 4, red). For a given gene all transition rates between cell states are assumed to be the same (gon).The decay rate of the final state (here: state 4) can differ (goff). Variations of gon/goff are exemplified in a 4 state model. The fraction of cells in a given state is shown as a function of time. The model assumes all cells to be in state 1 at t=0. The fraction of cells expressing the gene of interest (being in state 4) increases with time. The two panels on the right show the effects of increasing either gon (top) or goff (bottom).

131

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Our model assumes that each cell transits through different states during differentiation, and that each gene is expressed only in one speciﬁc state (Fig 7.1F). Application of the model to cell population averages implies that the gene expression data it predicts are determined by how many cells are in a state where the gene is in the “switched-on” state. To make the model mathematically tractable, we base it on three assumptions. (i) For each gene, the cell has to transit through n prior states before it enters the state n+1 where the gene is activated. (ii) Cells transit through the ﬁrst n states with a rate g1, and they leave the active state with a rate g2. (iii) For each gene, the parameters n, g1 and g2 can be different, i.e. each gene is modeled separately (Fig. 7.1F). The basic working of the model may be exemplified for an n=3 situation, in which the gene of interest is expressed in state 4 (final state) (Fig. 7.1F, bottom, left). On the population level, an increase of g1 in this situation would lead to earlier and more peaked gene expression (Fig. 7.1F, top, right). An increase of g2 would lead to an overall attenuated expression (Fig. 7.1F, bottom, right). An intuitive explanation for this is that few cells ever oc- cupy state 4 at the same time, because the off-rate g2 is high compared to the on-rate g1.

Modelling of dynamic transcriptome disturbances by VPA

In In order to extend the model to VPA-treated cells, additional data were required. We used transcriptome data published earlier (Waldmann et al. 2014), in which the effect of eight additional VPA concentrations (from 25 µM to 1 mM) was measured after 6 days of differentiation (Fig. 7.2A). When we inspected concentration-response curves for different genes in this data set, we noticed a high heterogeneity: for example gene expression was perturbed at different VPA concentration thresholds (NEFL is very sensitive while HPGD is only perturbed at high concentrations of VPA). Also, we found the deregulation to reach a saturation (SIX3) or not (KLF5). Finally, we noticed that some genes showed a non-monotonous concentration-response (e.g. GAD1 or NR2F2) (Fig 7.2B). In order to set up a model for the VPA effect that reflects the molecular features of the neurodevelopmental program, we reasoned that VPA was unlikely to affect the number of cell states. We assumed the drug to alter the parameters gon and goff. Therefore, we made gon and goff monotonic functions of the VPA concentration (modelled using a Hill function). The resulting model was then ﬁtted (i) to the expression data at t=144 h for different VPA concentrations and (ii) to the time series data for unperturbed differentiation and differentiations at 0.6 mM VPA (Fig. 7.2C). This procedure was applied for each gene, i.e. training each model with 12 parameters on data from 86 samples for 25 conditions.

132

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure 7. 2: Establishment of a mathematical gene expression model of VPA-disturbed neurogenesis.

(A) Data matrix used to fit the combined kinetic and VPA effect model. For each condition marked in black, data from 2 (time-dependency) or 4 (concentration response) microarrays were used. All numbers refer to independent differentiations. (B) Curves exemplifying various types of concentration-response behavior of individual genes (means ± moderated standard deviation, n = 4). Data for the effect of nine VPA concentrations on day 6 (t=144 h) are shown. (C) Exemplification of the VPA effect model constraints (for goff and an example gene down-regulated by VPA): It is assumed that the two rate constants of the kinetic model (gon and goff) depend on the concentration of VPA, and can be described by a Hill model (shown in the first column). The Hill function “goff(VPA)” is parametrized in such a way that the multi-step model fits both the endpoint data (expression levels for different VPA concentrations measured at 144 h, column 2) and the kinetic data (measurements of gene expression at different times ≤ 144 h, column 3). Three examples illustrate how the combination of constraints by endpoint and kinetic data lead to model optimization for goff (VPA). Upper panel: the kinetics are correctly fitted but the endpoint is wrong for small concentrations because the Hill curve is not steep enough. Middle panel: both the endpoint and kinetic data (after VPA treatment) are not correctly fitted because the turning point of the Hill function is at too high VPA concentrations. Bottom panel: correct function goff (VPA), fitting both the endpoint and the kinetic data. (D) Example fits for the genes PAX6 (top) and LGI1 (bottom). The left two panels show the fits for the Hill model, with the blue and brown points indicating the VPA concentrations of 0 and 0.6 mM, respectively. The third panel shows data and fit for the gene expression kinetics in the absence (blue) or presence of VPA (0.6 mM, brown). The fourth panel shows the data for the 9 VPA concentrations and one untreated sample at the endpoint (144 h) of the differentiation together with the fit line from the model. (E) Amplitude shift and time shift with respect to fitted gene expression kinetics are indicated by the offset of brown and blue dotted lines. (F) Distribution of genes with respect to the amplitude shift triggered by 0.6 mM VPA compared to the untreated differentiation. (G) Boxplots showing the distribution of time shifts for genes peaking in the indicated interval in cells differentiating under control conditions. Boxes indicate the interquartile distribution, horizontal bars the median. 133

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Example fits are shown for the genes PAX6 and LGI1 (Fig. 7.2D). The plots for the input functions for gon and goff illustrate how different concentration-dependencies of the rates gon and goff can lead to distinct concentration-response behavior at the endpoint. We found that the multi- step model with VPA concentration-dependent rates correctly described the observed changes in gene expression amplitude and timing caused by VPA treatment (Fig. 7.2D). To judge the general ability of the model to fit the kinetic and concentration-response data, we computed the correlation between data and fits. We found a correlation coefficient larger than 0.9 for 88% of the 1500 fitted genes for the unperturbed differentiation. For the concentration-response data, we compared the fits to the data for genes with a minimum log2 fold change of 1 at the endpoint. For this group of 558 genes, 91% had a correlation coefficient higher than 0.9. We concluded that the model correctly describes expression patterns for the large majority of regulated genes in the presence and absence of VPA. In order to characterize the overall effect of VPA on regulated genes, we computed the change in peak-to-trough expression (amplitude shift) and the change in peak timing (time shift) (Fig. 7.2E) from the model fits. We observed a bias towards a decreased amplitude upon VPA treatment (69% of all regulated genes) (Fig. 7.2F). We analyzed the change in peak timing separately for groups of genes with similar peak time in untreated cells because there is a bias towards positive shifts for early peaking genes and a bias towards negative shifts for late peaking genes. When binning genes according to the day of maximum expression in untreated differentiation, we observed that median time shifts were close to zero (Fig. 7.2G). The model fits thus indicate that expression changes over time are mostly attenuated by VPA but not delayed. Taken together, we established computational models for hundreds of genes that capture gene expression kinetics during differentiation and model the influence of VPA on the kinetics.

Explanation of non-monotonic concentration-response curves by gene expression model.

We set out to validate the gene expression kinetics predicted by the model at previously untested concentrations. For this purpose, we selected genes that showed qualitatively different concentration-response curves with respect to VPA at t=144 h (Fig. 7.3A). Saturating concentration-responses, were exemplified by three genes (PAX6, DAZL, OTX2) that are down-regulated and two genes (BMP5, SOX9) that a re-up-regulated by VPA. POU4F1 was included as a gene that shows a slight increase in expression at low concentrations, and a pronounced decrease at high concentrations.

134

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure 7. 3: Multi-state model-based predictions of gene expression under the influence of VPA.

(A) Model fit and experimental validation for genes with different types of concentration-response behavior. The left panel shows examples of genes with largely monotonic concentration-response behavior. The right panel exempli- fies non-monotonic responses. For each gene, the first column (from left to right) shows the experimental data (means ± moderated standard deviation) together with the model fit for the time-dependent expression in the presence of 0 (blue) or 0.6 mM (brown) VPA. Next, the concentration-response data for 6 days of VPA exposure are shown together with the model fit. The third column indicates the model predictions for gene expression kinetics at VPA concentrations of 0, 0.35, 0.6 and 0.8 mM. The fourth column shows expression kinetics data from a validation experiment using RT-PCR (means ± SD, n = 3). The concentrations were 0, 0.3, 0.6, 0.8 mM (blue to orange). (B, C) Detailed exemplification of non-monotonic concentration-response behavior predicted by the model for the GAD1 gene. The graphs in B show how the VPA concentration affects the production rate (gon) and the decay rate (goff). Both rates change monotonically with the VPA concentration. (C) The left plot shows the computed model prediction of gene expression over time, at VPA concentrations increasing from blue to orange (range: 0-1 mM). The peak concentration not only increases with increasing VPA concentration, but also moves to earlier time points. The plot to the right shows fold changes in gene expression based on the data on the left, for a time of 72h, 96h and 144 h and for various VPA concentrations vs control.

135

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Most importantly, we also selected genes that show a pronounced non-monotonic behavior, where expression is either reduced or increased at intermediate, but not at high concentration. (Fig. 7.3A, right column). Using the model ﬁts, we simulated expression kinetics for different VPA concentrations (0.35 mM, 0.6 mM and 0.8 mM VPA). The model predicted that treatment with VPA can dramatically alter the amplitude of peak expression (see e.g. OTX2), as well as timing of peak expression (see e.g. DAZL). BMP5 exemplified the case of changes in both peak expression and peak timing. These predictions were then tested by measuring expression kinetics with quantitative RT-PCR. When comparing the peak timing and changes in amplitude, we observed good qualitative agreement with the model predictions (Fig. 7.3A, right- most panel for each gene). Quantitatively, the 12 tested genes have an average correlation between predicted and measured expression values of 0.86 (Pearson’s r). Our observation that non-monotonic concentration-response curves were predicted with high accuracy was surprising, given the fact that the model represents the effect of VPA as monotonic functions

(gon(VPA) and goff(VPA)). We therefore analyzed the model simulations with regard to the question how such behavior can occur. An intuitive understanding of this model feature can be obtained from a detailed examination of the GAD1 gene (Fig. 7.3B/C). In this example, gon increases with increasing concentration and goff decreases with increasing concentration. This leads to an increase in amplitude and earlier peaking times as the VPA concentration is increased (Fig. 7.3C). However, because goff is more sensitive to changes in concentration than gon, the increase in amplitude precedes the change in timing. This explain why, at t=144 h expression is ﬁrst increased, then decreased. We concluded that the gene expression model we developed correctly predicts differentiation kinetics at VPA concentrations on which the model was not trained. It furthermore helps to explain non-monotonic concentration-response curves, as monotonic changes in activation and de-activation kinetics can in combination lead to non-monotonic fold change at a given timepoint (e.g. t=144 h), while peak time and maximal expression change monotonically.

Identification of Wnt signaling as major driver for divergent differentiation

Due to non-monotonic and transient effects of VPA, analysis of single time points or at single concentrations might be misleading. We therefore used our gene expression model to predict gene expression for various times and VPA concentrations, and used these predictions to ask which cellular processes are perturbed by VPA during differentiation.

136

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure 7. 4: Up-regulation of Wnt and TGFbeta pathway genes in cells exposed early in the differentiation time course to VPA.

(A) GSEA using curated gene sets from the MsigDB on the fold change between VPA-treated and untreated cells at the indicated time points and concentrations. The significance of enrichment is shown as a heatmap for the top enriched term “Neural crest stem cell up” (q<10-4) The fold change data were measured for different VPA concentrations at the time point 144 h and for different time points at the concentration of 0.6 mM VPA (grey indicates no measurement available). Fold changes for all other times and concentrations are model predictions. (B) GSEA as in A using KEGG pathway gene sets available in the MsigDB. The enrichment scores are shown as heatmaps for the five top enriched pathways (q<0.05), with red indicating enrichment and blue indicating de-richment (grey indicates no measurement available). (C) The fold change between cells treated with 0.6 mM VPA and untreated cells is shown as a heatmap for Wnt-induced and Wnt-repressed genes separately at the indicated time points. Signifi- cance is indicated below the time point (*, p<0.05, Fisher’s test) (D) The average fold change between cells treated with 0.6 mM VPA and untreated cells is shown for the following groups of genes: Genes up-regulated in neural crest cells (Neural crest), genes induced by Wnt signaling (Wnt) and genes annotated as being part of the TGFbeta pathway (TGFb). The symbols indicate the significance of the up-regulation at the respective time point (*, p<0.05; **, p<0.01, ***, p<0.001, Wilcoxon rank sum test) (E) Average concentration-response curves at the differentiation endpoint (t=144 h) for gene groups from D are shown. The symbols indicate significance as in (D). (F) H9-hPSC were differentiated towards neuroepithelial cells and treated with 0.6 mM VPA during different time windows. Two early windows, 24 h-72 h (blue) and 24 h-96 h (green) as well as two late windows, 72 h-144 h (magenta) and 96 h-144 h (orange) were compared to the continuous treatment 0 h-144 h (red). GSEA on the fold change of VPA treated cells compared to untreated cells for neural crest up-regulated genes for different treatment windows: Col- oured lines indicate the enrichment score for the respective treatment window and coloured vertical bars indicate the fold change rank of all neural crest up-regulated genes with respect to the fold change of all genes for the respective treatment window.

137

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Specifically, we predicted fold changes for regulated genes at densely spaced time points (0- 144 h) and VPA concentrations (0-600 µM). Subsequently, we performed gene set enrichment analysis (GSEA) on the predicted fold changes. The term with the maximum enrichment score was “Neural Crest Stem Cell-up (NCSC)”. For the VPA concentration of 0.6 mM, which is associated with clinical birth defects (Waldmann et al. 2014), neural crest stem cell (NCSC) signature genes were enriched from about 50 h onwards. Even at much lower drug concentrations (down to 0.1 mM), enrichment was found, but it occurred at later time points (Fig 7.4A). The model predictions were further used to identify possible mechanisms by which VPA induces NCSC genes. We found 18 KEGG pathways showing a sufficient overlap with our set of regulated genes to compute meaningful p-values. Five of them were significantly enriched: Focal adhesion, P53 signaling, MAPK signaling, TGFβ signaling and Wnt signaling. The time-concentration matrix of the enrichment scores of p53 signaling indicated activation of this pathway at minimal VPA concentrations, and increasing concentration did not enhance this effect. For focal adhesion, TGFβ and MAPK signaling, 300 µM VPA led to an up-regulation, which per- sisted until at least 144 h. The regulation of Wnt-associated genes was particularly interesting, with only a transient up-regulation for about 50 h. Increased VPA concentrations shifted the time frame of up-regulation towards earlier times, so that it was not enriched for at the 144 h data points (Fig 7.4B).The predicted up-regulation of Wnt signaling and TGFβ signaling is re- markable, because both pathways are known to play a role in neural crest development (Leung et al. 2016; Patthey et al. 2009). To further investigate the connection between Wnt/TGFβ signaling and NCSC induction, we derived a Wnt target gene signature (Supplementary table S1) from a transcriptome study in human embryonic stem cells (Mica et al. 2013). When examining the VPA-induced fold changes for these genes at the relevant time period (48-96 h), we observed that Wnt induced genes tended to be up-regulated by VPA, while Wnt-repressed genes were rather down-regulated. This trend is consistent with an activation of Wnt signaling by VPA (Fig. 7.4C).To confirm this hypothesis, we used a more robust approach: average fold- changes caused by VPA treatment for Wnt-activated genes, for TGFβ pathway genes, and NCSC signature genes were calculated for measured time points and VPA concentrations. This data showed that VPA treatment indeed lead to a transient response for Wnt-induced genes. In contrast, NCSC signature genes and TGFβ pathway genes showed increased up- regulation over time (Fig. 7.4D). Predictions from our model suggested that NCSC signature genes were induced already at very low VPA concentrations (Fig. 7.4E). A steep increase in regulation by VPA was observed in the 200-400 µM range, while the average fold change hardly increased at concentrations above 0.5 mM. Such data are of high interest for systems pharmacology models and toxicity prediction algorithms. With respect to our model, it was also 138

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation very interesting to observe that Wnt signaling target genes were not significantly altered, when analyzed at the usual model endpoint (144 h) (Fig. 7.4E). They were only clearly identified by modeling of data for earlier time points. This re-emphasizes the need to provide and analyze kinetic data.The transient activation of Wnt signaling targets could mean that differentiating cells are susceptible to altered Wnt signaling only for a limited duration. This would explain the frequent observation in developmental toxicology of particular windows-of-sensitivity (Balmer et al. 2014; Dreser et al. 2020). To investigate this further, we treated cells with VPA at different times and with variable treatment duration and reanalyzed a published dataset (Dreser et al. 2020; Waldmann et al. 2017) (Fig.7.4F). GSEA for NCSC genes showed a marked difference between early and late treatment windows, with late treatments failing to efficiently upregulate NCSC genes. Brief early treatments led to a very similar up-regulation as continuous treatment. This implies that differentiation disturbances by VPA are most efficient/toxic during early neurodevelopmental periods (Fig.7.4F). Taken together, these applicability tests of our mathematical gene expression model show that this form of data analysis can unveil complex regulation patterns, exemplified here by a transient dysregulation in Wnt signaling associated with a permanent up-regulation of the NCSC fate.

Altered VPA-induced neural crest formation by interference with Wnt signaling

Finally, we tested in how far the toxicant-induced shift of differentiation from neuroepithelial cells towards neural crest cells may be prevented by interference with Wnt signaling. First, a diverse set of Wnt target genes relevant for early neurodevelopment (DACT, SP5, SIX3, GAD) was analyzed to compare the response to VPA with the one triggered by the well-established Wnt activator CHIR (CHIR99021; a specific inhibitor of glycogen synthase kinase-3) (Balmer et al. 2014; Selenica et al. 2007). The analysis was performed after three days of differentiation (DoD3), a time period relevant to fate switching. Both VPA and CHIR triggered similar response patterns (Fig. S7.3A). Moreover, analysis of the established neuroepithelial test endpoints (expression of PAX6 and of OTX2) on DoD6 confirmed that Wnt activation during the differentiation process by CHIR indeed disturbed the correct differentiation. The drug attenuated the up- regulation of PAX6 and of OTX2; in other words, CHIR triggered a relative downregulation of the neuroepithelial markers compared to untreated control cells (Fig. S7.3B). Having established that Wnt signaling may disturb neuroepithelial differentiation in our experimental system, we used the beta-catenin inhibitor ICRT3 as tool to attenuate this pathway. In a proof-of-concept experiment we found that ICRT3 attenuated the expression of typical neural crest genes (TFAP2, PAX3, MSX1) in cells treated with CHIR (Fig. S7.5A).

139

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure 7. 5: Interference with toxicant-induced Wnt activation and neural crest formation.

(A) Pluripotent stem cells were differentiated within 6 days to neuroectodermal precursor cells; in some experiments, these were further differentiated until DoD15, when they formed neural rosettes. The developing cells were either exposed to valproic acid (VPA; 400 µM; from DoD0-DoD6) or they were co-treated with VPA + ICRT3 (beta-catenin inhibitor; 30/60 µM; from DoD2-6). For analysis, gene expression was measured by quantitative real-time PCR (qPCR), marker proteins were evaluated by immunocytochemistry or rosette formation was quantified by immunostaining followed by automated imaging. (B) The mRNA levels of three neural crest marker genes (TFAP2, PAX3, MSX1) was measured after treatment with VPA alone or VPA + ICRT3. Data is shown as fold change relative to the untreated control. (C) Cells were fixed and immunostained for SOX10 and nestin, after treatment with VPA or VPA + ICRT3 (30 µM). (D) Quantification of SOX10 staining in DoD6 cells. Stained cultures were imaged and SOX10-positive pixel were counted. They were normalized to the nuclear area (H-33342-positive pixels). (E) Cells were differentiated and treated as described in A until DoD15. Immunostaining was performed for ISL1 and p75, and quantification was performed as in D. Data is shown relative to untreated controls.

140

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

(F) Cells were treated as described in A until DoD15. Then, the tight junction marker ZO1 (zona occludens 1; red), the golgi marker GM130 (Golgi matrix protein 130 kDa; green) and nuclei (blue, H-33342 DNA-intercalating dye) were stained, and rosette formation was quantified. The left images show untreated cells that form neural rosettes. Each red dot surrounded by a green halo represents one rosette. The two dashed white rings (80 µm diameter) indicate exemplary rosettes. (G) The neural rosette formation assay (RoFA) software was used to quantify rosettes (normalized to the number of cells in a well). Data are given relative to the untreated control. All data are means ± SD of at least 3 replicates, *: p < 0.05; **: p < 0.01; ***: p < 0.001.

Thus, we proceeded to a series of experiments investigating the effect of ICRT3 on cells treated with VPA (Fig. 7.5A). On the gene expression level, 30 µM of ICRT3 already attenuated the VPA-induced up-regulation of TFAP2 and MSX1, but not PAX3 (Fig. 7.5B); at higher concentrations (60 µM), ICRT3 had pronounced attenuating effects on all three neural crest marker genes (Fig. S7.4B). More importantly, the beta-catenin inhibitor also prevented protein expression of the neural crest master regulator SOX10 in the presence of either 400 µM (Fig. 7.5C) or 600 µM VPA (Fig. S7.5A). This phenotypic effect was persistent, as the intermittent treatment with ICRT3 (from DoD2-DoD6) also prevented the up-regulation of neural crest marker proteins ISL1 and p75 (also named TNFRSF16, CD271, p75NTR) at DoD15 (Fig. 7.5D, Fig. S7.5B).These data strongly suggest that the intermittent up-regulation of Wnt pathway genes contributes to the differentiation switch away from neuroepithelial cells and towards neural crest. The rescue from disturbed neuro-differentiation by ICRT3 was further investigated in a rosette formation assay (Dreser et al. 2020; Waldmann et al. 2017). Self-organized formation of rosettes is a hallmark of the neuroepithelial phenotype, and has been considered as functional correlate of neural tube formation (Balmer et al. 2012; Elkabetz et al. 2008). The cells generated by our standard protocol (Dreser et al. 2020) showed extensive formation of neural rosettes, which was completely disturbed in cells exposed to VPA during the pivotal differentiation period of DoD0-DoD6 (Fig. 7.5F). Co-treatment with ICRT3 (from DoD2-DoD6) fully restored the rosette formation capacity of the cells (Fig. 7.5F, G). Thus, interference with Wnt signaling at a critical period of differentiation prevented the neurodevelopmental phenotype triggered by VPA.

141

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.6. Discussion

Analyzing perturbations of differentiating cells is particularly difficult, compared to similar approaches in more or less static cell cultures. For instance, a matrix of time- and concentration- resolved data needs to be analyzed. Here, we used interpolation of a limited amount of data by a kinetic model to make such an analysis feasible. The modeling principle may be applied to any other differentiation process in cell biology and organism development that is affected by an additional factor (signaling molecule, drug or gene dose). Previous genome-wide modeling was mainly focused on the description of gene expression dynamics in terms of RNA production and turnover rates (e.g. (Uhlitz et al. 2017)). This allowed excellent modeling of transcriptome changes triggered by defined signaling events in static cell cultures. Extensions of these models allowed very precise predictions of kinetic rates, particularly when combined with metabolic labeling (de Pretis et al. 2015; Jurges et al. 2018; Rabani et al. 2014). Recent progress also allowed definition of cell trajectories in single cell data (La Manno et al. 2018). A major principle that distinguishes our current work from previous transcriptome descriptions is our model assumption of distinct cellular states in which particular genes are expressed, and that the transition between these states is slow relative to gene expression kinetics. This feature allowed us to model long time series as typically observed during differentiation of human cells. An important finding of the model applications was that a small set of simple assumptions

(e.g. monotonic modeling of gon and goff) can lead to predictions of complex behavior, such as non-monotonic concentration-response curves. Another important conclusion is that few experimental samples may be used to predict unobserved points in a relatively large data matrix. Vice versa, the model may be used in the future to predict the most efficient experimental design to allow robust predictions in particular sub-areas of the differentiation-disturbance matrix. With the help of model simulations, we studied the effect of VPA on differentiating stem cells during the first six days of a well-established neural induction protocol. This model was chosen to exemplify processes where dynamic natural transcriptome changes are overlaid with those triggered by an external stimulus (2-dimensional response matrix). This particular biological model was also used as it is supported by robust clinical data (Blotiere et al. 2019; Bromley et al. 2017). The use of VPA during pregnancy can lead to a complex phenotype, often described as fetal valproate syndrome. Notably, different windows of exposure need to be distinguished. If the drug is given during the period of neural network formation (pregnancy months 2-3 onwards), an autism-related phenotype may arise. This is associated with the endophenotype of altered neurite connectivity, and it has recently been shown to be related to the down-regulation of the cytoskeletal/microfilament regulator MARCKSL1 (Chanda et al. 2019). Drug exposure during earlier phases of development (about 4 weeks after conception)

142

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation is better documented, and it leads to neural tube defects and alterations related to neural crest function (e.g. cleft palate) (Blotiere et al. 2019; Bromley et al. 2017). Spina bifida and hydrocephalus are the clinically best-documented malformations in this case, and the test system used here (STOP-tox(UKN)) refers to this neuro-teratological problem (Balmer et al. 2012). In- terestingly, the upstream signaling disturbances triggered by VPA may be similar in early and late phases: HDAC inhibition and block of GSK-3 activity. The activation of the Wnt pathway, and a related shift from central neural development to neural crest development was found here as a main pathological pathway for early-stage drug exposure. In exposed rats, it was shown that VPA induced beta-catenin and phospho-GSK3 in brain tissue (Gould et al. 2004; Qin et al. 2016). Although such studies suggest that VPA treatment affects Wnt signaling, an exact molecular explanation is still lacking. VPA itself is a poor inhibitor of GSK3, and other molecules that trigger similar responses as VPA (e.g. trichostatin A) do not affect this kinase at all. Possibly, the main primary target of VPA is a broad spectrum inhibition of HDACs (ex- cluding HDAC6), which then leads indirectly to a Wnt response (De Sarno et al. 2002; Wiltse 2005). It has been shown that the direct effects of VPA on histone acetylation have an indirect, but permanent effect on methylations (Balmer et al. 2014). In line with this, it has been suggested that VPA increases the expression of Wnt target genes in prenatally exposed rats by altering demethylation state of Wnt-activators (Wang et al. 2010). Dissecting the transcriptional changes by which VPA affects Wnt signaling is complicated by the dynamic nature of the problem. The initial signaling phase is quickly followed by secondary effects, and signaling pathways often contain feedbacks that limit signal duration. Thus, the timing of the observations after a perturbation can be critical when trying to identify direct effects on signaling, and concentration-dependent effects complicate the analysis. Our approach tries to overcome these limitations by augmenting the limited experimental resolution in time-concentration space, using a dynamic model. Further insights could be gained in the future by combining the model with single cell transcriptome data. This would allow anchoring of the model states to transcriptome states that are observed in cells during differentiation (Farrell et al. 2018; Jang et al. 2017; Saelens et al. 2019; Trapnell 2015). Under such conditions, we would be able to use the same states and their transition rates for all genes, instead of using independent state models for each gene. This would further constrain the parameters and thus provide robustness and interpretability of the model. Consequently, the transition rate parameters could yield typical half lifes of transcriptome states that are currently inaccessible by models using a single cell snapshot. It would also complement pseudotime reconstruction algorithms that lack absolute time scales (Haghverdi et al. 2016; Trapnell et al. 2014).Taken together, our modeling framework allowed us to model the effect of a perturbation on a complex developmental transition

143

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation and helped us to unveil the molecular pathway that mediates dysregulation due to the perturbation. We believe that this framework is broadly applicable for the analysis of transcriptome changes in complex differentiation processes that are subjected to perturbations.

7.7. Availability

The R scripts to generate Figures 1-4 are publicly available in the GitHub repository https://github.com/johannesmg/kinetic_modeling_stem_cell_transcriptome_dynamics. The R scripts used for fitting the dynamic model are also publicly available in the GitHub repository https://github.com/johannesmg/BatemanDiff.

7.8. Accession numbers

The raw CEL files of the Affymetrix chips that have been used are publicly available in the Gene Expression Omnibus (GEO) database under the accession number GSE147270.

7.9. Acknowledgement

This work was supported by the BMBF, EFSA, the DK-EPA (MST-667-00205), and the DFG (Konstanz Research School of Chemical Biology; KoRS-CB). It has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreements No. 681002 (EU-ToxRisk) and No. 825759 (ENDpoiNTs).

144

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

7.10. Supplementary information

Figure S7. 1: Primary antibodies that were used for the study.

Figure S7. 2: Primers that were used for the study.

Figure S7. 3: Wnt signaling and its consequences in the STOP-tox(UKN) assay.

(A) Cells were differentiated until DoD3 using the standard STOP-tox(UKN) assay protocol (= untreated controls), or they were exposed during the differentiation to valproic acid (VPA) or the Wnt activator CHIR. Then, cells were lysed, the mRNA was isolated, and the expression of the Wnt marker genes DACT, SP5, SIX3 and GAD1 (REF?) was investigated (reverse transcription real-time PCR). Data are means ± SD from three experiments. (B) Cells were differentiated as in A, but until DoD6. They were left otherwise untreated (control) or exposed to CHIR from DoD0-DoD6. Then, gene expression of the standard STOP-tox(UKN) assay endpoints PAX6 and OTX2 was measured by PCR. Data (means ± SD, n = 3) is depicted relative to the untreated control. The data confirm previous findings (8) that Wnt activation affects the assay outcome.

145

Results- manuscript 3: Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed ectodermal differentiation

Figure S7. 4: Attenua- tion of neural crest differentiation by the Wnt pathway inhibitor ICRT3.

Cells were differentiated until DoD6 using the standard STOP- tox(UKN) assay protocol, modified by the addition of CHIR (2 µM) or VPA (400 µM), two toxicants shifting the differentiation from the neuroepithelial lineage towards neural crest (8,9). On DoD6, the expression of marker genes was analyzed by PCR. (A) Cells were treated with CHIR alone, or the beta-catenin inhibitor ICRT3 (60µM) was added in addition from DoD2-6. The expression of the neural crest marker genes TFAP2, PAX3 and MSX1 was investigated, and data are indicated as CHIR+ICRT3 vs CHIR alone. Negative numbers on the abscissa indicate a relative downregulation (= attenuation of the CHIR effect) by ICRT3. (B) Cells were treated with VPA alone, or the beta-catenin inhibitor ICRT3 (30 µM and 60µM) was added in addition from DoD2-6. The expression of the neural crest marker genes TFAP2, PAX3 and MSX1 was investigated, and data are indicated as VPA+ICRT3 vs VPA alone. Negative numbers on the abscissa indicate a relative downregulation (= attenuation of the VPA effect) by ICRT3. All data are means ± SD, n=3.

Figure S7. 5: Rescue from toxicant-induced neural crest formation by the beta- catenin inhibitor ICRT3.

(A) Pluripotent stem cells were differentiated until DoD6, and exposed to solvent or valproic acid (VPA, 600 µM) from DoD0-DoD6. They were co- treated or not with ICRT3 (beta-catenin inhibitor; 60 µM) from DoD2-DoD6. Then, SOX10 protein was immunostained, and the number of positive cells was quantified (SOX10 positive pixels were counted automatically and they were normalized to the nuclear area, as identified by H-33342 staining). Data are means ± SD; n=3. The untreated control (Con; differentiation without toxicants or inhibitors) is used as reference to calculate fold-changes of treated cells. (B) Pluripotent stem cells were differentiated until DoD15, and exposed to valproic acid (VPA, 600 µM) from DoD0-DoD6. They were co-treated with ICRT3 from DoD2-DoD6. On DoD15, cells were fixed and immunostained for ISL1 and p75 (the low affinity nerve growth factor receptor NGFR, aliases: TNFRSF16, CD271, p75NTR). The number of positive cells was quantified as in A in 7-16 wells. Data are means ± SD. Data is shown relative to untreated control (Con).

146

Concluding discussion and perspectives

8. Concluding discussion and perspectives

The number of available in vitro DNT test systems is constantly increasing and many developmental processes can be covered already (e.g. neurite outgrowth (Delp et al. 2018; Hoelting et al. 2016; Krug et al. 2013a), neural crest cell migration (Dreser et al. 2015; Nyffeler et al. 2017b), PNS/CNS neuronal degeneration (Scholz et al. 1999). The test method discussed in this thesis, the UKN1 was already used for many studies to investigate important developmental mechanisms and it bears a lot of potential. This thesis consists of three manuscripts that lead to the improvement of the UKN1 test method as a DNT test. Moreover, in manuscript 2 and 3 we investigated its applicability for mode of action studies of compounds influencing gene expression. The first manuscript addresses the improvement of the concentration finding to avoid unspecific toxicity in the system and to distinguish between general toxicity and DNT effects. The use of relevant concentrations is especially important when exploring pathways of toxicity of an interfering compound. Manuscript 1 represents the first that addressed this issue and serves therefore as an important basis for future study planning and execution. The second manuscript addresses and improves the reliability and the predictive capacity. This was accomplished by the addition of a second endpoint to distinguish between irreversible DNT effects and a transient gene expression deregulation. The ability of the cells to form an “in vitro neural tube” was used as a phenotypic endpoint, which connects disturbances in gene expression with later effect in development. The neural rosettes resemble neural tube formation in vitro impressively and provides an endpoint that is strikingly close to the in vivo situation. It was shown, that the phenotype is clearly linked to gene expression changes and gene expression can be used to predict an outcome vice versa. Furthermore, the data were used to determine the distinct time windows of sensitivity for developmental disturbances caused by HDACi. The third manuscripts uses computational analysis based on the gene expression data of developing cells and gives an outlook how the future application of such a test may look like. Here, a statistical model was developed to simulate and predict gene expression changes at unknown concentrations and unknown time points during differentiation. Additionally, it correctly identified pathways that are only transiently up- and down-regulated. As a proof of concept, the pathway prediction based on the classical gene expression endpoint was also supported in this study by the rosette formation endpoint (manuscript 2). The combination proofed to be of value to investigate the pathway of toxicity of VPA.

147

Concluding discussion and perspectives

The following discussion and the demonstration of the evolving perspectives will focus on the UKN1 test method regarding its potential (i) as a “standard” DNT test method as well as (ii) its potential to investigate underlying developmental pathways and mechanisms. The following section will provide an overview of the history and the development of the test method.

8.1. History of the UKN1 test method

The UKN1 assay was initially published in the frame of the ESNATS (Embryonic Stem cell- based Novel Alternative Testing Strategies) project by (Krug et al. 2013b). In this projects, five stem cell based assays using transcriptomics as functional readout were combined into a first pilot test battery. Since this appeared to be a promising approach, a follow-up project was launched and the assay was constantly improved during the last decade. Initially, hESC were differentiated to NEPs within 6 days. However, during the years the differentiation was adjusted to different iPS cell lines to fulfil ethical regulations. A video protocol of the differentiation was published (Shinde et al. 2015), the assay was rated on its readiness for validation (Bal-Price et al. 2018a) and was transferred to another laboratory, recently. Ultimately, in frame of this work, a phenotypic readout serving as an anchor point for the transcriptome changes, was developed. Rosette formation was detected and analysed at DoD15 (Dreser et al. 2020); Re- sults chapter 2). Besides the development of the assay as a tool for testing for DNT compounds, the UKN1 assay was used for the investigation of key processes of development. Furthermore, it played an important role in defining important issues and solutions in toxicity testing strategy. However, the assay has not only improved technically, but also broadened its applicability by enabling the investigation of underlying developmental mechanisms. It was used to investigate how epigenetics are involved in toxicity of a compound (valproic acid). In a first study the system was used to gain a first epigenetic understanding about complex transcriptome changes. The study mainly addressed how transcription factors are downregulated (Balmer et al. 2012). In a second epigenetic study, the switch from reversible to irreversible histone marks was addressed. A “point of no return” was assessed as an epigenetic switch form acetylated (active) to methylated (repressive) histone marks changing differentiation track irreversibly (Balmer et al. 2014). The mechanism playing a role in DNA accessibility and gene expression was investigated by chromatin immunoprecipitation of histone marks. Furthermore, transcriptome data were used to generate interpretation strategies how to handle big data and how to use them for practical applications. The system served to investigate how different compound concentrations affected the transcriptome and helped to define proper con-

148

Concluding discussion and perspectives centrations for DNT testing strategy (BMC10). Concentration-dependent effects on the deregulation of probe sets as well as the effect on biological motifs and pathways was assessed and used to define criteria for a testing strategy (Waldmann et al. 2014). In the frame of this thesis, a pattern for adaptive response and cytotoxicity was found, and biomarkers to distinguish between this two effects were introduced (Waldmann et al. 2017); manuscript 1). A toxicity index was suggested, that analyses the transcriptome based on the DNT genes and the toxicity genes. It addresses the question if a compound affects neurodevelopment and to which extent (Shinde et al. 2016a). Additionally, two groups of compounds (HDACi and mercurial) were applied to the system and a pattern was identified to classify compounds and match them with their respective compound groups (Rempel et al. 2015). The study by Rempel et al allowed a first outlook if and how it will be possible to generate compound class-specific patterns and what is necessary to distinguish between groups. The authors of the study suggested 8 marker genes that were capable to distinguish between the two groups (Rempel et al. 2015) Finally, In frame of this thesis, a computational model was trained based on the transcriptome data of a time course of the unperturbed differentiation as well as gene expression perturbation induced by different VPA concentrations assessed at DoD6 (manuscript 3). The model is able to predict gene expression changes at all time points for different valproic acid concentration. Overall, the assay evolved and improved continuously over the last decade, broadening its applicability domain and significantly contributed to the understanding of the key requirements of DNT test systems. The findings and the development of the assay is summarized in figure 10.

149

Concluding discussion and perspectives

150

Concluding discussion and perspectives

151

Concluding discussion and perspectives

Figure 10: History of the UKN1 test method. Milestone papers were collected and sorted chronologically. An exemplary figure of a graphical summary was selected from each publication. The text on the right indicates the major findings and/or the scientific advances brought forward by the respective publications (red script). Green boxes indicate further development of the UKN1 assay. The following publications were included: [1] (Chambers et al. 2009) [2] (Balmer et al. 2012), [3] (Krug et al. 2013a), [4] (Balmer et al. 2014), [5] (Waldmann et al. 2014), [6] (Shinde et al. 2015), [7] (Shinde et al. 2016a), [8] (Rempel et al. 2015), [9] (Waldmann et al. 2017), manuscript 1, [10] (Bal-Price et al. 2018a), [11] (Dreser et al. 2020) manuscript 2, [12] manuscript 3.

8.2. Relevance and uniqueness of the test system

As described above, the test method has a long history and was used as a tool to investigate mechanisms that drive early neurodevelopment. Reproducibility and robustness were shown by intra and inter-laboratory transferability underlying the quality of the test. However, some shortcomings have not yet been addressed. For example, UKN1 possesses rather low throughput compared with other developed DNT models, e.g.(Krug et al. 2013a), (Delp et al. 2018), (Dreser et al. 2015; Nyffeler et al. 2017b). However, tests based on differentiating cells as cell model are sparse. Comparable tests, such as the mouse EST (Seiler and Spielmann 2011), which can be used as high throughput, possess the disadvantage that the test could not be transferred so far to a human EST. Other similar approaches include e.g. the modeling of the “germ layer formation” (de Leeuw et al. 2020; Shinde et al. 2016b). These tests, in contrast to the UKN1 differentiation, use undirected differentiation. This means that the pathway driving the development is not triggered by externally added compounds like in the UKN1. On one hand, the cells are not “manipulated” and will react on a compound stimuli without interference of other small molecules. On the other hand, undirected differentiation might also not reflect the in vivo situation one-to-one since in vivo many small molecule gradients guide the development in a complex way. Due to the mixed cell population, the cells likely only partly respond to a toxicant, which might lead minor cellular changes that cannot be recognized. Additionally, in such systems the assessment of the mode of action of a compound is more difficult since the underling pathways leading to the differentiation are unknown. Using rosette formation as a readout for a DNT test method has been previously proposed by Colleoni at al. in 2011, but it was only used for low scaling without automatization and detection effects on single rosettes (Colleoni et al. 2011; Colleoni et al. 2012). Another highly complex model addressed the neural rosette dynamics by investigation of interkinetic nuclear migration during development of early neural rosettes and early radial glia rosettes. The dynamics were also suggested for drug screening purposes (Ziv et al. 2015). Recent evidences showed that DNT high throughput tests not always are able to detect DNT compounds. It has been shown that compounds, which disturb the UKN1 differentiation such 152

Concluding discussion and perspectives as HDACi (which are also known in vivo to induce NTD (Omtzigt et al. 1992; Vajda et al. 2004) are missed in test batteries without differentiation systems (unpublished data: EFSA). This can be explained by the fact that the UKN1 test method detects long term (secondary) effects rather than direct toxicity. This makes it on one hand unique compared to other (fast) tests. On the other hand difficulties arise to interpret data and to follow direct compound effects. By capturing the effect at DoD6 (gene expression) or even later at DoD15 (rosette formation) the secondary effect manifests as a result of many different processes (toxicity and compensatory pathways) that accumulate. The differentiating cell culture probably consist of different cell populations. These conditions make it difficult to follow and to track the compounds’ mode of action. However, at the same time this represents an interesting and challenging research question, which will be/can be addressed by single cell sequencing in future studies. These studies will allow the tracking of different cell populations over time and follow the changes that are introduced by compounds. Single cell sequencing has for example been used to track development in gastruloids (van den Brink et al. 2020). The single cell data might be used to define unique marker patterns for different cell populations and then allow to monitor shifts in populations e.g. by using reporter or staining. In the past, using transcriptome changes as a toxicological endpoint has been criticized in terms of data interpretation since the significance, relevance and the reversibility of the expression deregulation is unknown (Corvi et al. 2016). By the addition of the phenotypic anchor to the transcriptome changes such criticism can be circumvented. Moreover, as demonstrated in manuscript 2, gene expression and phenotypic endpoint showed a clear correlation and the number of relevant genes could be reduced to a minimal marker set that may serve as a fingerprint. The identification and testing of appropriate concentrations is also currently under debate. This issue is of high relevance, and it was therefore addressed for the UKN1 system as well (manuscript 1). To identify appropriate testing concentrations, different strategies might be followed. The standard strategy for UKN1 is to determine the highest non-cytotoxic concentration, which includes performing a general toxicity test (resazurin assay). In a second step, the EC10 concentration or if no cytotoxicity is observed, the highest plausible concentration (e.g. cmax or higher) is tested in the assay. By this approach threshold concentrations for risk assessment can be investigated and specific DNT compounds are filtered, while unspecific cytotoxic effects are avoided. Other concepts, that first seem logical and easier to follow such as using fixed concentrations like the cmax or 20 fold cmax (Hengstler et al. 2020; Sjogren et al. 2018) are not applicable for the UKN1 assay since compounds that show general cytotoxicity would also be classified as hit regardless of the mode of action.

153

Concluding discussion and perspectives

Thinking in terms of high throughput, compound concentration that induced general toxicity markers could then be excluded from further analysis as proposed in manuscript 1.

8.3. Readiness for validation

The UKN1 test method has shown in various study designs that it is a highly promising approach. It justified its existence in the light of other tests and has also been part of studies in collaboration with pharmaceutical industry. Yet, to allow a wider implementation for regulatory or industrial purposes, all test systems must undergo a validation process. Official validation of a test method is used to proof that a test can be applied for a specific purpose. The key characteristics are assessed (precision, limit of detection, accuracy, specificity, sensitivity, robustness and transferability) (Leist et al. 2012a). In Europe, the European Center for Validation of Alternative Methods (ECVAM) is responsible for the organization of validation of submitted alternative methods. This process includes the mediation of experts, stakeholders, laboratories and the organization of huge ring trials. However, the main focus for validation lies on the reproducibility and the predictivity in comparison to conventional animal based tests that remain as gold standard. This represents a promising strategy for validation of tests that can be compared 1:1 to a traditional animal test (e.g. eye irritation, skin irritation (Liebsch et al. 2000; Roberts and Patlewicz 2014)). Unfortunately, this strategy is not applicable to a DNT-test method that does not aim to replace one specific animal experiment. Therefore, a purpose oriented, mechanistic based validation strategy was suggested (Hartung et al. 2013a). The emphasis is moved to mechanism, causality and biological relevance. However, to make use of mechanistic validation, knowledge of the mechanism in the test method is required. This should be addressed in a straight forward process by applying the Koch/Dale postulates to toxicity testing (“show it, block it, induce it to confirm it”). In complex systems, this might be not so easy to follow. Additionally, Bradford/Hill criteria (strength, consistency, specificity, tempo- rality, biological gradient, plausibility, coherence, experiment (similar factors)), have also to be taken into consideration. In a practical approach, compounds or better compound groups with known mechanism should be selected and used for mechanistic characterization and validation of the test that identifies these toxicants. Models are proposed that lack complexity and address only one biological and DNT relevant pathway. The authors suggest to “move away from correlation of phenomena towards molecular description of pathways” (Hartung and McBride 2011). The following step by step strategy was suggested to mechanistically validate (Hartung et al. 2013a):

“…– Condense the knowledge of biological/mechanistic circuitry (in the absence of xenobiotic challenge) underlying the hazard in question”. 154

Concluding discussion and perspectives

This has been addressed already before the UKN1 test method was set up. Early neurodevelopment includes the differentiation of the zygote to the neuroectoderm and the folding and formation of the neural tube, which further develops into the spinal cord and the brain. The hazard represents misguided development that results in congenital malformations like spina bifida or even more severe outcomes.

“– Compile evidence that reference chemicals leading to the hazard in question perturb the biology in question, i.e., mainly pathway identification by using reference substances in valid(ated) models and experimental proof of their role”. HDACi bear the hazard of causing congenital malformations in vivo. This has been shown in animal model as well as in epidemiological studies (Omtzigt et al. 1992; Vajda et al. 2004)

“– Develop a test that purports to reflect this biology” The UKN1 reflects the early neurodevelopment in vitro (Balmer et al. 2012; Chambers et al. 2009)

“- Verify that toxicants shown to employ this mechanism also do so in the model” HDACi deregulate genes during differentiation, as an outcome, HDACi prevent rosette formation (manuscript 2).

“- Verify that interference with this mechanism hinders positive test results…” HDACi induce the TGFbeta/Wnt pathway. By blocking it, rosette formation can be restored (manuscript 3).

In manuscript 2 a mechanistic validation is discussed since the newly developed endpoint of rosette formation enables now the correlation between the two endpoints. This strategy has been followed before for example by the Piersma group on whole rat embryo culture and zebrafish models (Hermsen et al. 2013; Tonk et al. 2013). The validation process description presents a biological confirmation for the test development. However, it does not consider all the technical questions that have to be answered to validate such a test. These issues will be addressed in the following section in more detail.

8.3.1. Readiness criteria: How far did we get?

A roadmap to assess new approach method (NAM) readiness for validation has been recently provided (Bal-Price et al. 2018a). In this publication, readiness criteria were defined and 17 test systems were challenged and rated based on a point system to evaluate its readiness for

155

Concluding discussion and perspectives validation. The rational for such a study was to define pre-validation criteria that allow research- ers to judge if a test system can be submitted for the validation process. The categories were split in three phases: - Phase I: evaluation of the test system, exposure scheme, documentation/SOP, main endpoints, cytotoxicity, test method controls and data evaluation. - Phase II: evaluation of the testing strategy, the robustness, test benchmarks, the prediction model and the applicability domain. - Phase III: evaluation of the screening hits. The UKN1 test system was rated according to the corresponding questionnaire and reached readiness marks in Phase I: A, Phase II: B, Phase III: A (overall A- to B+). This indicates that minor improvement is required to consider this assay as ready for regulatory approval, especially in Phase II. However, in the meantime the assay has improved and several key features were added. The following section will address the category II questions that did not reach full score in the former rating:

Category/Subcategory/ Question/ ”old” score New rating

Testing strategy- role in battery: Full score for Test is not standalone. To gain information about rela- standalone tests/ information on their relation to other tion to other tests it has to be included in a test battery. tests in a battery is required. (0/1) (0/1)

Robustness- inter-lab: Data available on transferability Assay was successfully transferred to another lab. /reproducibility in another lab. (0/1)  (1/1)

Prediction model-confirmation: Experimental testing of Prediction model was set up and confirmed in manu- prediction model; confirmation of function/predictivity. script 2.  (1/1) (0/1)

Prediction model- limitations: information on limitations Borderline range was defined (manuscript 2). Further of prediction model and how exceptions and special testing for borderline compounds required. However, cases have to be handled. (0/1) no special handling procedures defined.  (0.5/1)

Applicability domain - AOP: Information contributed to Has been set up (AOP wiki 275) and improved in frame an AOP KE/MIE; element of a KE testing battery. (0/1) of this thesis (next chapter).  (1/1)

The new rating of the UKN1 test method resulted in16.5 points for Category II (A mark). This indicates that the readiness criteria for a test method are fulfilled (overall A mark). Its readiness provides an important information for the use in industry and also for regulators. Other NAMs that reached A scores were addressing neuronal differentiation (hNPC3), oligodendrocyte differentiation (hNPC5), hNPC proliferation (hNPC1), hNPC migration (hNPC2) (Baumann et al. 2014), NCC proliferation and migration (UKN2-(Nyffeler et al. 2017b)), neurite outgrowth of CNS neurons (UKN4 (Krug et al. 2013a)), neurite outgrowth of PNS neurons (UKN5 (Hoelting

156

Concluding discussion and perspectives et al. 2016)), astrocytes, oligodendrocytes, myelination, microglia in 3D rat (3Dr (Zurich et al. 2013)). Most of the mentioned methods show a high throughput degree. However, some of the readouts also require longer differentiation processes. For example the hNPC1 uses as readout BrDU incorporation at DoD3 as well as DoD14 neurosphere diameter. For the UKN1 still some technical issues have to be solved to allow a higher throughput and a tiered testing strategy has to be designed that would include testing of already prioritized compounds. The incorporation of the test in a comprehensive test battery will allow to further define its limitations and its unique selling points.

8.4. Biological mode of action of HDAC inhibitors

8.4.1. NEPs differentiate towards neural crest cells

The effect of HDAC inhibitor exposure during development is well known and was investigated extensively. Also the effect of this chemical on the UKN1 model has been shown in several studies and has also been summarized in the description of an adverse outcome pathway (Chapter 3.5). Even if the mechanism of HDACi leading to congenital malformations remains unknown, many publications suggested a shift towards neural crest fate as a final outcome of HDACi intoxication during neural tube formation ((Murko et al. 2013); manuscript 1, 2, 3). This could also be confirmed on several levels in this thesis: (I) the investigation of neural crest marker genes on protein level (manuscript 2, figure 2, chapter 3, Suppl), on gene expression level (manuscript 2 and 3) as well as the expression based computational prediction supported the theory of neural crest cell development (manuscript 3, figure 7.4). Furthermore, the analysis of perturbed GOs and KEGG confirmed these findings (Waldmann et al. 2014). Neural crest cells arise from the neural tube border and the first waves of neural crest cells start delaminat- ing even before the neural tube fuses (Hörstadius 1950). However, the neural plate border is part of the neural plate mediating the final neural tube fusion. Additionally, it represents the contact area to the non-neuronal ectoderm that fuses in parallel to the neural tube and sepa- rates from the neural tube after fusion. It seems reasonable that the NEP differentiation track could easily be changed to a neural plate border fate and then towards neural crest cell differentiation by the HDAC inhibitor. Why this occurs remains unclear, it might just be mediated by a tipping on the scales - changing the balance and therefore inducing the differentiation towards the nearest relative. On the same time, the cell lineage commitment of the NEPs might also play an important role. The cells may already be restricted to a certain fate no matter what external trigger is applied.

157

Concluding discussion and perspectives

In vivo, the developing neural tube consist of many different cell types (regionally specific) that react specifically to different signaling molecules. However, neural rosettes are way smaller structures consisting of less cells and they are not exposed to different molecule gradients. The in vitro situation might therefore be different: while in vivo only the neural tube border contributes to NCC generation, in vitro there might not exist such a regional specificity, but the cells probably consist of a mixed cell population. If gene expression is deregulated, the most obvious result is a change of differentiation track towards a closely related cell type. This might even be strengthened by the fact that only the expression of two transcriptions factors (PAX3 and ZIC1) have been claimed to be both necessary and sufficient to induce a neural crest fate (Pla and Monsoro-Burq 2018). Until now, the whole cell population was used for transcriptome analysis. In future, single cell sequencing will enable to clarify if only particular cell types are affected and which are the underlying mechanisms. The computational modeling of the differentiation (manuscript 3, Figure 7.4) revealed that the signaling pathways TGFbeta and Wnt are activated at an intermediate time point by VPA. For dorsal patterning during neural tube formation and the induction of the non-neuronal ectoderm and neural plate border, BMP and TGFbeta are the main signaling molecules that are responsible (Figure 11). Wnt is activated thereby. This strengthens the evidence for a “dorsalizing” effect induced by VPA (Figure 11). Upon HDCAi treatment, also neural crest cell-derived sensory neurons have been detected (manuscript 2, figure 6.2), additionally to epithelial like structures (manuscript 1 and 2). These structures may resemble the non-neuronal ectoderm, further underlining the “dorsalizing” effect on the neuroectodermal cells.

Figure 11 Dorsalizing effect of HDACi exposure durng neural tube development.

If this dorsalizing effect would be a common feature of the system, a new endpoint could even be defined and assessed by specific markers. Other compounds such a retinoic acid and BIO

158

Concluding discussion and perspectives have been used to generate potential markers for testing purposes (Manuscript 2, suppl). How- ever, they were not yet analyzed for their effect on transcription on a GO/KEGG pathway level. Such an analysis might help to understand if a dorsalizing effect is a common mechanism or even represents the only general outcome of the UKN1 assay. Interestingly, GOs and KEGG pathways have been analyzed for varying valproic acid concentrations. It was reported that the severity of transcriptome changes amplified with increasing concentrations of VPA until a plat- eau is reached. However, apparently a further change in differentiation track does not occur. At some point, the concentration just becomes cytotoxic and marker/pathways for cytotoxicity become active. It remains open if other compounds would induce changes of the differentiation track, leading to the development of a completely different cell type or if the cells at that differentiation state only have the choice between neural crest or neural progenitors. Former studies have shown that valproic acid also leads to the induction of e.g. cardiac markers since genes of the GO “cardiac development” were significantly overrepresented, but also GOs related to “renal system development” or “bone morphogenesis” were detected (Balmer et al. 2014). This would indicate that cells try to differentiate also to other germ layers. However, also neural crest cells differentiate into parts of the heart as well as to cells contributing to various organs of the whole body, which may serve as an alternative explanation for this. A comprehensive characterization of this issue will need to be further elucidated by analysis of the effect induced by compounds interfering with various pathways.

8.4.2. Development of an AOP

By investigation of the effects caused by valproic acid in the UKN1 test, a huge background knowledge was gained. The gathered knowledge can now be used to set up a detailed adverse outcome pathway, which might be linked to other AOPs. Since HDACi exert part of their effect by changing the DNA accessibility, a massive secondary effect in gene expression is induced. The primary effect is hard to distinguish and to capture. To study the primary effects of HDACi, the first time points after treatment are of interest: HDACi inhibit deacetylation of histones, which in consequence leads to more acetylation or to a stabilization of histones that are acetylated already. Acetylation is known to act as an “active state” histone mark, that leads to a DNA-relaxation, making it accessible for transcription (Hezroni et al. 2011; Kamieniarz et al. 2012; Nightingale et al. 2007). In theory, the primary effect of HDACi is to maintain a high expression of already highly expressed genes. This is shown by looking at the modeling curves of manuscript 3, figure 7.2 a rough pattern. If genes

159

Concluding discussion and perspectives are already upregulated before 50 hours, addition of VPA leads to even more/faster upregulation. A secondary effect can be observed shortly after this 50 hours. It has been shown before (Balmer et al. 2014; Balmer and Leist 2014; Balmer et al. 2012), that the increase of the first acetylation is reversed back to baseline after 48 hours. Moreover, an epigenetic switch occurs, which results in histone methylation (silencing marks) of the transcription binding sites. There- fore downregulation of e.g. PAX6 or OTX2 does not represent the primary, but only the secondary VPA-effect. However, in this special case the secondary effect is hard to separate from the primary effect. At later time points the computational model proved to be able to capture and predict whole pathway patterns (manuscript 3, Figure 7.4). This approach allowed to identify the activation of intermediate pathways-mainly the TGF beta/Wnt pathway. This leads to a change in differentiation track in (parts of) the cell population resulting in the final adverse outcome of rosette disturbance (neural tube defects in vivo). This theory based AOP is also represented in figure 12.

Figure 12 Setup of a new „theory based adverse outcome pathway”. A theory based AOP was generated based on the findings of this study in combination with evidence from earlier studies ((Balmer et al. 2014; Balmer et al. 2012), AOPwiki no. 275)

AOPs for other compound classes might be generated in a similar manner:

Does a compound disturb rosette formation? ( no: safe)  Yes How is gene expression affected?

160

Concluding discussion and perspectives

- Apply computational model on defined time points and concentrations  investigate pathways that become activated and confirm by blocking/activating

- How is the cell population affected? Apply markers for subpopulation (described in chapter 8.2)

The AOP concept represents a promising, straight forward approach to summarize the critical events from an initiating event to the final (disease) outcome in general. However, the approach has also limitations. For example, the structure of hitherto existing AOPs are (i) far too specific to combine several AOPs or even finding common key events will be hard to facilitate (Bal-Price et al. 2015; Leist et al. 2017; Smirnova et al. 2014; Spinu et al. 2019). (ii) Often also the evidence for a link to a disease outcome is rather low and additionally (iii) a certain key event 1 that leads to a certain key event 2 in one AOP could in another AOP not lead to a key event or to a different 2nd key event. In a worst-case scenario the AOPs may not even share an adverse outcome. Still, if all the vulnerable PoTs/AOPs of a certain test method are known, it might help to classify new unknown compounds by e.g. quantitative structural analysis relationship (QSAR).

8.5. Limitations and outlook

The aim for new approach methods (NAM) for DNT testing is to gain knowledge about all vulnerable pathways that drive development. If a compound was detected as a hit in the respective test, the mode of action would be much easier to understand and also QSAR and other predictive approaches could be used with a higher accuracy based on such knowledge. For the UKN1 test method, many steps have been taken already to gain understanding. Nev- ertheless, it represents still a rather complex test system that is hard to capture in every detail. The interdisciplinary work in systems biology projects has proven to be of high value to solve such complex questions. With increasing knowledge about gene expression and pathway based analysis, also the quality of the UKN1 test assay increased constantly. Single cell sequencing will probably deliver a more detailed understanding of the real-life situation and probably pave the way for further development of the test. Until now, mainly HDAC inhibitors have been used for investigation. This might have led to an underestimation of other important pathways. For example, retinoic acid exhibits a crucial effect on UKN1 differentiation and rosette formation. Similarly, endocrine disrupting chemicals might also be an interesting target to look at since the accumulating effects in the assay might show consequences that will be missed in other fast assays. Furthermore, compounds that

161

Concluding discussion and perspectives affect shh (e.g. cyclopamine (Feutz and De Geyter 2019)) or notch signaling (Main et al. 2013) might also lead to a breakdown of rosette formation. Hence, including other compound classes and tool compound sets will be essential in the further development of the test and significantly contribute to the understanding of its strengths and weaknesses. Furthermore, the UKN1 test represents a directed differentiation. Differentiation towards the neuronal lineage is initiated and directed by externally added morphogens (smad inhibition). These strong stimuli could cause the system to be confined to its developmental path. Unlike for systems using undirected differentiation and monitoring changes thereof, compounds causing minor changes in essential pathways could be neutralized by the morphogens being present in high concentrations. The consequence could be a rather low sensitivity, because the compounds might need to have a strong effect to change the differentiation track of the developing cells. This could also serve as an explanation, why the differentiation track is often pushed towards a neural crest cell fate-the nearest relative- and not towards a completely different cell fate of another germ layer e.g. mesoderm. However, if this extrinsic controlled differentiation has a higher or smaller compensatory potential, as its in vivo counterpart, has not been addressed yet. This needs further clarification by testing more tool compounds to define the predictive power of the assay and determine sensitivity as well as the specificity. For future studies, the possibility to transfer the assay to iPSC lines is of high interest. It was shown that the assay can be transferred also to other cell lines, even if the protocol had to be adjusted for every cell line respectively. The use of iPS cell lines will further intensify since the use of embryonic stem cells for DNT testing is restricted by ethical regulations. With varying regulations between different countries, a validation of such a test method would use rather iPSC lines, to facilitate international use and applicability of the method. In the meanwhile, other side projects or even side assays have been arising from the classical UKN1 test method. Another promising DNT test, which uses pre-differentiated neuroectodermal precursors and monitors rosette formation only, was developed. Such an assay assesses not the differentiation process itself, but rather the physical process of the rosette formation. It could be used as in a high throughput, since the NEPs can be pre-differentiated and stored frozen. Rosette formation is done in a small well format (96well). It has been shown in time window experiments that compounds that affect the classical UKN1, do not affect this late rosette formation assay. It requires further clarification whether such an assay adds up the value of a test battery or if it gives the same results as a neural crest migration assay (Nyffeler et al. 2017b). This work showed that the UKN1 assay is ready to be validated from a biological as well as from a technical point of view. Since many projects are still ongoing, also the sensitivity and

162

Concluding discussion and perspectives the specificity can be defined soon. Apart from that it is constantly improving regarding technical details which make it less time and money consuming. Considering this, the UKN1 assay is expected to be ready to use in terms of commercial purposes in the near future.

163

Abbreviations

9. Abbreviations

AOP Adverse outcome pathway Benchmark concentration leading to 10% / 25% de- BMC10/25 crease compared to control BPA Bisphenol A

BR Borderline range

CNS Central nervous system

CsA Cyclosporin A

DMSO Dimethylsulfoxide

DNT Developmental neurotoxicity

DoD Day of differentiation

FC Fold change

FGF Fibroblast growth factor

GO Gene onthology

HDAC Histone acetyl transferase

HDACi Histone deacetylase inhibitor hESC Human embryonic stem cells

IFN beta Interferon beta iPSC Induced pluripotent stem cells

KE Key events

KNDP Key neurodevelopmental processes mEST Mouse embryonic stem cell test

NAM New approach methods

NCAM Neural cell adhesion molecule

NDD Neurodevelopmental distance

NEP Neuroepithelial/Neuroectodermal precursor Organization for economic co-operation and develop- OECD ment 164

Abbreviations

oGO: Overrepresented GO

PCA Poly component analysis

PCMB P-chloromercuribenzoic acid

PMA Phorbol 12-myristate 13-acetate

PS Probesets

PSC Pluripotent stem cells

PoT Pathway of toxicity

RA Retinoic acid Registration, evaluation, authorisation and restriction REACH of chemicals (EC 1907/2006) RoFA Rosette formation assay

SD Standard deviation Stem cell-based teratogenic omics prediction-UKN STOP-tox(UKN) toxicity assay (Shinde et al. 2016a), previously named UKN1 T Threshold

TEP Toxicological endophenotype

TSA Trichostatin A

U(T) Uncertainty of threshold

UKN1 University of Konstanz (1) assay (Krug et al. 2013b)

VPA Valproic acid

Wnta Wnt activator

165

Record of contribution

10. Record of contribution

Chapter 1:

For the manuscript „ Stem cell transcriptome responses and corresponding biomarkers that indicate the transition from adaptive responses to cytotoxicity” my contribution was to develop a phenotypic endpoint to anchor the transcriptome findings to a relevant outcome. This was done by assessment of rosette formation. Furthermore, I participated in the manuscript development by discussions together with collaborators and co-authors.

Chapter 2:

For the manuscript “Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances” the major part of the laboratory work as well as the writing process was performed by myself. I developed the rosette formation assay and I performed all differentiations and experiments related to the assay. My co -authors contributed as follows: The KNIME workflow was set up together with Timo Trefzer and Marcel Wiedemann, some images were taken by Simon Gutbier, Petra Kranaster and Christoph van Thriel. Some of the affymetrix samples were obtained by Stefanie Klima, affymetrix samples were measured in Cologne by Agapios Sachinidis and Margit Henry and the statistical analysis of the transcriptome data was performed by Katrin Madjar and Jörg Rahnenführer. I was supervised by Tanja Waldmann and Marcel Leist, who also contributed to the writing process.

Chapter 3:

For the manuscript “Kinetic modeling of stem cell transcriptome dynamics to identify regulatory modules of normal and disturbed neuroectodermal differentiation” all the experimental data were obtained by me. I differentiated the cells and provided all the transcriptome samples. I performed all the thesis supporting qPCR experiments as well as all experiments that are related to the rosette formation assay. The rescue experiments were as well planned and performed by myself. Johannes Meisig (Berlin) developed the computational model based on my data which I further confirmed by qPCR analysis. Johannes Meisig did the gene enrichment analyses for hypothesis generation, which I verified by phenotypical rescue experiments. I was

166

Record of contribution

supervised by Marcel Leist and Tanja Waldmann. Nils Blüthgen, Johannes Meisig, Tanja Waldmann and Marcel Leist were involved in the writing process.

167

Danksagung

11. Danksagung

Meine Doktorandenzeit war etwas außergewöhnlicher als bei meinen Kollegen und natürlich stark geprägt durch die Geburt meiner beiden Kinder Josha und Remy. Daher möchte ich mich an diese Stelle nicht nur für die übliche Betreuuung, den Input und die Förderung während meiner Doktorandenzeit bedanken, sondern vor allem auch für die Menschlicheit, die unendliche Geduld, Hilfe und Unterstützung die ich während dieser (doch etwas längeren Zeit als zunächst geplant) von euch allen, erfahren habe.

Dabei möchte ich zunächst meinem Doktorvater Marcel Leist danken, der mich trotz der Verzögerung nie unter Druck gesetzt hat und mir ermöglicht hat, meine Doktorarbeit und meine Publikationen so abzuschließen und trotzdem für meine Kinder da sein zu können. Die außergweöhnliche Flexibiliät, die du mir gewährt hast, kann man von den wenigsten Chefs erwarten.

Außerdem danke ich Alexander Bürkle und Daniel Dietrich für die Übernahme des Zweitgutachtens und des Prüfungsvorsitzes.

Besonders bedanken möchte ich mich bei der ganzen AG Leist für die Freundschaft und Hilfbereitschaft. Insbesondere möchte ich Tanja Waldmann, Stefanie Klima, Giorgia Pallocca und Marion Kapitza danken! Ihr habt mich immer mit Rat und Tat unterstützt und habt mir insbesondere während meiner Schwangerschaften enorm geholfen! Vielen vielen Dank!

Meinem Mann und Ex-Kollege Simon danke ich für seine Liebe, für unsere Zeit, für seine Unterstützung im Labor, dafür, dass er mich immer wieder aufbaut, wenn ich an mir zweifle, dafür, dass er mit mir gemeinsam das Abenteuer Doktorand und Familie gewagt hat, für sein Verständnis und für so vieles mehr. Du bist auch nach über zehn Jahren noch mein bester Freund. Natürlich danke ich auch unseren gemeinsamen Kindern Josha und Remy, die die Fertigstellung meiner Doktorarbeit zwar enorm aufgehalten, mein Leben aber auch mit unendlicher Freude gefüllt haben.

Ich danke außerdem meiner ganzen Familie, vor allem meiner Mama und meiner Oma, dafür dass ihr immer hinter mir gestanden seid und mir ermöglicht habt, all diese Jahre zu investieren. Für eure Hilfe mit den Kindern, für euer Interesse und für eure Liebe. (Ohni euch hätt ich des nemols gschafft  ).

168

References

12. References

Abranches E, Silva M, Pradier L, Schulz H, Hummel O, Henrique D, Bekman E (2009) Neural Differentiation of Embryonic Stem Cells In Vitro: A Road Map to Neurogenesis in the Embryo. PloS one 4(7) doi:ARTN e6286 10.1371/journal.pone.0006286 Ahlborn GJ, Nelson GM, Ward WO, Knapp G, Allen JW, Ouyang M, Roop BC, Chen Y, O'Brien T, Kitchin KT, Delker DA (2008) Dose response evaluation of gene expression profiles in the skin of K6/ODC mice exposed to sodium arsenite. Toxicol Appl Pharmacol 227(3):400-16 doi:10.1016/j.taap.2007.10.029 Ahlstrom JD, Erickson CA (2009) The neural crest epithelial-mesenchymal transition in 4D: a 'tail' of multiple non-obligatory cellular mechanisms. Development 136(11):1801-12 doi:10.1242/dev.034785 Alexa A, Rahnenfuhrer J, Lengauer T (2006) Improved scoring of functional groups from gene expression data by decorrelating GO graph structure. Bioinformatics 22(13):1600-7 doi:10.1093/bioinformatics/btl140 Alsdorf R, Wyszynski DF (2005) Teratogenicity of sodium valproate. Expert Opin Drug Saf 4(2):345-53 Amarnath S, Agarwala S (2017) Cell-cycle-dependent TGFbeta-BMP antagonism regulates neural tube closure by modulating tight junctions. J Cell Sci 130(1):119-131 doi:10.1242/jcs.179192 Amit M, Carpenter MK, Inokuma MS, Chiu CP, Harris CP, Waknitz MA, Itskovitz-Eldor J, Thomson JA (2000) Clonally derived human embryonic stem cell lines maintain pluripotency and proliferative potential for prolonged periods of culture. Dev Biol 227(2):271-8 doi:10.1006/dbio.2000.9912 Andersen ME, Clewell HJ, 3rd, Bermudez E, Willson GA, Thomas RS (2008) Genomic signatures and dose-dependent transitions in nasal epithelial responses to inhaled formaldehyde in the rat. Toxicol Sci 105(2):368-83 doi:10.1093/toxsci/kfn097 Aschner M, Ceccatelli S, Daneshian M, Fritsche E, Hasiwa N, Hartung T, Hogberg HT, Leist M, Li A, Mundi WR, Padilla S, Piersma AH, Bal-Price A, Seiler A, Westerink RH, Zimmer B, Lein PJ (2017) Reference compounds for alternative test methods to indicate developmental neurotoxicity (DNT) potential of chemicals: example lists and criteria for their selection and use. Altex 34(1):49-74 doi:10.14573/altex.1604201 Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G (2000) Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 25(1):25-9 doi:10.1038/75556 Attarwala H (2010) TGN1412: From Discovery to Disaster. J Young Pharm 2(3):332-6 doi:10.4103/0975-1483.66810 Baker N, Boobis A, Burgoon L, Carney E, Currie R, Fritsche E, Knudsen T, Laffont M, Piersma AH, Poole A, Schneider S, Daston G (2018) Building a developmental toxicity ontology. Birth Defects Res 110(6):502-518 doi:10.1002/bdr2.1189 Bal-Price A, Crofton KM, Leist M, Allen S, Arand M, Buetler T, Delrue N, FitzGerald RE, Hartung T, Heinonen T, Hogberg H, Bennekou SH, Lichtensteiger W, Oggier D, Paparella M, Axelstad M, Piersma A, Rached E, Schilter B, Schmuck G, Stoppini L, Tongiorgi E, Tiramani M, Monnet-Tschudi F, Wilks MF, Ylikomi T, Fritsche E (2015) International STakeholder NETwork (ISTNET): creating a developmental neurotoxicity (DNT) testing road map for regulatory purposes. Arch Toxicol 89(2):269-87 doi:10.1007/s00204-015-1464-2

Bal-Price A, Fritsche E (2018) Editorial: Developmental neurotoxicity. Toxicol Appl Pharmacol 354:1-2 doi:10.1016/j.taap.2018.07.016 Bal-Price A, Hogberg HT, Crofton KM, Daneshian M, FitzGerald RE, Fritsche E, Heinonen T, Hougaard Bennekou S, Klima S, Piersma AH, Sachana M, Shafer TJ, Terron A, Monnet-Tschudi F, Viviani B, Waldmann T, Westerink RHS, Wilks MF, Witters H, Zurich MG, Leist M (2018a) Recommendation on test readiness criteria for new approach methods in toxicology: Exemplified for developmental neurotoxicity. Altex 35(3):306- 352 doi:10.14573/altex.1712081 Bal-Price A, Pistollato F, Sachana M, Bopp SK, Munn S, Worth A (2018b) Strategies to improve the regulatory assessment of developmental neurotoxicity (DNT) using in vitro methods. Toxicol Appl Pharmacol 354:7-18 doi:10.1016/j.taap.2018.02.008 Balmer NV, Klima S, Rempel E, Ivanova VN, Kolde R, Weng MK, Meganathan K, Henry M, Sachinidis A, Berthold MR, Hengstler JG, Rahnenfuhrer J, Waldmann T, Leist M (2014) From transient transcriptome responses to disturbed neurodevelopment: role of histone acetylation and methylation as epigenetic switch between reversible and irreversible drug effects. Arch Toxicol 88(7):1451-68 doi:10.1007/s00204-014-1279-6 Balmer NV, Leist M (2014) Epigenetics and transcriptomics to detect adverse drug effects in model systems of human development. Basic Clin Pharmacol Toxicol 115(1):59-68 doi:10.1111/bcpt.12203 Balmer NV, Weng MK, Zimmer B, Ivanova VN, Chambers SM, Nikolaeva E, Jagtap S, Sachinidis A, Hescheler J, Waldmann T, Leist M (2012) Epigenetic changes and disturbed neural development in a human embryonic stem cell-based model relating to the fetal valproate syndrome. Hum Mol Genet 21(18):4104-14 doi:10.1093/hmg/dds239 Barenco M, Brewer D, Papouli E, Tomescu D, Callard R, Stark J, Hubank M (2009) Dissection of a complex transcriptional response using genome-wide transcriptional modelling. Mol Syst Biol 5 doi:ARTN 327 10.1038/msb.2009.84 Barenys M, Gassmann K, Baksmeier C, Heinz S, Reverte I, Schmuck M, Temme T, Bendt F, Zschauer TC, Rockel TD, Unfried K, Watjen W, Sundaram SM, Heuer H, Colomina MT, Fritsche E (2017) Epigallocatechin gallate (EGCG) inhibits adhesion and migration of neural progenitor cells in vitro. Arch Toxicol 91(2):827-837 doi:10.1007/s00204-016- 1709-8 Baumann J, Barenys M, Gassmann K, Fritsche E (2014) Comparative human and rat "neurosphere assay" for developmental neurotoxicity testing. Curr Protoc Toxicol 59:12 21 1-24 doi:10.1002/0471140856.tx1221s59 Baumann J, Gassmann K, Masjosthusmann S, DeBoer D, Bendt F, Giersiefer S, Fritsche E (2016) Comparative human and rat neurospheres reveal species differences in chemical effects on neurodevelopmental key events. Arch Toxicol 90(6):1415-27 doi:10.1007/s00204-015-1568-8 Beaudin AE, Stover PJ (2009) Insights into metabolic mechanisms underlying folate- responsive neural tube defects: a minireview. Birth Defects Res A Clin Mol Teratol 85(4):274-84 doi:10.1002/bdra.20553 Beccari L, Moris N, Girgin M, Turner DA, Baillie-Johnson P, Cossy AC, Lutolf MP, Duboule D, Arias AM (2018) Multi-axial self-organization properties of mouse embryonic stem cells into gastruloids. Nature 562(7726):272-276 doi:10.1038/s41586-018-0578-0 Benjamini Y, Hochberg Y (1995) Controlling the false discovery rate: a practical and powerful approach to multiple testing J R Stat Soc 57(1):289 - 300 Benjamini Y, Krieger AM, Yekutieli D (2006) Adaptive linear step-up procedures that control the false discovery rate. Biometrika 93 (3):491-507 Bentué-Ferrer D, Tribut O, Verdier MC (2010) [Therapeutic drug monitoring of valproate]. Therapie 65(3):233-40 doi:10.2515/therapie/2010029

Berthold MR, Cebron N, Dill F, Gabriel TR, Kötter T, Meinl T, Ohl P, Sieb C, Thiel K, Wiswedel B (2007) KNIME: The konstanz information miner. In: Preisach C, Burkhardt H, Schmidt-Thieme L, Decker R (eds) Data Analysis, Machine Learning and Applications. Springer, Berlin Heidelberg, p 319 - 326 Bessonnard S, De Mot L, Gonze D, Barriol M, Dennis C, Goldbeter A, Dupont G, Chazaud C (2014) Gata6, Nanog and Erk signaling control cell fate in the inner cell mass through a tristable regulatory network. Development 141(19):3637-3648 doi:10.1242/dev.109678 Bhinge A, Poschmann J, Namboori SC, Tian XF, Loh SJH, Traczyk A, Prabhakar S, Stanton LW (2014) MiR-135b is a direct PAX6 target and specifies human neuroectoderm by inhibiting TGF-beta/BMP signaling. Embo J 33(11):1271-1283 doi:10.1002/embj.201387215 Blaauboer BJ, Boekelheide K, Clewell HJ, Daneshian M, Dingemans MM, Goldberg AM, Heneweer M, Jaworska J, Kramer NI, Leist M, Seibert H, Testai E, Vandebriel RJ, Yager JD, Zurlo J (2012) The use of biomarkers of toxicity for integrating in vitro hazard estimates into risk assessment for humans. ALTEX 29(4):411-25 Blotiere PO, Raguideau F, Weill A, Elefant E, Perthus I, Goulet V, Rouget F, Zureik M, Coste J, Dray-Spira R (2019) Risks of 23 specific malformations associated with prenatal exposure to 10 antiepileptic drugs. Neurology 93(2):e167-e180 doi:10.1212/WNL.0000000000007696 Boles NC, Hirsch SE, Le S, Corneo B, Najm F, Minotti AP, Wang QJ, Lotz S, Tesar PJ, Fasano CA (2014) NPTX1 Regulates Neural Lineage Specification from Human Pluripotent Stem Cells. Cell Rep 6(4):724-736 doi:10.1016/j.celrep.2014.01.026 Bottini AA, Hartung T (2009) Food for thought... on the economics of animal testing. Altex 26(1):3-16 doi:10.14573/altex.2009.1.3 Boyles AL, Hammock P, Speer MC (2005) Candidate gene analysis in human neural tube defects. Am J Med Genet C Semin Med Genet 135C(1):9-23 doi:10.1002/ajmg.c.30048 Breiman L (2001) Random Forests. Machine Learning 45(1):5-32 doi:10.1023/a:1010933404324 Bromley RL, Weston J, Marson AG (2017) Maternal Use of Antiepileptic Agents During Pregnancy and Major Congenital Malformations in Children. JAMA 318(17):1700-1701 doi:10.1001/jama.2017.14485 Brynczka C, Labhart P, Merrick BA (2007) NGF-mediated transcriptional targets of p53 in PC12 neuronal differentiation. BMC Genomics 8:139 doi:10.1186/1471-2164-8-139 Burren KA, Savery D, Massa V, Kok RM, Scott JM, Blom HJ, Copp AJ, Greene ND (2008) Gene-environment interactions in the causation of neural tube defects: folate deficiency increases susceptibility conferred by loss of Pax3 function. Hum Mol Genet 17(23):3675-85 doi:10.1093/hmg/ddn262 Campbell CT, Sampathkumar SG, Yarema KJ (2007) Metabolic oligosaccharide engineering: perspectives, applications, and future directions. Mol Biosyst 3(3):187-94 doi:10.1039/b614939c Carlson M hgu133plus2.db: Affymetrix Human Genome U133 Plus 2.0 Array annotation data (chip hgu133plus2). R package version 322 Carreras Puigvert J, von Stechow L, Siddappa R, Pines A, Bahjat M, Haazen LC, Olsen JV, Vrieling H, Meerman JH, Mullenders LH, van de Water B, Danen EH (2013) Systems biology approach identifies the kinase Csnk1a1 as a regulator of the DNA damage response in embryonic stem cells. Science signaling 6(259):ra5 doi:10.1126/scisignal.2003208 Casati S (2018) Integrated Approaches to Testing and Assessment. Basic Clin Pharmacol Toxicol 123 Suppl 5:51-55 doi:10.1111/bcpt.13018 Chambers SM, Fasano CA, Papapetrou EP, Tomishima M, Sadelain M, Studer L (2009) Highly efficient neural conversion of human ES and iPS cells by dual inhibition of SMAD signaling. Nat Biotechnol 27(3):275-80 doi:10.1038/nbt.1529 Chambers SM, Mica Y, Studer L, Tomishima MJ (2011) Converting human pluripotent stem cells to neural tissue and neurons to model neurodegeneration. Methods Mol Biol 793:87-97 doi:10.1007/978-1-61779-328-8_6

Chanda S, Ang CE, Lee QY, Ghebrial M, Haag D, Shibuya Y, Wernig M, Sudhof TC (2019) Direct Reprogramming of Human Neurons Identifies MARCKSL1 as a Pathogenic Mediator of Valproic Acid-Induced Teratogenicity. Cell Stem Cell 25(1):103-+ doi:10.1016/j.stem.2019.04.021 Chandran V, Coppola G, Nawabi H, Omura T, Versano R, Huebner EA, Zhang A, Costigan M, Yekkirala A, Barrett L, Blesch A, Michaelevski I, Davis-Turak J, Gao FY, Langfelder P, Horvath S, He ZG, Benowitz L, Fainzilber M, Tuszynski M, Woolf CJ, Geschwind DH (2016) A Systems-Level Analysis of the Peripheral Nerve Intrinsic Axonal Growth Program. Neuron 89(5):956-970 doi:10.1016/j.neuron.2016.01.034 Chateauvieux S, Morceau F, Dicato M, Diederich M (2010) Molecular and Therapeutic Potential and Toxicity of Valproic Acid. J Biomed Biotechnol doi:Artn 479364 10.1155/2010/479364 Chickarmane V, Peterson C (2008) A Computational Model for Understanding Stem Cell, Trophectoderm and Endoderm Lineage Determination. PloS one 3(10) doi:ARTN e3478 10.1371/journal.pone.0003478 Cho A, Tang YT, Davila J, Deng SH, Chen L, Miller E, Wernig M, Graef IA (2014) Calcineurin Signaling Regulates Neural Induction through Antagonizing the BMP Pathway. Neuron 82(1):109-124 doi:10.1016/j.neuron.2014.02.015 Choi J, Lee S, Mallard W, Clement K, Tagliazucchi GM, Lim H, Choi IY, Ferrari F, Tsankov AM, Pop R, Lee G, Rinn JL, Meissner A, Park PJ, Hochedlinger K (2015) A comparison of genetically matched cell lines reveals the equivalence of human iPSCs and ESCs. Nat Biotechnol 33(11):1173-81 doi:10.1038/nbt.3388 Chow M, Boheler KR, Li RA (2013) Human pluripotent stem cell-derived cardiomyocytes for heart regeneration, drug discovery and disease modeling: from the genetic, epigenetic, and tissue modeling perspectives. Stem Cell Res Ther 4(4):97 doi:10.1186/scrt308 Chu YS, Eder O, Thomas WA, Simcha I, Pincet F, Ben-Ze'ev A, Perez E, Thiery JP, Dufour S (2006) Prototypical type I E-cadherin and type II cadherin-7 mediate very distinct adhesiveness through their extracellular domains. The Journal of biological chemistry 281(5):2901-10 doi:10.1074/jbc.M506185200 Ciruna B, Jenny A, Lee D, Mlodzik M, Schier AF (2006) Planar cell polarity signalling couples cell division and morphogenesis during neurulation. Nature 439(7073):220-4 doi:10.1038/nature04375 Colleoni S, Galli C, Gaspar JA, Meganathan K, Jagtap S, Hescheler J, Sachinidis A, Lazzari G (2011) Development of a neural teratogenicity test based on human embryonic stem cells: response to retinoic acid exposure. Toxicol Sci 124(2):370-7 doi:10.1093/toxsci/kfr245 Colleoni S, Galli C, Gaspar JA, Meganathan K, Jagtap S, Hescheler J, Sachinidis A, Lazzari G (2012) Characterisation of a neural teratogenicity assay based on human ESCs differentiation following exposure to valproic acid. Curr Med Chem 19(35):6065-71 Conti L, Cattaneo E (2010) Neural stem cell systems: physiological players or in vitro entities? Nat Rev Neurosci 11(3):176-87 doi:10.1038/nrn2761 Copp AJ (2005) Neurulation in the cranial region--normal and abnormal. J Anat 207(5):623-35 doi:10.1111/j.1469-7580.2005.00476.x Copp AJ, Adzick NS, Chitty LS, Fletcher JM, Holmbeck GN, Shaw GM (2015) Spina bifida. Nat Rev Dis Primers 1:15007 doi:10.1038/nrdp.2015.7 Copp AJ, Greene ND (2010) Genetics and development of neural tube defects. J Pathol 220(2):217-30 doi:10.1002/path.2643 Copp AJ, Greene ND, Murdoch JN (2003) Dishevelled: linking convergent extension with neural tube closure. Trends Neurosci 26(9):453-5 doi:10.1016/S0166-2236(03)00212- 1 Corvi R, Vilardell M, Aubrecht J, Piersma A (2016) Validation of Transcriptomics-Based In Vitro Methods. Adv Exp Med Biol 856:243-257 doi:10.1007/978-3-319-33826-2_10

Daston GP (2008) Gene expression, dose-response, and phenotypic anchoring: applications for toxicogenomics in risk assessment. Toxicol Sci 105(2):233-4 doi:10.1093/toxsci/kfn138 de Leeuw VC, Hessel EVS, Pennings JLA, Hodemaekers HM, Wackers PFK, van Oostrom CTM, Piersma AH (2020) Differential effects of fluoxetine and venlafaxine in the neural embryonic stem cell test (ESTn) revealed by a cell lineage map. Neurotoxicology 76:1- 9 doi:10.1016/j.neuro.2019.09.014 De Los Angeles A, Ferrari F, Xi R, Fujiwara Y, Benvenisty N, Deng H, Hochedlinger K, Jaenisch R, Lee S, Leitch HG, Lensch MW, Lujan E, Pei D, Rossant J, Wernig M, Park PJ, Daley GQ (2015) Hallmarks of pluripotency. Nature 525(7570):469-78 doi:10.1038/nature15515 de Pretis S, Kress T, Morelli MJ, Melloni GE, Riva L, Amati B, Pelizzola M (2015) INSPEcT: a computational tool to infer mRNA synthesis, processing and degradation dynamics from RNA- and 4sU-seq time course experiments. Bioinformatics 31(17):2829-35 doi:10.1093/bioinformatics/btv288 De Sarno P, Li X, Jope RS (2002) Regulation of Akt and glycogen synthase kinase-3 beta phosphorylation by sodium valproate and lithium. Neuropharmacology 43(7):1158-64 doi:10.1016/s0028-3908(02)00215-0 Delli Carri A, Onorati M, Lelos MJ, Castiglioni V, Faedo A, Menon R, Camnasio S, Vuono R, Spaiardi P, Talpo F, Toselli M, Martino G, Barker RA, Dunnett SB, Biella G, Cattaneo E (2013) Developmentally coordinated extrinsic signals drive human pluripotent stem cell differentiation toward authentic DARPP-32+ medium-sized spiny neurons. Development 140(2):301-12 doi:10.1242/dev.084608 Delp J, Gutbier S, Klima S, Hoelting L, Pinto-Gil K, Hsieh JH, Aichem M, Klein K, Schreiber F, Tice RR, Pastor M, Behl M, Leist M (2018) A high-throughput approach to identify specific neurotoxicants/ developmental toxicants in human neuronal cell function assays. Altex 35(2) doi:10.14573/altex.1712182 Dhara SK, Hasneen K, Machacek DW, Boyd NL, Rao RR, Stice SL (2008) Human neural progenitor cells derived from embryonic stem cells in feeder-free cultures. Differentiation 76(5):454-64 doi:10.1111/j.1432-0436.2007.00256.x Dhara SK, Stice SL (2008) Neural differentiation of human embryonic stem cells. J Cell Biochem 105(3):633-40 doi:10.1002/jcb.21891 Dolmetsch R, Geschwind DH (2011) The Human Brain in a Dish: The Promise of iPSC-Derived Neurons. Cell 145(6):831-834 doi:10.1016/j.cell.2011.05.034 Dragunow M (2008) Opinion - The adult human brain in preclinical drug development. Nat Rev Drug Discov 7(8):659-U10 doi:10.1038/nrd2617 Dreser N, Madjar K, Holzer AK, Kapitza M, Scholz C, Kranaster P, Gutbier S, Klima S, Kolb D, Dietz C, Trefzer T, Meisig J, van Thriel C, Henry M, Berthold MR, Bluthgen N, Sachinidis A, Rahnenfuhrer J, Hengstler JG, Waldmann T, Leist M (2020) Development of a neural rosette formation assay (RoFA) to identify neurodevelopmental toxicants and to characterize their transcriptome disturbances. Arch Toxicol 94(1):151-171 doi:10.1007/s00204-019-02612-5 Dreser N, Zimmer B, Dietz C, Sugis E, Pallocca G, Nyffeler J, Meisig J, Bluthgen N, Berthold MR, Waldmann T, Leist M (2015) Grouping of histone deacetylase inhibitors and other toxicants disturbing neural crest migration by transcriptional profiling. Neurotoxicology 50:56-70 doi:10.1016/j.neuro.2015.07.008 Dulac C (2010) Brain function and chromatin plasticity. Nature 465(7299):728-35 doi:10.1038/nature09231 Dunn SJ, Martello G, Yordanov B, Emmott S, Smith AG (2014) Defining an essential transcription factor program for naive pluripotency. Science 344(6188):1156-1160 doi:10.1126/science.1248882 Dupin E, Creuzet S, Le Douarin NM (2006) The contribution of the neural crest to the vertebrate body. Advances in experimental medicine and biology 589:96-119 doi:10.1007/978-0- 387-46954-6_6

Elkabetz Y, Panagiotakos G, Al Shamy G, Socci ND, Tabar V, Studer L (2008) Human ES cell- derived neural rosettes reveal a functionally distinct early neural stem cell stage. Genes Dev 22(2):152-65 doi:10.1101/gad.1616208 Eom DS, Amarnath S, Fogel JL, Agarwala S (2012) Bone morphogenetic proteins regulate hinge point formation during neural tube closure by dynamic modulation of apicobasal polarity. Birth Defects Res A Clin Mol Teratol 94(10):804-16 doi:10.1002/bdra.23052 Escuin S, Vernay B, Savery D, Gurniak CB, Witke W, Greene ND, Copp AJ (2015) Rho-kinase- dependent actin turnover and actomyosin disassembly are necessary for mouse spinal neural tube closure. J Cell Sci 128(14):2468-81 doi:10.1242/jcs.164574 Farrell JA, Wang Y, Riesenfeld SJ, Shekhar K, Regev A, Schier AF (2018) Single-cell reconstruction of developmental trajectories during zebrafish embryogenesis. Science 360(6392) doi:10.1126/science.aar3131 Fasano CA, Phoenix TN, Kokovay E, Lowry N, Elkabetz Y, Dimos JT, Lemischka IR, Studer L, Temple S (2009) Bmi-1 cooperates with Foxg1 to maintain neural stem cell self- renewal in the forebrain. Genes Dev 23(5):561-74 doi:10.1101/gad.1743709 Fattahi F, Steinbeck JA, Kriks S, Tchieu J, Zimmer B, Kishinevsky S, Zeltner N, Mica Y, El- Nachef W, Zhao H, de Stanchina E, Gershon MD, Grikscheit TC, Chen S, Studer L (2016) Deriving human ENS lineages for cell therapy and drug discovery in Hirschsprung disease. Nature 531(7592):105-9 doi:10.1038/nature16951 Fenech M (2000) The in vitro micronucleus technique. Mutat Res 455(1-2):81-95 doi:10.1016/s0027-5107(00)00065-8 Fenech M, Morley AA (1985) Measurement of micronuclei in lymphocytes. Mutat Res 147(1- 2):29-36 doi:10.1016/0165-1161(85)90015-9 Feutz AC, De Geyter C (2019) Accuracy, discriminative properties and reliability of a human ESC-based in vitro toxicity assay to distinguish teratogens responsible for neural tube defects. Arch Toxicol 93(8):2375-2384 doi:10.1007/s00204-019-02512-8 Finn MA, Walker ML (2007) Spinal lipomas: clinical spectrum, embryology, and treatment. Neurosurg Focus 23(2):E10 doi:10.3171/FOC-07/08/E10 Forreryd A, Johansson H, Albrekt AS, Borrebaeck CA, Lindstedt M (2015) Prediction of chemical respiratory sensitizers using GARD, a novel in vitro assay based on a genomic biomarker signature. PLoS One 10(3):e0118808 doi:10.1371/journal.pone.0118808 Franco R, Cidlowski JA (2012) Glutathione efflux and cell death. Antioxid Redox Signal 17(12):1694-713 doi:10.1089/ars.2012.4553 Frank CL, Brown JP, Wallace K, Mundy WR, Shafer TJ (2017) From the Cover: Developmental Neurotoxicants Disrupt Activity in Cortical Networks on Microelectrode Arrays: Results of Screening 86 Compounds During Neural Network Formation. Toxicol Sci 160(1):121-135 doi:10.1093/toxsci/kfx169 Frank CL, Brown JP, Wallace K, Wambaugh JF, Shah I, Shafer TJ (2018) Defining toxicological tipping points in neuronal network development. Toxicol Appl Pharmacol 354:81-93 doi:10.1016/j.taap.2018.01.017 Fritsche E, Barenys M, Klose J, Masjosthusmann S, Nimtz L, Schmuck M, Wuttke S, Tigges J (2018a) Development of the Concept for Stem Cell-Based Developmental Neurotoxicity Evaluation. Toxicol Sci 165(1):14-20 doi:10.1093/toxsci/kfy175 Fritsche E, Crofton KM, Hernandez AF, Hougaard Bennekou S, Leist M, Bal-Price A, Reaves E, Wilks MF, Terron A, Solecki R, Sachana M, Gourmelon A (2017) OECD/EFSA workshop on developmental neurotoxicity (DNT): The use of non-animal test methods for regulatory purposes. Altex 34(2):311-315 doi:10.14573/altex.1701171 Fritsche E, Grandjean P, Crofton KM, Aschner M, Goldberg A, Heinonen T, Hessel EVS, Hogberg HT, Bennekou SH, Lein PJ, Leist M, Mundy WR, Paparella M, Piersma AH, Sachana M, Schmuck G, Solecki R, Terron A, Monnet-Tschudi F, Wilks MF, Witters H, Zurich MG, Bal-Price A (2018b) Consensus statement on the need for innovation, transition and implementation of developmental neurotoxicity (DNT) testing for regulatory purposes. Toxicol Appl Pharmacol 354:3-6 doi:10.1016/j.taap.2018.02.004 Gammill LS, Bronner-Fraser M (2003) Neural crest specification: migrating into genomics. Nat Rev Neurosci 4(10):795-805 doi:10.1038/nrn1219

Gao C, Weisman D, Lan J, Gou N, Gu AZ (2015) Toxicity mechanisms identification via gene set enrichment analysis of time-series toxicogenomics data: impact of time and concentration. Environmental science & technology 49(7):4618-26 doi:10.1021/es505199f Gaspard N, Bouschet T, Hourez R, Dimidschstein J, Naeije G, van den Ameele J, Espuny- Camacho I, Herpoel A, Passante L, Schiffmann SN, Gaillard A, Vanderhaeghen P (2008) An intrinsic mechanism of corticogenesis from embryonic stem cells. Nature 455(7211):351-U10 doi:10.1038/nature07287 Gaspard N, Vanderhaeghen P (2010) Mechanisms of neural specification from embryonic stem cells. Curr Opin Neurobiol 20(1):37-43 doi:10.1016/j.conb.2009.12.001 Geelen JA, Langman J (1977) Closure of the neural tube in the cephalic region of the mouse embryo. Anat Rec 189(4):625-40 doi:10.1002/ar.1091890407 Geelen JA, Langman J (1979) Ultrastructural observations on closure of the neural tube in the mouse. Anat Embryol (Berl) 156(1):73-88 doi:10.1007/BF00315716 Gilbert SF, Singer SR, Tyler MS, Kozlowski RN (2006) Developmental biology. In: Sinauer Associates,. Companion website http://www.devbio.com Table of contents http://www.loc.gov/catdir/toc/ecip068/2006004780.html Gottesman, II, Gould TD (2003) The endophenotype concept in psychiatry: etymology and strategic intentions. The American journal of psychiatry 160(4):636-45 Gould TD, Chen G, Manji HK (2004) In vivo evidence in the brain for lithium inhibition of glycogen synthase kinase-3. Neuropsychopharmacol 29(1):32-38 doi:10.1038/sj.npp.1300283 Grandjean P, Landrigan PJ (2006) Developmental neurotoxicity of industrial chemicals. Lancet 368(9553):2167-78 doi:10.1016/S0140-6736(06)69665-7 Grandjean P, Landrigan PJ (2014a) Neurobehavioural effects of developmental toxicity. The Lancet Neurology 13(3):330-8 doi:10.1016/S1474-4422(13)70278-3 Grandjean P, Landrigan PJ (2014b) Neurodevelopmental toxicity: still more questions than answers--authors' response. The Lancet Neurology 13(7):648-9 doi:10.1016/S1474- 4422(14)70121-8 Greene ND, Copp AJ (2009) Development of the vertebrate central nervous system: formation of the neural tube. Prenat Diagn 29(4):303-11 doi:10.1002/pd.2206 Greene ND, Copp AJ (2014) Neural tube defects. Annu Rev Neurosci 37:221-42 doi:10.1146/annurev-neuro-062012-170354 Grenier J, Teillet MA, Grifone R, Kelly RG, Duprez D (2009) Relationship between neural crest cells and cranial mesoderm during head muscle development. PloS one 4(2):e4381 doi:10.1371/journal.pone.0004381 Grinberg M, Stober RM, Edlund K, Rempel E, Godoy P, Reif R, Widera A, Madjar K, Schmidt- Heck W, Marchan R, Sachinidis A, Spitkovsky D, Hescheler J, Carmo H, Arbo MD, van de Water B, Wink S, Vinken M, Rogiers V, Escher S, Hardy B, Mitic D, Myatt G, Waldmann T, Mardinoglu A, Damm G, Seehofer D, Nussler A, Weiss TS, Oberemm A, Lampen A, Schaap MM, Luijten M, van Steeg H, Thasler WE, Kleinjans JC, Stierum RH, Leist M, Rahnenfuhrer J, Hengstler JG (2014) Toxicogenomics directory of chemically exposed human hepatocytes. Arch Toxicol 88(12):2261-87 doi:10.1007/s00204-014-1400-x Groves AK, LaBonne C (2014) Setting appropriate boundaries: fate, patterning and competence at the neural plate border. Dev Biol 389(1):2-12 doi:10.1016/j.ydbio.2013.11.027 Haghverdi L, Buttner M, Wolf FA, Buettner F, Theis FJ (2016) Diffusion pseudotime robustly reconstructs lineage branching. Nat Methods 13(10):845-8 doi:10.1038/nmeth.3971 Hales CN, Barker DJ (1992) Type 2 (non-insulin-dependent) diabetes mellitus: the thrifty phenotype hypothesis. Diabetologia 35(7):595-601 doi:10.1007/BF00400248 Hall BK (2008) The neural crest and neural crest cells: discovery and significance for theories of embryonic organization. Journal of biosciences 33(5):781-93 Harbron C, Chang KM, South MC (2007) RefPlus: an R package extending the RMA Algorithm. Bioinformatics 23(18):2493-4 doi:10.1093/bioinformatics/btm357

Harrill JA, Freudenrich T, Wallace K, Ball K, Shafer TJ, Mundy WR (2018) Testing for developmental neurotoxicity using a battery of in vitro assays for key cellular events in neurodevelopment. Toxicol Appl Pharmacol 354:24-39 doi:10.1016/j.taap.2018.04.001 Harris MJ, Juriloff DM (2007) Mouse mutants with neural tube closure defects and their role in understanding human neural tube defects. Birth Defects Res A 79(3):187-210 doi:10.1002/bdra.20333 Hartung T (2008) Food for thought ... on alternative methods for cosmetics safety testing. Altex 25(3):147-62 doi:10.14573/altex.2008.3.147 Hartung T (2009) Toxicology for the twenty-first century. Nature 460(7252):208-12 doi:10.1038/460208a Hartung T, Blaauboer B, Leist M (2009) Food for thought ... on education in alternative methods in toxicology. Altex 26(4):255-63 doi:10.14573/altex.2009.4.255 Hartung T, Hoffmann S, Stephens M (2013a) Mechanistic validation. Altex 30(2):119-30 doi:10.14573/altex.2013.2.119 Hartung T, Leist M (2008) Food for thought ... on the evolution of toxicology and the phasing out of animal testing. Altex 25(2):91-102 Hartung T, Luechtefeld T, Maertens A, Kleensang A (2013b) Integrated testing strategies for safety assessments. Altex 30(1):3-18 Hartung T, McBride M (2011) Food for Thought ... on mapping the human toxome. Altex 28(2):83-93 doi:10.14573/altex.2011.2.083 Heise T, Schug M, Storm D, Ellinger-Ziegelbauer H, Ahr HJ, Hellwig B, Rahnenfuhrer J, Ghallab A, Guenther G, Sisnaiske J, Reif R, Godoy P, Mielke H, Gundert-Remy U, Lampen A, Oberemm A, Hengstler JG (2012) In vitro - in vivo correlation of gene expression alterations induced by liver carcinogens. Current medicinal chemistry 19(11):1721-30 Hengstler JG, Sjogren AK, Zink D, Hornberg JJ (2020) In vitro prediction of organ toxicity: the challenges of scaling and secondary mechanisms of toxicity. Arch Toxicol 94(2):353- 356 doi:10.1007/s00204-020-02669-7 Herberg M, Roeder I (2015) Computational modelling of embryonic stem-cell fate control. Development 142(13):2250-2260 doi:10.1242/dev.116343 Hermsen SA, Pronk TE, van den Brandhof EJ, van der Ven LT, Piersma AH (2013) Transcriptomic analysis in the developing zebrafish embryo after compound exposure: individual gene expression and pathway regulation. Toxicol Appl Pharmacol 272(1):161-71 doi:10.1016/j.taap.2013.05.037 Herwig R, Gmuender H, Corvi R, Bloch KM, Brandenburg A, Castell J, Ceelen L, Chesne C, Doktorova TY, Jennen D, Jennings P, Limonciel A, Lock EA, McMorrow T, Phrakonkham P, Radford R, Slattery C, Stierum R, Vilardell M, Wittenberger T, Yildirimman R, Ryan M, Rogiers V, Kleinjans J (2015) Inter-laboratory study of human in vitro toxicogenomics-based tests as alternative methods for evaluating chemical carcinogenicity: a bioinformatics perspective. Arch Toxicol doi:10.1007/s00204-015- 1617-3 Herzog JI, Schmahl C (2018) Adverse Childhood Experiences and the Consequences on Neurobiological, Psychosocial, and Somatic Conditions Across the Lifespan. Front Psychiatry 9 doi:ARTN 420 10.3389/fpsyt.2018.00420 Hessel EVS, Staal YCM, Piersma AH (2018) Design and validation of an ontology-driven animal-free testing strategy for developmental neurotoxicity testing. Toxicol Appl Pharmacol 354:136-152 doi:10.1016/j.taap.2018.03.013 Heyer DB, Meredith RM (2017) Environmental toxicology: Sensitive periods of development and neurodevelopmental disorders. Neurotoxicology 58:23-41 doi:10.1016/j.neuro.2016.10.017 Hezroni H, Sailaja BS, Meshorer E (2011) Pluripotency-related, valproic acid (VPA)-induced genome-wide histone H3 lysine 9 (H3K9) acetylation patterns in embryonic stem cells. J Biol Chem 286(41):35977-88 doi:10.1074/jbc.M111.266254

Hibbard BM (1964) The Role of Folic Acid in Pregnancy; with Particular Reference to Anaemia, Abruption and Abortion. J Obstet Gynaecol Br Commonw 71:529-42 doi:10.1111/j.1471-0528.1964.tb04317.x Hoelting L, Klima S, Karreman C, Grinberg M, Meisig J, Henry M, Rotshteyn T, Rahnenfuhrer J, Bluthgen N, Sachinidis A, Waldmann T, Leist M (2016) Stem Cell-Derived Immature Human Dorsal Root Ganglia Neurons to Identify Peripheral Neurotoxicants. Stem Cells Transl Med 5(4):476-87 doi:10.5966/sctm.2015-0108 Honkela A, Girardot C, Gustafson EH, Liu YH, Furlong EEM, Lawrence ND, Rattray M (2010) Model-based method for transcription factor target identification with limited data. P Natl Acad Sci USA 107(17):7793-7798 doi:10.1073/pnas.0914285107 Honkela A, Peltonen J, Topa H, Charapitsa I, Matarese F, Grote K, Stunnenberg HG, Reid G, Lawrence ND, Rattray M (2015) Genome-wide modeling of transcription kinetics reveals patterns of RNA production delays. P Natl Acad Sci USA 112(42):13115-13120 doi:10.1073/pnas.1420404112 Hörstadius S (1950) The neural crest : its properties and derivatives in the light of experimental research. Oxford University Press, London Hunt PR, Olejnik N, Bailey KD, Vaught CA, Sprando RL (2018) C. elegans Development and Activity Test detects mammalian developmental neurotoxins. Food Chem Toxicol 121:583-592 doi:10.1016/j.fct.2018.09.061 Igarashi Y, Nakatsu N, Yamashita T, Ono A, Ohno Y, Urushidani T, Yamada H (2015) Open TG-GATEs: a large-scale toxicogenomics database. Nucleic Acids Res 43(Database issue):D921-7 doi:10.1093/nar/gku955 Islam MM, Zhang CL (2015) TLX: A master regulator for neural stem cell maintenance and neurogenesis. Biochim Biophys Acta 1849(2):210-6 doi:10.1016/j.bbagrm.2014.06.001 Ito T, Ando H, Handa H (2011) Teratogenic effects of thalidomide: molecular mechanisms. Cell Mol Life Sci 68(9):1569-1579 doi:10.1007/s00018-010-0619-9 Jagtap S, Meganathan K, Gaspar J, Wagh V, Winkler J, Hescheler J, Sachinidis A (2011) Cytosine arabinoside induces ectoderm and inhibits mesoderm expression in human embryonic stem cells during multilineage differentiation. Br J Pharmacol 162(8):1743- 56 doi:10.1111/j.1476-5381.2010.01197.x Jang S, Choubey S, Furchtgott L, Zou LN, Doyle A, Menon V, Loew EB, Krostag AR, Martinez RA, Madisen L, Levi BP, Ramanathan S (2017) Dynamics of embryonic stem cell differentiation inferred from single-cell transcriptomics show a series of transitions through discrete cell states. Elife 6 doi:10.7554/eLife.20487 Janvilisri T (2015) Omics-based identification of biomarkers for nasopharyngeal carcinoma. Dis Markers 2015:762128 doi:10.1155/2015/762128 Jennings P (2013) Stress response pathways, toxicity pathways and adverse outcome pathways. Arch Toxicol 87(1):13-4 doi:10.1007/s00204-012-0974-4 Jennings P, Limonciel A, Felice L, Leonard MO (2013) An overview of transcriptional regulation in response to toxicological insult. Arch Toxicol 87(1):49-72 doi:10.1007/s00204-012- 0919-y Johansson H, Albrekt AS, Borrebaeck CA, Lindstedt M (2013) The GARD assay for assessment of chemical skin sensitizers. Toxicol In Vitro 27(3):1163-9 doi:10.1016/j.tiv.2012.05.019 Johansson H, Lindstedt M, Albrekt AS, Borrebaeck CA (2011) A genomic biomarker signature can predict skin sensitizers using a cell-based in vitro alternative to animal tests. BMC Genomics 12:399 doi:10.1186/1471-2164-12-399 John B, Lewis KR (1966) Chromosome variability and geographic distribution in insects. Science 152:711-721 Johnson MP, Gerdes M, Rintoul N, Pasquariello P, Melchionni J, Sutton LN, Adzick NS (2006) Maternal-fetal surgery for myelomeningocele: neurodevelopmental outcomes at 2 years of age. Am J Obstet Gynecol 194(4):1145-50; discussion 1150-2 doi:10.1016/j.ajog.2006.01.072

Johnson WE, Li C, Rabinovic A (2007) Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 8(1):118-27 doi:10.1093/biostatistics/kxj037 Judson R, Houck K, Martin M, Richard AM, Knudsen TB, Shah I, Little S, Wambaugh J, Setzer RW, Kothiya P, Phuong J, Filer D, Smith D, Reif D, Rotroff D, Kleinstreuer N, Sipes N, Xia M, Huang R, Crofton K, Thomas RS (2016) Analysis of the Effects of Cell Stress and Cytotoxicity on In Vitro Assay Activity Across a Diverse Chemical and Assay Space. Toxicol Sci 153(2):409 doi:10.1093/toxsci/kfw148 Jurges C, Dolken L, Erhard F (2018) Dissecting newly transcribed and old RNA using GRAND- SLAM. Bioinformatics 34(13):i218-i226 doi:10.1093/bioinformatics/bty256 Kadereit S, Zimmer B, van Thriel C, Hengstler JG, Leist M (2012a) Compound selection for in vitro modeling of developmental neurotoxicity. Front Biosci 17:2442-60 Kadereit S, Zimmer B, van Thriel C, Hengstler JG, Leist M (2012b) Compound selection for in vitro modeling of developmental neurotoxicity. Frontiers in bioscience 17:2442-60 Kamieniarz K, Izzo A, Dundr M, Tropberger P, Ozretic L, Kirfel J, Scheer E, Tropel P, Wisniewski JR, Tora L, Viville S, Buettner R, Schneider R (2012) A dual role of linker histone H1.4 Lys 34 acetylation in transcriptional activation. Genes Dev 26(8):797-802 doi:10.1101/gad.182014.111 Kanehisa M, Goto S (2000) KEGG: kyoto encyclopedia of genes and genomes. Nucleic Acids Res 28(1):27-30 Kaufman DS, Hanson ET, Lewis RL, Auerbach R, Thomson JA (2001) Hematopoietic colony- forming cells derived from human embryonic stem cells. Proc Natl Acad Sci U S A 98(19):10716-21 doi:10.1073/pnas.191362598 Kibar Z, Salem S, Bosoi CM, Pauwels E, De Marco P, Merello E, Bassuk AG, Capra V, Gros P (2011) Contribution of VANGL2 mutations to isolated neural tube defects. Clin Genet 80(1):76-82 doi:10.1111/j.1399-0004.2010.01515.x Kiefer JC (2007) Back to basics: Sox genes. Dev Dyn 236(8):2356-66 doi:10.1002/dvdy.21218 Kirby ML, Hutson MR (2010) Factors controlling cardiac neural crest cell migration. Cell adhesion & migration 4(4):609-21 Kleinstreuer NC, Yang J, Berg EL, Knudsen TB, Richard AM, Martin MT, Reif DM, Judson RS, Polokoff M, Dix DJ, Kavlock RJ, Houck KA (2014) Phenotypic screening of the ToxCast chemical library to classify toxic and therapeutic mechanisms. Nat Biotechnol 32(6):583-91 doi:10.1038/nbt.2914 Koch P, Opitz T, Steinbeck JA, Ladewig J, Brustle O (2009) A rosette-type, self-renewing human ES cell-derived neural stem cell with potential for in vitro instruction and synaptic integration. Proc Natl Acad Sci U S A 106(9):3225-30 doi:10.1073/pnas.0808387106 Kondo A, Kamihira O, Ozawa H (2009) Neural tube defects: Prevalence, etiology and prevention. Int J Urol 16(1):49-57 doi:10.1111/j.1442-2042.2008.02163.x Koren G, Berkovitch M, Ornoy A (2018) Dose-Dependent Teratology in Humans: Clinical Implications for Prevention. Pediatr Drugs 20(4):331-335 doi:10.1007/s40272-018- 0294-0 Krebs A, Nyffeler J, Rahnenfuhrer J, Leist M (2018) Normalization of data for viability and relative cell function curves. Altex 35(2):268-271 doi:10.14573/1803231 Krishnaswamy A, Yamagata M, Duan X, Hong YK, Sanes JR (2015) Sidekick 2 directs formation of a retinal circuit that detects differential motion. Nature 524(7566):466-70 doi:10.1038/nature14682 Krug AK, Balmer NV, Matt F, Schonenberger F, Merhof D, Leist M (2013a) Evaluation of a human neurite growth assay as specific screen for developmental neurotoxicants. Arch Toxicol 87(12):2215-31 doi:10.1007/s00204-013-1072-y Krug AK, Gutbier S, Zhao L, Poltl D, Kullmann C, Ivanova V, Forster S, Jagtap S, Meiser J, Leparc G, Schildknecht S, Adam M, Hiller K, Farhan H, Brunner T, Hartung T, Sachinidis A, Leist M (2014) Transcriptional and metabolic adaptation of human neurons to the mitochondrial toxicant MPP(+). Cell death & disease 5:e1222 doi:10.1038/cddis.2014.166 Krug AK, Kolde R, Gaspar JA, Rempel E, Balmer NV, Meganathan K, Vojnits K, Baquie M, Waldmann T, Ensenat-Waser R, Jagtap S, Evans RM, Julien S, Peterson H, Zagoura

D, Kadereit S, Gerhard D, Sotiriadou I, Heke M, Natarajan K, Henry M, Winkler J, Marchan R, Stoppini L, Bosgra S, Westerhout J, Verwei M, Vilo J, Kortenkamp A, Hescheler J, Hothorn L, Bremer S, van Thriel C, Krause KH, Hengstler JG, Rahnenfuhrer J, Leist M, Sachinidis A (2013b) Human embryonic stem cell-derived test systems for developmental neurotoxicity: a transcriptomics approach. Arch Toxicol 87(1):123-43 doi:10.1007/s00204-012-0967-3 Kruse JP, Gu W (2009) Modes of p53 regulation. Cell 137(4):609-22 doi:10.1016/j.cell.2009.04.050 Kuegler PB, Zimmer B, Waldmann T, Baudis B, Ilmjarv S, Hescheler J, Gaughwin P, Brundin P, Mundy W, Bal-Price AK, Schrattenholz A, Krause KH, van Thriel C, Rao MS, Kadereit S, Leist M (2010) Markers of murine embryonic and neural stem cells, neurons and astrocytes: reference points for developmental neurotoxicity testing. Altex 27(1):17-42 La Manno G, Soldatov R, Zeisel A, Braun E, Hochgerner H, Petukhov V, Lidschreiber K, Kastriti ME, Lonnerberg P, Furlan A, Fan J, Borm LE, Liu Z, van Bruggen D, Guo J, He X, Barker R, Sundstrom E, Castelo-Branco G, Cramer P, Adameyko I, Linnarsson S, Kharchenko PV (2018) RNA velocity of single cells. Nature 560(7719):494-498 doi:10.1038/s41586-018-0414-6 Lancaster MA, Corsini NS, Wolfinger S, Gustafson EH, Phillips AW, Burkard TR, Otani T, Livesey FJ, Knoblich JA (2017) Guided self-organization and cortical plate formation in human brain organoids. Nat Biotechnol 35(7):659-666 doi:10.1038/nbt.3906 Laurence KM (1981) Recurrence risk of neural tube defects. J Med Genet 18(4):322-3 doi:10.1136/jmg.18.4.322 Le Douarin N, Kalcheim C (1999) The neural crest, 2nd edn. Cambridge University Press, Cambridge Lee H, Shamy GA, Elkabetz Y, Schofield CM, Harrsion NL, Panagiotakos G, Socci ND, Tabar V, Studer L (2007) Directed differentiation and transplantation of human embryonic stem cell-derived motoneurons. Stem Cells 25(8):1931-9 doi:10.1634/stemcells.2007- 0097 Legler J, van den Brink CE, Brouwer A, Murk AJ, van der Saag PT, Vethaak AD, van der Burg B (1999) Development of a stably transfected estrogen receptor-mediated luciferase reporter gene assay in the human T47D breast cancer cell line. Toxicol Sci 48(1):55- 66 doi:10.1093/toxsci/48.1.55 Leist M, Ayrton AD, Ioannides C (1992) A cytosolic oxygenase activity involved in the bioactivation of 2-aminofluorene. Toxicology 71(1-2):7-20 Leist M, Efremova L, Karreman C (2010) Food for thought ... considerations and guidelines for basic test method descriptions in toxicology. Altex 27(4):309-17 doi:10.14573/altex.2010.4.309 Leist M, Ghallab A, Graepel R, Marchan R, Hassan R, Bennekou SH, Limonciel A, Vinken M, Schildknecht S, Waldmann T, Danen E, van Ravenzwaay B, Kamp H, Gardner I, Godoy P, Bois FY, Braeuning A, Reif R, Oesch F, Drasdo D, Hohme S, Schwarz M, Hartung T, Braunbeck T, Beltman J, Vrieling H, Sanz F, Forsby A, Gadaleta D, Fisher C, Kelm J, Fluri D, Ecker G, Zdrazil B, Terron A, Jennings P, van der Burg B, Dooley S, Meijer AH, Willighagen E, Martens M, Evelo C, Mombelli E, Taboureau O, Mantovani A, Hardy B, Koch B, Escher S, van Thriel C, Cadenas C, Kroese D, van de Water B, Hengstler JG (2017) Adverse outcome pathways: opportunities, limitations and open questions. Arch Toxicol 91(11):3477-3505 doi:10.1007/s00204-017-2045-3 Leist M, Hartung T (2013) Inflammatory findings on species extrapolations: humans are definitely no 70-kg mice. Archives of toxicology 87(4):563-7 doi:10.1007/s00204-013- 1038-0 Leist M, Hartung T, Nicotera P (2008) The dawning of a new age of toxicology. ALTEX 25(2):103-14 Leist M, Hasiwa N, Daneshian M, Hartung T (2012a) Validation and quality control of replacement alternatives – current status and future challenges. Toxicology Research 1(1):8-22 doi:10.1039/C2TX20011B

Leist M, Hasiwa N, Rovida C, Daneshian M, Basketter D, Kimber I, Clewell H, Gocht T, Goldberg A, Busquet F, Rossi AM, Schwarz M, Stephens M, Taalman R, Knudsen TB, McKim J, Harris G, Pamies D, Hartung T (2014) Consensus report on the future of animal-free systemic toxicity testing. Altex 31(3):341-56 doi:10.14573/altex.1406091 Leist M, Lidbury BA, Yang C, Hayden PJ, Kelm JM, Ringeissen S, Detroyer A, Meunier JR, Rathman JF, Jackson GR, Jr., Stolper G, Hasiwa N (2012b) Novel technologies and an overall strategy to allow hazard assessment and risk prediction of chemicals, cosmetics, and drugs with animal-free methods. Altex 29(4):373-88 doi:10.14573/altex.2012.4.373 Leontaridou M, Gabbert S, Landsiedel R (2019) The impact of precision uncertainty on predictive accuracy metrics of non-animal testing methods. Altex 36(3):435-446 doi:10.14573/altex.1810111 Leontaridou M, Urbisch D, Kolle SN, Ott K, Mulliner DS, Gabbert S, Landsiedel R (2017) The borderline range of toxicological methods: Quantification and implications for evaluating precision. Altex 34(4):525-538 doi:10.14573/altex.1606271 Leung AW, Murdoch B, Salem AF, Prasad MS, Gomez GA, Garcia-Castro MI (2016) WNT/beta-catenin signaling mediates human neural crest induction via a pre-neural border intermediate. Development 143(3):398-410 doi:10.1242/dev.130849 Li H, Bai J, Zhong G, Lin H, He C, Dai R, Du H, Huang L (2019) Improved defined approaches for predicting skin sensitization hazard and potency in humans. Altex 36(3):363-372 doi:10.14573/altex.1809191 Li HH, Hyduke DR, Chen R, Heard P, Yauk CL, Aubrecht J, Fornace AJ, Jr. (2015) Development of a toxicogenomics signature for genotoxicity using a dose-optimization and informatics strategy in human cells. Environmental and molecular mutagenesis 56(6):505-19 doi:10.1002/em.21941 Liebsch M, Traue D, Barrabas C, Spielmann H, Uphill P, Wilkins S, McPherson JP, Wiemann C, Kaufmann T, Remmele M, Holzhutter HG (2000) The ECVAM Prevalidation Study on the Use of EpiDerm for Skin Corrosivity Testing. Altern Lab Anim 28(3):371-401 doi:10.1177/026119290002800309 Livak KJ, Schmittgen TD (2001) Analysis of relative gene expression data using real-time quantitative PCR and the 2(-Delta Delta C(T)) Method. Methods 25(4):402-8 doi:10.1006/meth.2001.1262 Lock EF, Abdo N, Huang R, Xia M, Kosyk O, O'Shea SH, Zhou YH, Sedykh A, Tropsha A, Austin CP, Tice RR, Wright FA, Rusyn I (2012) Quantitative high-throughput screening for chemical toxicity in a population-based in vitro model. Toxicol Sci 126(2):578-88 doi:10.1093/toxsci/kfs023 London L, Beseler C, Bouchard MF, Bellinger DC, Colosio C, Grandjean P, Harari R, Kootbodien T, Kromhout H, Little F, Meijster T, Moretto A, Rohlman DS, Stallones L (2012) Neurobehavioral and neurodevelopmental effects of pesticide exposures. Neurotoxicology 33(4):887-96 doi:10.1016/j.neuro.2012.01.004 López-Tobón A, Villa CE, Cheroni C, Trattaro S, Caporale N, Conforti P, Iennaco R, Lachgar M, Rigoli MT, Marcó de la Cruz B, Lo Riso P, Tenderini E, Troglio F, De Simone M, Liste-Noya I, Macino G, Pagani M, Cattaneo E, Testa G (2019) Human Cortical Organoids Expose a Differential Function of GSK3 on Cortical Neurogenesis. Stem Cell Reports doi:https://doi.org/10.1016/j.stemcr.2019.09.005 Low Y, Sedykh A, Fourches D, Golbraikh A, Whelan M, Rusyn I, Tropsha A (2013) Integrative chemical-biological read-across approach for chemical hazard classification. Chem Res Toxicol 26(8):1199-208 doi:10.1021/tx400110f Low Y, Uehara T, Minowa Y, Yamada H, Ohno Y, Urushidani T, Sedykh A, Muratov E, Kuz'min V, Fourches D, Zhu H, Rusyn I, Tropsha A (2011) Predicting drug-induced hepatotoxicity using QSAR and toxicogenomics approaches. Chem Res Toxicol 24(8):1251-62 doi:10.1021/tx200148a Maatz H, Jens M, Liss M, Schafer S, Heinig M, Kirchner M, Adami E, Rintisch C, Dauksaite V, Radke MH, Selbach M, Barton PJ, Cook SA, Rajewsky N, Gotthardt M, Landthaler M, Hubner N (2014) RNA-binding protein RBM20 represses splicing to orchestrate cardiac pre-mRNA processing. J Clin Invest 124(8):3419-30 doi:10.1172/JCI74523

Main H, Radenkovic J, Jin SB, Lendahl U, Andersson ER (2013) Notch signaling maintains neural rosette polarity. PloS one 8(5):e62959 doi:10.1371/journal.pone.0062959 Malcoe LH, Shaw GM, Lammer EJ, Herman AA (1999) The effect of congenital anomalies on mortality risk in white and black infants. Am J Public Health 89(6):887-92 doi:10.2105/ajph.89.6.887 Marchetto MC, Hrvoj-Mihic B, Kerman BE, Yu DX, Vadodaria KC, Linker SB, Narvaiza I, Santos R, Denli AM, Mendes APD, Oefner R, Cook J, McHenry L, Grasmick JM, Heard K, Fredlender C, Randolph-Moore L, Kshirsagar R, Xenitopoulos R, Chou G, Hah N, Muotri AR, Padmanabhan K, Semendeferi K, Gage FH (2019) Species-specific maturation profiles of human, chimpanzee and bonobo neural cells. Elife 8 doi:ARTN e37527 10.7554/eLife.37527 Marchetto MCN, Narvaiza I, Denli AM, Benner C, Lazzarini TA, Nathanson JL, Paquola ACM, Desai KN, Herai RH, Weitzman MD, Yeo GW, Muotri AR, Gage FH (2013) Differential L1 regulation in pluripotent stem cells of humans and apes. Nature 503(7477):525-+ doi:10.1038/nature12686 McShane SG, Mole MA, Savery D, Greene ND, Tam PP, Copp AJ (2015) Cellular basis of neuroepithelial bending during mouse spinal neural tube closure. Dev Biol 404(2):113- 24 doi:10.1016/j.ydbio.2015.06.003 Mica Y, Lee G, Chambers SM, Tomishima MJ, Studer L (2013) Modeling Neural Crest Induction, Melanocyte Specification, and Disease-Related Pigmentation Defects in hESCs and Patient-Specific iPSCs. Cell Rep 3(4):1140-1152 doi:10.1016/j.celrep.2013.03.025 Miller C, Schwalb B, Maier K, Schulz D, Dumcke S, Zacher B, Mayer A, Sydow J, Marcinowski L, Dolken L, Martin DE, Tresch A, Cramer P (2011) Dynamic transcriptome analysis measures rates of mRNA synthesis and decay in yeast. Mol Syst Biol 7 doi:ARTN 458 10.1038/msb.2010.112 Minguez P, Al-Shahrour F, Dopazo J (2006) A function-centric approach to the biological interpretation of microarray time-series. Genome informatics International Conference on Genome Informatics 17(2):57-66 Minoux M, Rijli FM (2010) Molecular mechanisms of cranial neural crest cell migration and patterning in craniofacial development. Development 137(16):2605-21 doi:10.1242/dev.040048 Moskalev AA, Smit-McBride Z, Shaposhnikov MV, Plyusnina EN, Zhavoronkov A, Budovsky A, Tacutu R, Fraifeld VE (2012) Gadd45 proteins: relevance to aging, longevity and age-related pathologies. Ageing research reviews 11(1):51-66 doi:10.1016/j.arr.2011.09.003 Mundy WR, Robinette B, Radio NM, Freudenrich TM (2008) Protein biomarkers associated with growth and synaptogenesis in a cell culture model of neuronal development. Toxicology 249(2-3):220-9 doi:10.1016/j.tox.2008.05.012 Murdoch JN, Copp AJ (2010) The relationship between sonic Hedgehog signaling, cilia, and neural tube defects. Birth Defects Res A Clin Mol Teratol 88(8):633-52 doi:10.1002/bdra.20686 Murko C, Lagger S, Steiner M, Seiser C, Schoefer C, Pusch O (2013) Histone deacetylase inhibitor Trichostatin A induces neural tube defects and promotes neural crest specification in the chicken neural tube. Differentiation 85(1-2):55-66 doi:10.1016/j.diff.2012.12.001 Nasu M, Takata N, Danjo T, Sakaguchi H, Kadoshima T, Futaki S, Sekiguchi K, Eiraku M, Sasai Y (2012) Robust formation and maintenance of continuous stratified cortical neuroepithelium by laminin-containing matrix in mouse ES cell culture. PloS one 7(12):e53024 doi:10.1371/journal.pone.0053024 Nau H (1994) Valproic Acid-Induced Neural-Tube Defects. Ciba F Symp 181:144-152 New DA (1978) Whole-embryo culture and the study of mammalian embryos during organogenesis. Biol Rev Camb Philos Soc 53(1):81-122

Newgreen D, Gibbins I (1982) Factors controlling the time of onset of the migration of neural crest cells in the fowl embryo. Cell and tissue research 224(1):145-60 Newgreen DF, Gibbins IL, Sauter J, Wallenfels B, Wutz R (1982) Ultrastructural and tissue- culture studies on the role of fibronectin, collagen and glycosaminoglycans in the migration of neural crest cells in the fowl embryo. Cell and tissue research 221(3):521- 49 Nightingale KP, Gendreizig S, White DA, Bradbury C, Hollfelder F, Turner BM (2007) Cross- talk between histone modifications in response to histone deacetylase inhibitors: MLL4 links histone H3 acetylation and histone H3K4 methylation. J Biol Chem 282(7):4408- 16 doi:10.1074/jbc.M606773200 Nikolopoulou E, Galea GL, Rolo A, Greene ND, Copp AJ (2017) Neural tube closure: cellular, molecular and biomechanical mechanisms. Development 144(4):552-566 doi:10.1242/dev.145904 Nishimura T, Honda H, Takeichi M (2012) Planar cell polarity links axes of spatial dynamics in neural-tube closure. Cell 149(5):1084-97 doi:10.1016/j.cell.2012.04.021 Nyffeler J, Dolde X, Krebs A, Pinto-Gil K, Pastor M, Behl M, Waldmann T, Leist M (2017a) Combination of multiple neural crest migration assays to identify environmental toxicants from a proof-of-concept chemical library. Arch Toxicol doi:10.1007/s00204- 017-1977-y Nyffeler J, Karreman C, Leisner H, Kim YJ, Lee G, Waldmann T, Leist M (2017b) Design of a high-throughput human neural crest cell migration assay to indicate potential developmental toxicants. Altex 34(1):75-94 doi:10.14573/altex.1605031 O'Rahilly R, Muller F (2002) The two sites of fusion of the neural folds and the two neuropores in the human embryo. Teratology 65(4):162-70 doi:10.1002/tera.10007 Ogata H, Goto S, Sato K, Fujibuchi W, Bono H, Kanehisa M (1999) KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res 27(1):29-34 Omtzigt JG, Los FJ, Grobbee DE, Pijpers L, Jahoda MG, Brandenburg H, Stewart PA, Gaillard HL, Sachs ES, Wladimiroff JW, et al. (1992) The risk of spina bifida aperta after first- trimester exposure to valproate in a prenatal cohort. Neurology 42(4 Suppl 5):119-25 Orlandi A, Ferlosio A, Arcuri G, Scioli MG, De Falco S, Spagnoli LG (2010) Flt-1 expression influences apoptotic susceptibility of vascular smooth muscle cells through the NF- kappaB/IAP-1 pathway. Cardiovascular research 85(1):214-23 doi:10.1093/cvr/cvp288 Ornoy A (2009) Valproic acid in pregnancy: How much are we endangering the embryo and fetus? Reproductive Toxicology 28(1):1-10 doi:10.1016/j.reprotox.2009.02.014 Otsu N (1979) A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 9:62-66 doi:10.1109/TSMC.1979.4310076 Ou JX, Ball JM, Luan YZ, Zhao TT, Miyagishima KJ, Xu YF, Zhou HZ, Chen JG, Merriman DK, Xie Z, Mallon BS, Li W (2018) iPSCs from a Hibernator Provide a Platform for Studying Cold Adaptation and Its Potential Medical Applications. Cell 173(4):851-+ doi:10.1016/j.cell.2018.03.010 Pallocca G, Grinberg M, Henry M, Frickey T, Hengstler JG, Waldmann T, Sachinidis A, Rahnenfuhrer J, Leist M (2016) Identification of transcriptome signatures and biomarkers specific for potential developmental toxicants inhibiting human neural crest cell migration. Arch Toxicol 90(1):159-80 doi:10.1007/s00204-015-1658-7 Pangilinan F, Molloy AM, Mills JL, Troendle JF, Parle-McDermott A, Signore C, O'Leary VB, Chines P, Seay JM, Geiler-Samerotte K, Mitchell A, VanderMeer JE, Krebs KM, Sanchez A, Cornman-Homonoff J, Stone N, Conley M, Kirke PN, Shane B, Scott JM, Brody LC (2012) Evaluation of common genetic variants in 82 candidate genes as risk factors for neural tube defects. BMC Med Genet 13:62 doi:10.1186/1471-2350-13-62 Patthey C, Edlund T, Gunhaga L (2009) Wnt-regulated temporal control of BMP exposure directs the choice between neural plate border and epidermal fate. Development 136(1):73-83 doi:10.1242/dev.025890 Perris R, Perissinotto D (2000) Role of the extracellular matrix during neural crest cell migration. Mechanisms of development 95(1-2):3-21 Piersma AH, Bosgra S, van Duursen MB, Hermsen SA, Jonker LR, Kroese ED, van der Linden SC, Man H, Roelofs MJ, Schulpen SH, Schwarz M, Uibel F, van Vugt-Lussenburg BM,

Westerhout J, Wolterbeek AP, van der Burg B (2013) Evaluation of an alternative in vitro test battery for detecting reproductive toxicants. Reprod Toxicol 38:53-64 doi:10.1016/j.reprotox.2013.03.002 Piersma AH, Genschow E, Verhoef A, Spanjersberg MQ, Brown NA, Brady M, Burns A, Clemann N, Seiler A, Spielmann H (2004) Validation of the postimplantation rat whole- embryo culture test in the international ECVAM validation study on three in vitro embryotoxicity tests. Altern Lab Anim 32(3):275-307 doi:10.1177/026119290403200307 Pla P, Monsoro-Burq AH (2018) The neural border: Induction, specification and maturation of the territory that generates neural crest cells. Dev Biol 444 Suppl 1:S36-S46 doi:10.1016/j.ydbio.2018.05.018 Qin LY, Dai XF, Yin YH (2016) Valproic acid exposure sequentially activates Wnt and mTOR pathways in rats. Mol Cell Neurosci 75:27-35 doi:10.1016/j.mcn.2016.06.004 Rabani M, Raychowdhury R, Jovanovic M, Rooney M, Stumpo DJ, Pauli A, Hacohen N, Schier AF, Blackshear PJ, Friedman N, Amit I, Regev A (2014) High-resolution sequencing and modeling identifies distinct dynamic RNA regulatory strategies. Cell 159(7):1698- 710 doi:10.1016/j.cell.2014.11.015 Radio NM, Breier JM, Shafer TJ, Mundy WR (2008) Assessment of chemical effects on neurite outgrowth in PC12 cells using high content screening. Toxicol Sci 105(1):106-18 doi:10.1093/toxsci/kfn114 Rempel E, Hoelting L, Waldmann T, Balmer NV, Schildknecht S, Grinberg M, Das Gaspar JA, Shinde V, Stober R, Marchan R, van Thriel C, Liebing J, Meisig J, Bluthgen N, Sachinidis A, Rahnenfuhrer J, Hengstler JG, Leist M (2015) A transcriptome-based classifier to identify developmental toxicants by stem cell testing: design, validation and optimization for histone deacetylase inhibitors. Arch Toxicol 89(9):1599-618 doi:10.1007/s00204-015-1573-y Richard AM, Judson RS, Houck KA, Grulke CM, Volarath P, Thillainadarajah I, Yang C, Rathman J, Martin MT, Wambaugh JF, Knudsen TB, Kancherla J, Mansouri K, Patlewicz G, Williams AJ, Little SB, Crofton KM, Thomas RS (2016) ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicology. Chem Res Toxicol 29(8):1225-51 doi:10.1021/acs.chemrestox.6b00135 Roberts DW, Patlewicz GY (2014) Integrated testing and assessment approaches for skin sensitization: a commentary. J Appl Toxicol 34(4):436-40 doi:10.1002/jat.2943 Roffers-Agarwal J, Gammill LS (2009) Neuropilin receptors guide distinct phases of sensory and motor neuronal segmentation. Development 136(11):1879-88 doi:10.1242/dev.032920 Romero AC, Del Rio E, Vilanova E, Sogorb MA (2015) Expression of biomarker genes of differentiation in D3 mouse embryonic stem cells after exposure to different embryotoxicant and non-embryotoxicant model chemicals. Data in brief 5:354-65 doi:10.1016/j.dib.2015.09.015 Russell WMS, Burch RL (1992) The principles of humane experimental technique. Universities Federation for animal welfare, Wheathampstead (UK) Saelens W, Cannoodt R, Todorov H, Saeys Y (2019) A comparison of single-cell trajectory inference methods. Nat Biotechnol 37(5):547-554 doi:10.1038/s41587-019-0071-9 Sand S, Parham F, Portier CJ, Tice RR, Krewski D (2016) Comparison of Points of Departure for Health Risk Assessment Based on High-Throughput Screening Data. Environ Health Perspect doi:10.1289/EHP408 Sauer JM, Hartung T, Leist M, Knudsen TB, Hoeng J, Hayes AW (2015) Systems Toxicology: The Future of Risk Assessment. International journal of toxicology 34(4):346-8 doi:10.1177/1091581815576551 Sauka-Spengler T, Bronner-Fraser M (2008) Evolution of the neural crest viewed from a gene regulatory perspective. Genesis 46(11):673-82 doi:10.1002/dvg.20436 Savagner P (2001) Leaving the neighborhood: molecular mechanisms involved during epithelial-mesenchymal transition. BioEssays : news and reviews in molecular, cellular and developmental biology 23(10):912-23 doi:10.1002/bies.1132

Schaap MM, Wackers PF, Zwart EP, Huijskens I, Jonker MJ, Hendriks G, Breit TM, van Steeg H, van de Water B, Luijten M (2015) A novel toxicogenomics-based approach to categorize (non-)genotoxic carcinogens. Arch Toxicol 89(12):2413-27 doi:10.1007/s00204-014-1368-6 Schmidt BZ, Lehmann M, Gutbier S, Nembo E, Noel S, Smirnova L, Forsby A, Hescheler J, Avci HX, Hartung T, Leist M, Kobolak J, Dinnyes A (2017) In vitro acute and developmental neurotoxicity screening: an overview of cellular platforms and high- throughput technical possibilities. Arch Toxicol 91(1):1-33 doi:10.1007/s00204-016- 1805-9 Scholz G, Genschow E, Pohl I, Bremer S, Paparella M, Raabe H, Southee J, Spielmann H (1999) Prevalidation of the Embryonic Stem Cell Test (EST)-A New In Vitro Embryotoxicity Test. Toxicol In Vitro 13(4-5):675-81 Scholz S, Fischer S, Gundel U, Kuster E, Luckenbach T, Voelker D (2008) The zebrafish embryo model in environmental risk assessment--applications beyond acute toxicity testing. Environ Sci Pollut Res Int 15(5):394-404 doi:10.1007/s11356-008-0018-z Schulpen SH, de Jong E, de la Fonteyne LJ, de Klerk A, Piersma AH (2015a) Distinct gene expression responses of two anticonvulsant drugs in a novel human embryonic stem cell based neural differentiation assay protocol. Toxicol In Vitro 29(3):449-57 doi:10.1016/j.tiv.2014.12.001 Schulpen SH, Pennings JL, Piersma AH (2015b) Gene expression regulation and pathway analysis after valproic acid and carbamazepine exposure in a human embryonic stem cell based neuro-developmental toxicity assay. Toxicol Sci doi:10.1093/toxsci/kfv094 Seiler AE, Spielmann H (2011) The validated embryonic stem cell test to predict embryotoxicity in vitro. Nat Protoc 6(7):961-78 doi:10.1038/nprot.2011.348 Selenica ML, Jensen HS, Larsen AK, Pedersen ML, Helboe L, Leist M, Lotharius J (2007) Efficacy of small-molecule glycogen synthase kinase-3 inhibitors in the postnatal rat model of tau hyperphosphorylation. Brit J Pharmacol 152(6):959-979 doi:10.1038/sj.bjp.0707471 Sequerra EB, Goyal R, Castro PA, Levin JB, Borodinsky LN (2018) NMDA Receptor Signaling Is Important for Neural Tube Formation and for Preventing Antiepileptic Drug-Induced Neural Tube Defects. The Journal of Neuroscience 38(20):4762-4773 doi:10.1523/jneurosci.2634-17.2018 Shah I, Setzer RW, Jack J, Houck KA, Judson RS, Knudsen TB, Liu J, Martin MT, Reif DM, Richard AM, Thomas RS, Crofton KM, Dix DJ, Kavlock RJ (2015) Using ToxCast Data to Reconstruct Dynamic Cell State Trajectories and Estimate Toxicological Points of Departure. Environ Health Perspect doi:10.1289/ehp.1409029 Shibuya M, Yamaguchi S, Yamane A, Ikeda T, Tojo A, Matsushime H, Sato M (1990) Nucleotide sequence and expression of a novel human receptor-type tyrosine kinase gene (flt) closely related to the fms family. Oncogene 5(4):519-24 Shinde V, Hoelting L, Srinivasan SP, Meisig J, Meganathan K, Jagtap S, Grinberg M, Liebing J, Bluethgen N, Rahnenfuhrer J, Rempel E, Stoeber R, Schildknecht S, Forster S, Godoy P, van Thriel C, Gaspar JA, Hescheler J, Waldmann T, Hengstler JG, Leist M, Sachinidis A (2016a) Definition of transcriptome-based indices for quantitative characterization of chemically disturbed stem cell development: introduction of the STOP-Tox and STOP-Tox tests. Arch Toxicol doi:10.1007/s00204-016-1741-8 Shinde V, Klima S, Sureshkumar PS, Meganathan K, Jagtap S, Rempel E, Rahnenfuhrer J, Hengstler JG, Waldmann T, Hescheler J, Leist M, Sachinidis A (2015) Human Pluripotent Stem Cell Based Developmental Toxicity Assays for Chemical Safety Screening and Systems Biology Data Generation. J Vis Exp(100):e52333 doi:10.3791/52333 Shinde V, Perumal Srinivasan S, Henry M, Rotshteyn T, Hescheler J, Rahnenfuhrer J, Grinberg M, Meisig J, Bluthgen N, Waldmann T, Leist M, Hengstler JG, Sachinidis A (2016b) Comparison of a teratogenic transcriptome-based predictive test based on human embryonic versus inducible pluripotent stem cells. Stem Cell Res Ther 7(1):190 doi:10.1186/s13287-016-0449-2

Sim M (2002) Case studies in the use of toxicological measures in epidemiological studies. Toxicology 181-182:405-9 doi:10.1016/s0300-483x(02)00478-x Sjogren AK, Breitholtz K, Ahlberg E, Milton L, Forsgard M, Persson M, Stahl SH, Wilmer MJ, Hornberg JJ (2018) A novel multi-parametric high content screening assay in ciPTEC- OAT1 to predict drug-induced nephrotoxicity during drug discovery. Arch Toxicol 92(10):3175-3190 doi:10.1007/s00204-018-2284-y Sletten EM, Bertozzi CR (2009) Bioorthogonal chemistry: fishing for selectivity in a sea of functionality. Angew Chem Int Ed Engl 48(38):6974-98 doi:10.1002/anie.200900942 Smirnova L, Harris G, Leist M, Hartung T (2015) Cellular resilience. ALTEX 32(4):247-60 doi:10.14573/altex.1509271 Smirnova L, Hogberg HT, Leist M, Hartung T (2014) Developmental neurotoxicity - challenges in the 21st century and in vitro opportunities. Altex 31(2):129-56 doi:http://dx.doi.org/10.14573/altex.1403271 Smith JL, Schoenwolf GC (1989) Notochordal induction of cell wedging in the chick neural plate and its role in neural tube formation. J Exp Zool 250(1):49-62 doi:10.1002/jez.1402500107 Smyth GK (2005) Limma: Linear models for microarray data. In: Gentleman R, Carey V, Dudoit S, Irizarry W, Huber W (eds) Bioinformatics and Computational Biology Solutions using R and Bioconductor. 2006/05/02 edn. Springer, New York, p 397 - 420 Spate AK, Busskamp H, Niederwieser A, Schart VF, Marx A, Wittmann V (2014) Rapid labeling of metabolically engineered cell-surface glycoconjugates with a carbamate-linked cyclopropene reporter. Bioconjug Chem 25(1):147-54 doi:10.1021/bc4004487 Spector AA, Grayson WL (2017) Stem Cell Fate Decision Making: Modeling Approaches. Acs Biomater Sci Eng 3(11):2702-2711 doi:10.1021/acsbiomaterials.6b00606 Spinu N, Bal-Price A, Cronin MTD, Enoch SJ, Madden JC, Worth AP (2019) Development and analysis of an adverse outcome pathway network for human neurotoxicity. Arch Toxicol 93(10):2759-2772 doi:10.1007/s00204-019-02551-1 Stebbings R, Findlay L, Edwards C, Eastwood D, Bird C, North D, Mistry Y, Dilger P, Liefooghe E, Cludts I, Fox B, Tarrant G, Robinson J, Meager T, Dolman C, Thorpe SJ, Bristow A, Wadhwa M, Thorpe R, Poole S (2007) "Cytokine storm" in the phase I trial of monoclonal antibody TGN1412: better understanding the causes to improve preclinical testing of immunotherapeutics. Journal of immunology 179(5):3325-31 Stiegler NV, Krug AK, Matt F, Leist M (2011) Assessment of chemical-induced impairment of human neurite outgrowth by multiparametric live cell imaging in high-density cultures. Toxicol Sci 121(1):73-87 doi:10.1093/toxsci/kfr034 Sturla SJ, Boobis AR, FitzGerald RE, Hoeng J, Kavlock RJ, Schirmer K, Whelan M, Wilks MF, Peitsch MC (2014) Systems toxicology: from basic research to risk assessment. Chem Res Toxicol 27(3):314-29 doi:10.1021/tx400410s Sutton LN, Sun P, Adzick NS (2001) Fetal neurosurgery. Neurosurgery 48(1):124-42; discussion 142-4 doi:10.1097/00006123-200101000-00023 Suzuki IK, Vanderhaeghen P (2015) Is this a brain which I see before me? Modeling human neural development with pluripotent stem cells. Development 142(18):3138-3150 doi:10.1242/dev.120568 Sztajnkrycer MD (2002) Valproic acid toxicity: overview and management. J Toxicol Clin Toxicol 40(6):789-801 doi:10.1081/clt-120014645 Takahashi K, Yamanaka S (2006) Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. Cell 126(4):663-76 doi:10.1016/j.cell.2006.07.024 Tawk M, Araya C, Lyons DA, Reugels AM, Girdler GC, Bayley PR, Hyde DR, Tada M, Clarke JD (2007) A mirror-symmetric cell division that orchestrates neuroepithelial morphogenesis. Nature 446(7137):797-800 doi:10.1038/nature05722 Teodorczyk M, Schmidt MH (2014) Notching on Cancer's Door: Notch Signaling in Brain Tumors. Frontiers in oncology 4:341 doi:10.3389/fonc.2014.00341 Terron A, Bennekou SH (2018) Towards a regulatory use of alternative developmental neurotoxicity testing (DNT). Toxicol Appl Pharmacol 354:19-23 doi:10.1016/j.taap.2018.02.002

Theunissen PT, Robinson JF, Pennings JL, de Jong E, Claessen SM, Kleinjans JC, Piersma AH (2012) Transcriptomic concentration-response evaluation of valproic acid, cyproconazole, and hexaconazole in the neural embryonic stem cell test (ESTn). Toxicological sciences : an official journal of the Society of Toxicology 125(2):430-8 doi:10.1093/toxsci/kfr293 Thomas RS, Clewell HJ, 3rd, Allen BC, Wesselkamper SC, Wang NC, Lambert JC, Hess- Wilson JK, Zhao QJ, Andersen ME (2011) Application of transcriptional benchmark dose values in quantitative cancer and noncancer risk assessment. Toxicol Sci 120(1):194-205 doi:10.1093/toxsci/kfq355 Thomas RS, Wesselkamper SC, Wang NC, Zhao QJ, Petersen DD, Lambert JC, Cote I, Yang L, Healy E, Black MB, Clewell HJ, 3rd, Allen BC, Andersen ME (2013) Temporal concordance between apical and transcriptional points of departure for chemical risk assessment. Toxicol Sci 134(1):180-94 doi:10.1093/toxsci/kft094 Thomson JA, Itskovitz-Eldor J, Shapiro SS, Waknitz MA, Swiergiel JJ, Marshall VS, Jones JM (1998) Embryonic stem cell lines derived from human blastocysts. Science 282(5391):1145-7 doi:10.1126/science.282.5391.1145 Tomson T, Battino D, Perucca E (2016) Valproic acid after five decades of use in epilepsy: time to reconsider the indications of a time-honoured drug. Lancet Neurology 15(2):210-218 doi:10.1016/S1474-4422(15)00314-2 Tonk EC, Pennings JL, Piersma AH (2015) An adverse outcome pathway framework for neural tube and axial defects mediated by modulation of retinoic acid homeostasis. Reprod Toxicol 55:104-13 doi:10.1016/j.reprotox.2014.10.008 Tonk EC, Robinson JF, Verhoef A, Theunissen PT, Pennings JL, Piersma AH (2013) Valproic acid-induced gene expression responses in rat whole embryo culture and comparison across in vitro developmental and non-developmental models. Reprod Toxicol 41:57- 66 doi:10.1016/j.reprotox.2013.06.069 Trapnell C (2015) Defining cell types and states with single-cell genomics. Genome Res 25(10):1491-8 doi:10.1101/gr.190595.115 Trapnell C, Cacchiarelli D, Grimsby J, Pokharel P, Li S, Morse M, Lennon NJ, Livak KJ, Mikkelsen TS, Rinn JL (2014) The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells. Nat Biotechnol 32(4):381-386 doi:10.1038/nbt.2859 Turnbull DM, Rawlins MD, Weightman D, Chadwick DW (1983) Plasma concentrations of sodium valproate: their clinical value. Ann Neurol 14(1):38-42 doi:10.1002/ana.410140107 Uehara T, Ono A, Maruyama T, Kato I, Yamada H, Ohno Y, Urushidani T (2010) The Japanese toxicogenomics project: application of toxicogenomics. Molecular nutrition & food research 54(2):218-27 doi:10.1002/mnfr.200900169 Uhlitz F, Sieber A, Wyler E, Fritsche-Guenther R, Meisig J, Landthaler M, Klinger B, Bluthgen N (2017) An immediate-late gene expression module decodes ERK signal duration. Mol Syst Biol 13(5) doi:ARTN 928 10.15252/msb.20177554 Vajda FJ, O'Brien T J, Hitchcock A, Graham J, Cook M, Lander C, Eadie MJ (2004) Critical relationship between sodium valproate dose and human teratogenicity: results of the Australian register of anti-epileptic drugs in pregnancy. J Clin Neurosci 11(8):854-8 doi:10.1016/j.jocn.2004.05.003 van de Leemput J, Boles NC, Kiehl TR, Corneo B, Lederman P, Menon V, Lee C, Martinez RA, Levi BP, Thompson CL, Yao SY, Kaykas A, Temple S, Fasano CA (2014) CORTECON: A Temporal Transcriptome Analysis of In Vitro Human Cerebral Cortex Development from Human Embryonic Stem Cells. Neuron 83(1):51-68 doi:10.1016/j.neuron.2014.05.013 van den Brink SC, Alemany A, van Batenburg V, Moris N, Blotenburg M, Vivie J, Baillie- Johnson P, Nichols J, Sonnen KF, Martinez Arias A, van Oudenaarden A (2020) Single- cell and spatial transcriptomics reveal somitogenesis in gastruloids. Nature 582(7812):405-409 doi:10.1038/s41586-020-2024-3

van der Burg B, van der Linden S, Man H, Winter R, Jonker L, van Vugt‐Lussenburg B, Brouwe A (2013) A Panel of Quantitative Calux® Reporter Gene Assays for Reliable High- Throughput Toxicity Screening of Chemicals and Complex Mixtures High‐Throughput Screening Methods in Toxicity Testing. p 519-532 van der Linden SC, von Bergh AR, van Vught-Lussenburg BM, Jonker LR, Teunis M, Krul CA, van der Burg B (2014) Development of a panel of high-throughput reporter-gene assays to detect genotoxicity and oxidative stress. Mutat Res Genet Toxicol Environ Mutagen 760:23-32 doi:10.1016/j.mrgentox.2013.09.009 van Thriel C, Westerink RH, Beste C, Bale AS, Lein PJ, Leist M (2012) Translating neurobehavioural endpoints of developmental neurotoxicity tests into in vitro assays and readouts. Neurotoxicology 33(4):911-24 doi:10.1016/j.neuro.2011.10.002 van Vliet E, Kuhnl J, Goebel C, Martinozzi-Teissier S, Alepee N, Ashikaga T, Blomeke B, Del Bufalo A, Cluzel M, Corsini E, Delrue N, Desprez B, Gellatly N, Giese C, Gribaldo L, Hoffmann S, Klaric M, Maillere B, Naisbitt D, Pallardy M, Vocanson M, Petersohn D (2018) State-of-the-art and new options to assess T cell activation by skin sensitizers: Cosmetics Europe Workshop. Altex 35(2):179-192 doi:10.14573/altex.1709011 Waldmann T, Grinberg M, Konig A, Rempel E, Schildknecht S, Henry M, Holzer AK, Dreser N, Shinde V, Sachinidis A, Rahnenfuhrer J, Hengstler JG, Leist M (2017) Stem Cell Transcriptome Responses and Corresponding Biomarkers That Indicate the Transition from Adaptive Responses to Cytotoxicity. Chem Res Toxicol 30(4):905-922 doi:10.1021/acs.chemrestox.6b00259 Waldmann T, Rempel E, Balmer NV, Konig A, Kolde R, Gaspar JA, Henry M, Hescheler J, Sachinidis A, Rahnenfuhrer J, Hengstler JG, Leist M (2014) Design principles of concentration-dependent transcriptome deviations in drug-exposed differentiating stem cells. Chem Res Toxicol 27(3):408-20 doi:10.1021/tx400402j Walker RL, Ramaswami G, Hartl C, Mancuso N, Gandal MJ, De La Torre-Ubieta L, Pasaniuc B, Stein JL, Geschwind DH (2019) Genetic Control of Expression and Splicing in Developing Human Brain Informs Disease Mechanisms. Cell 179(3):750-+ doi:10.1016/j.cell.2019.09.021 Walmsley RM (2008) GADD45a-GFP GreenScreen HC genotoxicity screening assay. Expert Opin Drug Metab Toxicol 4(6):827-35 doi:10.1517/17425255.4.6.827 Wang ZP, Xu L, Zhu XP, Cui WG, Sun Y, Nishijo H, Peng YW, Li RX (2010) Demethylation of Specific Wnt/beta-Catenin Pathway Genes and its Upregulation in Rat Brain Induced by Prenatal Valproate Exposure. Anat Rec 293(11):1947-1953 doi:10.1002/ar.21232 Weigt S, Huebler N, Braunbeck T, von Landenberg F, Broschard TH (2010) Zebrafish teratogenicity test with metabolic activation (mDarT): effects of phase I activation of acetaminophen on zebrafish Danio rerio embryos. Toxicology 275(1-3):36-49 doi:10.1016/j.tox.2010.05.012 Weng MK, Zimmer B, Poltl D, Broeg MP, Ivanova V, Gaspar JA, Sachinidis A, Wullner U, Waldmann T, Leist M (2012) Extensive transcriptional regulation of chromatin modifiers during human neurodevelopment. PloS one 7(5):e36708 doi:10.1371/journal.pone.0036708 Wilde JJ, Petersen JR, Niswander L (2014) Genetic, epigenetic, and environmental contributions to neural tube closure. Annu Rev Genet 48:583-611 doi:10.1146/annurev-genet-120213-092208 Wilmes A, Limonciel A, Aschauer L, Moenks K, Bielow C, Leonard MO, Hamon J, Carpi D, Ruzek S, Handler A, Schmal O, Herrgen K, Bellwon P, Burek C, Truisi GL, Hewitt P, Di Consiglio E, Testai E, Blaauboer BJ, Guillou C, Huber CG, Lukas A, Pfaller W, Mueller SO, Bois FY, Dekant W, Jennings P (2013) Application of integrated transcriptomic, proteomic and metabolomic profiling for the delineation of mechanisms of drug induced cell stress. Journal of proteomics 79:180-94 doi:10.1016/j.jprot.2012.11.022 Wiltse J (2005) Mode of action: Inhibition of histone deacetylase, altering WNT-dependent gene expression, and regulation of beta-catenin - Developmental effects of valproic acid. Crit Rev Toxicol 35(8-9):727-738 doi:10.1080/10408440591007403

Wolpert L (2007) Principles of development, Third edition. edn. Oxford University Press, Oxford Wyszynski DF (2006) Neural tube defects : from origin to treatment. Oxford University Press, Oxford Yamaguchi Y, Shinotsuka N, Nonomura K, Takemoto K, Kuida K, Yosida H, Miura M (2011) Live imaging of apoptosis in a novel transgenic mouse highlights its role in neural tube closure. J Cell Biol 195(6):1047-60 doi:10.1083/jcb.201104057 Yoo YD, Huang CT, Zhang XQ, Lavaute TM, Zhang SC (2011) Fibroblast Growth Factor Regulates Human Neuroectoderm Specification Through ERK1/2-PARP-1 Pathway. Stem Cells 29(12):1975-1982 doi:10.1002/stem.758 Yu R, Wang X, Moazed D (2018) Epigenetic inheritance mediated by coupling of RNAi and histone H3K9 methylation. Nature 558(7711):615-619 doi:10.1038/s41586-018-0239- 3 Zhang Q, Bhattacharya S, Conolly RB, Clewell HJ, Kaminski NE, Andersen ME (2014) Molecular signaling network motifs provide a mechanistic basis for cellular threshold responses. Environ Health Perspect 122(12):1261-70 doi:10.1289/ehp.1408244 Zhang SC, Wernig M, Duncan ID, Brustle O, Thomson JA (2001) In vitro differentiation of transplantable neural precursors from human embryonic stem cells. Nat Biotechnol 19(12):1129-33 doi:10.1038/nbt1201-1129 Zhang X, Huang CT, Chen J, Pankratz MT, Xi J, Li J, Yang Y, Lavaute TM, Li XJ, Ayala M, Bondarenko GI, Du ZW, Jin Y, Golos TG, Zhang SC (2010) Pax6 is a human neuroectoderm cell fate determinant. Cell Stem Cell 7(1):90-100 doi:10.1016/j.stem.2010.04.017 Zimmer B, Kuegler PB, Baudis B, Genewsky A, Tanavde V, Koh W, Tan B, Waldmann T, Kadereit S, Leist M (2011a) Coordinated waves of gene expression during neuronal differentiation of embryonic stem cells as basis for novel approaches to developmental neurotoxicity testing. Cell Death Differ 18(3):383-395 doi:10.1038/cdd.2010.109 Zimmer B, Kuegler PB, Baudis B, Genewsky A, Tanavde V, Koh W, Tan B, Waldmann T, Kadereit S, Leist M (2011b) Coordinated waves of gene expression during neuronal differentiation of embryonic stem cells as basis for novel approaches to developmental neurotoxicity testing. Cell Death Differ 18(3):383-95 doi:10.1038/cdd.2010.109 Ziv O, Zaritsky A, Yaffe Y, Mutukula N, Edri R, Elkabetz Y (2015) Quantitative Live Imaging of Human Embryonic Stem Cell Derived Neural Rosettes Reveals Structure-Function Dynamics Coupled to Cortical Development. PLoS Comput Biol 11(10):e1004453 doi:10.1371/journal.pcbi.1004453 Zurich MG, Stanzel S, Kopp-Schneider A, Prieto P, Honegger P (2013) Evaluation of aggregating brain cell cultures for the detection of acute organ-specific toxicity. Toxicol In Vitro 27(4):1416-24 doi:10.1016/j.tiv.2012.06.018