Tumor-Stroma Interactions Differentially Alter Drug Sensitivity Based on the Origin of Stromal Cells

University of Massachusetts Medical School eScholarship@UMMS

GSBS Dissertations and Theses Graduate School of Biomedical Sciences

2018-10-25

Tumor-stroma interactions differentially alter drug sensitivity based on the origin of stromal cells

Benjamin D. Landry University of Massachusetts Medical School

Let us know how access to this document benefits ou.y Follow this and additional works at: https://escholarship.umassmed.edu/gsbs_diss

Part of the Cancer Biology Commons, Cell Biology Commons, and the Neoplasms Commons

Repository Citation Landry BD. (2018). Tumor-stroma interactions differentially alter drug sensitivity based on the origin of stromal cells. GSBS Dissertations and Theses. https://doi.org/10.13028/rx10-jp17. Retrieved from https://escholarship.umassmed.edu/gsbs_diss/1011

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License. This material is brought to you by eScholarship@UMMS. It has been accepted for inclusion in GSBS Dissertations and Theses by an authorized administrator of eScholarship@UMMS. For more information, please contact [email protected].

Tumor-stroma interactions differentially alter drug sensitivity based on the origin of stromal cells

A Dissertation Presented

BENJAMIN D. LANDRY

Submitted to the Faculty of the University of Massachusetts Graduate School of Biomedical Sciences, Worcester in partial fulfillment of the requirements for the degree of

DOCTOR OF PHILOSOPHY

October 25th, 2018

CANCER BIOLOGY

Tumor-stroma interactions differentially alter drug sensitivity based on the origin of stromal cells

A Dissertation Presented

BENJAMIN D. LANDRY

This work was undertaken in the Graduate School of Biomedical Sciences Cancer Biology Program

Under the mentorship of

______Michael Lee, Ph.D. - Thesis Advisor

______Leslie Shaw, Ph.D. - Member of Committee

______Amir Mitchell, Ph.D. - Member of Committee

______Robert Brewster, Ph.D. - Member of Committee

______Suzanne Gaudet, Ph.D. - External Member of Committee

______Jennifer Benanti, Ph.D. - Chair of Committee

______Mary Ellen Lane, Ph.D., Dean of the Graduate School of Biomedical Sciences

October 25th, 2018 iii

ACKNOWLEDGEMENTS

I thank Michael Lee for all his support, advice, mentoring, patience, and teachings that have helped make this project possible. Mike’s scientific insight and creativity is second to none and I am honored to have trained in his lab.

I thank my committee for the valuable advice, feedback, and mentoring that has helped shaped my project. I would not be where I am today without your support and insightful scientific comments.

I thank my lab mates, both past and present, for contributing to my scientific and personal growth. I will always treasure the wisdom, laughter, conversation, and experiences we shared during my time at UMASS.

I thank my friends who have supported me through my trials, tribulations, successes, and failures over the past 5 years. Jen, I will miss our Friday lunch chats over Indian food. Kurtis, you were an amazing roommate and a spectacular baker. Jose, I will miss grabbing coffee with you to blow off steam. Joseph, you are the most chaotic boardgame player I know, and you always reminded me there was life outside the lab. Tan, some of my happiest moments were spent with you and I cannot thank you enough for listening to my lab stories and cheering me up whenever I was down.

And finally, I wish to thank my family for providing me with the love and support necessary to make it through graduate school. To my mom Denise, who would gallantly make it through the first paragraph of my papers before falling asleep, who always had a spare turkey or roast hidden in the freezer for a iv

delicious Sunday meal, and whose ability to see humor in the smallest things would always lift my spirits. To my dad Scott, who would readily chat science with me and always seem to have an interesting NPR article to discuss, whose computer support has saved me from countless moments of panic, and who has always been there cheering me on. To my brother Pat, whose mechanical skills kept my precious 99 ford contour running long past when it should have died and always ensured I could make it into lab. To my sister Sara, whose witty sarcasm at family dinners would always have me returning to lab in a good mood. To my sister Katie, who will forever associate my work with “yeast worms” and whose cheery disposition would always make me smile. v

ABSTRACT

Tumor heterogeneity observed between patients has made it challenging to develop universal or broadly effective cancer therapies. Therefore, an ever- growing movement within cancer research aims to tailor cancer therapies to individual patients or specific tumor subtypes. Tumor stratification is generally dictated by the genomic mutation status of the tumor cells themselves.

Importantly, non-genetic influences – such as interactions between tumor cells and other components of the tumor microenvironment – have largely been ignored. Therefore, in an effort to increase treatment predictability and efficacy, we investigated how tumor-stroma interactions contribute to drug sensitivity and drug resistance.

I designed a high throughput co-culture screening platform to measure how tumor-stroma interactions alter drug mediated cell death. I identified tumor- stroma interactions that strongly desensitize or sensitize cancer cells to various drug treatments. The directionality of these observed phenotypes was dependent on the stromal cell tissue of origin. Further study revealed that interactions between tumor cells and fibroblasts modulate apoptotic priming in tumor cells to mediate sensitivity to chemotherapeutics. The principles uncovered in this study have important implications on the use of drugs that are designed to enhance apoptosis. For example, based on our screening data, I hypothesized and experimentally validated that the effectiveness of BH3 mimetic compounds would be strongly dependent on the fibroblast growth environment. Taken together, our vi

study highlights the importance of understanding how environmental interactions alter the drug responses of cancer cells and reveals a mechanism by which stromal cells drive broad spectrum changes in tumor cell sensitivities to common chemotherapeutics.

vii

TABLE OF CONTENTS

ACKNOWLEDGEMENTS ...... iii ABSTRACT ...... v LIST OF TABLES ...... xi LIST OF FIGURES ...... xii PREFACE ...... xiv CHAPTER I: INTRODUCTION ...... 1 Traditional Methods of Cancer Treatment ...... 2 Surgery ...... 2 Radiation Therapy ...... 4 Cytotoxic Chemotherapies ...... 5 Mechanisms of Cell Death ...... 10 Apoptosis ...... 10 Breast Cancer ...... 13 Tumor Microenvironment ...... 15 CAFS ...... 15 CHAPTER II: A GFP-based co-culture methodology identifies fibroblast mediated desensitization of cancer cells to erlotinib and but fails to measure cell death...... 19 Abstract: ...... 19 Introduction: ...... 20 Results: ...... 22 Discussion: ...... 28 Materials and Methods: ...... 30 Data Availability: ...... 32 CHAPTER III: Tumor‐stroma interactions differentially alter drug sensitivity based on the origin of stromal cells ...... 33 Abstract: ...... 33 Introduction: ...... 34 Results: ...... 36 Cell‐intrinsic sensitivity to commonly used chemotherapy is similar for basal‐ like and mesenchymal‐like TNBC cells...... 36 viii

Coculture screen optimized to monitor cell death reveals widespread stromal influence on TNBC drug sensitivity...... 41 Primary fibroblasts grown in coculture with TNBC cells display an activated phenotype and modulate drug sensitivity similar to cancer‐associated fibroblasts (CAFs)...... 54 TNBC–fibroblast interactions are sufficient for inducing differential drug sensitivity between basal‐like and mesenchymal‐like TNBCs...... 56 Divergent interactions between TNBCs and fibroblasts based on the anatomical origin of fibroblast cells...... 60 Fibroblast‐dependent, and drug‐independent, variation in drug sensitivity occurs through modulation of the mitochondrial apoptotic priming state of TNBC cells...... 64 Discussion: ...... 72 Materials and Methods: ...... 76 Data Availability: ...... 84 Acknowledgements: ...... 84 Author contributions: ...... 84 Funding: ...... 85 CHAPTER IV: Fibroblast secretion of IL-8 sensitizes TNBC cells to DNA damage independent of altering apoptotic priming...... 86 Abstract: ...... 86 Introduction: ...... 87 Results: ...... 88 Discussion: ...... 94 Materials and Methods: ...... 96 CHAPTER V: DISCUSSION ...... 98 Matching your research question with the right cell death assay...... 100 Cancer:fibroblast interactions in vitro are sufficient to replicate clinically observed differences in drug sensitivity between subclasses of TNBC cells. .. 104 Fibroblast mediated alterations in drug sensitivity are diverse in not just the magnitude of the response but also the directionality...... 107 Fibroblasts derived from metastatic and non-metastatic tissues correlate with the desensitizing and sensitizing phenotypes...... 109 Are TNBC preferentially enriched for sensitizing interactions? ...... 110 ix

Do the sensitizing and desensitizing phenotypes validate in vivo? ...... 111 Exploring beyond fibroblasts, adding immune cells to the environment...... 112 Diving into the data, directions for future research...... 113 What are the molecular mechanisms underlying the change in apoptotic priming? ...... 115 Therapeutic implications of fibroblasts modulating apoptotic priming...... 116 What are the mechanisms involved in the drug specific and environmental independent interactions observed? ...... 118 Advances in understanding the microenvironmental regulation of TNBC drug response...... 118 APPENDIX A: Studying Cellular Signal Transduction with OMIC Technologies ... 121 Abstract: ...... 121 Introduction: ...... 122 Methods for High-Throughput Collection of Quantitative Signaling Data: ...... 125 Measurement of protein expression and post-translational modification levels...... 126 Mass spectrometry ...... 132 Immunoblotting ...... 134 Protein lysate microarrays ...... 135 Multiplexed flow cytometry, mass cytometry, and bead-based assays ...... 138 Caveats in antibody-based proteomics ...... 141 Methods to measure protein–protein interaction...... 142 Yeast 2-hybrid (Y2H) assays ...... 142 Membrane Y2H, MaMTH, and other protein complementation assays ...... 143 LUMIER ...... 145 Protein chips ...... 146 Quantitative measurement of protein localization or protein activity...... 147 Quantitative microscopy ...... 147 In vitro kinase and other in vitro enzyme assays ...... 149 Inference of pathway activity through transcription or functional genomics ...... 151 Experimental design for OMIC experiments: Basic principles and practical considerations...... 153 Conceiving OMIC experiments: Hypotheses Or discovery? ...... 158 x

Considerations in discovery science...... 159 Strong hypotheses for OMIC-scale systems biology...... 161 Approaches to Modeling Signal Transduction Processes: ...... 163 Mechanistic and semi-mechanistic approaches ...... 166 Data-driven modeling approaches ...... 169 Considerations for modeling signaling networks...... 173 Parameter estimation ...... 173 Choice of perturbations, observables, and sampling times ...... 176 Optimizing experimental design for modeling...... 177 Data pre-processing, normalization, and statistical analysis ...... 179 Conclusion: ...... 182 The philosophy of OMIC experiments: Will they get us closer to the truth? .. 182 Acknowledgements: ...... 185 REFERENCES ...... 186

LIST OF TABLES

Table 1.1: Traditional Chemical Chemotherapeutics ...... 6

Table 3.1: Biological reagents used in this study ...... 45

Table 3.2: Chemical reagents used in this study ...... 46

Table 3.3: Co-culture screening data ...... 47

Table 3.4: Co-culture screen statistics ...... 50

Table 5.1: Comparison of cell death assays ...... 100 xii

LIST OF FIGURES

Figure 1.1: The DNA Damage Response (DDR) network ...... 7

Figure 1.2: The intrinsic and extrinsic apoptotic signaling networks ...... 11

Figure 1.3: Apoptotic threshold is a function of the balance between pro and anti- apoptotic proteins ...... 12

Figure 2.1: Cancer-fibroblast interactions mediate the cytostatic effect of the targeted agent erlotinib ...... 23

Figure 2.2: Lung fibroblast mediated erlotinib desensitization is not consistent between lung cancer cell lines ...... 25

Figure 2.3: GFP-based measurements do not accurately report cell death ...... 27

Figure 3.1: TNBC cell sensitivity to topoisomerase inhibition is not well predicted by basal‐like versus mesenchymal‐like gene expression status ...... 38

Figure 3.2: Variable sensitivity to common DNA damaging agents in TNBC cells ...... 40

Figure 3.3: Coculture screen to identify tumor–stroma interactions reveals divergent interactions between fibroblasts and TNBC cells ...... 43

Figure 3.4: Statistical analysis of co-culture screen to identify TNBC-fibroblast interactions that significantly alter drug sensitivity ...... 48

Figure 3.5: Fibroblasts influence drug sensitivity across TNBC cell lines ...... 51

Figure 3.6: Nearly all classes of drugs are altered by TNBC-fibroblast interactions ...... 52

Figure 3.7: Fibroblasts specific modulation of drug responses in TNBC cells .... 53

Figure 3.8: Primary fibroblasts grown in culture display an activated phenotype, similar to Cancer Associated Fibroblasts (CAFs) ...... 55

Figure 3.9: PCA of TNBC drug responses in co-culture ...... 57

xiii

Figure 3.10: Drug responses of TNBC grown in coculture—but not in monoculture—are sufficient to distinguish basal‐like and mesenchymal‐ like TNBC subclasses ...... 58

Figure 3.11: Divergent interactions between TNBCs and fibroblasts based on fibroblast tissue of origin ...... 61

Figure 3.12: The first principle component captures variation associated with fibroblast tissue of origin ...... 62

Figure 3.13: Conceptual mechanisms by which fibroblasts could alter broad- spectrum drug sensitivities of associated cancer cells ...... 65

Figure 3.14: Fibroblasts alter drug sensitivity through modulation of mitochondrial apoptotic priming ...... 66

Figure 3.15: Modulation of mitochondrial priming by fibroblasts alters sensitivity to BH3 mimetic drugs ...... 71

Figure 3.16: Growth in matrigel desensitizes TNBC sensitivity to doxorubicin regardless of the fibroblast co-culture environment ...... 75

Figure 4.1: Conditioned media modulates cancer cell drug sensitivity ...... 89

Figure 4.2: IL-8 cytokine increases cancer cell drug sensitivity independent of apoptotic priming ...... 92

Figure A.1: Common mechanisms of signal transduction ...... 124

Figure A.2: Overview of OMIC approaches used in signaling research ...... 127

Figure A.3: Information overload ...... 163

Figure A.4: Methods of modeling signaling data ...... 165 xiv

PREFACE

Figure 2.3 and Chapter III was taken almost verbatim from:

Landry BD, Leete T, Richards R, et al. Tumor‐stroma interactions differentially alter drug sensitivity based on the origin of stromal cells. Mol Syst Biol. 2018;14(8):e8322. doi:10.15252/msb.20188322.

With author contributions as listed:

This project was conceived by BDL and MJL. Experiments were designed by BDL, HRS, ADS, SRP, and MJL. Coculture screen was designed and executed by BDL. All other experiments were executed by BDL, TL, RR, PC‐G, HRS, and GR. Analysis was conducted by BDL, TL, RR, HRS, MEH, and MJL. Manuscript was written and edited by BDL, ADS, SRP, and MJL.

Specifically:

Thomas Leete performed the experiments in Fig 3.3F-I and analysis in 3.14D Ryan Richards performed the experiments in Fig 3.3F-I and Fig. 3.14H and analysis of 3.1C Megan Honeywell performed analysis in Fig. 3.15B-C Peter Cruz-Gordillo contributed data to Fig 3.1A-B and Fig 3.2A-C Hannah Schwartz performed the experiment in Appendix Figure S2E in the above which was excluded from this thesis. Michael Lee was instrumental in performing the analyses in Fig 3.1, 3.2, 3.4, 3.9, 3.10, 3.12 and contributed to experimental design

Appendix A was taken almost verbatim from:

Landry BD, Clarke DC, Lee MJ. Studying Cellular Signal Transduction with OMIC Technologies. J Mol Biol. 2015;427(21):3416-3440. doi:10.1016/j.jmb.2015.07.021.

CHAPTER I: INTRODUCTION

Cancer is one of the leading ailments worldwide, accounting for roughly 1 in 7 deaths according to recent data from the American Cancer Society.

Throughout the 20th and 21st century, great strides have been made both in our understanding of the underlying biology of cancer cells, and in the development of therapies to treat a range of tumor types. However, cancer is a heterogeneous disease, in which patient-to-patient variability prevents the creation of a single universal cure. Therefore, it is likely necessary to tailor therapies to the features of an individual cancer type, or even an individual patient tumor. Yet, an important unanswered question remains: what information should be considered for optimal disease stratification? My thesis research is focused on understanding the contribution of the tumor microenvironment to treatment outcomes. Currently, disease stratification is dominated by the identification of specific genomic mutations or aberrations, with little regard for environmental contributions. In contrast, I aimed to test how different normal cells that reside within tumors influence cancer cell drug sensitivity. This information may ultimately be valuable for predicting optimal therapies and improving patient survival rates.

In the remainder of my thesis introduction, I will explore the history of chemotherapy, from the advent and use of traditional cytotoxic agents, to more modern targeted therapies. A key argument will be that our pursuit of novel therapeutic targets needs to evolve beyond genomic identification of “oncogenic 2

drivers”, and should also include cell non-autonomous sources, such as the tumor microenvironment. I have developed a high-throughput assay optimized to measure cell death to test the effect of environmental perturbations on cancer drug sensitivity, using triple-negative breast cancer (TNBC) as a model system, and fibroblasts derived from different anatomical locations as an environmental variable. A statistical and bioinformatic analysis of this study revealed an understanding of the interactions that control tumor drug response, and I investigated the molecular mechanisms responsible for fibroblast dependent changes in TNBC drug sensitivity. The insight gained from my data will advance the field’s understanding how cell death is regulated in the context of cancer treatment and may result in an improved ability to create more effective drug combinations for treating aggressive tumors.

Traditional Methods of Cancer Treatment

Surgery

The earliest treatment of cancer was the use of surgical excision, dating back to ancient times, with records detailing Roman physicians removing the breast tissue of women with malignant tumors1. Although physicians would excise the tumorous tissue, they would later observe the recurrence of cancerous tumors in patients, with the Roman physician Celsus noting “After excision, even when a scar has formed, none the less the disease has returned”2. As a result,

Roman physician Galen promoted the idea of rarely performing surgery as it 3

often accelerated death due to surgical complications rather than cure the disease1,3,4.

Advances in aseptic techniques reduced the number of complications from surgery allowing physicians to remove more and more tissue around the tumor mass. These advances culminated in a procedure called Halsted’s radical mastectomy4. This procedure initially included only removal of the entire breast tissue, but then advanced to include removal of the regional lymphatics, and eventually also included removal of the pectoralis muscles. However, a key issue that was not addressed by even the most aggressive surgical methods was the existence of distant growths of tumor cells, called metastases.

By the late 1950’s, Bernard and Edwin Fisher began to investigate the biology of tumor metastasis and would eventually invalidate Halsted’s use of radical surgery as the ideal cancer treatment. Fisher proposed that cancer was both a regional and systemic disease and that primary tumors had the ability to yield cancer cells that could disseminate throughout the body using the lymphatic and circulatory systems4. To test this contradicting hypothesis, Fisher conducted a series of experiments that included tracking radiolabeled cancer cells traveling through the blood stream, as well as measuring the impact of removing regional lymph nodes on the occurrence of metastases5–7. Based on these data, Fisher concluded that radical surgeries to remove large amounts of tissue could not prevent tumor metastasis, which prompted a clinical trial to compare Halsted’s mastectomy to less invasive procedures. Surprisingly, this study found no 4

significant difference between the treatment groups in 25 years of follow up appointments8. These findings shifted the standard surgical treatment of cancer to less invasive surgeries.

Although there have been great advances in the field of cancer surgery, surgical removal of the tumor by itself is generally not sufficient to cure cancer due to tumor cells leaving the primary site4. Cancer metastasis has spurred the search for finding new treatment strategies to pair with cancer surgery that can target cells that were left behind after surgical excision.

Radiation Therapy

Despite surgery being the main technique to treat cancers in the late 19th century and early 20th century, some cancer types, such as head and neck cancers, were inoperable leading to the critical need for alternative treatment methods. The field of radiation oncology began after the discoveries of X-rays by

Wilhelm Roentgen in 1895 and the discovery of radium by Marie and Pierre

Curie9,10. Early treatments consisted of giving patients a single large dose of radiation which would kill the tumor but at the cost of heavy damage to the surrounding normal tissue, which prompted the development of frequent smaller doses – called fractionated radiotherapy – to minimize harm to surrounding normal tissue9,11,12. During the same time period it was shown that radiation exposure induced mutations in DNA and that rapidly dividing cells and undifferentiated cells, such as the gametes, were more susceptible to radiation 5

treatments13–15. This prompted the study of how DNA damage could be used to preferentially kill cancer cells. However, this topic is still a matter of great debate, and a consensus has not been reached regarding the reasons why DNA damage preferentially kills tumor cells. Currently roughly 50% of patients receive radiation therapy as part of the treatment of cancer, whether alone or in conjunction with surgery16. Yet, radiation therapy remains a local cancer therapy and the need to treat metastatic tumors has led to the development of more systemic treatment options.

Cytotoxic Chemotherapies

The need for a systemic treatment to eradicate both the primary tumor tissue and metastatic cells cultivated the use of chemical compounds as a treatment method. One of the first chemical compounds used as a cancer therapeutic was the use of nitrogen mustard in treating non-Hodgkin’s lymphoma.

Alfred Gilman and Louis Goodman showed that nitrogen mustard gas could be used to reduce tumor burden a rodent model of lymphoid tumors and Gustaf

Lindskog showed the efficacy in lymphomas in human patients17–19.

Nitrogen mustards induce DNA damage in cancer cells by forming bulky

DNA adducts as well as intra or inter-strand crosslinks20. Crosslinks on DNA lead to the disruption of the transcription machinery and, during DNA replication, double strand breaks as well as replication lesions21. The observation that DNA damage induced by nitrogen mustards kills cancer cells, while minimizing 6

damage to the surrounding normal tissue, led to the creation of a wide variety of

DNA damaging chemotherapeutics which can be summarized in table 1.1.

Damage to DNA activates a signaling pathway called the DNA Damage

Response (DDR), which coordinates several cell behaviors, including cell growth arrests, DNA repair, and, if the damage is severe, activation of programmed cell death22. There are three layers to the DDR, proteins that sense DNA damage, proteins that transmit and amplify the signal, and effector proteins that coordinate the cellular processes of cell cycle arrest, repair or death (see Figure 1.1 for overview). For single strand breaks (SSBs), RPA coated DNA recruits and 7

activates the ataxia telangiectasia RAD3-related (ATR) kinase. ATR in turn phosphorylates and activates the checkpoint kinase CHK1, which phosphorylates many targets, including CDC25 causing cell cycle arrest23. In the case of double strand breaks (DSBs), the MRE11-RAD50-NDS1 (MRN) complex is recruited to the damage site and recruits the ataxia telangiectasia mutated (ATM) kinase.

ATM is activated by autophosphorylation and propagates the damage signal by phosphorylating and activating the checkpoint kinase CHK2. This kinase then phosphorylates and activates effector proteins such as CDC25A, resulting in cell 8

cycle arrest, BRCA1, resulting in activation of repair machinery to fix the DSBs as well as S-phase arrest, and p53, resulting in cell cycle arrest or induction of apoptosis24.

Chemical compounds that induce DNA damage are thought to leverage inherent differences in DNA damage sensitivity between cancerous and non- cancerous cells. In some cases, the sources of this difference are well understood. For example, some cancer cells have compromised DNA repair machinery. In most cases however, the sources of these differences are not clear. A common interpretation is that DNA damage targets all rapidly dividing cells. This speculation was recently tested by Tony Letai and colleagues who found it to be a poor explanation of their data, as they observed that many slow growing cancers are sensitive to DNA damage and many fast growing tissues are insensitive25. They later demonstrated that chemosensitivity was more closely correlated to the intrinsic apoptotic priming state of the cancer cells.

In addition to chemical compounds that cause direct DNA damage, many effective compounds function by interfering with cell cycle progression and cell division. During studies concerning nutrient requirements for proper development, it was observed that folic acid deficiencies in humans led to reduced bone marrow, similar to effects of mustard gas exposure. This observation spurred Sydney Farber to develop a series of folate antagonists, such as methotrexate, for the treatment of children with leukemia26. Methotrexate binds to dihydrofolate reductase (DHFR), outcompeting folate, and prevents the 9

formation of tetrahydrofolate, an important co-factor in the synthesis of thymidine26. The resulting decrease in nucleotide pools activates the G1/S cell cycle checkpoint and arrests proliferation27. However, cells that are progressing through S-phase or cancers that are checkpoint deficient incur DNA damage due to deficient nucleotide pools which results in cell death27,28. Recently, it has been also been shown that 5-FU and FUDR, chemotherapeutics that also affect thymidine synthesis, induce cell death through bacterial ribonucleotide metabolism in an in vivo environment using C. elegans as a model organism29.

The G1/S checkpoint and entry into the cell cycle was not the only targetable cell cycle phase as drugs were developed to disrupt cell progression through mitosis and division.

During the 1960s the National Cancer Institute set up a chemical screening platform to test the efficacy of compounds submitted by scientists and companies. One such screen tested extracts of over 1,000 plant species and observed that a compound in bark from the Taxus brevifolia tree was toxic to cancer cells30. This compound, later named paclitaxel, was found to target the protein tubulin and stabilize microtubules from disassembling31,32. Stabilization of microtubules results in the inability to properly form of the mitotic spindle causing cell cycle arrest during mitosis. Checkpoint deficient cells that progress through mitosis incur DNA damage through improper segregation of the chromosomes, resulting in cell death, while cells that experience prolonged activation of the mitotic checkpoint activate apoptosis, also resulting in cell death33. The 10

development of the aforementioned compounds, which act on pathways that are intrinsic to all cells and are able to circulate throughout the body, allowed for the treatment cancer in a systemic manner. These compounds, along with surgery and radiotherapy, have greatly improved outcomes for many cancer patients.

Unfortunately, for many subtypes of cancer – and for essentially all types of metastatic cancer – effective therapies are still lacking, underscoring the need for new therapeutic concepts.

Mechanisms of Cell Death

Apoptosis

The ability of a cell to commit death, when needed, is a necessary and highly regulated process. One of the most common forms of programmed cell death is apoptosis, a non-inflammatory process that kills cells using mitochondrial rupture and caspase activation. Apoptosis is controlled by two pathways, the intrinsic apoptotic pathway, which responds to internal cellular stresses like DNA damage, and the extrinsic apoptotic pathway, which responds to extracellular signals, such as TNF34–37.

In brief, the intrinsic apoptotic pathway is initiated by stress signals, which lead to the formation of pores in the mitochondrial outer membrane. A critical event following mitochondrial outer membrane permeabilization (MOMP) is release of cytochrome C, which catalyzes formation of the apoptosome. The critical function of the apoptosome is cleavage and activation of “executioner” 11

caspases, such as caspase-3. Historically, activation of executioner caspases was considered the mechanism of cell death (active caspases are proteases that

can cleave a wide variety of protein substrates, leading to cell death); however, it is now generally accepted that cells cannot survive following MOMP. Thus, downstream activation of effector caspases is more likely involved in suppressing the release of inflammatory signals from apoptotic corpses. The extrinsic apoptotic pathway is initiated by the binding of cognate signaling molecules to a 12

class of cell surface proteins called death receptors. Ligand binding leads to oligomerization, recruitment of adaptor proteins, and ultimately to the recruitment and activation of the initiator caspases 8/10. These initiator caspases in turn activate the effector caspases to induce protein cleavage as well as cleave the protein BID to facilitate Mitochondrial Outer Membrane Permeabilization (MOMP) and release of cytochrome C (See Figure 1.2 for an overview).

In the context of drug therapy, the degree of cell death is a function of two properties: the degree to which the drug activates the intrinsic or extrinsic pathways, and the degree to which the cell is “primed” for death. Apoptotic priming is thought to be controlled by the local ratio of pro- and anti-apoptotic proteins at the mitochondrial membrane (see figure 1.3 for an overview). When pro-apoptotic signals outweigh the anti-apoptotic signals, the cell passes the apoptotic threshold and is committed to cell death36. Identification and 13

measurement of the apoptotic threshold has been challenging as it is a function of protein expression, location, and a complex network of interactions. Recently an assay has been developed, called BH3 profiling, that determines the apoptotic threshold of cells by measuring the amount of synthetic pro-apoptotic peptide required to rupture a cells mitochondria38.

Modulation of apoptosis can occur either by altering the drug induced signaling dynamics or by altering the levels of pro or anti-apoptotic proteins. It has been shown that interaction with the microenvironment, in this instance cancer associated fibroblasts, is able to change cancer cell drug response through mediating apoptotic priming39. However, it is unclear if alternate signals from the tumor microenvironment change how damage signaling is processed to alter the onset of apoptosis.

Breast Cancer

Breast cancer is the most frequently diagnosed cancer among American and European women40,41. Despite improvements in therapy that have increased survival rates, breast cancer remains the second leading cause of cancer deaths in women worldwide42. Breast cancer is a heterogeneous disease comprising of several subtypes with distinct histological, molecular, and clinical features. Triple negative breast cancer (TNBC) is a subtype of breast cancer that accounts for nearly 20% of patient cases42. TNBC is characterized by the absence of the amplification of the estrogen receptors, progesterone receptors, and excess 14

HER2 protein. Due to the lack of targetable receptors, TNBCs have limited therapeutic options and are commonly treated with DNA damaging agents.

Breast cancers cause death by metastasizing to secondary organs and interfering with essential biological functions. TNBC is one of the most aggressive subtypes of breast cancer and metastasizes mainly to the brain, lungs, liver or bone marrow. Once TNBCs have metastasized to a secondary site they become less sensitive to chemotherapeutics that were effective at treating the primary tumor. This fact has spurred the need for a better understanding on what factors contribute to drug resistance at secondary sites. In this thesis I aim to explore how microenvironmental factors, primarily fibroblasts, contribute to this change in drug sensitivity.

When designing my screening platform, I took care to choose fibroblasts from common metastatic sites of TNBC as well as non-common metastatic sites.

This decision was made not only to increase the diversity of the fibroblast panel but to explore whether the microenvironment of metastatic sites differentially altered the sensitivity of TNBC. I hypothesized that fibroblasts derived from common metastatic sites would desensitize TNBC cell lines to DNA damaging agents while fibroblasts from non-metastatic sites would have no effect on drug sensitivity.

Tumor Microenvironment

It is becoming well appreciated that tumor-stroma interactions can alter many behaviors of tumor cells, including growth rate, gene expression state, and even drug responses. Therefore, to effectively treat cancer, it is crucial to understand the nature of tumor-stroma interactions, and in particular, which interactions cause changes in drug sensitivity. Most studies that investigate the tumor microenvironment use in vivo systems. While these studies demonstrate the importance of the tumor microenvironment for various tumor phenotypes, they are generally unable to dissect which specific interactions contribute to the phenotype due to the complexity of the tumor microenvironment. In my thesis, I aim to use an in vitro co-culture system, in which I can selectively vary one component of the microenvironment, to investigate the role of individual environmental components on cancer cell drug sensitivity and cell death. For my initial study I chose to investigate the role of cancer associated fibroblast on cancer cell drug sensitivity and cell death.

CAFS

Normal fibroblasts maintain normal tissue structure and function through the deposition of matrix and secretion of both cytokines and growth factors.

During tumorigenesis, cancer cells hijack normal fibroblasts whose gene expression pattern changes to promote tumor, growth, angiogenesis, invasion and metastasis. Cancer associated fibroblasts (CAFs) adopt a gene expression 16

state similar to myofibroblasts, a cell type rarely observed in healthy tissue but involved in wound healing43. CAFs are identified by the expression of distinct marker proteins such as α smooth muscle actin (α SMA), fibroblast activation protein (FAP), and fibroblast specific protein 1 (FSP1)43. Resident normal fibroblasts primarily differentiate into CAFs by responding to signals from cancer cells, such as TGF-b, miR-31, miR-214, miR155, and CXCL12, but other cell types have also been observed to adopt a CAF like gene expression state44,45.

CAFs in the tumor microenvironment secrete a wide range of signaling molecules that support cancer cells during tumorigenesis. CAFs promote cell proliferation by secreting growth factors such as the hepatocyte growth factor

(HGF), fibroblast growth factor (FGFs) and insulin-like growth factor 1 (IGF1)46–48.

Invasion and metastasis are promoted by the secretion of metalloproteases that degrade the surrounding extracellular matrix as well as TGF-b that aids in the

EMT transitions of metastatic cancer cells43,46. CAFs not only create an environment for cancer cells to thrive but they are able to modulate cancer cells’ response to chemotherapeutic drugs. Studies have shown that CAFs desensitize cancer cells to targeted agents, however there is a dearth of knowledge in how interactions with fibroblasts mediate DNA damage induced death39,49. My thesis research will focus on studying the interactions between fibroblasts and triple- negative breast cancer cells and identifying tumor-stroma co-culture conditions that alter cancer cell drug sensitivity. I will explore the extent of the role these 17

interactions have in modulating apoptotic signaling dynamics as wells as the apoptotic threshold.

There have been few studies that have investigated the effect of the tumor microenvironment on cancer drug sensitivity, specifically with a focus on fibroblasts. Straussman et al. was one of the first studies that explored these complex interactions between tumor cells and fibroblasts. In this study, the authors developed a high throughput co-culture system to screen the interactions between 23 fibroblasts and 45 cancer cells when treated with 35 chemotherapeutics. The study demonstrated that fibroblast mediated desensitization to chemotherapeutics is common, especially to targeted agents, and further identified an interaction where fibroblast secretion of HGF decreased the effeteness of RAF inhibition in BRAF-mutant melanoma49. Marusyk et al. further advanced the field with their study showing that CAFs derived from breast tissue or brain metastases increased the resistance of HER2+ breast cancer cells to the targeted inhibitor lapatinib39. The study also demonstrated that normal fibroblasts derived from breast or brain tissue had the same effect on cancer cell drug response as CAFs.

Despite the previous studies laying a strong foundation, there are gaps in the data that I aim to address in this thesis. First, both studies focused on targeted cancer therapies and did not determine how fibroblasts effect of DNA damaging agents in a comprehensive manner. Second, neither study had a focus on TNBC, with the Straussman et al. study spanning seven different cancer types 18

and the Marusyk et al. study focusing on HER2+ breast cancer. Futhermore, neither study investigated whether there was a difference in fibroblast mediated changes in chemosensitvity based on whether the fibroblast was derived from a common metastatic or non-common metastatic site for that cancer type. Marusyk et al. focused on CAFs derived from either the primary breast tissue or metasteses from the brain, a common metastatic environment for breast cancer.

The diversity of the cancer types in the Staussman et al. study prevented them from investigaing the division between metastaitic and non-metastatic fibroblasts, with the study mainly spanning three different fibrobalst types focusing on skin, lung, and breast. In this thesis, I aim to screen TNBCs treated with a comprehensive list of chemotherapeutics co-cultured with either fibroblasts derived from common or non-common metastatic sites using an assay that specifically measure cell death. I aim to address some of the gaps in the literature presented above and advance the knowledge of the field in order to create more effective therapies for the treatment of TNBC. 19

CHAPTER II: A GFP-based co-culture methodology identifies fibroblast

mediated desensitization of cancer cells to erlotinib and but fails to

measure cell death.

Abstract:

The tumor microenvironment plays a pivotal role in the growth, progression, and survival of cancer cells. Cancer associated fibroblasts (CAFs) are one of the main components of the tumor microenvironment and have recently been shown to alter cancer cell drug sensitivity to targeted agents. Yet, it remains unclear how these interactions mediate cancer cell drug sensitivity to cytotoxic agents. Identification of novel fibroblast-tumor interactions that modulate drug sensitivity, as well as understanding the underlying molecular mechanisms, will aid in the development of more effective strategies to treat cancer patients. However, identifying these interactions is challenging due to the need to capture robust cell viability measurements, specific to the cancer cell population, in co-culture. Here, I demonstrate that a GFP-based co-culture assay captures fibroblast mediated cancer cell proliferation effects but fails to accurately measure cancer cell death. Additionally, co-culture induces a pro- growth effect in cancer cells and reduces the effectiveness of the targeted inhibitor erlotinib. The magnitude of these phenotypes is dependent on the fibroblast:cancer ratio, with the magnitude increasing as the ratio increases, as well as the tissue identity that the fibroblast was obtained from. This study 20

highlights the need to develop a new assay, optimized in measuring cancer cell specific death in co-culture, to investigate the effect of fibroblast-cancer interactions on cancer cell drug sensitivity to cytotoxic agents.

Introduction:

Drug resistance remains a challenging hurdle to overcome in the treatment of cancer patients. Complete tumor remission following exposure to either cytotoxic or targeted agents is rare, and a growing body of evidence suggests that environmental interactions play a role in mediating cancer drug resistance50. One of the most abundant non-malignant cell types that reside within tumors are cancer associated fibroblasts (CAFs). These normal stromal cells play an active role in remodeling the tumor microenvironment to promote tumor growth and progression43,51–53. CAFs are a major source of growth factors, cytokines, and matrix deposition. Each of these factors has been shown to alter cancer cell sensitivity to targeted therapies39,49. However, there remains a lack of understanding of how CAF mediated interactions with tumor cells modulate drug response to cytotoxic agents.

A mechanistic understanding of how interactions between fibroblasts and cancer cells modulate the behaviors of tumors has been a challenge due to heterogeneity within and between patients/animals. Primary tumors can metastasize to a variety of locations in the body and interact with tissue specific 21

fibroblasts. Therefore, characterization of these interactions requires the study of fibroblasts obtained from a variety of anatomical locations.

A successful assay to measure the effect of fibroblasts on cancer cell drug response contains the ability to measure the viability of the cancer cell population independent of the fibroblast cells. A high-throughput co-culture assay has been developed, which uses GFP-labeled cancer cells and unlabeled fibroblasts.

Measuring GFP fluorescence is used to specifically quantify the life/death of cancer cells regardless of the existence of other non-cancer cell types.

Furthermore, GFP is amenable to measurement using a fluorescent plate reader, facilitating high-throughput and time-resolved measurements. This technique was recently used to test a large number of fibroblast-cancer-drug interactions and revealed a mechanism for fibroblast mediated resistance to targeted RAF- inhibitors49. In this chapter, I demonstrate that the GFP based screening assay is proficient in capturing fibroblast-cancer-drug interactions that effect cell proliferation but is not accurate in measuring cell death. I show that fibroblasts generally induce a pro-proliferative phenotype when co-cultured with lung and breast cancer cell lines, as well as describe a tissue specific drug desensitization interaction in the lung cancer line A549 when treated with the EGFR inhibitor erlotinib. When testing the assay with cytotoxic agents, I found that the use of

GFP-labeled cells was not accurate in measuring cell death.

Results:

To test the ability of a recently published in vitro GFP-based assay to measure tumor-fibroblast mediated effects on drug sensitivity, I piloted a small screen designed to measure assay sensitivity in capturing proliferation and cell death phenotypes. I selected a panel of primary fibroblasts to be co-cultured, at varied ratios, with either triple-negative breast cancer (TNBC) or lung cancer cell lines that were stably transfected to constitutively express a soluble cytoplasmic

GFP molecule. Each co-culture condition was treated with the targeted agent erlotinib, an EGFR inhibitor, or the cytotoxic drug camptothecin, a topoisomerase inhibitor (Fig. 2.1A). Included in the panel was a TNBC cell line, BT20, whose interactions with both erlotinib and camptothecin in mono-culture have previously been described54. To determine the basal effect of co-culture on the growth rate of tumor cells in the absence of any drug, I compared the growth rates of cancer cells grown in mono-culture or in co-culture with fibroblasts. Relative growth rate was approximated using the area under the curve (AUC) of the growth time course (Fig 2.1B). Generally, co-culture of tumor cells with fibroblasts promoted increased cancer cell growth, regardless of fibroblast tissue identity. This growth phenotype scaled with the fibroblast:cancer ratio, with a greater proliferative advantage seen as the fibroblast:cancer ratio is increased (Fig 2.1B and D).

Surprisingly, I observed a growth disadvantage in the TNBC line MDA-MB-231 when cultured with fibroblasts derived from the tissue of origin, the primary breast

fibroblast HMF, as well as a common metastatic tissue, the lung derived fibroblasts IMR90 and WI38 (Fig 2.1B and C). This phenotype once again scaled with fibroblast:cancer ratio, with tumor growth reaching only 50% of mono-culture in the 2:1 fibroblast:cancer ratio (Fig 2.1C).

Next, I wanted to determine if co-culture of tumor cells with fibroblasts alters the drug response to the EGFR inhibitor erlotinib. Inhibition of EGFR impairs growth factor signaling and induces a cytostatic proliferation phenotype, as seen in BT20 and A549 tumor cell lines (Fig 2.1E). Subsequently, co-culture of tumor cells with fibroblasts generally had a neutral effect on cancer cell drug response, displayed as white, or a drug desensitization effect, displayed as red

(Fig 2.1F). Fibroblast mediated desensitization to erlotinib varied by tumor cell line, with only the bone fibroblast Hs-5 and the brain fibroblast HBVAF desensitizing all 4 lines. The degree of erlotinib desensitization also varied between each fibroblast co-culture condition. The strongest desensitization phenotypes were observed from fibroblasts that were derived from either the brain, bone marrow, or lungs, all common sites of breast cancer metastasis. As seen with the proliferation phenotype, the degree to which erlotinib was inhibited depended on the fibroblast:cancer ratio (Fig 2.1G).

One of the strongest erlotinib desensitizing effects was observed in the lung cancer cell line A549 when co-cultured with the lung derived fibroblast

IMR90 (Fig 2.2A). This was surprising as targeted EGFR inhibitors, such as

erlotinib, were developed and have been successful in treating non-small cell lung cancer (NSCLC) in patients that have mutations in EGFR. The observation that A549 is desensitized to erlotinib in a lung environment is concerning for the creation of patient therapies, so I wanted to investigate whether this phenotype was conserved among other lung cancer cell lines. I selected a panel of lung cancer cell lines, varying EGFR, KRAS, and p53 mutation status, to test the effect of fibroblast:cancer co-culture on erlotinib efficacy (Fig 2.2B). Once again co-culture with fibroblasts generally increased tumor cell proliferation compared to mono-culture, although the lung line H727 displayed a neutral effect that transitioned into inhibitory as the fibroblast:cancer ratio increased (Fig 2.2C).

Fibroblast mediated altering of drug response was variable across the lung cancer cell lines (Fig 2.2D). Generally, co-culture with fibroblasts desensitized cancer cells to erlotinib. However, there was little to no effect observed in the lung lines H727 and H1650 (except in the lung environment), the only two lines that contained mutated p53 signaling. The greatest degree of erlotinib desensitization was observed in the co-culture conditions with either brain, bone marrow or lung derived fibroblasts. Importantly, the strong lung fibroblast mediated desensitization phenotype observed in the A549 cell line was not consistent across the expanded lung cancer cell line panel. The cell lines KP7B and H358 displayed a similar trend as A549, and both cell lines contained WT

EGFR and p53. However, to determine if the lung specific desensitization to

erlotinib is significant or enriched in a subtype of lung cancer, the panel of lung cancer cell lines would have to be further expanded.

After observing that the GFP assay was adept at measuring proliferation related phenotypes involving targeted agents, I wanted to test the assay’s sensitivity in capturing cell death in response to cytotoxic agents. Analysis of

GFP microscopy images revealed that co-culture with the adrenal gland derived fibroblast HADF desensitized BT20 cells to the cytotoxic therapy camptothecin

(Fig 2.3A and B). Interestingly, analysis of this same condition using a fluorescence plate reader failed to capture the difference in death between cancer cells in monoculture and co-culture to camptothecin (Fig 2.3C). In addition to not capturing the fibroblast influence on camptothecin sensitivity, plate reader‐ based analysis also failed to capture cell death, with all fluorescence measurements recording higher values than the initial pre‐drug measurement

(Fig 2.3C). The inability of the plate reader-based approach to capture cell death was likely due to the stability of GFP fluorescence, which is still fluorescent even after the cells have died (Fig 2.3D). These data indicate that measuring GFP fluorescence using a fluorescence plate reader is not sensitive in quantifying the degree of cell death in a population of cells.

Discussion:

In this chapter I explored the ability of a high-throughput assay utilizing

GFP expressing cancer cells and a fluorescent plate reader to capture 29

proliferation and cell death phenotypes in an in vitro co-culture system. Using this

GFP based assay, I was able to observe a pro-proliferation phenotype in both

TNBC and lung cancer cell lines when co-cultured with primary fibroblasts in vitro. When comparing the proliferation phenotypes for one cancer cell line across the 8 fibroblasts conditions, there was a general pro-proliferation trend that lacked tissue specificity. This suggests that the mechanism that increases cancer cell growth is a general fibroblast feature and does not rely on tissue specific differentiation.

Additionally, I demonstrated that co-culture of cancer cells with fibroblasts desensitizes cancer cells to the targeted EGFR inhibitor erlotinib. This is consistent with the overall finding from Golub and colleagues, which suggested that fibroblasts facilitate resistance to growth factor inhibition through secretion of growth factors49. Surprisingly, one of the strongest desensitizing environments, among the 4 cell lines tested in Fig 2.1F as well as among the different fibroblast environments, was the co-culture of A549 with the lung fibroblast IMR90. Despite this strong phenotype not being conserved across a variety of lung cancer cell lines, the observation of this tissue specific desensitization of A549 to erlotinib has clinical implications. Erlotinib was developed for the treatment of lung cancers and the observation that there are strong stromal specific interactions that lower the efficacy of the drug indicates that clinicians must treat patients on a case by case basis rather than prescribing one general therapy. Understanding these environmental interactions and discovering novel biomarkers to indicate 30

when the standard treatment of care will not be effective will dramatically increase our ability to treat cancer.

Finally, my data showed that the GFP assay, which was proficient in identifying proliferation related phenotypes, could not accurately measure cell death. The stability of the GFP molecule inflates the fluorescent signal and does not decrease when a cell dies. When the wells were washed with PBS to remove any soluble GFP molecules, the GFP signal decreased. However, I noticed that the number of healthy cells attached to the plate had also decreased indicating that the wash step was too harsh. These data indicate that washing to remove

GFP would not be viable in accurately measuring cell death. Given this, our data call into question prior studies that concluded that cytotoxic drugs are not affected by tumor-fibroblast interactions. Future efforts will rely on the development of new high-throughput screening approaches to investigate the effects of fibroblast cancer cell interactions on cytotoxic chemotherapeutics. One potential fluorescent based assay that would be compatible with the plate reader would be the use of the mitochondrial membrane dye JC-1. JC-1 undergoes a fluorescent shift when the mitochondrial membrane becomes depolarized, one of the initial steps of apoptosis. This method is able to report cell death without relying on the degradation of a fluorescent molecule.

Materials and Methods:

Cell lines and reagents

Cell lines BT‐20, MDA‐MB‐231, MDA‐MB‐468, H358, H727, H1650, H1975,

KP7B, A549, Hs-5, Hs27A, IMR90, WI38, and WS-1 were obtained from

American Type Culture Collection (ATCC, Manassas, VA). All cell lines were grown in 10% FBS (Thermofisher Hyclone cat# SH30910.03 lot# AYG161519), 2 mM glutamine, and penicillin/streptomycin. BT‐20, IMR90, WI38, and WS‐1 were cultured in Memα + Earle's salts. H358, H727, H1650, and H1975 were cultured in RPMI 1640 media. MDA‐MB‐231, MDA‐MB‐468, KP7B, Hs27A, and HS‐5 were cultured in Dulbecco's modified Eagle's medium (DMEM). Primary fibroblasts HMF, HADF, and HBVAF were purchased from ScienCell (Carlsbad,

CA). Primary fibroblast cells purchased from ScienCell were cultured in

Fibroblast Medium (cat# 2301) for four doublings before being transitioned to

DMEM. All cells were cultured at 37°C in a humidified incubator supplied with 5%

CO2 and maintained at a low passage number (< 20 passages for cancer). Prior to expansion and freezing, a small sample of each primary fibroblast was expanded to determine each cell's Hayflick limit to ensure that experiments could be performed prior to the onset of replicative senescence. Erlotinib was purchased from LC Labs (cat# E-4007) and suspended at 10mM in DMSO

Plate Reader Based GFP Assay

To determine cell proliferation and cell death using a fluorescence plate reader,

TNBC and Lung cancer cells were stably transfected with GFP (pRetroQ‐ 32

AcGFP1‐N1). Transfected cells were treated with puromycin and sorted to select for a pool of cells of similar GFP fluorescence. Cells were selected until a parallel non‐transformed plate exposed to puromycin was completely dead. For coculture experiments, fibroblast cell lines were plated at an 8:1, 4:1, 2:1, 1:1, and 1:2 ratio to cancer cells in a Greiner 96‐well plate in 100 μl of FluoroBrite media and allowed to adhere for 3 h. Following adherence of fibroblast cells, cancer cells constitutively expressing GFP were plated at a concentration of 5,000 cells per

100 μl of FluoroBrite media and allowed to adhere overnight. Drug was added the next day and cell measurements were measured every 24 h for 96 h using a

Tecan M1000 Pro Plate reader.

Fluorescence microscopy

For imaging of GFP‐labeled cells, cells were plated as describe in the plate reader methodology. Following drug exposure (various times as indicated in figures), images were collected using an EVOS FL‐AUTO automated fluorescence microscope. Analysis was performed using a custom CellProfiler pipeline.

Data Availability:

The CellProfiler script used for the analyses shown in Fig 2.3 can be viewed at

DOI 10.15252/msb.20188322 as Code EV1. 33

CHAPTER III: Tumor‐stroma interactions differentially alter drug sensitivity based on the origin of stromal cells

Chapter III was taken almost verbatim from: Landry BD, Leete T, Richards R, et al. Tumor‐stroma interactions differentially alter drug sensitivity based on the origin of stromal cells. Mol Syst Biol. 2018;14(8):e8322. doi:10.15252/msb.20188322.

Abstract:

Due to tumor heterogeneity, most believe that effective treatments should be tailored to the features of an individual tumor or tumor subclass. It is still unclear, however, what information should be considered for optimal disease stratification, and most prior work focuses on tumor genomics. Here, we focus on the tumor microenvironment. Using a large‐scale coculture assay optimized to measure drug‐induced cell death, we identify tumor–stroma interactions that modulate drug sensitivity. Our data show that the chemo‐insensitivity typically associated with aggressive subtypes of breast cancer is not observed if these cells are grown in 2D or 3D monoculture, but is manifested when these cells are cocultured with stromal cells, such as fibroblasts. Furthermore, we find that fibroblasts influence drug responses in two distinct and divergent manners, associated with the tissue from which the fibroblasts were harvested. These divergent phenotypes occur regardless of the drug tested and result from modulation of apoptotic priming within tumor cells. Our study highlights unexpected diversity in tumor–stroma interactions, and we reveal new principles that dictate how fibroblasts alter tumor drug responses.

Introduction:

A central challenge in medicine is selecting which drug or drug combination will be the most beneficial for a given patient. In cancer therapy, this decision has typically been based on the anatomical origin of the disease, in combination with drug screening to empirically identify the most efficacious compounds. In most cases, drug response rates vary considerably, and the causes of this response variability remain unclear. Thus, for ongoing efforts to improve precision/personalized medicine it is critical to identify features that contribute to the observed drug response variability.

Several studies now exist that have explored the relationship between tumor genetics or tumor gene expression and drug response55–61. Many insights have been gained from these and other studies, but even collectively, these studies fail to create a clear understanding of the variable levels of sensitivity to commonly used chemotherapeutics62,63. An important consideration is that substantial non‐genetic heterogeneity exists within tumors, and these influences are generally missed in studies that focus exclusively on tumor genomics. For instance, several classes of normal cells typically reside within tumors. It is increasingly recognized that many tumor phenotypes, including tumor initiation, epithelial‐to‐mesenchymal transition (EMT), metastatic potential, and drug sensitivity, are influenced by interactions between cancer cells and the normal cells residing within or near tumors64,65. Prior studies have typically highlighted the variable and unpredictable nature of tumor–stroma interactions, in which the 35

drug sensitivity of cancer cells appears to depend on the particular combination of tumor cell, stromal cell, and drug used66. Thus, for efforts to improve precision medicine, a critical unmet need is to learn “rules” dictating how stromal cells influence the drug sensitivity of cancer cells.

Here, we develop a mixed coculture assay optimized to specifically quantify cell death, rather than cell proliferation, and we use this assay to characterize functional interactions between tumor cells, stromal cells, and anticancer chemotherapeutic agents. We report that stromal fibroblasts influence tumor drug response in two distinct and divergent manners: Some interactions result in drug resistance, while others cause drug sensitization. Importantly, these divergent influences were associated with the anatomical tissue from which the fibroblasts were harvested. Surprisingly, our data show that these distinct fibroblast‐dependent phenotypes are conserved regardless of the identity or molecular target of the drug. These broad‐spectrum changes to drug sensitivity result from modulation of mitochondrial apoptotic “priming”, which changes the threshold for initiation of apoptosis in cancer cells. Taken together, our study highlights previously unappreciated principles, which dictate how stromal fibroblasts alter a tumor cell's drug response. 36

Results:

Cell‐intrinsic sensitivity to commonly used chemotherapy is similar for basal‐like and mesenchymal‐like TNBC cells.

To explore cell non‐autonomous regulation of drug sensitivity, we began by focusing on a tumor subclass that displays notable response heterogeneity, without clear mechanisms that underlie these differences. “Triple‐negative” breast cancers (TNBCs) are the most chemosensitive subtype of breast cancer, but also the subtype with the shortest disease‐free survival and lowest overall survival rates67,68. This paradox is thought to result from heterogeneity within the

TNBC subclass69. Additionally, although TNBC can be further stratified into several definable groups which differ in chemosensitivity69, it remains unclear which features are responsible for creating the variable drug sensitivity that is observed clinically.

To highlight this variability in the drug response, we selected a panel of ten TNBC cells from either the “basal‐like” or “mesenchymal‐like” expression classes69–71. Basal‐like (BL) cells—sometimes referred to as “basal A”, “basal‐like

1”, and “basal‐like 2”—are defined by expression of basal or myoepithelial genes.

These cells are highly proliferative, tend to have elevated expression of DNA damage response genes, and generally respond at higher rates to cytotoxic chemotherapies in vivo69. Mesenchymal‐like (ML) TNBCs—which includes

“mesenchymal”, “mesenchymal stemlike”, and “claudin‐low” expression classes—are enriched for expression of genes related to EMT, and genes 37

associated with stemness. These cells are more “aggressive” clinically, more de‐ differentiated, more metastatic, and more chemoresistant in vivo69,72. Thus, we initially reasoned that identifying mechanisms which account for the variability in

DNA damage sensitivity between the BL and ML subclasses may aid in patient stratification or help to identify new strategies for improving responses to these agents.

To identify features that contribute to differential DNA damage sensitivity between BL and ML cells, we began by profiling the response of TNBC cells to doxorubicin (also called Adriamycin), a topoisomerase II inhibitor that is commonly used in the treatment of TNBC. We suspected that if the observed clinical patterns of aggressiveness were due to intrinsic differences in drug sensitivity associated with these gene expression states, different levels of sensitivity to doxorubicin should be observed in vitro. Indeed, the least sensitive cells were HCC‐1395, a TNBC of the ML expression state; the most sensitive cells were MDA‐MB‐468, a TNBC in the chemosensitive BL category (Fig 3.1A).

In contrast, however, the rest of the cell lines tested were similarly sensitive to doxorubicin, regardless of their gene expression state. To see whether this was unique to doxorubicin, we also profiled responses to other topoisomerase inhibitors in this panel of cells. Overall, these data reveal relatively similar levels of drug sensitivity across these 10 cell lines (Fig 3.2A). To more rigorously determine whether the patterns of sensitivity to these drugs could be used to distinguish BL versus ML cells, we performed hierarchical clustering using either 38

the EC50 or the maximum effect observed for each drug. This analysis also failed to correctly separate BL and ML cells based on their observed drug sensitivity profile (Fig 3.1B and Fig 3.2B).

The results from our in vitro analysis suggest that these BL and ML cells have similar sensitivity to commonly used chemotherapeutics. This, of course, is not in line with the expected observation that ML tumors respond at lower rates than BL tumors in vivo73. One explanation could be bias within our samples, as our dataset was comprised of a relatively small number of TNBCs. To address this, we also analyzed data publicly available through the LINCS consortium, which include drug sensitivities for a larger panel of 24 BL or ML cell lines (11 BL and 13 ML) and a larger panel of common anticancer drugs74. These data also show that BL and ML cells have similar levels of sensitivity to topoisomerase inhibitors, specifically, or to all anticancer drugs, generally (Fig 3.1C and Fig

3.2D). Thus, taken together, these data highlight that the subtype‐dependent differences in drug sensitivity, which may be expected given responses observed in patients, are generally not observed when these cells are grown in standard in vitro cell culture conditions.

Another potential explanation for the discrepancy between our data and the relative drug sensitivities that were expected could be that our cells were grown in 2D, rather than using 3D culturing conditions. It has generally been found that many cell behaviors differ when cells are grown in 2D versus 3D, and that 3D culture is in many ways a more accurate representation of the in vivo 40

environment75,76. To test whether growth in 3D recapitulates the expected distinction between BL and ML cells, we retested sensitivity to 10 topoisomerase inhibitors for TNBC cells grown as 3D colonies in a Matrigel growth environment.

Growth of these TNBC cells in 3D colonies strongly altered drug sensitivity (Fig

3.1D). In some cases, a modest trend was observable in which ML cells appear less sensitive to drugs (e.g., camptothecin), but these trends were not statistically significant. The dominant trend was an overall desensitization to these drugs, without further refining the distinction between BL and ML cells (P > 0.05 for all drugs in 2D and 3D culture; Fig 3.1D and Fig 3.2C). This finding is consistent with prior studies, which demonstrated that growth in 3D induces a general resistance to drug‐induced apoptosis77. Thus, taken together, our data highlight that these BL and ML TNBC cells are similarly sensitive to commonly used chemotherapeutics, at least when cultured in standard 2D or 3D monoculture conditions. This finding raises the possibility BL and ML subtype‐specific responses to treatment are not cell‐intrinsic properties, but rather a product of subtype‐specific interactions between tumor cells and microenvironmental features.

Coculture screen optimized to monitor cell death reveals widespread stromal influence on TNBC drug sensitivity.

Based on the results of our in vitro drug screen of TNBC cells grown in monoculture, we aimed to test the hypothesis that differences between the 42

chemosensitivity of BL and ML cells are induced, in part, by cell non‐autonomous influences. Several studies have suggested that interactions between tumor cells and components of the tumor microenvironment—including extracellular matrix, growth factors, and other stromal cell types—can alter sensitivity to chemotherapy49,77,78. We focused on interactions between cancer cells and stromal fibroblasts, which are often the predominant stromal type found within the tumor microenvironment79.

To identify tumor–stroma interactions that alter drug response, we initially tested an in vitro coculture system that was successfully used to evaluate tumor– stroma–drug interactions49 (Fig 2.3C and D). However, these data indicate that measurements of GFP fluorescence using a fluorescence plate reader were not sufficiently sensitive for quantifying the degree of cell death in a population of cells. Based on these results, we modified our coculture screen to optimize measurement of drug‐induced cell death. We used JC‐1, a dye that accumulates within mitochondria and is often used as a surrogate measure of apoptotic cell death80 (Fig 3.3A and B). At low concentrations, JC‐1 exists as a monomer and yields green fluorescence; however, when accumulated at high concentrations within mitochondria, this dye forms aggregates, which yield red fluorescence.

Thus, the red fluorescence of JC‐1 reports cellular mitochondrial integrity, which is lost when cells activate apoptosis. To assess the suitability of JC‐1 to quantify changes in the degree of cell death in coculture, we again piloted this assay on

BT‐20 cells treated with camptothecin in the presence or absence of HADF. 43

Images of these cells taken prior to drug exposure confirm punctate red fluorescence in BT‐20, but not HADF, confirming that the dye is not exchanged between cells in coculture (Fig 3.3B). 96 h after exposure to camptothecin, the majority of BT‐20 cells had significantly reduced JC‐1 red fluorescence, suggesting that mitochondrial integrity has been compromised (Fig 3.3C).

Importantly, JC‐1 red fluorescence measured using a fluorescence plate reader was sufficiently sensitive for observing both the potent cell death of BT‐20 cells in monoculture and the protective effect of HADF cells in coculture (Fig 3.3D).

To evaluate the role of stromal fibroblasts in DNA damage sensitivity, we selected six TNBC cell lines (three BL and three ML) that have relatively similar levels of sensitivity to DNA damage. These JC‐1‐labeled TNBC cells were grown in monoculture or in coculture with each of a panel of 16 primary human fibroblasts. Each culture was exposed to a four‐point dose range of 42 anticancer drugs, which included at least one drug per class for all current FDA‐approved breast cancer drugs (Tables 3.1 and 3.2). JC‐1 red fluorescence was quantified at 8‐hour intervals for 72 hours. In total, we collected more than 300,000 measurements of drug sensitivity (Fig 3.3E, Fig 3.4A–C, and Table 3.3). We found a strong overall correlation among biological replicates, indicating that the stromal influences observed were not due to measurement noise (Fig 3.4D). To identify TNBC–fibroblast interactions that significantly altered sensitivity, we used a statistical fold‐change cutoff of 3× the standard deviation observed among replicates. This analysis identified 5,039 significantly changed drug responses 45

(Fig 3.4E). As may be expected, this list of “hits” was significantly depleted for responses at early times (i.e., 8 h), low doses (0.1 μM), and responses to anti‐ estrogen drugs (Tables 3.3 and 3.4). Non‐response to anti‐estrogen compounds is expected as TNBCs do not express estrogen or progesterone receptors.

The majority of TNBC cell–fibroblast interactions did not alter drug sensitivity (Fig 3.4A–C). Nonetheless, our screen revealed many striking phenotypes of strongly altered drug responses, and overall, our data cover nearly the entire landscape of possible positive and negative interactions (Fig 3.3E). To determine the reliability of these measurements, we selected a subset of these interactions to validate by flow cytometry. For example, our screen identified that palbociclib killed more than 80% of HCC‐1143 cells, a basal‐like TNBC, if applied to these cells in monoculture. However, this drug was rendered ineffective when

HCC‐1143 cells were cocultured with the fibroblast cell, HCPF, resulting in only a

20–40% decrease in cell viability (orange dots, Fig 3.3E). A flow cytometry‐based analysis of cell death recapitulated this drug desensitization phenotype (Fig 3.3F and G). Additionally, our coculture screen identified instances in which the efficacy of etoposide is significantly improved in coculture conditions. For example, etoposide was ineffective in killing mesenchymal‐like Hs578T cells in monoculture but killed more than 50% of these cells grown in coculture with skin fibroblast cells, WS1 (purple dots, Fig 3.3E). This phenotype was interesting because prior studies have found that etoposide is minimally active in monoculture, which was surprising given the clinical utility of this compound54. 50

Flow cytometry‐based analysis of cell death confirmed that etoposide‐induced cell death in Hs578T is significantly enhanced by coculture with WS1 fibroblasts

(Fig 3.3H and I). Thus, our coculture drug screen identified a spectrum of TNBC– 51

fibroblast interactions that modulate the drug response of TNBC cells in both positive and negative directions (Figs 3.5–3.7).

Primary fibroblasts grown in coculture with TNBC cells display an activated phenotype and modulate drug sensitivity similar to cancer‐associated fibroblasts (CAFs).

Our coculture drug screen was comprised mainly of normal primary fibroblasts. Cancer‐associated fibroblasts (CAFs, also called “activated” fibroblasts, myoepithelial cells, or myofibroblasts) are thought to be largely distinct from primary fibroblasts in their ability to promote aspects of tumorigenesis and tumor progression43,64. Thus, to address the relevance of our findings we also aimed to compare the behaviors of primary fibroblasts and

CAFs, and in particular, the ability of these different cell types to modulate drug responses of TNBCs in coculture. We first asked whether our primary fibroblasts adopt an activated phenotype in culture. Prior studies have found that the unnatural environmental stiffness of in vitro culture can cause primary fibroblasts to spontaneously adopt the activated phenotype81. Using immunofluorescence microscopy, we determined α−smooth muscle actin (SMA) expression in fibroblasts, a marker of the activated fibroblast expression state that is commonly observed in CAFs. Robust but variable SMA positivity was observed for all primary fibroblasts tested (Fig 3.8A). To more precisely quantify percentages of

SMA+ fibroblasts, and to determine whether coculture modulated SMA positivity, we measured the percent of SMA+ cells using flow cytometry (Fig 3.8B). The percentage of SMA+ fibroblasts was relatively high, even when fibroblasts were grown in monoculture (median 63%; range 10–85%), and was consistently 55

increased when fibroblasts were cocultured with TNBC cells (Fig 3.8C). Thus, primary fibroblasts grown in coculture with TNBC cells generally adopt the activated expression state that is commonly observed for CAFs. 56

To compare the degree to which primary or CAF cells modulate drug sensitivity of TNBC cells, we repeated a small portion of our coculture drug screen using one primary mammary fibroblast (HMF) and two breast cancer‐ associated fibroblasts (Hs343T and Hs578BST). Overall, the drug sensitization/desensitization profile revealed a similar pattern of relative drug sensitivity when compared to our initial coculture screen (Fig 3.8D). Additionally, primary and CAF cells modulated sensitivity to common anticancer drugs in a significantly correlated manner (r = 0.46; P < 0.0001). This suggests that, although primary fibroblasts and CAFs likely differ in many substantial ways, these different cell types were similar in the manner in which they alter drug sensitivity of associated cancer cells. This finding is consistent with a previous study by Polyak and colleagues, which found that patient‐derived primary or CAF cells from breast or brain specimens similarly desensitized HER2‐positive breast cancer cells to the HER2 inhibitor lapatinib39.

TNBC–fibroblast interactions are sufficient for inducing differential drug sensitivity between basal‐like and mesenchymal‐like TNBCs.

Our profiling of TNBC cell lines grown in 2D or 3D monoculture failed to identify robust differences in the drug sensitivities between BL and ML cells (Fig

3.1). Since our coculture screen revealed strong positive and negative changes in the drug responses of these cells (Figs 3.5–3.7), we asked whether these altered drug sensitivity profiles improved the resolution between BL and ML cells. 57

To test this, we performed principal component analysis (PCA) on our coculture screening data. We reasoned that PCA would be beneficial due to the high dimensionality of our data (i.e., multiple drugs, cells, coculture environments, doses, times). PCA uses the covariation structure of the data to reduce data

dimensionality to a smaller number of “principal components”, with each component being comprised of related information82.

Principal component analysis of our coculture data reduced these complex observations to 10 principal components, with the first two components capturing

53% of the overall variation in drug sensitivity (Fig 3.9A). The projection of our data onto PC1 and PC2 revealed a clear separation of BL and ML cells (Fig

3.10A). Notably, this expected pattern was not visible in drug response data collected on these TNBC cells grown in monoculture (Figs 3.1B and 3.10B).

Thus, our screening data reveal that coculture of TNBCs with fibroblast cells 58

modulates drug sensitivity in a manner that enhances the distinction between BL and ML cells.

Because each principal component captures unrelated/orthogonal variation in the data, we also sought to determine which components were capturing variation associated with the BL/ML dichotomy. For example, the 59

BL/ML distinction could be captured independently on all PCs. This would suggest that BL and ML cells respond differently to all drugs, in all coculture environments. Alternatively, it is also possible that the BL/ML distinction is restricted to one or a few PCs, suggesting that BL and ML cells respond in distinct ways only in the context of a subset of drugs and/or environmental conditions. Thus, to investigate this further, we quantified the degree of separation of BL/ML cells across all 10 PCs (Fig 3.10C). A strong association was identified for PC2, which captured approximately 13% of the data. Modest but significant associations with BL/ML were also found for PC3, 5, and 6, which each of which captured relatively small amounts of the dataset (Fig 3.9B). Thus, we focused on finding features in our data that were driving the association between PC2 and BL/ML cells. To achieve this, we inspected the PCA vector loadings, which report the degree to which each variable contributes to a given component. We noticed strong positive loading coefficients on PC2 for conventional chemotherapeutics, such as etoposide and paclitaxel (Fig 3.10D).

To determine whether these observations were statistically robust, we quantified the statistical enrichment of each class of drugs on PC2, finding the strongest enrichments for topoisomerase inhibitors and microtubule poisons, two drug classes that are commonly used in the treatment of breast cancer (Fig 3.10E).

We selected a subset of drugs to re‐evaluate that had strong positive loading coefficients on PC2, suggesting that BL and ML cells would have clearly distinct responses to these drugs. Consistent with these PCA generated insights, we 60

found that BL cell sensitivity and ML cell sensitivity to etoposide and paclitaxel were clearly distinct in coculture, but not in monoculture, with BL cells generally being more sensitive to these conventional chemotherapies (Fig 3.10F and G).

Taken together, these data demonstrate that coculture with fibroblasts was sufficient to enhance distinctions between BL and ML TNBC cells.

Divergent interactions between TNBCs and fibroblasts based on the anatomical origin of fibroblast cells.

Our coculture screen identified environmentally modulated drug responses that differed between BL and ML subtypes of TNBC. Notably, however, these distinctions were seen only for a small fraction of the drug responses tested in our screen, and primarily only in the context of drugs that are currently used in the treatment of TNBC. This distinction suggests that most of the response variation found in our screen is not associated with the BL/ML dichotomy and that opportunities may be found in our dataset that sensitize both BL and ML cells to conventional chemotherapeutics. Thus, we reasoned that deeper insights into the other sources of drug response variation in our data may reveal new strategies for potentiating drug responses in TNBC. PC1, which captured approximately

40% of the overall variation in drug sensitivity, was revealing a dominant pattern in our data that is unrelated to the BL/ML dichotomy (Fig 3.10A). To gain a deeper understanding of the sources of drug response variability in our data, we focused on identifying features of our data that were associated with PC1. We 61

observed a noticeable clustering on PC1 for scores related to each fibroblast cell type (Fig 3.11A and Fig 3.12). To determine whether this pattern was revealing fibroblast cell‐specific, or fibroblast tissue‐specific variation, we computed the correlation between fibroblasts from the same tissue, versus between fibroblasts derived from different tissues. We found a significantly higher correlation between fibroblasts that were derived from the same anatomical tissue (Fig

3.11B and Fig 3.12). These data suggest that fibroblasts harvested from the same tissue modulate the drug responses of associated TNBC cells in similar ways.

Fibroblasts from different tissues could be modulating one of several aspects of TNBC drug response, including the response rate, magnitude of response (e.g., Emax or EC50), or the directionality of the influence (e.g., increased or decreased sensitivity). Because our dataset included a surprising level of directional variability in fibroblast influence, we began by asking whether the directional variance in our screen could be attributed to distinct directionally specific influences of fibroblast from different tissues (i.e., Is a fibroblast from a given tissue intrinsically more likely to induce drug sensitization/de‐ sensitization?). To test this, we calculated the ratio of drug responses for each

TNBC–drug combination, in coculture versus monoculture. To facilitate visual inspection of the relative influences induced by each fibroblast type, we organized the data by dose and time, in order to highlight conserved fibroblast‐ dependent influences (Fig 3.11C). Each data tile was then subsequently grouped by stromal location and drug, and a map of this type was created for each TNBC cell line. From this analysis, clear differences between fibroblast lines were visible, with each fibroblast promoting either drug sensitization or desensitization

(Fig 3.11D). These patterns were similar between BL and ML subtypes, although

ML TNBCs had more strongly polarized responses, in both positive and negative directions (Fig 3.11E). Some drug‐specific responses, such as the desensitization of TNBCs to sunitinib, occurred regardless of which fibroblast was used in coculture. Interestingly, however, these drug‐specific interactions were rare, and most fibroblast‐induced changes in drug sensitivity were observed 64

in a similar manner, across essentially all drugs, and in both BL and ML subtype

TNBCs. Thus, rather than finding unpredictable or idiosyncratic interactions, specific to precise TNBC–fibroblast–drug combinations, our analysis reveals a dominant and relatively straightforward pattern: Fibroblasts modulate the drug response of TNBC cells in distinct and divergent manners, which are largely dependent on the anatomical origin from which the cells were harvested.

Fibroblast‐dependent, and drug‐independent, variation in drug sensitivity occurs through modulation of the mitochondrial apoptotic priming state of

TNBC cells.

An unexpected phenotype from our screen was the degree to which a given fibroblast's influence over TNBC drug sensitivity was consistent regardless of which drug was applied (Fig 3.11D and E). Thus, we next aimed to determine the mechanism by which fibroblasts could interact with TNBC cells to produce these divergent—and largely drug identity independent—changes in TNBC drug response. The simplest mechanism that is consistent with this observation would be that these fibroblasts cause a direct modulation of TNBC cell growth or survival, independent of the drugs added (Fig 3.13, example I and Fig 3.14A). To test this, we used GFP‐tagged TNBC cells to monitor TNBC‐specific growth/survival phenotypes, in the absence of any drug. We found that most fibroblast cells either did not alter the growth rate of TNBC cells, or induced a modest growth rate increase in TNBC cells grown in coculture (Fig 3.14B). 65

Furthermore, in the context of fibroblasts that consistently sensitized drug response of all TNBC cell lines (e.g., WS1, C12385, or HUF), coculture did not significantly decrease growth or survival, suggesting that a growth fitness or survival defect does not account for the broad‐spectrum drug sensitization seen in coculture with these cells. In rare instances, coculture conditions did result in a significant TNBC cell growth rate decrease, such as seen with MDA‐MB‐231 cells grown with H6013, a fibroblast derived from lung tissue (Fig 3.14B).

Notably, the M231‐H6013 interaction induced broad‐spectrum drug desensitization (i.e., enhanced survival). Thus, even in the rare instances in 66

which fibroblast cells mediated fitness defects, growth rate or survival modulation does not appear to account for the observed pattern of influences on TNBC drug response.

A second mechanism by which fibroblast cells could enhance drug efficacy could be by modulating drug bioavailability (Fig 3.14A, example ii). This could be achieved, for example, if fibroblast cells altered the levels of expression of drug efflux pumps in cancer cells, if fibroblasts sequestered drugs, or if fibroblasts metabolized the drugs, creating a more/less potent or bioavailable compound (Fig 3.13). This latter mechanism was recently reported to explain a microbiome–drug interaction that modulates toxicity of the chemotherapeutic 5‐

FU29. To test the role of fibroblast modulation of drug metabolism or availability, we focused on drugs that activate death through induction of DNA damage. For this set of compounds, drug potency should be proportional to level of γ−H2AX, which marks sites of DNA double‐stranded breaks. We quantified γ−H2AX nuclear intensity following exposure to DNA‐damaging agents, in the presence and absence of fibroblasts. These measurements were made at four time points following exposure to 1 μM teniposide, cisplatin, etoposide, or camptothecin. We used GFP‐labeled TNBC cells to identify TNBC cell nuclei, and images were quantified using a CellProfiler‐based automated image analysis83 (Fig 3.14C).

We began by inspecting the most strongly sensitizing and desensitizing coculture environments to determine whether these extreme cases could be explained by differences in the apparent drug potency. TNBC nuclear γ−H2AX intensity was 68

similar in BT‐20 cells grown in monoculture or when cocultured with C12385, a uterine fibroblast that strongly sensitized TNBC drug response (Fig 3.14D).

Similarly, differences in γ−H2AX between mono‐ and cocultures were also not observed when BT‐20 cells were cocultured with Hs27A, a bone fibroblast that strongly desensitized drug responses (Fig 3.14E). To more comprehensively determine whether modulation of drug potency could account for the observed pattern of drug sensitization/de‐sensitization, we compared these changes in

γ−H2AX intensity from quantitative image analysis, to the relative changes in drug sensitivity from our coculture screen. Overall, there was a low correlation between the degree to which γ−H2AX was modulated by fibroblasts and the phenotypic influence of these fibroblasts, which was not significant when compared to a randomized dataset (r = 0.134, P > 0.05; Fig 3.14F). Thus, it does not appear that fibroblast influences on drug sensitivity generally occur through modulation of drug availability or potency.

The insights gained from γ−H2AX intensity analysis are also consistent with our general observation that fibroblast cells influence drug sensitivity in similar ways across diverse and unrelated classes of drugs. In other words, it does not appear that the mechanisms by which fibroblast cells influence the drug responses in TNBC cells are specific to the drug compounds themselves or drug‐ specific responses of TNBC cells. Drug‐induced cell death is the product of at least two independent influences: the drug‐specific cell response (i.e., the ability of a drug to change a cell from a healthy to a dead state) and the degree to 69

which the cell is “primed” for death (i.e., how “close” the healthy cell is to dying)80,84. Thus, a third mechanism that we tested was whether fibroblasts alter the degree of mitochondrial apoptotic priming (Fig 3.13A, example iii). The state of apoptotic priming is thought to relate to the relative local concentration of pro‐ and anti‐apoptotic proteins on the surface of mitochondria85. Thus, the priming state of a cell is not easily determined from gene or protein expression levels, but relative changes in priming can be empirically determined using the BH3 profiling technique86. This assay was recently used to demonstrate that normal or cancer‐ associated fibroblasts (CAFs) from mammary and brain tissue induce resistance to the HER2 inhibitor, lapatinib, in HER2 overexpressing breast cancers39.

Similarly, mammary‐ and brain‐derived fibroblast also induced desensitization to lapatinib in our screen (Fig 3.11D). Thus, we used the BH3 profiling assay to assess the degree to which fibroblasts alter the mitochondrial priming state of

TNBC cells. We selected five fibroblast cells that produced the strongest and most consistent modulation of drug sensitivity. The mitochondrial response to

BIM peptide was quantified by monitoring cytochrome c retention within cancer cells by flow cytometry (Fig 3.14G). BH3 profiling revealed that fibroblast coculture significantly altered the mitochondrial priming state in both basal‐like

(BT‐20) and mesenchymal‐like (MDA‐MB‐231) TNBC cells (Fig 3.14H and I).

Furthermore, the degree to which mitochondrial priming was increased or decreased was also highly correlated with relative drug sensitivity observed in our coculture screen (Fig 3.14I). Thus, in instances where fibroblasts strongly 70

alter broad‐spectrum drug sensitivities of TNBC cells, both the positive and negative changes in drug sensitivity are induced by modulation of mitochondrial priming.

We next aimed to determine whether the fibroblast‐induced changes in

TNBC mitochondrial priming translated to different levels of sensitivity to drugs designed to modulate the apoptotic threshold. For example, BH3 mimetic drugs function by inhibiting anti‐apoptotic proteins, such as BCL2 and other related family members. Because these agents inhibit apoptotic inhibitors, but do not in and of themselves generate pro‐apoptotic activating signals, their efficacy relies on the native priming state of cancer cells87. Thus, we reasoned that fibroblasts that enhance or suppress the priming state of cancer cells may differentially alter sensitivity to BH3 mimetic drugs. To test this, we exposed GFP‐labeled BT‐20 cells to varying concentrations of the topoisomerase II inhibitor teniposide and/or

ABT‐737, a broad‐spectrum BCL2 family inhibitor. Images were collected following 48 hours of drug exposure (Fig 3.15A). To quantify cells, we used a

CellProfiler‐based automated image analysis. Contrary to what is commonly seen for many hematopoietic cancers, ABT‐737 did not kill BT‐20 cells when applied as a single agent (Fig 3.15B, top); however, combinations of ABT‐737 and teniposide were generally synergistic when applied to BT‐20 cells grown in monoculture (Fig 3.15C, top). The synergistic interaction between ABT‐737 and teniposide was further enhanced when BT‐20 cells were grown in coculture with

C12385, a uterine fibroblast that enhanced apoptotic priming (Fig 3.15B and C, 71

middle). In contrast, when BT‐20 cells were grown in coculture with HADF, an adrenal fibroblast that potently decreased apoptotic priming, the drug synergy between these two agents was largely blocked, with most dose combinations resulting in additivity, or even modest drug antagonism (Fig 3.15B and C, bottom). These data confirm the role of fibroblasts in modulating apoptotic 72

priming of associated tumor cells. Additionally, our data demonstrate that tumor– stroma interactions have the capacity to modulate, not only drug sensitivity, but also drug–drug interactions, further highlighting that these interactions are critical features that dictate how cancer cells respond to drugs.

Discussion:

In this study, we explored interactions between tumor cells and stromal cells to identify those that modulate sensitivity to commonly used chemotherapeutics. We found that fibroblasts alter drug sensitivity of tumor cells, and that the responses were highly variable, both in magnitude and in direction.

Our statistical analysis clarified that the directional variability in fibroblast influence is predominantly associated with the anatomical tissue from which the fibroblast cells were harvested. This is an important observation, particularly considering that prior studies have typically only observed that stromal influences were unpredictable and/or very specific to the particular tumor–fibroblast–drug combination66. Surprisingly, we found that fibroblast‐dependent changes in drug response were consistently observed, regardless of which drug was applied, which was driven by fibroblast‐dependent modulation of mitochondrial apoptotic priming within cancer cells.

A major surprise from our work is the substantial directional variability in fibroblast influence and in particular the large proportion of tumor–fibroblast interactions that result in drug sensitization. Prior studies that have interrogated 73

fibroblast–tumor cell–drug interactions have found that these interactions generally result in drug resistance, with only rare instances in which stromal cell interactions lead to drug sensitization39,49,66. The differences observed in our screen likely resulted, in part, from the experimental scale used in our study. For example, some recent smaller scale studies have concluded that fibroblasts induce therapeutic resistance regardless of their tissue of origin39. As noted above, our study successfully replicated the drug resistance phenotypes found in prior studies, and in our data, these drug resistance phenotypes are found in the context of many drug‐sensitizing phenotypes that were not previously tested. An additional possibility is that our screening methodology, which was designed to exclusively monitor drug‐induced cell death, contributed to the enhanced resolution of cell death sensitization. In fact, this feature is likely to have played a major role, considering the limited ability of other common approaches to quantify differences in the degree of cell death (Fig 2.3). A third possibility is that fibroblast‐mediated drug sensitization is a more common phenotype in TNBC, as this cancer subtype was not deeply profiled in prior studies. A focused interrogation of other cancer subtypes may help to determine the extent to which fibroblast influences vary across different cancers.

Taken together, the findings from our study have potentially important implications that warrant consideration in the context of “personalized” or

“precision” medicine, particularly for efforts to improve efficacy of commonly used therapies, such as cytotoxic chemotherapies or other pro‐apoptotic agents. For 74

instance, as we did not see differences in chemosensitivity between BL and ML cells grown in monoculture, it is possible that differences in chemosensitivity between these subtypes of TNBC may not be cell‐intrinsic features, but instead may be the product of interactions between these cells and stromal cells. Thus, it is unclear whether detailed studies on the genomics of TNBC cells will be informative for identifying strategies to enhance chemosensitivity. Additionally, our experiments exploring the efficacy of BH3 mimetic compounds in combination with conventional chemotherapy reveal that the nature of drug–drug interactions (i.e., synergy or antagonism) depends strongly on the growth environment and not only on the cancer genotype. Furthermore, our ability to predict this drug–drug–environment interaction was facilitated by a mechanistic understanding of how fibroblasts modify drug responses in TNBC cells. Thus, future efforts to predict effective drug combinations in vivo will likewise require a greater understanding of how stromal cells from different tissues modulate the drug sensitivity of cancer cells.

Our findings highlight two potentially new opportunities to improve therapeutic responses in TNBC. First, our data show that the relative drug insensitivity of ML subtype TNBCs is restricted to only a few drug classes, which were generally strong apoptotic agents that are commonly used in treatment today. The drug specificity of the BL/ML dichotomy suggests that opportunities may exist for improving the responses of ML subtype cancers, perhaps in the context of other classes of drugs. Indeed, some drugs—which were equally 75

efficacious in BL and ML cells grown in monoculture—were in fact more effective in the ML subtype when these cells were grown in coculture (see for example, bortezomib, Fig 3.11D). The ability to directly test these hypotheses in vivo is currently limited, as coculture xenograft models typically require the use of external matrix (e.g., Matrigel) to support the implantation of fibroblasts, which obscures the tumor–fibroblast interaction (Fig 3.16). Second, the strong drug‐

independent influences of fibroblasts suggest that a more generalizable strategy may be to block interactions with drug desensitizing fibroblasts or to mimic the interactions of drug‐sensitizing fibroblasts. Future studies should therefore aim to gain a comprehensive understanding of the mechanisms of interaction between 76

fibroblasts and cancer cells, and in particular, the mechanisms by which fibroblasts alter the priming state of cancer cells.

Materials and Methods:

Cell lines and reagents

Cell lines BT‐20, HCC‐1143, Hs578T, Hs578BST, MDA‐MB‐231, MDA‐MB‐436,

MDA‐MB‐468, HCC‐2157, HCC‐1806, HCC‐1395, Hs27A, HS‐5, WI‐38, IMR‐90,

Hs343T, WS‐1 were obtained from American Type Culture Collection (ATCC,

Manassas, VA), and cell line CAL‐120 was obtained from Deutsche Sammlung von Mikroorganismen und Zellkulturen GmbH (DSMZ). All cell lines were grown in 10% FBS (Thermofisher Hyclone cat# SH30910.03 lot# AYG161519), 2 mM glutamine, and penicillin/streptomycin. BT‐20, CAL‐120, and WS‐1 were cultured in Memα + Earle's salts. HCC‐1143, HCC‐2157, HCC‐1806, and HCC‐1395 were cultured in RPMI 1640 media. Hs578T, MDA‐MB‐231, MDA‐MB‐436, MDA‐MB‐

468, Hs27A, HS‐5, and Hs343T were cultured in Dulbecco's modified Eagle's medium (DMEM). Hs578T was further supplemented with 10 μg/ml insulin.

Hs578BST was supplemented with 30 ng/ml EGF. Primary fibroblasts, H‐6231,

H‐6201, H‐6076, H‐6019, and H‐6013, were purchased from Cell Biologics

(Chicago, IL); HCPF, HPF‐a, HHSteC, HMF, HAdF, HUF, and HCF‐a were purchased from ScienCell (Carlsbad, CA); and C‐12385 was purchased from

Promocell (Heidelberg, Germany). Primary fibroblast cells purchased from Cell 77

Biologics, ScienCell, and Promocell were cultured in the media (ScienCell—

Fibroblast Medium cat# 2301; Cell Biologics—Complete Fibroblast Medium/w Kit cat# M2267; Promocell—Fibroblast Growth Medium 2 cat# C‐23020) for four doublings before being transitioned to DMEM. All cells were cultured at 37°C in a humidified incubator supplied with 5% CO2 and maintained at a low passage number (< 20 passages for cancer). Prior to expansion and freezing, a small sample of each primary fibroblast was expanded to determine each cell's Hayflick limit to ensure that experiments could be performed prior to the onset of replicative senescence. A complete list of drugs used in this study is included in

Table 3.2.

Coculture screen using JC1 dye

Fibroblast cell lines were grown to 80% confluence before being trypsinized and stained with 5 μM CellTrace Violet Proliferation dye (Thermofisher #C34557) in

PBS at a concentration of 1 × 106 cell/ml for 15 min at 37°C. 1,500 stained cells were plated in 40 μl FluoroBrite media (Thermofisher # A1896701), supplemented with 10% FBS, 2 mM glutamine, and penicillin/streptomycin, in a

Greiner clear 384‐well plate (#781986) and allowed to adhere for 3 h. Cancer cell lines were then trypsinized and stained with 1.5 μg/ml (final concentration) JC‐1

(Thermofisher # T3168) in FluoroBrite at a concentration of 1 × 106 cell/ml for 20 min at 37°C. Cancer cells were then plated at 1,500 cells in 40 μl FluoroBrite per well in the 384‐well plate. For monoculture conditions, unlabeled cancer cells 78

were added to each well, in order to keep the cell density consistent with coculture conditions. Cells were allowed to adhere overnight. The following morning, 8 μl of a 10× drug stock was added to the wells using a VIAFLO 96

Electronic 96‐channel pipetting robot. JC‐1 fluorescence was then read at five spots across each well using a Tecan M1000 Pro Plate Reader at the excitation wavelength of 535 nM ± 17 nM and an emission wavelength of 590 nM ± 17 nM every 8 h for 72 h. Background fluorescence was determined by treating labeled cells with alamethicin, a membrane permeabilizing agent that punctures plasma membrane and mitochondrial membranes. Fluorescence measurements were normalized relative to pre‐drug treatment values for each well.

Cell viability and cell death assays

Cell viability assays were performed either using CellTiter‐Glo (cat# G7570), for cells grown in monoculture, or flow cytometry, for coculture assays (other than the coculture screen, described above). For CellTiter‐Glo, which measures viability as a function of ATP concentration, cells were plated in Greiner 96‐well plates (cat# 655 090) at 5,000 cells per well in 100 μl of their respective growth media and allowed to adhere overnight. 10 μl of a 10× drug stock, diluted in PBS, was added to each well. Cells were subsequently allowed to grow at 37°C for 72 h. At 72 h post‐drug addition, 33 μl of CellTiter‐Glo reagent was added to each well. The CellTiter‐Glo assay was performed according to manufacturer's directions, with the reagent diluted 1:3 (relative to media volume). Luminescence 79

was read using a Tecan M1000 Pro Plate Reader. Cell death measurements to validate the JC‐1 screen data were collected using the Live/Dead Violet reagent

(Thermofisher cat# L34963) and analyzed by flow cytometry. Cancer cells and fibroblast cells were plated at a 1:1 ratio in DMEM and allowed to adhere overnight. Drugs were added from a 1,000× stock, and cells were exposed for the specified times. Cells were trypsinized at the specified times, suspended in

PBS at a concentration of 1 × 106 cells/ml and stained with a 1:1,000 dilution of the Live/Dead Violet reagent for 30 min on ice. Cells were then fixed with 4% formaldehyde for 10 min at room temperature and run on an LSR II FACS machine with a laser excitation of 405 nm and emission of 450 nm.

Drug sensitivity of cells grown in 3D culture conditions

Culturing of TNBC cells in 3D colonies was performed using the “3D on top” method developed by Bissell and colleagues88. Briefly, a thick layer of cold

Matrigel (corning cat#356235) was applied to the bottom of 96 well plates, which were subsequently heated to 37°C for 30 min to promote solidification. Cancer cells were plated at a concentration of 10,000 cells in 100 μl complete media +

2% Matrigel and were grown for 72 h to induce 3D colony formation. After 72 h of growth, the media were aspirated and media containing drug + 2% Matrigel were added to each well. At 72 h post‐drug addition, cell viability was measured by adding 100 μl of CellTiter‐Glo reagent to each well. The CellTiter‐Glo assay was 80

performed according to the manufacturer's directions, and luminescence was read using a Tecan M1000 Pro Plate Reader.

Growth rate measurements using GFP‐labeled cells

To determine cell proliferation rate using a fluorescence plate reader, TNBC cells were stably transfected with GFP (pRetroQ‐AcGFP1‐N1). Transfected cells were selected with puromycin (BT‐20 at 1.5 μg/ml, 468 at 0.5 μg/ml, and 231 at 2

μg/ml). Cells were selected until a parallel non‐transformed plate exposed to puromycin was completely dead. The selected population was subsequently sorted by FACS to collect cells with similar levels of GFP fluorescence. For coculture experiments, fibroblast cell lines were plated at an 8:1, 4:1, 2:1, 1:1, and 1:2 ratio to cancer cells in a Greiner 96‐well plate in 100 μl of FluoroBrite media and allowed to adhere for 3 h. Following adherence of fibroblast cells,

TNBC cells constitutively expressing GFP were plated at a concentration of

10,000 cells per 100 μl of FluoroBrite media and allowed to adhere overnight.

Cell measurements were measured every 24 h for 96 h using a Tecan M1000

Pro Plate reader.

Fluorescence and immunofluorescence microscopy

For quantitative analysis of p‐H2AX nuclear intensity, fibroblast cells were plated at a density of 1,500 cells per 25 μl in DMEM in a 384‐well plate and allowed to adhere for 3 h. Cancer cells were stained with 5 μM CellTrace CFSE dye 81

(Thermofisher cat# C34554) at a concentration of 1 × 106 cells/ml in PBS for 15 min at 37°C. Labeled cells were plated at 1,500 cells per 25 μl DMEM and allowed to adhere overnight. Drugs were added from a 10× stock solution in

PBS, and cells were exposed for 1, 6, and 18 h before being fixed with 4% formaldehyde for 10 min at room temperature. Cells were washed twice in PBS, then permeabilized with 0.5% Triton X‐100 for 10 min at room temperature. Cells were washed twice with PBS; blocked in 10% goat serum (Thermofisher cat#

16210064) for one hour; stained with the p‐Histone H2A.X (Ser139) antibody

(Cell Signaling Technologies #9718S) in 1% goat serum in PBS overnight at 4C; stained with Alexa‐647 antibody (1:250 dilution, Thermofisher A21244) in 1% goat serum in PBS for 2 h at room temperature. Imaging was performed using an

IXM‐XL high‐throughput automated microscope. Analysis was performed using a custom CellProfiler pipeline (available upon request).

SMA analysis by microscopy and FACS

Fibroblast cells were stained with 1uM CellTracker Deep Red Dye (Thermofisher cat#C34565) at a concentration of 1 × 106 cells/ml for 15 min at 37°C and subsequently plated in 96‐well plates at a concentration of 10,000 cells per well and allowed to adhere for 3 h. Cancer cells were then stained with 5 μM

CellTrace Violet Proliferation Dye at a concentration of 1 × 106 cells/ml for 15 min at 37°C and plated at a concentration of 10,000 cells per well and allowed to adhere for 24 h. After 24 h, cells were fixed with 4% formaldehyde for 10 min at 82

room temperature and washed twice with PBS. Cells were then permeabilized with 0.5% Triton X for 10 min at room temperature, washed twice with PBS, and then blocked in 10% goat serum for one hour. Cells were stained with the αSMA antibody (1:75 dilution, Cell Signaling Technology cat#19245S) in 1% goat serum for 8 hours at room temperature and then stained with the Alexa‐488 antibody

(1:100 dilution Thermofisher cat#A11008) in 1% goat serum for 1 h at room temperature. Imaging was performed using an EVOS FL‐AUTO automated fluorescence microscope. A similar process was used for analysis of SMA expression by FACS, except following formaldehyde fixation, and cells were exposed to 100% methanol for 2 h at −20°C and washed twice with PBS. Cells were stained with the αSMA antibody (1:75 dilution, Cell Signaling Technology cat#19245S) in 50% Odyssey Blocking Buffer (LI‐COR cat# 927‐40000) 50%

PBS‐T for 8 hours at room temperature and then stained with the Alexa‐488 antibody (1:100 dilution Thermofisher cat#A11008) for 1 h at room temperature in

50% Odyssey Blocking Buffer 50% PBS‐T. FACS was run on an LSR II machine with a laser excitation of 488nm and emission of 530 nm.

Mitochondrial priming assay

Mitochondrial priming assays were performed according to a previously published iBH3 protocol89. For monoculture conditions, 1 × 106 cancer cells were plated in a 10‐cm dish and allowed to adhere overnight. For coculture conditions, fibroblasts were stained with Cell Trace Violet plated at a 1:1 ratio with cancer 83

cells. Twenty‐four hours post‐plating cells were trypsinized and arrayed in a 384‐ well plate. A dose series of BIM peptide (GenScript) was added (100, 33, 10, 3.3,

1, 0.33 μM) along with either DMSO (vehicle control) or alamethicin (Enzo, BML‐

A150‐0005), a mitochondrial depolarizing agent, which was used at a final concentration of 25 μM as a positive control. Plasma membrane permeabilization was achieved by the addition of digitonin at a final concentration of 20 μg/ml.

Cells were incubated with BIM peptide at room temperature for 1 hour before being fixed and stained for cytochrome c retention (Fisher cat# BDB560263).

Samples were analyzed on an LSRII flow cytometer.

Data analysis and statistics

All statistical analyses were performed using GraphPad Prism and/or MATLAB, generally using pre‐built functions (Fisher's exact test, t‐test, etc.). PCA was performed using SIMCA, and data were z‐scored (mean centered and unit variance scaled). Hierarchical clustering was performed using Spotfire using the default settings (UPGMA clustering method; Euclidean distance measure; average value ordering weight; z‐score calculation normalization method; empty value replacement: NA). Analysis of flow cytometry data was performed using

FlowJo. Combination Index (CI) was calculated by dividing observed drug sensitivity by the expected drug response given Bliss Independence (for two drugs, A and B, expected = A + B − [A*B]).

Data Availability:

The complete coculture screening data can be viewed at DOI

10.15252/msb.20188322 as Table EV3. The CellProfiler scripts used for the analyses shown in Figs 3.14, and 3.15 can be viewed at DOI

10.15252/msb.20188322 Code EV2, and Code EV3, respectively.

Acknowledgements:

We thank members of PSB and the DFCI CCSB for valuable discussions during the design and execution of this study. Additionally, we thank the UMass Med

School Flow Cytometry Core and High‐Throughput Imaging Core for training and advice during the execution of this study. Research reported in this publication was supported by the National Institute of General Medical Sciences of the

National Institutes of Health under Award Number R01GM127559 (MJL) and by grants to MJL from the Richard and Susan Smith Family Foundation, Newton,

MA; the Breast Cancer Alliance; the American Cancer Society (RSG‐17‐011‐01).

BDL, PCG, and RR were supported by a NIH training grant (Translational Cancer

Biology Training Grant, T32‐CA130807). Additional support was provided by the

UMass President's S&T Fund to MJL and SRP.

Author contributions:

This project was conceived by BDL and MJL. Experiments were designed by

BDL, HRS, ADS, SRP, and MJL. Coculture screen was designed and executed 85

by BDL. All other experiments were executed by BDL, TL, RR, PC‐G, HRS, and

GR. Analysis was conducted by BDL, TL, RR, HRS, MEH, and MJL. Manuscript was written and edited by BDL, ADS, SRP, and MJL.

Funding:

HHS | NIH | National Institute of General Medical Sciences (NIGMS)

R01GM127559

American Cancer Society (ACS) RSG‐17‐011‐01

HHS | NIH | National Cancer Institute (NCI) T32‐CA130807

Richard and Susan Smith Family Foundation (Smith Family Foundation)

Breast Cancer Alliance (BCA)

UMass President's S&T Fund 86

CHAPTER IV: Fibroblast secretion of IL-8 sensitizes TNBC cells to DNA

damage independent of altering apoptotic priming.

Abstract:

Fibroblasts secrete cytokines, chemokines, and growth factors into the tumor microenvironment that promote tumor growth and drug resistance. Pro- inflammatory cytokines activate pro-proliferative and survival signaling networks that aid in preventing the onset of apoptotic cell death in cancer cells. However, it remains unclear if fibroblast secreted factors are able to modulate dug response by altering the priming state of cancer cells. Here, I demonstrated that conditioned media (CM) collected from fibroblasts is capable of sensitizing or desensitizing TNBC cells to DNA damaging chemotherapeutics. I identified the pro-inflammatory cytokine IL-8 as being sufficient in inducing the sensitizing phenotype. However, I continued to observe an attenuated sensitization phenotype using IL-8 depleted CM, suggesting that other soluble factors contribute to drug sensitization. Interestingly, the mechanism behind IL-8 mediated drug sensitization is independent of the apoptotic priming mechanism observed in the co-culture screen from chapter 3. Treatment of TNBC cells with recombinant IL-8 protein showed no change in the apoptotic threshold of the cells. This study demonstrates the ability of fibroblasts to mediate cancer cell drug sensitivity through the secretion of soluble factors as well as highlights that 87

there are distinct mechanisms that fibroblasts utilize to sensitize TNBC cells to

DNA damage.

Introduction:

As highlighted in chapter 3, my co-culture screening approach identified novel cancer:fibroblast interactions that alter drug sensitivity through modulating apoptotic priming. However, the molecular mechanisms underlying the change in the apoptotic threshold of cancer cells was not defined, necessitating a deeper characterization of the cancer:fibroblast signaling interactions to better understand the sensitizing phenotype. Interactions between fibroblasts and cancer cells can be mediated through a variety of methods, such as the secretion of soluble factors (e.g. growth factors, cytokines, etc.), the deposition of matrix proteins, and/or direct cell-cell contact. Given the broad range of molecules involved in each of these different modes of interaction, it is necessary to limit the investigation to one type of cancer:fibroblast interaction.

Fibroblasts are responsible for the secretion of a myriad number of cytokines, chemokines, and growth factors in the tumor microenvironment.

Fibroblast mediated secretion of pro-inflammatory cytokines promotes tumor cell proliferation and survival through multiple signaling axes, such as IL-6 mediated

JAK/STAT3 activation as well as the activation of NF-kB90–93. Secretion of growth factors by fibroblasts, such as HGF, also contribute to cancer cell proliferation as well as mediating drug desensitization to targeted chemotherapeutics49. 88

However, the ability of soluble factors to alter the apoptotic threshold of cancer cells has not been characterized. Therefore, I chose to focus on testing the ability of soluble factors to mediate the apoptotic priming dependent drug sensitizing phenotype seen in the C12385, HUF, and WS1 cancer:fibroblast interactions.

Results:

To investigate the molecular mechanisms that dictate the drug sensitizing and desensitizing phenotypes observed in co-culture of TNBC cells and fibroblasts, I tested whether conditioned media (CM) from fibroblast cells modulates drug response. To measure the rate of death as well as the lethal fraction of TNBC cells, I adapted a fluorescent based assay that utilizes the dye sytox green (Fig 4.1A). Sytox green is a cell impermeable dye that fluoresces only when bound to DNA. As cells die, their membrane structure is disrupted which allows sytox green to enter the nucleus and bind to DNA. Conditioned media was collected following 48 hours of culture with fibroblasts and added to

TNBC cells prior to addition of sytox green and teniposide, a topoisomerase inhibitor. Sytox green fluorescence was measured at 12-hour time points as a proxy for cell death. At the end of the time course, triton-x (0.02% final concentration) was added to each well to permeabilize the membranes of any sytox green excluding viable cells. Each well was read again, with 100% of the cells now sytox green positive, and the resulting fluorescence measurement was 89

used as a proxy for the total number of cells in a well. These pre and post- permeabilization measurements were used to calculate the lethal fraction of each condition.

The lethal fraction of each CM condition was compared to the lethal fraction of control cells and then Z-scored for each cell line. A positive Z-score, displayed as red and seen in the Hs27A and Hs5 conditions, indicates a CM mediated drug desensitization to teniposide. A negative Z-score, displayed as blue and seen in the C12385, HUF, and WS1 conditions, indicates a CM mediated drug sensitization to teniposide. To determine if CM was altering other features of death besides the lethal fraction, such as onset time or rate, I focused on collecting kinetic death measurements of three strong sensitizing cancer:fibroblast interactions. These interactions included the two uterine fibroblasts C12385 and HUF as well as the skin fibroblast WS1. I found that conditioned media from WS1, HUF, and C12385 increased the rate of death in all six TNBC cell lines, compared to control cells, in response to teniposide treatment (Fig 4.1C). In contrast, conditioned media from Hs27A, a bone fibroblast cell that caused a desensitizing phenotype, generally did not alter

TNBC drug response, although these data were variable across the six TNBC cell lines (Fig 4.1B and C). These data indicate that secreted factors from tissue specific fibroblasts are sufficient to sensitize TNBC cells to DNA damage.

To identify secreted factors that are responsible for these phenotypes, I profiled conditioned media obtained from sensitizing and desensitizing fibroblasts 91

for the presence of 45 common cytokines, chemokines, and growth factors (Fig

4.2A). To identify factors that are involved in drug sensitization, I plotted the relative protein levels of each factor against the relative survival ratio calculated from the co-culture screen in chapter 3 and calculated the correlation. A negative correlation between these two values would reveal potential hits, since sensitizing environments produce a low relative survival ratio and factors that control the sensitizing phenotype would be expected to be expressed at relatively high levels. Conversely, a positive correlation would indicate hits involved in drug desensitization, since the relative survival ratio will be high as well as the relative protein level of the factor. Of the 45 cytokines profiled, I observed a strong negative correlation for IL8 (Fig 4.2 B and C). I also observed a positive correlation between relative drug sensitivity and secretion of IL6, a cytokine that is already known to induce resistance to DNA damaging chemotherapies (Fig

4.2C)90,94.

To test if IL8 was sufficient to alter sensitivity to DNA damage, I treated

BT20 cells with recombinant IL8 and measured the rate of death in response to camptothecin, a topoisomerase inhibitor. Addition of IL8 increased the rate of

DNA damage induced cell death by approximately 2.5-fold, whereas cells treated with recombinant IL6 decreased the rate of cell death by approximately 2-fold

(Fig 4.2D). To further determine the role of IL8 in conditioned media induced drug sensitization, I treated conditioned media obtained from WS1, HUF, or C12385 fibroblasts with an IL8 neutralizing antibody. In all three conditions, IL8 92

neutralizing antibodies inhibited the drug sensitization phenotype induced by conditioned media (Fig 4.2E). Interestingly, the use of IL8 neutralizing antibodies failed to restore BT20 drug response to the level of control cells (Fig 4.2 E black line), arguing for the fact that other secreted factors in conditioned media also contribute to the drug sensitization phenotype. Thus, cytokine analysis of CM derived from sensitizing fibroblasts identified and demonstrated that the pro- inflammatory cytokine IL-8 plays an important role in sensitizing TNBC cells to

DNA damage.

To test if IL8 is mediating drug sensitivity in cancer cells by altering the apoptotic threshold, I utilized the iBH3 assay to measure the priming state of cancer cells exposed to recombinant IL8 protein. In brief, cancer cells are treated with digitonin to permeabilize the cell membrane and exposed to increasing concentrations of recombinant BIM peptide before measuring the amount of cytochrome c protein retained in the mitochondria by FACS. Cells that are more primed for death will undergo MOMP at a lower BIM concentration and lose most of their mitochondrial cytochrome c protein. Cells treated with DMSO but no BIM protein or the mitochondrial pore forming agent alamethecin serve as negative and positive controls (Fig 4.2F). BT20 cells were exposure to IL8 for increasing amounts of time before apoptotic priming was tested. However, there was no observable difference in the BIM dose response curves or EC50 values between control cells and cells exposed to IL8 protein (Fig 4.2G). These data argue that

IL8 is not acting to alter drug response through modulating apoptotic priming and 94

that other interactions are responsible for the priming mediated sensitizing phenotype observed in the co-culture screen performed in chapter 3.

Discussion:

In this chapter I explored a possible mechanism underlying the change in apoptotic priming observed in the sensitizing and desensitizing phenotypes previous characterized in the co-culture screen from chapter 3. I demonstrated the ability of conditioned media obtained from fibroblasts cells to alter the drug sensitivity of TNBC cells. Specifically, I showed C12385, HUF and WS1 conditioned media sensitized TNBC cells to DNA damage while Hs27A conditioned media desensitized cells to damage. It is interesting to note that although the general trend in altered drug sensitivity was consistent whether lethal fraction or the rate of death was measured, there was a moderate amount of variability in the phenotype depending on which measurement was used. For instance, measuring lethal fraction of cancer cells treated with Hs27A conditioned media showed a strong desensitizing phenotype but there was little change in the rate of death compared to control cells. Similarly, there was a more variable response in the drug sensitivity phenotype when measuring lethal fraction of

C12385, HUF, and WS1 treated cells compared to the observed universal increase in the rate of cell death. These disparities highlight the complexity of measuring cell death related phenotypes and the need to validate the data using multiple assays. 95

To further elucidate the factors responsible for the conditioned media sensitizing phenotype, I used Luminex cytokine analysis to identify proteins whose levels were anti-correlated to viability measurements. I identified the inflammatory cytokine IL8 as a mediator of drug sensitivity in TNBC cells.

Previous research has revealed that IL8 modulates cell death, however the studies have demonstrated IL8 as having both a positive and negative effect95–97.

I found that IL8 increased the rate of death in TNBC cells independent of changing apoptotic priming. IL8 signals through the CXCR1/2 receptors which control the activation of many downstream effector molecules, including NF-KB.

The timing and dynamics of these signaling axes can lead to either pro-survival or pro-death outcomes in the cell98. As indicated by the data presented in this chapter as well as the previous studies detailing the pro-survival roles of IL8 in cancer cells, there needs to be a more comprehensive understanding of the crosstalk between inflammatory and apoptotic signaling networks. Deeper analysis will clarify the molecular mechanisms that modulate the context dependent pro-survival vs pro-apoptotic phenotypes associated with IL8 signaling.

It should also be noted that while I identified IL8 as a mediator of drug sensitivity, I only screened a small portion of known cytokines, chemokines and growth factors. The fact that neutralizing IL8 signaling using antibodies still shows an increase in the rate of cell death as well as IL8 not altering the apoptotic threshold of cancer cell indicates that there are other interactions that 96

contribute to the observed phenotypes. A more detailed analysis of the composition of fibroblast conditioned media as well as the investigation of the effect of simultaneous exposure to multiple cytokines on drug sensitivity will yield valuable information in the underlying molecular mechanisms and signaling pathways responsible for the drug sensitization phenotype.

Materials and Methods:

Conditioned media assays

Fibroblasts were plated in a 10 cm dish at a concentration of 750,000 cells in 10mL of DMEM and allowed to adhere. Cultures were incubated at 37C for 48 hours and conditioned media was collected and filtered through a 0.45um syringe filter. To test whether conditioned media altered drug sensitivity, cancer cell lines were plated in the respective conditioned media at a concentration of 3000 cells per 50uL conditioned media in a 384 well plate and allowed to adhere overnight.

24 hours post plating, 5uL of a 10x drug stock was added to each well along with the Sytox green reagent (Fisher Scientific, S7020). Sytox Green was used at a

5uM final concentration, and fluorescence was measured using a Tecan Plate reader. The Sytox Green fluorescence reading is proportional to the number of dead cells in the well. To determine numbers of live cells and the percent viability, Triton x100 was added to each well (0.2% for 3 hours at 37C) to induce permeabilization of remaining live cells. 97

Cytokine analysis

Conditioned media for the selected fibroblast lines were collected 48 hours post plating and arrayed in biological replicates with two technical replicates each. Cytokine and chemokine analysis was performed according to the manufacturer’s instructions for the Cytokine/Chemokine/Growth Factor 45-Plex

Human ProcartaPlex™ Panel 1 kit obtained from Thermofisher (cat# EPXR450-

12171-901).

Mitochondrial priming assay

Mitochondrial priming assays were performed according to the iBH3 protocol from Ryan et al89. For priming, 1x10^6 cancer cells were plated in a

10cm dish and allowed to adhere overnight. Recombinant IL-8 cytokine was added for the specified time and cells were then trypsinized and arrayed in a 96 well plate. A dose series of BIM peptide was added (100uM, 33uM, 10uM,

3.3uM, 1uM, 0.33uM) along with either DMSO (vehicle control) or alamethicin, a mitochondrial depolarizing agent, which was used at a final concentration of

25uM as a positive control. Plasma membrane permeabilization was achieved by the addition of digitonin at a final concentration of 20 µg/mL. Cells were incubated with BIM peptide at room temperature for 1 hour before being fixed and stained for cytochrome c retention (Fisher cat# BDB560263). Samples were analyzed on an LSRII flow cytometer. 98

CHAPTER V: DISCUSSION

An ongoing challenge in the treatment of cancer is choosing the optimal drug or drug combination for a particular patient. Drug efficacy varies substantially between patients, even within a single cancer subtype, which necessitates a better understanding of the factors that alter cancer drug sensitivity. Several studies have investigated the link between cancer genetics and drug response, which accounts for a fraction of the observed variability, but even with comprehensive genomic information, predicting therapeutic responses tends to be beyond our current grasp. One limitation is the dearth in understanding how non-genetic variations alter drug response. Components from the tumor microenvironment, for example CAFs, have been shown to influence the drug sensitivity of cancer cells. However, we currently lack the knowledge necessary to predict which fibroblast-cancer-drug interactions will alter drug sensitivity or the directionality of the response. A deeper understanding of the factors that govern tumor-stroma interactions will reveal potential biomarkers or

“rules” that would function to increase our ability to predict therapeutic responses in a variety of environmental contexts.

In my thesis, I use an optimized high throughput screening method to investigate how cancer-fibroblast interactions mediate response to DNA damaging chemotherapeutics. I focused primarily on solid epithelial tumors, including those in the “triple-negative” breast cancer class, and non-small cell 99

lung cancer class. I demonstrated the ability of a JC-1 dye-based assay to accurately measure cell death in an in vitro co-culture system and performed a screen to determine the effect of 4,032 drug-cancer-fibroblast interactions on cancer cell drug sensitivity. Analysis of the screen data showed that a bulk of the interactions observed had minimal effect on drug sensitivity, however, a small subset of interactions either greatly sensitized or desensitized cancer cells to

DNA damaging agents. PCA of the data also revealed that co-culture of cancer cells with fibroblasts was sufficient to capture the clinically observed difference in basal drug response between two TNBC sub-classes, Basal-Like (BL) and

Mesenchymal-Like (ML).

My study revealed two interesting patterns: 1) For an individual cancer- fibroblast interaction, the change in drug sensitivity was generally independent of which drug was tested, and 2) For an individual cancer-drug interaction, the direction of the effect was specific to the tissue that the fibroblast was obtained from. I further demonstrated that the drug independent sensitivity effect was caused by a change in apoptotic priming induced by cancer-fibroblast interactions. Using this knowledge, I was able to predict the efficacy of a combination therapy consisting of a DNA damaging agent and a pan-BCL-2 inhibitor based on the environmental context.

100

Matching your research question with the right cell death assay.

Measuring cell death and proliferation is a complex process, with multiple assays available that yield different types of information. Throughout my thesis I used four different assays to measure cell death and proliferation. The assays are split into two main categories, assays that can be used to measure death and proliferation in one population in mono-culture or assays that can be used to measure death and proliferation in one population in both mono and co-culture

(an overview of each assay can be viewed in Table 5.1).

GFP CellTiter Glo Sytox Green JC-1 Dye Fluorescence

Mono-culture Mono-culture Culture Mono-culture Mono-culture or or Conditions Co-culture Co-culture

Cost $$$ $$ $$ $

High High High Low (20 min to Throughput (add to plate (add to plate (need to make stain then then measure) with drug) each cell line) plate)

Plate Reader Instrument Plate Reader Plate Reader or Plate Reader Used Microscopy

Measures Proliferation Proliferation (Using a plate Proliferation Cell Death Cell Death Cell Death reader)

Yes Kinetics (need multiple Yes Yes Yes plates)

Table 5.1: Comparison of cell death assays. A detailed comparison of the four cell death assays that were used in the thesis. 101

The first mono-culture only assay is CellTiter Glo, which works by lysing the cells in each well and measuring the total ATP as a marker of viable cells.

CellTiter Glo is very accurate and sensitive, having a detection limit of 10 cells, but is quite expensive. CellTiter Glo can be used to measure either cell death or proliferation by comparing the number of viable cells in a treated well compared to a control well. To measure the kinetics of death or proliferation, however, a separate plate must be used for each timepoint since the cells are lysed as part of the measurement. This process can introduce noise into the experiment as the kinetic measurements consist of multiple separate wells instead of following one well through time. Also, throughput is limited due to the base cost of CellTiter Glo and the need for multiple plates for kinetic measurements. This assay is best suited for observing cell death at a single timepoint. In my thesis I used this assay to measure dose curves of cancer cell lines grown in mono-culture treated with drug for 72 hours either in 2D or 3D culturing conditions.

The second mono-culture only assay is the use of the Sytox Green dye, which works by entering dead cells with disrupted cell membranes and binding to

DNA. When bound to DNA, Sytox Green produces a fluorescent signal which can then be read by a plate reader to determine the number of dead cells present in the well. Sytox Green is quite accurate and moderately priced. Sytox Green can be used to measure the kinetics of cell death by reporting the number of dead cells in the well over time. Sytox Green can also be used to measure cell proliferation and lethal fraction at a single timepoint by first measuring the 102

number of dead cells in a well, then adding an agent to kill the remaining viable cells, and then measuring the total number of cells in the well to calculated the number of viable cell that were in the well. Due to the ease of use, adding the

Sytox dye to the well at the beginning of the experiment, throughput is quite high.

This assay is best suited for observing cell death over a range of timepoints as well as measuring the lethal fraction at a single timepoint. In my thesis I used this assay to measure the rate of death or lethal fraction of cancer cells treated with media conditioned by fibroblast cells.

The first assay that works in either mono-culture or co-culture is measuring GFP fluorescence. This assay works by integrating a cassette coding a soluble cytoplasmic GFP molecule controlled by a constitutive promoter into a cancer cell line. Each transfected cell line is then selected for the cassette using puromycin and sorted by FACS to select for a population that of cells with similar

GFP fluorescence. Cytoplasmic GFP fluorescence scales with cell number and can be used to determine the number of viable cells in a well. Since only one population of cells are labeled, this assay can be used to measure death and proliferation of one population when cultured with other cell types. The GFP fluorescence assay is accurate and moderately priced, given the time and reagents needed to create each cell line. GFP fluorescence can be used to measure cell death or proliferation using automated microscopy but only cell proliferation using a plate reader. The reason behind this is that when a cell dies the cytoplasmic GFP molecules form small long-lasting bright puncta that are 103

seen by a plate reader and inflate the total well fluorescence, i.e. the number of viable cells. These same puncta are observed when using automated microscopy but are filtered out using a size threshold in CellProfiler when determining the number of viable cells, allowing for the calculation of cell death.

The throughput of this assay is limited by the need to create each cell line as well as the use of automated microscopy to measure cell death. Therefore, this assay is best suited for measuring the effect of drugs on cell proliferation using a plate reader. In my thesis I used this assay to determine the effect of co-culture with fibroblasts on cancer cell proliferation with and without the addition of targeted growth factor inhibitors.

The second assay that works in either mono-culture or co-culture is the

JC-1 dye, which works by staining healthy mitochondria in viable cells with red fluorescent puncta. When the cell begins to die, healthy mitochondria are subject to mitochondrial outer membrane permeabilization (MOMP) and lose the red fluorescent puncta which can be observed using a plate reader. Since only one population of cells are labeled with the JC-1 dye, this assay can be used to measure death of one specific population when cultured with other cell types.

The JC-1 dye is inexpensive and accurate in measuring cell death, although there is a relatively high basal background signal. The JC-1 assay can measure cell death kinetics in a single well over time but cannot measure cell proliferation.

The throughput of this assay is quite high as the cells only need to be stained for

20 minutes before being plated and they can be measured using a plate reader. 104

Therefore, this assay is best suited for co-culture experiments involving multiple cell lines. In my thesis I used this assay to measure death in a large co-culture screen between TNBC cell lines and fibroblast cells.

In terms of predicting patient outcomes, the information obtained from the

JC-1 assay would be the most informative. The JC-1 assay is the only assay I used that can measure the cell death specifically of the cancer cell population when co-cultured with fibroblasts. Since this co-culture condition is the closest to mimicking an in vivo environment, the measurements observed from the JC-1 assay would hold the most weight in predicting drug efficacy in patients. The kinetic measurements of death obtained from the JC-1 assay could be used in determining a viable patient dosing strategy while the terminal viability measurement at 72 hours would inform clinicians which drugs would be the most efficacious in treating the patient’s tumor.

Cancer:fibroblast interactions in vitro are sufficient to replicate clinically observed differences in drug sensitivity between subclasses of TNBC cells.

Studying cancer in vitro using cell lines derived from patient tumors has revealed a tremendous amount of information about cancer cell biology, including how cancers induce cell death in response to DNA damage. Yet, there has been a persisting criticism of in vitro studies concerning how the environment is artificial and does not mimic the interactions the tumor is seeing in an actual cancer patient. Therefore, the gold standard for validating phenotypes in the field 105

of cancer biology has traditionally been to test the phenotype in vivo using an animal model. While in vivo animal models provide an environment that more closely mimics the physical and spatial cancer physiology observed in patients, in vivo systems introduce a range of complicating factors. The diversity of cell types creates a complex network of interactions between the cancer cells and the surrounding environment. This multi-faceted complexity reduces the ability to study the effect of individual interactions on cancer cell behavior. However, I demonstrated that an in vitro co-culture system can identify tumor-environment interactions that underlie a drug sensitivity phenotype observed in vivo.

Triple-Negative Breast Cancers are breast cancers that are characterized by the lack of the estrogen receptor (ER), progesterone receptor (PR), and amplifications in the HER2 receptor. TNBCs are divided into multiple subclasses, with the BL and ML subclasses being observed to have very different drug sensitivities clinically in patients. The BL subclass is more sensitive to DNA damaging chemotherapeutics, while the ML subclass is more resistant. However, when this feature is tested in vitro using a basic culturing system, I observe no statically significant difference in drug sensitivity between the two classes. These data argue that the subclass specific drug sensitivity phenotype is controlled by some characteristic intrinsic to in vivo physiology, which is lacking from in vitro systems (Fig 3.1A-C). 3D culturing of cancer cell lines in Matrigel better recapitulates in vivo morphologies while still offering the throughput of traditional in vitro systems. I tested the ability of 3D culture to induce the differential drug 106

response between the TNBC sub-classes and, while I observed a general desensitization to DNA damaging chemotherapeutics, I was not able to recover the sensitivity phenotype (Fig. 3.1D and 3.2D). However, when I performed PCA on my data set of drug responses from specific cancer-fibroblast-drug conditions measured using an in vitro co-culture system I was able to recapitulate the in vivo drug sensitivity phenotype (Fig 3.10A and F). This result demonstrates the ability of in vitro work to identify specific interactions in the in vivo environment that are responsible for altering drug sensitivity without the need to use animal models.

Being able to use in vitro generated data to understand in vivo biology, would allow for the characterization of important interactions using cheaper, higher throughput, and more flexible methods. Expanding on this concept, it would be possible to screen a variety of cell types, matrix proteins, and soluble factors to identify specific interactions that also recapitulate the BL vs ML drug sensitivity difference. Using this data, it may be possible to model the behaviors of cancer cells in a complex in vivo environment using only measurements collected in simpler in vitro environments. This approach would allow for the selection of the most promising combination therapies and in the end save on animal handling costs and time. One caveat is that generating complex models using co-culture based interactions would require relatively limited “higher order” interactions. Testing other environmental components, and the degree to which higher order influences exist, are important future efforts based on our work.

107

Fibroblast mediated alterations in drug sensitivity are diverse in not just the magnitude of the response but also the directionality.

The tumor microenvironment plays a diverse role in supporting cancer growth, survival, progression, and drug response. The most abundant non- malignant cell in the environment are CAFs, normal fibroblasts that have been reprogrammed to myofibroblast-like expression state. CAFs function to deposit and degrade matrix, secrete growth factors, cytokines, and chemokines, as well as mediate cell-cell signaling through integrin interactions. Several studies have begun to explore the complex relationship between CAFs and cancer cell drug response and reported that interactions with CAFs generally desensitize cancer cells to chemotherapeutics with rare instances of drug sensitization. This drug desensitization phenotype is thought to occur regardless of the origin of the CAF, as shown by Kornelia Polyak and colleagues who observed lapatinib resistance in both mammary and brain derived normal fibroblasts and CAFs29. Given these previously reported phenotypes, it was surprising to observe that the screening data produced using the JC-1 co-culture system revealed striking bidirectional effects on drug response. Interestingly I also observed an almost equal distribution between statistically significant desensitizing, 2296, and sensitizing,

1688, interactions (Fig 3.3E and Table 3.4). The magnitude of the responses also varied greatly, with some sensitizing and desensitizing interactions inducing a two-fold increase or decrease in cell viability (Fig 3.3E). 108

I hypothesize that I was able to find these sensitizing interactions because of the screen design and methodology used. As I demonstrated in chapter 2, the

GFP based assay that was used previously to measure these interactions was only sensitive in capturing proliferation and not cell death. Desensitizing drug interactions would be captured as the assay was adept at capturing increased

GFP fluorescence, however the stability of GFP complicates the capture of sensitizing interactions that rely on a decreasing signal. The JC-1 based assay is insensitive to protein degradation and therefore able to capture both types of interactions without sacrificing assay throughput. Moving forward, there needs to be a greater effort to be cognizant of precisely what the assay is measuring, especially when measuring cell death. Scott Dixon and colleagues have recently as highlighted this fact with the development of the STACK assay that measures single cell death kinetics in a population using high throughput fluorescence microscopy99. It is also important to note that the assay used did not sacrifice throughput for sensitivity, as only around 2% of interactions screened showed a statistically significant sensitizing or desensitizing phenotype. The rarity of these interactions requires that the screening space must be large in order to identify hits. Future use of the screening methodology designed and optimized in this study will lead to identification of novel cancer-fibroblast interactions and a deeper understanding of the tumor microenvironment.

109

Fibroblasts derived from metastatic and non-metastatic tissues correlate with the desensitizing and sensitizing phenotypes.

One of the aims of this thesis was to explore the relationship between the fibroblasts derived from metastatic sites of TNBCs versus non-metastatic sites.

The data presented in this thesis demonstrates that there is a divergent effect on

TNBC chemosensitivity based on the tissue that the fibroblast was derived from.

Fibroblasts derived from common metastatic tissues, such as brain or lung, desensitized TNBC cells to targeted therapies as well as DNA damaging agents.

I also found that these fibroblasts desensitized the more aggressive mesenchymal subtype of TNBC to a greater degree than the less aggressive basal-like subtype. This observation could explain why TNBC tumors that are able to metastasize, the mesenchymal subtype, become chemo-resistant after establishing a tumor at one of the four common metastatic tissues. Further study into the connection between the gene expression state of the mesenchymal subtype of TNBC and the secretory pattern of the fibroblasts derived from metastatic sites would aid in the development of better cancer therapies.

Combination therapies would be used to block the interactions between the

TNBC cells and fibroblasts which would then sensitize the cancer cells to approved chemotherapies and improve patient outcomes.

Another discovery I presented in my thesis was that fibroblasts derived from uncommon metastatic sites, particularly the skin and uterus, had a sensitizing effect on TNBC cells exposed to drugs. This was surprising since it 110

has been noted in the literature that fibroblasts generally have a desensitizing effect on cancer cells regardless of their origin. This data could help explain how the metastatic pattern of TNBC has been limited to four tissues since the advent of neoadjuvant chemotherapy, which is used to shrink primary tumors before surgery. Micro-metastases that have migrated to uncommon tissues would be sensitized to the neoadjuvant chemotherapeutic treatment and be killed before growing into metastatic tumors. Further understanding how the fibroblasts derived from non-metastatic tissues sensitize TNBC cells to drugs would improve combination therapies. Using a drug that mimics these interactions for tumors that have migrated to metastatic tissues would increase the effectiveness of DNA damaging agents and improve patient outcomes.

Are TNBC preferentially enriched for sensitizing interactions?

One caveat to the observed enrichment in sensitizing interactions is that I only screened cancer:fibroblast interactions using one cancer type. It is possible that the triple-negative class of breast cancer is preferentially enriched for these interactions compared to ER/PR or HER2 amplified breast cancers, or other types of cancer in general. Therefore, it would be advisable to expand the screening approach first to other breast cancer cell lines and finally to other cancer types. Expanding the cancer types screened would provide answers to two interesting questions: 1) Is the ratio of sensitizing to desensitizing interactions consistent among cancer types? and 2) Does the identity of the 111

fibroblasts the mediate these phenotypes change based on the cancer type being screened? Identification of sensitizing interactions that span multiple cancer cell types would increase the therapeutic options facing those patients as well as increasing the efficacy of already used DNA damaging agents.

Do the sensitizing and desensitizing phenotypes validate in vivo?

As previously mentioned, the gold standard for validating a cancer related phenotype is to test it using an in vivo animal model. Drug response could be measured in animals receiving a co-injection of luciferase expressing cancer cells and either sensitizing or desensitizing fibroblasts into the mammary fat pad.

The injection methodology would have to be optimized to increase fibroblast seeding at the site of tumor initiation, as our data suggest that Matrigel should not be used, due to competition between Matrigel and fibroblast mediated interactions. Specifically, I observed that drug desensitization mediated by culture in Matrigel is dominant over the cancer-fibroblast co-culture sensitizing phenotype (Fig 3.16). Replicating the phenotypes in vivo would also test how robust the CAF mediated drug response phenotype is. The co-culture system where I observed the large sensitization and desensitization phenotypes, up to a

2-fold change from control cells, confined the two cell types to a small area and forced cell-cell interactions. However, injection of the cells in an animal model would allow for more movement and introduce other cell-cell interactions besides the CAF-tumor interaction in the co-culture system. If the phenotype is robust, 112

the introduction of other cell-cell interaction would not change the CAF mediated alteration of drug response. I would also be able to measure the dominance of the novel sensitizing interactions I discovered. Murine mammary fibroblasts, which have been previously shown to desensitize breast cancer to chemotherapeutics, would infiltrate the tumor and compete with the sensitizing

CAFS to modulate cancer cell drug sensitivity39. Determining the dominance of the sensitizing phenotype would be important for its use as a therapy and aid in predicting what environments it can be used in.

Exploring beyond fibroblasts, adding immune cells to the environment.

The work presented in my thesis has shown that interactions between fibroblasts and cancer cells change drug response, but the in vivo tumor microenvironment is composed of many more elements than just fibroblasts. To further model how environmental interactions influence cancer cell chemosensitivity, the next step would be to add another component of the environment to the in vitro co-culture system and test if the observed sensitizing or desensitizing phenotypes are altered. The next component of the microenvironment that I would like to add would be immune cells. The tumor microenvironment is highly inflammatory and attracts many types of immune cells such as lymphocytes, macrophages and dendritic cells. One subset of immune cells that I would be interested in testing would be T-cells. T-cells are recruited to the inflammatory microenvironment and, once activated, target cancer cells and 113

induce death. I would like to add activated T-cells to a select number of drug- fibro-cancer combinations and measure whether the basal rate of death of the cancer-fibroblast combination is altered using the JC-1 assay. I hypothesize that the secretory pattern of fibroblasts derived from common metastatic tissues would reduce the effectiveness of activated T-cells resulting in the conserved desensitization to chemotherapeutics.

Diving into the data, directions for future research.

The screen performed in my thesis to investigate the effect of cancer- fibroblast interactions on the efficacy of chemotherapeutics yielded a large amount of data. In my thesis I performed PCA to determine large scale trends in the data across the six cancer cell lines, which resulted in the described difference between fibroblasts derived from metastatic versus non-metastatic sites. However, further analysis of the data can reveal other research questions that would aid in understanding interactions that alter cancer cell response to chemotherapeutics.

My thesis focused on fibroblast specific interactions, however preliminary analysis of the data also reveals drug specific and fibroblast independent interactions that alter therapeutic response. For example, the efficacy of microtubule poisons is decreased when cancer cells are co-cultured with fibroblasts regardless of the anatomical location they were derived from. RNAseq analysis comparing the changes in expression state between cancer cells in 114

mono-culture versus co-culture across the 16 fibroblast lines for one cancer cell line would elucidate the molecular mechanism behind the resistant phenotype.

Understanding the underlying mechanism would allow for the development of combination therapies to increase the effectiveness of microtubule poisons.

Another facet of the data that can be further explored are the divergent sensitivity phenotypes observed between fibroblasts derived from the same anatomical area. For example, there were two liver derived fibroblasts that had opposite drug independent sensitivity phenotypes. Comparing the expression profiles and secreted proteins between these two fibroblasts would reveal valuable insights into the molecular mechanisms underlying the change in apoptotic priming observed in cancer cells when co-cultured with these fibroblasts.

CAFs alter cancer cell drug response by modulating apoptotic priming.

When a cell is exposed to a drug, commitment to cell death is a function of two properties. First, the degree to which the apoptotic signaling pathway is activated and second, the basal priming state of the cell prior to drug exposure.

The priming state of the cell is dictated by the ratio of pro and anti-apoptotic proteins at the mitochondrial membrane. Interactions that modulate either the expression, levels, or localization of these proteins can therefore alter the commitment to apoptosis as a result of cellular damage. In this study I was able 115

to demonstrate that interactions with CAFs were capable of either raising or lowering the priming state of cancer cells, leading to altered drug sensitivity.

Furthermore, I was able to predict the therapeutic efficacy of a drug combination consisting of a DNA damaging agent and a BCL-2 inhibitor based on the environmental context. This has huge implications in the treatment of tumors as the environmental context, in this case the tissue identity of the CAFs, could be used as a biomarker to inform what drug or drug combinations should be used in that specific patient.

What are the molecular mechanisms underlying the change in apoptotic priming?

As previously mentioned in chapter 4, CAFs interact with cancer cells in a variety of methods, making it challenging to identify specific molecular mechanisms. I attempted to attribute the priming mediated sensitization phenotype to the expression of the cytokine IL8 but was unable to observe a change in the apoptotic threshold in cells treated with recombinant IL8 protein. I hypothesize the multiple factors are involved in altering the apoptotic threshold. I would test the ability of conditioned media derived from sensitizing fibroblasts to alter apoptotic priming. If conditioned media is not capable of altering priming, I would move to test the other methods of CAF interacting with cancer cells such and matrix protein depositions. Depending on the method of communication between the CAFs and cancer cells, novel therapies could be developed to treat cancer. For example, if a desensitizing interaction is mediated by binding to a 116

specific matrix protein, an antibody-based approach to block the interaction can be employed to sensitize the cancer cells to cytotoxic agents. The tissue specificity of the phenotype suggests that the molecular mechanisms might also be tissue specific. This would be beneficial in the treatment of metastatic tumors or primary tumors in a difference tissue type, since expressing a different tissue specific protein would preferentially sensitize the cancer cell over causing any gross changes to the normal tissue.

Therapeutic implications of fibroblasts modulating apoptotic priming.

BCL-2 inhibitors are a new class of drugs that are currently being investigated as potential chemotherapeutic agents. The rationale behind the development of BCL-2 inhibitors was the observation that many cancers upregulate the BCL-2 family of proteins. BCL-2 family proteins play an important role in regulating the initiation of apoptosis by inhibiting pro-apoptotic proteins. It was hypothesized that the upregulation of the BCL-2 family was leading to chemo-insensitivity in cancers and that inhibiting the function of these proteins would re-sensitize cancer to DNA damaging chemotherapeutics. It has been shown that BCL-2 inhibitors, such as ABT-737, have been effective in increasing the sensitivity of non-small cell lung cancer using in vivo xenograft mouse models.

In my thesis I see a similar sensitization of the TNBC cell line BT20 grown in mono-culture when treated with a combination of the DNA damaging agent 117

teniposide and the BCL-2 inhibitor ABT-737. This is shown by a synergistic combination index, indicating that the drugs in combination are more effective at killing the cells than expected based on the responses observed when the drugs are given individually. This phenotype is enhanced when the culturing environment is changed to co-culture with uterine derived fibroblasts that sensitize TNBC cells to DNA damage. However, when the TNBC cells are co- cultured with adrenal derived fibroblasts that induce a desensitizing effect to DNA damage, the synergistic effect between ABT-737 and teniposide is lost and turns antagonistic. This can be explained by the fibroblast mediated anti-apoptotic effect being greater in magnitude than the BCL-2 inhibitor mediated pro-apoptotic effect, rendering ABT-737 ineffective.

The observed environmental dependent effectiveness of ABT-737 is an important feature that should be noted when designing combination therapies for patients with TNBC. Combination therapies between ABT-737 and DNA damaging agents will be effective only in environments that have a neutral or sensitizing effect on apoptotic priming, such as a primary tumor in the breast environment or a metastatic tumor in the uterine environment. However, this combination therapy would be ineffective against tumors that follow the common metastatic pattern of TNBC and grow in drug desensitizing environments such as the lung, liver, or bone.

118

What are the mechanisms involved in the drug specific and environmental independent interactions observed?

Not all cancer-fibroblast-drug interactions that change drug response act through modulating apoptotic priming. There are striking drug specific but environment independent interactions, such as interactions involving vinblastine or mitoxantrone. The environmental indifference of these interactions suggests that the mechanism altering drug response is independent of the change in the apoptotic threshold. I hypothesize that these drugs are activating a cell death pathway besides apoptosis, therefore they are insensitive to changes in priming.

This would have potential therapeutic advantages as a combination therapy with these two classes of drugs kill cells that have become resistant to apoptosis.

Advances in understanding the microenvironmental regulation of TNBC drug response.

The work presented in this thesis has added to the field of microenvironmental regulation of cancer cell chemosensitivity. While previous studies have investigated the effect of fibroblasts on cancer cell chemosensitivity, the bulk of the data focused on targeted agents. In my thesis I demonstrated that the desensitizing effect that was previously observed with targeted agents is also observed with DNA damaging agents. The underlying mechanism of this change in sensitivity is due to modulation of apoptotic priming. I also demonstrated that fibroblasts can sensitize TNBC cells to chemotherapeutics, a surprising finding 119

as the current literature describes fibroblasts as being desensitizing influencers.

My data also shows that the direction fibroblasts alter drug response is correlated with whether the fibroblasts were derived from common TNBC metastatic tissues. This finding offers new thoughts into why TNBCs are constrained to four metastatic sites following neoadjuvant chemotherapy, the chermoresistant influence of the fibroblasts allow micrometastases to survive treatment, as well as opens new avenues for the development of breast cancer therapies.

The data presented in my thesis concerning the change in apoptotic priming and the use of BCL-2 inhibitors also adds another dimension to data previously described by the Polyak group. Marusyk et al. demonstrated that BCL-

2 inhibitors decreased the fibroblast mediated desensitization to chemotherapeutics and advocated the use of BCL-2 inhibitors in combination with the targeted agent lapatinib. However, when I investigated the use of BCL-2 inhibitors used in combination with DNA damaging agents I found that the combination therapy was antagonistic in conditions where cancer cells were co- cultured with fibroblasts derived from common metastatic tissues. This is an important distinction because the combination of BCL-2 inhibition and DNA damage would not be effective in treating metastatic TNBC tumors. One difference between the assays used in Marusyk et al. and my thesis is that

Marusyk et al. recovered the cancer cells from the 3D co-culture before performing the priming assay while I performed the priming assay on cancer cells in co-culture with fibroblasts. This difference in the protocol could indicate that 120

there is a rapid expression state change when cancer cells are removed from coculture with fibroblasts and that there needs to be constant signaling from fibroblasts to stay in the drug desensitized state. The data and conclusions presented in this thesis are important in the development of combination therapies that would be effective in treating metastatic TNBC.

121

APPENDIX A: Studying Cellular Signal Transduction with OMIC Technologies

Appendix A was taken almost verbatim from: Landry BD, Clarke DC, Lee MJ. Studying Cellular Signal Transduction with OMIC Technologies. J Mol Biol. 2015;427(21):3416-3440. doi:10.1016/j.jmb.2015.07.021.

Abstract:

In the gulf between genotype and phenotype exists proteins and, in particular, protein signal transduction systems. These systems use a relatively limited parts list to respond to a much longer list of extracellular, environmental, and/or mechanical cues with rapidity and specificity. Most signaling networks function in a highly nonlinear and often contextual manner. Furthermore, these processes occur dynamically across space and time. Because of these complexities, systems and “OMIC” approaches are essential for the study of signal transduction. One challenge in using OMIC-scale approaches to study signaling is that the “signal” can take different forms in different situations.

Signals are encoded in diverse ways such as protein-protein interactions, enzyme activities, localizations, or post-translational modifications to proteins.

Furthermore, in some cases signals may be encoded only in the dynamics, duration, or rates of change of these features. Accordingly, systems-level analyses of signaling may need to integrate multiple experimental and/or computational approaches. As the field has progressed, the non-triviality of integrating experimental and computational analyses has become apparent. 122

Successful use of OMIC methods to study signaling will require the “right” experiments and the “right” modeling approaches, and it is critical to consider both in the design phase of the project. In this review, we discuss common OMIC and modeling approaches for studying signaling, emphasizing the philosophical and practical considerations for effectively merging these two types of approaches to maximize the probability of obtaining reliable and novel insights into signaling biology.

Introduction:

A cell's ability to perceive and understand its surroundings, respond to changes in the environment, and/or adapt to defend homeostasis is governed by complex signal transduction systems. In plain terms, the basic purpose of signaling is to relay information about the outside of the cell to the nucleus. This signaling often results in alterations to the transcriptional state of the cell as to modify cellular behavior100. The state of the signaling network creates the background “context” that is often missing in genomic studies of behavior/drug response/disease outcome, which has substantially limited our ability to leverage genomic or transcriptomic insight for therapeutic benefit. Decades of research have identified signaling pathways that control essentially every cellular function

(growth, migration, death, differentiation, division, secretion, etc.). Owing to this fact and because these systems typically include many “drug-able” components, signaling has emerged as a common target in modern molecular medicine. 123

However, after a decade or so of optimism, it has become clear that the pharmacological targeting of signaling often results in unanticipated consequences, which manifest clinically as toxicity, acquired resistance, and/or limited efficacy. Progress in the area of targeted (or “personalized” or “precision”) medicine will likely require a more refined understanding of how signaling pathways function, how they are altered in disease states, and how they are augmented by drug therapies.

The complexity of signal transduction networks demands “systems-level” approaches, but answering this demand remains a major challenge. Signaling networks are densely interconnected, which facilitates signaling “crosstalk” (i.e., pathway activities influence each other), and most forms of crosstalk are incompletely understood. Furthermore, the “architecture” of a signaling process

(i.e., the connectivity or wiring of the signaling circuit) is regulated dynamically and can change through time. This rewiring or reprogramming of signaling networks is a concept that is only now being explored in a systematic way54,101.

Finally, the information content (i.e., the signal) takes different forms in different situations. Thus, no single OMIC method is appropriate for the comprehensive study of signaling. In canonical examples, the signal is encoded and transmitted by the absolute levels of a protein or post-translationally modified form of a protein. Additionally, signals are also encoded by the location of a protein; the interactions between proteins; the relative stoichiometry of positive and negative regulators; and even the rate, duration, or patterns of change between these 124

states (Fig. A.1). Moreover, if the future continues to reaffirm recent precedent, these “complex” forms of signaling will prove to be commonplace.

Because signaling has long been acknowledged as being complex, many systems biology approaches have been applied to their study. In the early days, the typical approach was to generate a large amount of data using the latest 125

OMIC technology (e.g., microarray) and to apply whatever computational approach was currently the “hottest”, followed by believing that new insights would inevitably emerge from the modeling. This philosophically problematic approach was exacerbated by the practical difficulties of these two steps often being performed by different people in different laboratories whose communication was likely impaired by differences in field-specific jargon and culture, as well as more fundamental differences in intellectual goals. Since these early efforts—which we call version 1.0 of systems biology of signal transduction—the field has matured considerably. Systems approaches featuring

“OMIC” data and computational modeling are and will continue to be instrumental in the study of signaling. However, it is clear that their successful implementation requires careful planning and consideration of the respective strengths and weaknesses of the experimental and computational strategies not only in isolation but also in relation to each other. In this review, we will cover different ways that a systems biologist can collect and model signaling data. We will cover some recent examples from the literature and discuss in detail some common pitfalls and practical considerations for successfully integrating OMIC data with computational approaches in signaling research.

Methods for High-Throughput Collection of Quantitative Signaling Data:

The creation of robust and precise models relies on the generation of accurate data sets that measure how signals in the cell are conveyed. High- 126

throughput methodologies have facilitated the generation of reliable large-scale data sets for a wider variety of signals. Because signals are encoded in different ways, many different technologies are necessary to measure the propagation of a signal throughout the cell. Methods include those that are based on “molecular profiling” of protein signaling states and those based on “molecular perturbation” and inference102. In this section, we will review various assays used to collect signaling data, discuss the advantages and limitations of each method, and also discuss practical considerations in the design of OMIC experiments for interrogating signaling. An overview of these approaches is also available in Fig.

A.2. This section will be formatted around four basic ways in which signals are encoded: the levels, interactions, localizations, and activities of various signaling proteins103.

Measurement of protein expression and post-translational modification levels.

In the simplest situations, the signal is encoded by the mere existence or non-existence of a post-translationally modified form of a protein or, more rarely, the total expression level of a protein. Several methods exist for quantifying protein levels, but only one, mass spectrometry, can be argued to provide comprehensive OMIC-scale depth. 127

128

129

130

131

132

Mass spectrometry

To many, the gold standard in proteomics is mass spectrometry. Unlike any other proteomic technologies, mass spectrometry can be used to measure complete proteomes and thousands of post-translationally modified proteins, rivaling the depth and richness of other genomic technologies. Furthermore, these measurements are precise and quantitative, and these obviate the need to rely on the existence of specific antibodies (although antibodies are often used in mass spectrometry experiments to enrich for protein fractions of interest). By using cellular fractionation techniques or enrichment of proteins by chromatographic or other methods, it is possible to comprehensively and quantitatively study the global protein landscape, specific subsets of proteins, or proteins belonging to certain cellular compartments. For example, mass spectrometry has been used to study the global ubiquitin proteome by immunoprecipitating ubiquitin104, phosphotyrosine proteomes by using pan- phosphotyrosine antibodies105, and autophagy-regulatory proteins by precipitating proteins involved in vesicle trafficking106. These insights have provided a comprehensive parts list for important regulatory functions in biology.

One particularly exciting application of mass spectrometry in signaling research is the use of multiplex kinase inhibitor beads107. In MIB-MS (mass spectrometry of multiplex kinase inhibitor beads), a cocktail of inhibitors that promiscuously bind to many kinases are conjugated to beads and mixed with protein lysate. Several inhibitors are used to target the majority of the expressed 133

kinome. The binding of a particular kinase to the bead occurs as a function of the kinase's native affinity for the inhibitor and its level of expression, as well as the level of kinase activity, since kinase inhibitors bind preferentially to activated forms of kinases. Thus, differential binding between experimental conditions can be inferred in most cases to reflect differences in kinase activity. Duncan et al. used MIB-MS to profile kinome activity in breast cancer cells revealing total kinome reprogramming following targeted MEK inhibition101. In addition to these multi-inhibitor conjugated beads, others have used covalent kinase inhibitors in a similar manner. Activity-based proteome profiling uses small-molecule probes designed to react with nucleophilic residues, most typically with cysteine, near the kinase active site to form irreversible covalent interactions. These activity- dependent interactions can be specific for target kinases, although recent research suggests that off-target binding to both kinases and non-kinase proteins can be problematic108.

While powerful, some issues do potentially exist for the use of mass spectrometry in the study of signal transduction. For example, decoding mass spectra to reveal peptide identity is non-trivial in general, and this may be worsened in signaling studies because many signaling proteins share similar domains, some of which have very similar and sometimes identical peptide sequences. Excellent software exists to aid in peptide identification; however, many of these statistical tools may be underestimating false discovery rates, sometimes dramatically so109,110. Furthermore, mass-spectrometry-based 134

proteomics is relatively slow, expensive, and does not scale well to the analysis of hundreds or thousands of samples that can easily emerge from studies involving many ligand doses, experimental time points, and other perturbations.

Undoubtedly, many of these issues will ultimately be resolved. Mass spectrometry is a mature but ever-changing technology with many different specific iterations. For example, while mass spectrometry typically subsamples the input protein, recent studies report deep, even full, proteome depth111. Also, when proteomes are subsampled, comparisons between samples are made non- trivial, but these comparisons can be made using labeling methods such as

SILAC112, using isobaric mass tags, or in label-free approaches with the aid of computational tools113. Recent advances in targeted mass spectrometry have increased the specificity and sensitivity of target identification and analysis114.

Immunoblotting

Traditional methodologies based on Western blotting continue to be tried and true tools for measuring total protein levels and levels of specific modified forms of proteins, assuming suitable antibodies exist (see note below on antibody-based proteomics). Using various biochemical separation and/or immunoprecipitation techniques, it is also possible to quantify protein localization to various subcellular compartments or interactions between proteins and other proteins or biomolecules. These tools are often comforting to biological scientists both because of their familiarity and because visualization of a band of predicted 135

size on a gel provides assurance that the antibody is reporting the correct protein. If paired with an appropriate analysis platform—one with a large dynamic range that is confirmed to be linear or whose non-linearities can be fit to a standard curve—these approaches can be quantitative115. Although the multiplexing capabilities on immunoblots are limited to two or three parallel measurements, the relatively low cost of these approaches allows throughput to be achieved through repetition, particularly when the lysate material is abundant.

Also, modern commercially available platforms have dramatically shortened the gel electrophoresis and transfer durations, such that it is feasible for a single operator to analyze hundreds of samples per day, each for dozens of specific analytes. Immunoblotting remains a common tool for validating and precisely studying single proteins or even small protein networks or as complementary methods to “fill in the gaps” left by other high-throughput techniques. For most applications, however, these approaches simply lack the throughput necessary for generating comprehensive OMIC-scale quantitative data sets.

Protein lysate microarrays

A higher-throughput approach that uses the same intuitive design as immunoblotting is lysate microarrays116. In these approaches, antibodies are still needed in order to quantify the levels of proteins or modified protein species.

These approaches utilize robotic tools to print material in regular patterns, typically on a smaller and denser scale that could be achieved by hand. Analysis 136

of protein lysates in this miniaturized format requires much less biological material than standard approaches, and it uses dramatically reduced amounts of expensive reagents such as antibodies. Two basic formats exist: forward- and reverse-phase arrays. Forward-phase arrays are constructed by printing capture antibodies on a flat surface and by flowing a cellular lysate over the printed antibody chip. In this approach, a second antibody is typically then applied to determine the level of protein captured by the immobilized antibody. While this approach negates the need to label the lysate material, multiple specific antibodies that can bind to the target protein without competition must exist.

In addition to the need for multiple target specific antibodies, forward- phase arrays also generally require relatively large amounts of protein lysate.

Reverse-phase protein arrays (RPPAs) alleviate this issue. Here, small amounts of protein lysates (approximately picogram range) are spotted and dried onto a rigid protein-binding surface such as nitrocellulose. The protein lysate array can then be probed and measured in a manner akin to a traditional immunoblot.

RPPA can rapidly measure many lysates in parallel, wherein “multiplexing” is achieved through repeated printing of lysates onto parallel arrays and probing with different antibodies. RPPAs are likely the most cost effective and the highest-throughput solution for experiments that include many samples

(approximately 100–10,000 +) with relatively few target analytes of interest

(dozens to hundreds). For example, a single array can be printed to include thousands (even 10,000–100,000 +) of individual lysates, which may represent 137

different cell types, drug treatments, time points, patient samples, and so on.

Once printed, RPPAs essentially function similarly to high-throughput “dot blots”, and they share many of the same benefits and limitations. For example, RPPAs omit the electrophoresis step used in immunoblotting. This omission reduces assay time but also can increase non-specific antibody binding, which limits dynamic range and assay sensitivity. Thus, antibodies used in RPPA need to be validated carefully117,118. A modification to the RPPA method that may improve sensitivity is the micro-Western array119. The central difference between these methods is that cell lysates are printed onto a gel and undergo a short semi-dry electrophoresis step before being transferred to a membrane and probed with primary and secondary antibodies. The short electrophoresis step may separate competing epitopes and may allow for resolution of the specific protein bands, assuming that non-specific binding proteins migrate differently (i.e., are of a different size)120.

In spite of some obvious limitations, many antibodies have been carefully validated and hundreds of protein targets can be studied currently using

RPPA121. Many researchers have used RPPA and other array-based methods to study signaling, particularly in cancer biology. For example, RPPAs are used to analyze primary human tumor samples as a part of the Tumor Cancer Genome

Atlas122. Considering the limited amounts of patient material available for each sample and the relevance of signaling proteins in targeted medicine, RPPA plays a critical role in this analysis. In another example, Federici et al. recently used 138

RPPA to interrogate signaling pathway activation states of all cell lines in the

NCI-60123. Their pathway activation mapping approach revealed several phenotypes that were common to subsets of the NCI-60 cell lines and transcended typical organ-type distinctions. By integrating these data with other

OMIC data for coding and non-coding RNAs and genomic features, they were able to identify functional networks that significantly improved prediction of drug sensitivity compared to traditional approaches that use only genomic or transcriptomic features.

Multiplexed flow cytometry, mass cytometry, and bead-based assays

For many signaling proteins, the population expression or activity levels measured in bulk lysates can be deceiving. For example, at the single-cell level, some signaling systems only exist in “on” or “off” states and graded population responses actually reflect differences in the percent of cells that are “on” and not the degree to which the pathway is “on”. This scenario is not only obviously true for binary signals or “all-or-none” processes such as cell death124 but also true for other processes that may intuitively appear more graded than binary. In fact, even in unimodal distributions, cell-to-cell heterogeneities can be substantial and can underlie divergent behaviors, including sensitivity and resistance to drugs125,126. Flow cytometry is probably the most widely used assay to quantitatively study signal transduction with single-cell resolution. The basic workflow is similar to other antibody-based techniques, although cells are first 139

fixed as individual cellular units, before permeabilization and labeling with specific antibodies. Labeled cells are passed in single file through a flow cell where fluorescence intensity, cell size, and cellular internal complexity can be measured. In OMIC applications, flow cytometry suffers from two basic weaknesses. First, this is not a high-throughput method, since sample preparation and read time can be quite extensive, particularly for adherent cells.

Second, multiplexing of measurements is limited by the available fluorescent spectra. Using advanced machinery and careful fluorescence compensation, one can multiplex more than a dozen measurements127, but most users measure 1–

12 signals per sample.

To improve upon the multiplexing limitations of traditional flow cytometry,

Scott Tanner and colleagues coupled the single-cell profiling capabilities of flow cytometry to the exquisite sensitivity of mass spectrometry128. Mass cytometry uses antibodies conjugated to rare-earth metals, typically those in the lanthanide series, instead of fluorophores129. Samples are first processed using flow cytometry techniques to create a stream of single cells, which are subsequently vaporized into single-cell droplets. The amount and identity of the bound antibody is quantified using a mass spectrometer to detect differences in mass corresponding to the different atomic masses of the rare-earth metals. Using mass cytometry, one could theoretically analyze more than 100 unique analytes in a single cell, but currently, 30–50 distinct signals are feasibly analyzed simultaneously. Mass cytometry is still nascent technology but has already 140

transformed our understanding of immune cell development. In a recent publication, Nolan and colleagues used mass cytometry to label 34 parameters in human hematopoietic cells, including surface and intercellular proteins130. Their analyses revealed previously unknown relationships between precise signaling states and well-characterized hematopoietic cell subtypes and revealed a more refined view of hematopoietic stem cell developmental trajectories.

While mass cytometry offers impressive multiplexing capabilities, other flow-based approaches can expand multiplexing capability by an order of magnitude although not with single-cell resolution. Multiplexed-bead-based assays (like those sold commercially by Luminex) combine antibody labeling with laser-based cytometry. Here, polystyrene or magnetic beads are filled with specific ratios of two or three dyes, creating a colorimetric signature analogous to a barcode used in genomic or RNA-sequencing methods. Each uniquely colored bead is then conjugated to an antibody against a specific protein such that measured fluorescence levels can be attributed to the specific protein of interest.

This technology allows for the concurrent measurement of 500 simultaneous signals per sample, although antibody cross-reactivity typically limits the multiplexing capability well below this threshold. Using this method, Du et al. recently profiled nearly all tyrosine kinases in hundreds of cancer cell lines, each treated with dasatinib, an inhibitor of the tyrosine kinase Src131. A similar strategy was also used to increase the throughput of transcriptome profiling. In this case, antibodies were replaced with oligonucleotide probes for each of 1000 141

“landmark” genes, which allowed inference of genome-wide transcriptional states at a fraction of the cost of traditional array or sequencing methods132.

Caveats in antibody-based proteomics

Essentially all of the abovementioned approaches rely on the existence of high-fidelity antibodies to detect the specific epitopes of interest, and a discussion of their use must consider the main limitation: the lack of reliable antibodies for most protein targets133. Also, with the exception of immunoblotting, the other methods all attempt to bind to a target epitope without an electrophoresis step to remove other competing off-target binding sites. Thus, the proportion of available antibodies that produce reliable signals in these high- throughput platforms is only a fraction of the antibodies that produce reliable signals in immunoblots. Estimates vary, but generally between 1% and 10% of antibodies that work in immunoblots will work in the other methods we discussed.

Also, because most methods do not produce a visible band on a gel, additional measures should be taken to validate that the signal is reporting the levels of the correct protein. When validating antibodies for RPPA, for example, we previously used a two-step protocol. We began first by testing antibodies in RPPA and identifying those that reported a dynamic range of at least 1.5-fold across 90 generic control conditions, which were designed to feature positive and negative controls for all targets of interest. In a second phase, we retested that small subset of antibodies in immunoblot format to confirm that the change in signal 142

occurred in a protein of correct predicted size and that the two methods produced correlated data54. Additionally, antibodies that are accurate when used in isolation may not produce reliable signals when multiplexed with other antibodies due to cross-reactivity; thus, the reliability of measurements derived from highly multiplexed platforms should be tested and validated rigorously.

Methods to measure protein–protein interaction.

Many post-translational modifications or activated signaling proteins cause conformational changes or the creation of new binding domains, resulting in regulated protein–protein interactions. Thus, measuring the extent to which proteins interact can indicate signaling pathway activity, and identifying new binding partners may reveal new proteins involved in signal regulation. Below is an overview of common tools used to measure protein–protein interaction in high throughput.

Yeast 2-hybrid (Y2H) assays

The yeast 2-hybrid system is a classic approach to identify protein–protein interactions134. This approach takes advantage of the modularity of many transcription factors, which have two separable functional domains that confer

DNA-binding and transcription-activating functions. In Y2H, a protein of interest is fused to the DNA-binding domain while a protein or library of proteins to be screened are fused to the corresponding transcriptional activation domain. 143

Protein–protein interaction can then drive transcription of a reporter gene135.

There are several caveats to this method. Non-physiological false-positive interactions can occur either due to overexpression of the screened proteins or because proteins are forced to co-exist in the nucleus that may not typically reside in a shared subcellular compartment. Also, false negatives can occur. Y2H will often miss dynamically regulated interactions or interactions that require a post-translational modification that may be lacking in yeast or not properly regulated in the assay. Also, since the readout for binding occurs in the nucleus, only proteins that can accommodate nuclear localization can be screened.

Y2H screening can be performed in high throughput, even on a comprehensive scale. Using Y2H, Vidal and colleagues have recently published the first full-proteome-scale map of the human protein–protein interaction network136. This systematic analysis has doubled the list of known/potential binary interactions, and strikingly, their analysis has revealed that the interactions identified by systematic Y2H are free from a bias toward well-studied proteins existing in the established interactome. Thus, this data set is certain to include clues into many unexplored functional relationships.

Membrane Y2H, MaMTH, and other protein complementation assays

The nuclear localization of traditional Y2H assays can be a severe limitation for some aspects of signal transduction. Transmembrane proteins, for example, play important roles in signaling as the primary signal relay between 144

the cell exterior and interior, and regulated protein–protein interaction among transmembrane proteins is a major mechanism by which signals are encoded at the receptor layer of a signaling pathway. For example, the primary means by which receptor tyrosine kinase autophosphorylation transmits a signal is through the recruitment of proteins containing phosphotyrosine-binding domains.

Determining binding interactions in high throughput for membrane proteins is challenging since they contain large stretches of hydrophobic residues and cannot be expressed readily in aqueous solutions. Membrane Y2H and a modified version that works in human cells, mammalian membrane two-hybrid

(MaMTH), were developed to circumvent this issue and allow for high-throughput screening of membrane protein interactions137,138. Both assays rely on the concept of split ubiquitin binding. A transmembrane protein of interest is tagged with both the C-terminal half of ubiquitin and a transcription factor that is used to drive transcription of a colorimetric or fluorescent reporter protein. Proteins that are to be screened for binding are tagged with the N-terminal half of ubiquitin.

Binding events between the membrane protein of interest and the screened protein yield the formation of a complete ubiquitin molecule, which is recognized by endogenous deubiquitinating enzymes, releasing the tethered transcription factor to drive reporter expression. Using MaMTH, Petschnigg et al. studied

EGFR signaling in non-small cell lung cancers containing EGFR mutations138. By screening binding partners of the mutated receptor and comparing them to a wild-type EGFR signature, the investigators were able to identify a novel binding 145

protein that promotes constitutive signaling. MaMTH is a new technology that will certainly be used in the coming years for the study of receptor signaling and drug mechanism of action in human disease.

In addition to these approaches, there exist several other methods that rely on the complementation of a split reporter protein, which can regain function when the two halves are in proximity. In addition to the use of ubiquitin or transcription factors, other similar strategies have used fragments fluorescent proteins (bimolecular fluorescence complementation)139 of the tobacco etch virus protease (split TEV)140, dihydrofolate reductase141, β-lactamase142, luciferase143,

LacZ144, and many other proteins, demonstrating the versatility and generalizability of this technique.

LUMIER

Another commonly used method for detecting protein–protein interactions is co-affinity purification. To adapt this generic strategy to high-throughput analyses, the Wrana laboratory developed LUMIER (luminescence-based mammalian interactome mapping). In this method, a protein of interest is fused to luciferase and co-expressed in mammalian cells with target proteins fused to an affinity purification tag. After signal stimulation and cell lysis, interaction/co- immunoprecipitation can be inferred through luminescence intensity. The main advantage of this approach over other co-immunoprecipitation-based methods is throughput, especially since these assays are sensitive enough to be performed 146

in microtiter plates. LUMIER was originally used to investigate transforming growth factor-β (TGF-β) signaling in human cells. Wrana and colleagues screened over 500 potential TGF-β binding partners and identified links between the TGF-β pathway and various other processes, such as epithelial tight junction regulation, which contextually help regulate TGF-β signaling145. More recently,

Lindquist and colleagues applied the LUMIER technique to comprehensively study the Hsp90 interactome146 and to study the specificity of kinase inhibitors in living cells147. In the latter example, the authors leveraged the fact that binding of small molecules to kinases results in thermodynamic stabilization and thus loss of interaction with chaperone proteins. This study revealed critical targets of kinase inhibitors, including many that were unexpected.

Protein chips

In addition to protein lysate microarrays described above, many other types of arrays can be built to study protein biology, some of which are powerful tools for studying protein–protein interaction rather than expression levels148.

Protein chips are engineered by arraying and immobilizing purified proteins or protein domains, instead of antibodies or lysates. Protein chips have many uses.

For example, a protein, lipid, nucleic acid, or small molecule of interest could be profiled against a near-proteome-wide protein chip to reveal all potential binding partners149. Huang et al. used this approach to identify new components of mTOR signaling150, and Jones et al. used protein chips featuring nearly all 147

human Src homology 2 (SH2) and phosphotyrosine-binding (PTB) domains to comprehensively study protein recruitment to activated ErbB receptors151.

Quantitative measurement of protein localization or protein activity.

For many signaling nodes, the activation status cannot be inferred from the levels of a post-translationally modified protein or from the interactions between proteins. In these cases, it may be necessary to directly measure enzymatic activity or localization of the signaling protein of interest. In general, available methods can be quantitative but are not easily performed in a high throughput. Here, we briefly survey quantitative methods for studying protein activity and the dynamics of protein localization.

Quantitative microscopy

The transmission of a signal is often driven by a translocation event of a protein, for example, from the cytoplasm to the nucleus, or the clustering of a protein to the plasma membrane. Thus, for some signaling nodes, quantitative microscopy can be a reliable approach to monitor pathway activation.

Additionally, microscopy may be the best approach when studying rapid signals or for those that occur in an oscillatory manner, as is the case for NF-κB signaling152 and p53 signaling153,154. Recent analyses of NF-κB signaling by

Covert and colleagues revealed substantial heterogeneity at the single-cell level.

These quantitative single-cell insights were instrumental in revealing the use of 148

both digital and analog features to differentially control activation of specific NF-

κB target genes152. Single-cell resolution has also been valuable for the study of signaling processes that are initiated stochastically. For example, a critical event in apoptotic cell death signaling activation is the formation of pores that permeabilize the mitochondria (called mitochondrial outer membrane permeabilization or MOMP). The timing of MOMP after cell death stimulation is variable but can be measured directly using fluorescence resonance energy transfer (FRET) sensors, fluorescent reporter constructs, or membrane potential dyes that localize to the mitochondria124,155. Single-cell analyses of MOMP dynamics have revealed a nearly invariant “switch-like” activation that is masked by the graded population-level responses, and computational modeling of these data have revealed critical features that contribute to this binary response and activation stochasticity156.

Quantitative, automated microscopy is becoming increasingly available as microscopes have become better and cheaper. Image processing and quantification, however, remain major challenges in the use of quantitative microscopy. Several freely available image analysis programs exist to aid in quantitative image analysis, including CellProfiler157, PhenoRipper158,

ImageRail159, and ImageJ160.

149

In vitro kinase and other in vitro enzyme assays

In many kinases, phosphorylation on the activation loop can be used safely to indicate kinase activity. Unfortunately, this feature is occasionally misinterpreted to assume that all phosphorylation events on kinases modulate their activity. Most phosphorylation on kinases do not alter their kinase activity and many kinases do not require activation-loop phosphorylation to become active. Furthermore, for many kinases, good activation-loop antibodies do not exist. Thus, the only universal method for determining the activity of all kinases is to measure it directly. In vitro kinase assays have been traditionally performed using purified proteins or cell lysates incubated with radioactive isotopes and purified substrate proteins or peptides. These assays are reasonably straightforward, do not require much specialized equipment, and can be performed in a medium to high throughput with very little material161. Because many kinases are promiscuous, these assays often need to be conducted in the presence of inhibitors or following enrichment or purification. Accordingly, these assays likely cannot be performed at a comprehensive OMIC scale. Thus, most practical applications will use in vitro enzyme assays to study single proteins and very small networks or as a complementary assay in conjunction with other approaches. For example, Janes et al. used in vitro kinase assays in conjunction with immunoblotting and antibody arrays to study contextual dependencies in cytokine-induced apoptosis162. Recently, the Gygi laboratory has developed high- throughput methods to directly monitor kinase activity against synthetic 150

substrates in vitro using mass spectrometry163. This approach, called KAYAK

(kinase activity assay for kinome profiling), uses synthetic peptides featuring motifs for known kinases. Arrayed or pooled peptides are mixed with cell lysates and the acquisition of peptide phosphorylation is quantified by mass spectrometry following some enrichment steps. This approach was used to quantify the activities of 90 kinases across various treatment conditions in many different cell lines.

Of course, kinases are not the only enzymes and most enzymes can be purified and studied biochemically in the cell-free in vitro environment. For example, in vitro biochemical assays are a common method for studying the activity of GTPases, which do not leave a physical “mark” on their substrates.

Methods that have sufficient throughput and coverage to be considered “OMIC” approaches are rare, but some exist and certainly others are in development. For example, assays have been developed to facilitate high-throughput quantification of phosphoprotein-specific phosphatase activity164, as well as matrix metalloprotease activity165.

For most in vitro enzyme assays, the main limitation to throughput is the need to purify specific proteins. Activity-based probes and chemical sensors provide an excellent alternative when available. One example is sensors that exploit the chelation-enhanced fluorescence of sulfonamido-oxine (Sox) chromophores, which increase in fluorescence when bound to Mg2 +166. These chromophores can be conjugated to peptides containing phosphorylation motifs 151

for a kinase of interest and engineered in such a way that Mg2 + binding requires peptide phosphorylation. In addition to Sox chromophores, there exist other fluorophores and dyes whose fluorescence properties are sensitive to the microenvironment in which they are bound167,168. These reagents are particularly valuable for studying signals transmitted through G proteins, which transmit signals only by virtue of a small conformational change upon GTP binding.

Because their activation cycles in vivo are short and their interactions with substrates are relatively weak, this large and important class of signaling proteins has been traditionally difficult to study in the context of living cells. FRET-based activity probes were used by Klaus Hahn and colleagues to visualize the location and dynamics of RhoA activation in living cells169. These studies revealed distinct spatiotemporal patterns of RhoA activity in response to different extracellular cues, revealing a previously unappreciated level of richness in signaling diversity.

Inference of pathway activity through transcription or functional genomics

When collecting signaling data, it is not always possible (and sometimes not necessary) to view the signal at each node as it is transmitted through a network. In these cases, an endpoint measurement of network activation may be a sufficient substitute. Most stimuli-dependent signaling events result in a change in transcriptional states, whether to activate or repress certain genetic programs.

A transcriptional readout can be used to infer the signaling state of a cell using microarray, RNA sequencing, or microfluidic technologies that allow for high- 152

throughput quantitative PCR132. These approaches are quantitative, are high throughput, and can be performed on a comprehensive scale. The main limitation is the difficulty in relating a particular transcriptional state to the activation or inhibition of a discrete signaling protein. Some computational approaches exist to aid in this effort. The most widely used would be GSEA (gene set enrichment analysis)170. GSEA uses a rank-ordered list of differentially expressed genes to identify enrichments or depletions of a priori defined genetic “signatures” (i.e., cohorts of genes that are somehow related to each other or to a biological process). This approach is not formally restricted to the study of signal transduction, but its application to signaling research is possible by using signatures that report signaling states. Additionally, the Califano laboratory has invented several approaches for reverse engineering regulatory networks using transcriptional profiles171–173. Most of these algorithms focus on transcription factor activity rather than on the activity of upstream signaling proteins, but in many cases, the “master regulators” identified are proteins with defined roles in a signaling circuit. Inference algorithms that use proteome data rather than transcriptome data also exist. The Linding laboratory has developed several approaches that use mass spectrometry data to infer the connectivity of signaling networks174.

In addition to transcription- and mass-spectrometry-based methods, functional genomics can be used to infer signaling mechanisms. Several screening approaches are common in biological research, including the use of 153

specific drugs or small molecules, shRNA/RNAi, or the clustered regularly interspaced short palindromic repeat-based genome editing CRISPR175,176.

These perturbations can be paired with simple measures of pathway activation

(even at a single node) to reveal functional signaling networks. Screens can be performed in a high-throughput manner but hits typically need to be validated by an independent method due to high potential for off-target effects. With all inference-based approaches, observed phenotypes will be a mixture of direct, indirect, and compensatory effects. Deconvoluting these effects can be challenging particularly because signaling proteins typically have multiple roles in different contexts and are shared in many different processes, creating extensive crosstalk. Functional genomics are ideally suited for revealing a molecular “parts list”. However, these approaches do not typically result in sufficient understanding of pathway structure, which may be needed to make reliable predictions about how cells will behave following pathway stimulation or inhibition at various nodes. Thus, these approaches are best used in combination with more direct measures of signaling activity and in close partnership with computational approaches.

Experimental design for OMIC experiments: Basic principles and practical considerations.

Experimental design is an important issue in all areas of research, but some special considerations are needed in OMIC experiments due to their size 154

and complexity. Additionally, OMIC experiments are expensive, with a single experiment often costing thousands of dollars. These experiments therefore demand effective and efficient designs and error-free execution because they are not feasibly redone. Several features that are common to OMIC experiments, like the use of multiwell plates, can create strong batch effects177. Proper experimental design is necessary to minimize and remove, or at least identify, these problems. In the early days of “big data” science, some practitioners held the view that OMIC experiments could be held to a different, maybe lower, standard of experimental rigor (leaving out some controls, not performing biological or technical replicates, etc.). These decisions were often made based on cost feasibility and also with the idea that data generated were just “a screen” and would be validated carefully afterwards. If these ideas did not seem crazy at the time, they certainly should now. Large signaling data sets are not being used solely as unvalidated resources but are also used by computational biologists to aid in model parameterization or by pharmaceutical companies to motivate drug development programs178–180. Implementing the basic principles of sound experimental design and including all necessary controls are essential steps for minimizing bias and maximizing interpretability and reliability of OMIC experiments. Experimental design principles for proteomics experiments have been reviewed in detail elsewhere181–184 and Glass has presented a valuable discussion of the various types of controls needed to interpret biological experiments185. In this section, we briefly highlight some key concepts. 155

A major challenge of biological experimentation is separating signal from noise. In statistical terms, the experimental signals represent biological sources of “variance”, that is, changes in the measured signal caused by the experimental treatments, which the experimenter seeks to quantify. The noise of an experiment will likely arise from two principal sources: bias and random error.

Bias is the systematic deviation of the measured effect from the “true effect”.

These deviations can be attributed to both technical and biological sources and presumably the experimenter only cares about the latter. Random error refers to the unsystematic errors that enter the experiment due to unknown sources, which reduce the precision of the measurements and the ability to detect treatment effects. In OMIC experiments, because the operator cannot optimize each analyte, data point, or assay individually, the signal-to-noise requirements will be likely more stringent than for a low-throughput assay that can be more easily fine-tuned per experiment.

The three universal principles of sound experimental design are randomization, blocking, and replication. Randomization is a process that is applied in two contexts: first, experimental units are allocated to treatment groups with equal probability; second, experimental units are treated (and the resulting samples processed) in a random order. Randomization helps to reduce bias by minimizing the probability that experimental units with similar characteristics are assigned to the same treatment group. Randomization also helps satisfy the 156

necessary condition of independence, which is a common assumption in most statistical tests.

Blocking refers to the allocation of treatment groups into experimental units or “blocks”. Blocks are subgroups of experimental units that exhibit less variance within groups than between groups. An example of an experimental block in an OMIC experiment is each 96-well assay plate within an experiment that features multiple plates or each sequencing lane in a RNA-sequencing experiment. In multiwell plates, even the well location represents a potentially confounding block. Despite the experimenter's best efforts to treat them equally, each block is processed at a different time, which may introduce a systematic variance. These block or “batch” effects are prevalent in OMIC experiments177.

Defending against batch effects requires the use of blocking strategies in which replicates from the same treatment group are distributed across different plates

(and preferably in different locations on a plate) rather than loaded together on a single plate, in order to ensure that the batch effect does not confound the biological effect.

Replication is the repeated treatment and/or assaying of multiple experimental units. Replication is needed for precise and accurate results, for generalizing the results of the sample to the population from which it was drawn, and for generating statistically powerful tests. Two replicate types are commonly featured in OMIC experiments: biological and technical. Some ambiguity exists in the community for what constitutes a biological versus technical replicate186, but 157

in most instances, a biological replicate represents an experimental unit that was treated independently of the other units. Examples include each animal of a mouse experiment or each dish of cells in a cell-culture-based experiment.

Technical replicates consist of samples drawn from the same experimental unit that are measured independently. While technical replicates are useful for assessing the variability of the assay, for most applications, biological replicates should be emphasized because they incorporate both the biological and the technical sources of variance. Furthermore, replication should occur across blocks that reflect principal technical sources of variance in order to avoid bias and to maximize generalizability. For example, an experiment should be performed purposefully in separate batches, even if not logistically required to do so, in order to obtain a more accurate effect estimate. The use of inappropriate replication strategies is called “pseudoreplication” and is a common problem in many reported studies187. An added complexity common to many OMIC approaches is the use of nested data, which are data featuring multiple observations derived from a single sample, as in the use of highly multiplexed signaling measurements. These nested data invalidate the presumption of independence. Recent reports have discussed methods to address some of these issues through optimal experimental design and altered strategies for replication188,189.

Additional design issues include the decision of whether to pool samples, which is usually considered when costs must be kept low, and in sample size 158

calculation. These figures are required to estimate the number of replicates that the experiment must have to detect meaningful treatment effects with acceptable probability, if they in fact exist. Often is the case, however, that sample size is fixed due to logistics (e.g., cost). In such cases, strategies exist to pick optimal type I and type II error pairs190.

Conceiving OMIC experiments: Hypotheses Or discovery?

An additional consideration is whether the OMIC approach will be used to facilitate a hypothesis-driven experiment or strictly for discovery. Since the inception of systems biology, the validity of discovery science has been a tough sell to grant panels and thesis committees, who generally expect hypotheses to be tested191. However, specifying sensible hypotheses for OMIC experiments is sometimes challenging, which has led to revisiting the debate that science is only valid when using Popper's hypothetico-deductive logical framework191,192. In practice, scientific advances are made in many ways, including deduction

(hypothesis falsification), induction (data-driven discovery)192, the question/model cycle191, and strong inference (propose multiple competing hypotheses and design experiments to discriminate them)193,194. Therefore, discovery science should be considered valid, at least in principle. Even so, all forms of science, from those that are purely discovery driven to those that are hypothesis driven, do need to start with a cogent and well-conceived scientific question195, and the 159

lack of a good question, not the lack of a hypothesis, likely caused the limitations of many studies that were based in discovery.

Considerations in discovery science.

We have observed that discovery science has pitfalls that can render experiments uninformative, thus wasting time, money, and effort. First, effective discovery experiments require just as much, if not more, careful thought in conception and design as do those seeking to test a hypothesis. The designs of these experiments are too often given insufficient thought, the reasons for which include a lack of training among biomedical researchers in experimental design and analysis178, as well as the strong temptation to just “do the experiment” because “we'll find something”. The latter attitude likely arises due to the desire to publish outcompeting the desire to seek the “truth”.

Second, effective discovery experiments demand that the measurement technology measures most, if not all, of the system in an unbiased manner, such that the results can be reproduced and generalized. For sequencing genomes, this ideal is arguably met196, whereas it is almost certainly not met for the other

OMIC methods that we discussed above, all of which are widely used in systems research. Therefore, the data from such experiments may be idiosyncratic to the specific experiment and not safely generalizable.

Third, having a well justified hypothesis is important for determining the prior probability of a result being true, which in turn determines the false-positive 160

and false-negative probabilities of the result once obtained197. Accordingly,

OMIC-based measurements based on weak or non-existent hypotheses should be expected to be rife with false results, which in some cases may exacerbate the problem of irreproducibility of biomedical science198,199.

Given these pitfalls, it is worth highlighting that systems biology was not, and is not, meant to be a discovery science200. Although the term “systems biology” was coined at least 50 years ago, the modern era of systems biology is typically said to begin with the Human Genome Project201. The Human Genome

Project and other endeavors that purely practice discovery science are commonly grouped with systems biology due to the use of similar OMIC technologies. While these fields are related, they are not the same. A relationship exists because discovery science aims to define the elements, whereas systems biology aims to define the functional relationships among these elements.

However, in another way, discovery science and systems biology are very different. Discovery science and hypothesis-driven science philosophically contrast each other, but the same discordance does not exist between hypothesis-driven science and systems biology. Since its initial definition, it was proposed that “the integration of these two approaches, discovery and hypothesis-driven science, is one of the mandates of systems biology”201. Thus, a systems-level study is meant to use OMIC-scale data to perform predictive, hypothesis-driven science. Generating OMIC-scale hypotheses is challenging, 161

and systems biologists continue to struggle with casting their work in a logically satisfying intellectual framework.

Strong hypotheses for OMIC-scale systems biology.

The following are the three basic types of hypotheses that apply to OMIC experiments in signaling research:

(1) The “discovery” hypothesis that at least one of the analytes will be affected; that is, the levels of a protein or its post-translationally modified forms will increase or decrease in response to at least one of the experimental perturbations;

(2) A systems-level hypothesis pertaining to a property of the system or to a mathematical or computational model of the system. For example, the hypothesis can focus on the network topology (e.g., that it is scale free or that some importance can be inferred from the structure), the estimation of parameter values, and/or model discrimination, that is, distinguishing whether one model explains the data better than alternative models; and

(3) Specific (reductionist) hypotheses about one or more of the measured analytes, for example, that c-Jun N-terminal kinase will be phosphorylated in response to TNF-α treatment. 162

The first type of hypothesis is the most prevalent in the systems biology literature, and the third type is the least prevalent. This ordering is likely due to the burden for finding evidence a priori to support the hypothesis. In the case of the first two types of hypotheses, relatively little evidence is needed for justification. In the case of the third type, one would have to justify distinct hypotheses for each analyte, which in order to do properly would require systematic reviewing of the relevant literature and databases for each analyte in order to determine how it is likely to respond to perturbations in the experiment.

The advantage of doing this initial work prior to conducting the experiment is that specific elements of the design might be better resolved. Furthermore, it may help to alleviate the bias that could be introduced by doing the same type of search after the experiment to rationalize the “hits”.

Each of these hypotheses is acceptable as long as they are exploited for specifying the experimental design and analysis, which is ultimately the raison d'être of hypotheses. In crafting a strong hypothesis, no matter the type, the variables must be operationalized, assumptions must be explicitly stated or considered, and thought must be given to the extent to which the conclusions can be generalized. Articulating hypotheses that contain these details facilitate decision making regarding the design of the experiment, the statistical analyses, and the appropriate interpretation of the data. In practice, bridging the gap between OMIC-driven discovery science and question- or hypothesis-driven 163

science often requires some form of computational or mathematical modeling. As with the experimental approaches discussed above, these methods each have distinct strengths, limitations, and design principles.

Approaches to Modeling Signal Transduction Processes:

For most signaling processes, a substantial amount of information has been accumulated already about the proteins involved (the parts list or “nodes”) 164

and the functional relationships or connections between these proteins (the

“edges”). Thus, it is worth considering whether adding another node or another edge will improve our understanding of the system. One such example is shown in Fig. A.3, a wiring diagram of the integrated signaling pathways that control cell death, including DNA damage, cytokine, growth factor, and other stress- response signals. The edges shown are all high confidence, meaning they

“probably” all exist. Even so, these “probably” do not exist in all contexts, or at all time points, and certainly not all with the same levels of importance. The dense interconnectivity and temporal/contextual dependence impairs our intuitive understanding of how the system operates. This situation worsens with each new insight, whether it is new nodes or new edges. Thus, once OMIC data are collected for a signaling process of interest, additional tools are typically necessary to facilitate the transition from “data” to “understanding”. Highly quantitative data of the type discussed above can be modeled using various approaches (Fig. A.4). Different approaches might be preferred in different scenarios. Considerations include the nature of the data available (amount, specificity, level of detail, depth of information within a signaling “pathway”, breadth of information across parallel pathways), the size of the process being modeled (i.e., the number of nodes and edges), and the question being asked.

Other more authoritative and comprehensive reviews exist for each of the methods that we will discuss82,202–204. Here, we survey common approaches that

165

166

can be used, with an emphasis on statistical approaches that may be optimal for studying large, complex, and imperfectly understood networks.

Mechanistic and semi-mechanistic approaches

The most commonly used framework for modeling signal transduction processes is kinetic modeling, which are typically expressed as ordinary differential equation (ODE) or partial differential equation. ODEs are attractive because they are a natural framework for dynamical systems such as signaling202. Each ODE represents the rate of change for one molecular component, which is a function of the rates of each biochemical transformation that influences the level of that component (i.e., enzyme-catalyzed reactions, protein–protein interactions, and transport between compartments). The rates of the biochemical transformations are functions of the concentrations of the reactants (the molecular components) and intrinsic kinetic parameters. The concentrations of the molecular components change over time but the kinetic parameter values are typically held constant. Because the molecular components are linked by their overlapping participation in a common set of biochemical reactions, the equations for the individual molecular components are organized into a system of interdependent ODEs. In most applications to date, the models were calibrated to experimental data primarily consisting of time series of protein levels, measured using immunoblotting, in response to experimental perturbations such as stimulation by receptor ligands (e.g., hormones or 167

cytokines). These models can be used to simulate the behaviors of a system over time and are particularly useful for testing conditions that would otherwise be infeasible to test (due to lack of reagents, cost, etc.). Kinetic modeling can be used to generate testable predictions, rule out possible mechanistic explanations or competing hypotheses, or identify the system parameters to which the model behavior is particularly sensitive or insensitive.

The dynamic nature of signal transduction makes kinetic modeling the ideal framework. Indeed, some flexibility exists such that the approaches can be used to model spatial information205, stochastic or deterministic biological processes, and even responses that occur on very different timescales (e.g., cell division, which occurs on the timescale of hours, in response to hormone signaling, facets of which occur on shorter timescales). However, these approaches are most effective for signaling systems that are relatively well understood and for which the rate constants are known or can be measured easily. These two factors typically represent major obstacles to implementation because our understanding of signaling network biochemistry is far from complete. In particular, our understanding of crosstalk is poor for most signaling processes, likely due to our long history of studying signaling using reductive approaches. Furthermore, even for systems in which crosstalk is known to exist, these connections are often context specific206, thus causing an added complication that is only recently being explicitly explored in a systematic fashion207. These problems scale with the size of the modeled system, such that, 168

in practical terms, kinetic modeling is typically best suited for the study of relatively small signaling circuits, composed of tens of proteins, not hundreds or thousands.

A reasonable compromise for many signaling systems is the use of logical models. These models lack the highly mechanistic insight generated from a kinetic model but do not require any rate parameters to be measured or estimated. As a result, logic models can be used generally to study much larger signaling networks, even in cases where connectivity is contextual or imperfectly understood, as is often the case in signaling biology. In a logic model, the signaling pathway is reconstructed as a series of logical operators—“AND,” “OR,”

“NOR”—that describe the relationship between the activities of protein intermediates. This approach happens to be similar to the way signaling biologists have been intuitively drawing wiring diagrams since the advent of signal transduction as a research topic, both conceptually and with respect to model's output. In the simplest forms, such as those that use Boolean “on/off”

(0,1) rules, logic models also suffer from some of the same shortcomings as static signal wiring diagrams. For example, the influences between proteins are equal in magnitude and binary, which would typically not be the case in a real biological system. Furthermore, these models do not traditionally provide as output temporal dynamics of signaling. In more sophisticated approaches, both of these shortcomings are addressed, and there now exist logic modeling frameworks that encode graded influences between pathway nodes and the time 169

dependencies of signaling activity states208,209. While these methods are still nascent, we predict that they will serve as important tools in the signaling biologist's toolkit.

Data-driven modeling approaches

Both kinetic and logic modeling methods are valuable and intuitive platforms; however, both methods require a reasonably mature understanding of the circuitry of a signaling process, the so-called “prior knowledge network”.

Because our existing knowledge of network circuitry is incomplete, particularly in the area of signaling crosstalk, research in signal transduction can still benefit greatly from approaches that require no prior knowledge and instead are completely empirical and data driven. Integration of empirical insights with prior knowledge can often form a strong foundation to infer function or regulatory relationships. Data-driven approaches generally produce more abstract insights than mechanistic modeling but are nevertheless well suited to enhance understanding about the basic topology (i.e., wiring) of a signaling network, the influences and connectivity between network components, and the likely routes of information flow through a network. These approaches share the common underlying assumption that the relationships between signaling proteins (i.e., their influences and interactions) manifest as statistically observable trends within the data210. For example, the activities of signaling proteins that control a cellular response such as proliferation or apoptotic cell death would be predicted to 170

covary with these responses when measured across many diverse conditions or system perturbations.

Many data-driven modeling methods exist and are becoming widely used in signaling research. For example, graphical network maps (protein–protein interaction networks, gene regulatory networks) have long been used to define the abstract structure of a system211. The edges that connect nodes in a graph can represent essentially any relationship and need not be mechanistic or physical in nature. Thus, networks can be a useful method for integrating disparate types of data212,213, as is often necessary in signaling research. When considered in the context of ontology or known functional groups, these networks can be used to make predictions about function or co-regulation.

Clustering approaches (hierarchical, K means, K-nearest neighbors, etc.) are commonly used to identify subsets of the data that appear to share a relationship. Co-clustering might imply that the entities are themselves parts of a functional unit, like a multi-subunit complex, or that they are all regulated by the same upstream protein, or that they are all simply parts of a shared regulatory network. When these approaches are integrated with variables such as time and experimental perturbations, more mechanistic insight can be often inferred about why the clusters emerged and/or about the function of certain components within a cluster214. In this way, clustering can be a powerful hypothesis-generating tool, but as is true with most data-driven approaches, these hypotheses should be validated using other techniques, both computational and experimental215. 171

Finally, several multivariate regression and dimensionality reduction methods exist, including principal component analysis (PCA), partial-least- squares regression (PLSR), stepwise regression, least absolute shrinkage and selection operator and elastic net regression, linear discriminant analysis, factor analysis, and so on. These are becoming common tools in biological research because of their power for describing signaling networks and their ease of use82.

These methods take advantage of the fact that many signaling nodes carry redundant information or are highly covariant in response to a given stimulus.

Dimensionality reduction approaches such as PCA can be used to simplify the data without losing much information. These techniques dramatically improve our ability to visualize complex networks and obtain a more intuitive understanding of how signals flow through the network. Other approaches, like PLSR, which also uses eigenvector decomposition, offer the ability to explicitly identify relationships between a set of input signals and output cellular responses.

One important caveat in the use of data-driven modeling is the selection of data to be modeled. Data-driven approaches are unbiased (or at least, they are not biased by prior knowledge); thus, they offer an opportunity to identify novel and even counterintuitive or paradoxical information, provided that those signals were chosen to be measured. This advantage is counterbalanced by the limitation that spurious relationships may be identified. This limitation is not applicable if the data generated are truly comprehensive, but such data are rare in signaling research, particularly for methods that rely on antibodies as 172

described above. Thus, some thought or strategy should be applied to select targets for analysis that are both likely to be involved but not already known. One effective strategy is to combine multiple OMIC approaches, some that allow comprehensive analysis, to identify and prioritize certain nodes for study and others that allow more detailed study of signaling activation states. For example, global transcriptomic analysis could be used to identify processes that are changed in response to stimuli or activities or subnetworks that appear to be enriched in the transcriptional signature. Based on pathway informatics, the activity of some or all proteins involved in those pathways could be directly measured in response to stimuli over time and these data could be modeled to identify which signaling behaviors were most predictive of a cell fate of interest.

This approach was used in our recent work, which studied mechanisms of drug synergy in triple-negative breast cancer54. Another point of caution with data- driven modeling approaches is that variance scaling is often used to make comparable disparate measurements. This can lead to artifactual “stretching” of invariant data (i.e., data with no changes across all conditions), such that the model will essentially be fitting or overemphasizing experimental noise.

There exist many excellent resources that discuss the mathematics that underlie each of these modeling formalisms and for describing how each can be used57,161,202–204,216. In the remainder of this review, we discuss considerations for applying these modeling methods on the types of OMIC data described and how their integration advances the study of signaling. 173

Considerations for modeling signaling networks.

One overarching goal of systems biology is to build predictive models of cell signaling networks. Achieving this goal requires special considerations with respect to what to measure and when or how to measure it, in addition to the basic issues described above for experimental design. In general, the quality of existing models of cell signaling networks is deficient. These models feature incomplete and uncertain topologies and large uncertainties in the parameter values217. Importantly, these models are often still useful and, in many cases, have made reliable predictions that led to biological insight.

Parameter estimation

As we stated above, most mathematical models of signaling are kinetic models in which signaling is a function of the concentrations of the proteins and their intrinsic rates of production and decay. For most systems, the concentrations of most molecular components are measurable, but typically, the kinetic parameter values are not and must be inferred from the data using model optimization methods (see Sun et al.204 for a comprehensive review of optimization methods in systems biology). Any OMIC techniques applied for the purposes of calibrating or validating models of signal transduction pathways will have to follow the same principles.

Two related but distinct concepts in parameter estimation are parameter identifiability and parameter uncertainty. The former refers to the theoretical 174

ability to learn a model parameter's true value given infinite data. Mathematically, an identifiable parameter will have finite confidence intervals218. Conversely, non- identifiability implies that the true values of one or more of a model's parameters can never be known. Two types of parameter non-identifiability exist: structural and practical. Structural non-identifiability refers to redundant off-setting parameters in the model whereas practical non-identifiability refers to the general lack of data given the number of parameters219. Structural non-identifiability can arise in several ways such as the model having excessive parameters, inadequate experimental design (choice of perturbations and sampling strategies), or the lack of measured observables. Some of these issues can be resolved through model reduction techniques, but the lack of measurements can only be resolved by collecting qualitatively different data, either by applying different perturbations or by measuring additional observables. Practical non- identifiability is alleviated by collecting data of sufficient quantity and quality.

Specifically, these data could be qualitatively different or could be collected from additional or different sampling times, from additional replicates, or using more reliable assays with less variance. Assuming that the parameters are identifiable, uncertainties in their estimated values will still exist. These uncertainties are typically expressed as confidence intervals. Parameter uncertainties lead to uncertainties in the predictive power of the model, although it is possible for a model's predictions to be well constrained even though the parameters are 175

not220. As with the solutions to non-identifiability, parameter uncertainties can be reduced with data of higher quality and quantity.

From the perspective of parameter identification, OMIC technologies will likely play an important role in creating better models of signal transduction.

OMICs can help address issues of structural non-identifiability by adding observables. OMICs can help to constrain model topologies by identifying protein–protein interactions in the system of interest. The role of OMICs in addressing practical non-identifiability and parameter uncertainties is less clear because these entities are a strong function of the experimental design and the quality rather than quantity of the data. In this regard, one solution may be to use many perturbations and sampling times, which limits the applicability of certain

OMIC techniques (such as mass spectrometry) due to the prohibitive cost of their use in measuring hundreds or thousands of samples. In such cases, many of the antibody-based “medium-throughput” proteomic techniques described above may represent a good compromise between the numbers of perturbations and observables. Accordingly, multiplexed-bead-based immunoassays, RPPAs, and micro-Western arrays—each of which can be applied to hundreds of samples— have been used to supply data for computational models of signaling networks54,119,203,208,221.

176

Choice of perturbations, observables, and sampling times

The power and efficiency of experimental designs for calibrating and validating models of cell signaling depend on the choices of what to measure and how to measure them, namely, the experimental perturbations (and perturbation sources), the observables, and the sampling times182. Of these, particularly important are the perturbations used to activate or deactivate signaling pathways.

Well-chosen perturbations can reveal spurious relationships and ultimately allow the parameters associated with their logical and/or biochemical relationships to be estimated and refined. Furthermore, most topology estimation methods leverage the correlations between the changes in node levels in response to perturbation. Accurate estimation of these correlations depends on the sufficiency and quality of the perturbations. The combinatorial application of perturbations is also critical for detecting interactions (i.e., synergism or antagonism), which can serve as evidence for connectivity222. We believe that most early systems biology efforts used insufficient numbers of perturbations and many did not apply perturbations in a combinatorial manner. Common perturbations include naturally occurring variation (e.g., different cell types); receptor ligands such as hormones or cytokines; genetic perturbations such as transient, stable, or inducible overexpression or knockdown of protein; and small molecules, which pharmacologically activate or inhibit protein activities. The concentrations or doses of these variables represent experimental design 177

choices, which are often useful for contributing variation in the observables, thus enabling effective parameter estimation.

Observables (i.e., what is measured) are typically the levels of total proteins and/or their post-translationally modified forms (e.g., phosphorylated, acetylated, sumoylated, etc.), as well as levels of coding and non-coding RNAs that are regulated by signaling pathways. Less common observables include the levels of protein complexes (e.g., measured by immunoprecipitation) and localization of proteins (e.g., protein levels in nuclear extracts). Observables are usually chosen on the basis of logistics, such as the availability and/or cost of detection reagents, but should be based more commonly on the biology (i.e., how is the signal encoded). Sampling times are chosen according to the timescale on which the signaling system operates, as well as its duration of action.

Optimizing experimental design for modeling.

All logistically feasible combinations of the abovementioned variables are possible, such that it is exceedingly difficult to choose the ones to use in a rational manner. In the past, these choices were made ad hoc, based on tradition

(e.g., three replicates per treatment condition) or feasibility (e.g., availability of antibodies). Today, computational methods are available to optimize experimental designs devoted to generating data for modeling signal transduction networks182. Work on optimal experimental designs for signaling 178

models is still in its infancy. Early work used sensitivity analysis to prioritize experiments for investigating the dynamics of TNF-α/NF-κB signaling223. Casey et al. demonstrated the use of optimal design for constraining the variability of predictions in a model of EGFR signaling224, and more recently, Tidor and colleagues have demonstrated that optimal selection of experiments can lead to accurate parameter estimates in models of EGF and NGF signaling in PC12 cells225,226. It is currently unclear as to whether computational optimization of experimental design is feasible for OMIC-sized models. In principle, these methods should be scalable but in practice may be limited by computational power.

While considerable important progress has been made, these methods of optimization have some important limitations. First, the methods are non-trivial to apply. The calculations must be customized to the experimental question, which precludes automation and requires specialized expertise182. Second, the topology of the model is typically assumed to be known225, which is almost never the case in practice. Third, uncertainties in the input functions representing the perturbations must be accounted for because they can propagate into uncertainties in the parameter values227. These uncertainties come in many forms. For example, the degree to which a siRNA can knock down protein levels varies widely, and this information is generally unknown and not included in the optimization. Also, it is well established that the potencies of receptor ligands such as hormones and cytokines vary over time but many signaling models do 179

not incorporate this knowledge. In early models of TGF-β signaling, for example, the level of TGF-β was assumed to be constant228, whereas it was later found that TGF-β depletion is a major factor in downstream signaling kinetics229.

Finally, experimental design optimization algorithms assume that only random error is present in the data despite bias being a major source of variance. Some biases, such as batch effects, can be eliminated by combining proper experimental design (blocking and randomization) with data normalization techniques.

Data pre-processing, normalization, and statistical analysis

All proteomic methods provide quantitative readouts that are assumed to be proportional to the abundance or activity of the signaling proteins of interest.

In the case of mass spectrometry, the readout can be spectral counts or spectral intensities. For antibody-based proteomics, the readouts are typically fluorescence (e.g., the median fluorescence intensity of each bead type in a multiplexed-bead-based immunoassay), chemiluminescence, or optical density

(e.g., from plate-based colorimetric assays such as ELISA). To provide interpretable data, these raw readouts must be translated into quantities that reflect the absolute or relative abundances of total or modified proteins, which are then further analyzed to infer the biology of the system.

Translating raw data into interpretable quantities typically involves transformation, normalization, and scaling, while the downstream analyses 180

involve statistical or mathematical modeling. Log transformation is a common means to render the data more interpretable and to stabilize the variance of the signal as a function of the signal magnitude. The data are then commonly normalized to facilitate fair comparisons across samples or batches.

Normalization is used to remove unwanted systematic variance from the data, typically that which is associated with technical sources of variance.

Numerous normalization methods have been developed for OMIC data.

Many of these are applicable to any type of high-throughput data, but some are idiosyncratic to a specific proteomic assay. Normalization methods can be classified into three basic types: (1) biological methods in which one or more separate measurements are made in order to interpret the measurement of interest, such as comparing the signal measure to that of an invariant

“housekeeping” protein; (2) unsupervised methods in which models are used to specify putative sources of variance that are not related to the variables describing the study; and (3) supervised methods, which utilize study-specific variables including batch and treatment effects230. Different normalization techniques have been evaluated for mass spectrometry231, multiplexed-bead- based flow cytometric immunoassays221, RPPAs232,233, two-dimensional gel electrophoresis234, antibody microarrays, immunoblotting235, and flow cytometry236,237. Data normalization methods continue to be developed such that it would be premature to declare that one approach is best, but we emphasize 181

that normalization is a critical step in the analysis of OMIC data that should be carefully considered.

In addition to normalization, analysis of large signaling data sets also often requires scaling to compare disparate analytical measurements. The need for scaling is particularly critical for antibody-based proteomics because the assay signal intensity will be a function of both the protein expression level and antibody-epitope binding affinity (i.e., how “good” is the antibody), and the latter must be accounted for in the analysis. One common approach is to unit variance scale the data238, but this procedure can substantially alter the data. Specifically, the procedure maintains of the general “shape” of the data but disregards its magnitude. Situations exist in which this approach is beneficial, but for others, it is counterproductive. Variance rescaling can lead to misinterpretations if the source data were of low variance to begin with. Proper quantification, transformation, normalization, and scaling are all required to gain reliable insight using statistical or data-driven models and the “right” ways to do each of these depends greatly on the nature of the data collected. Another more complex issue is the integration of very heterogeneous types of data239. It will likely be necessary to integrate genomic, proteomic, or clinical data of different types, some of which may be categorical, continuous, or discontinuous and may have different sizes, formats, and parameters. Some efforts are underway to improve standardization and improve the ability to compare disparate data across different assay platforms240. This issue is complex and represents a 182

contemporary challenge for systems biology to resolve. In the meantime, efforts involving the integration of OMIC data sets will benefit from carefully considering experimental design issues and planning devoted to merging experimental and computational approaches in order to maximize the potential for new insights.

Conclusion:

The philosophy of OMIC experiments: Will they get us closer to the truth?

To some, OMIC-based science represents an important way forward in generating comprehensive and unbiased knowledge, ultimately promising to reveal predictive cause-and-effect relationships about nature allowing us to diagnose, manage, treat, and prevent disease241. To others, the OMIC revolution has been a disappointment. Sydney Brenner, for example, has famously called systems biology a “low-input, high-throughput, no-output” science242. In light of these conflicting views and the roughly two decades of OMIC history on which to reflect, it is worth considering how far we have come and where we need to go moving forward.

As we stated in the introduction, there was indeed a nascent form of systems biology—a version 1.0—in which the apparent idea was to do science as it had always been performed, but to simply do it on a larger scale: rather than make a knockout, make all knockouts, rank order these, and take the best one to study. This approach quickly morphed into a realization that these mountains of data could be “hypothesis generating” rather than “hypothesis driven”. It is 183

perhaps a valid criticism that this approach produced relatively few successes199.

There are good reasons for this. For one, OMIC technologies, in and of themselves, did not address many of the pitfalls of biomedical science and in fact may have exacerbated some of these issues199. Also, early OMIC experiments were limited by cost feasibility, which occasionally resulted in “atypical” experimental designs missing some controls and/or replication. Furthermore, researchers had often fallen prey to the mistaken assumption that measuring more variables might overcome the limitations of these feasibility-limited experimental designs. Instead, OMIC experiments require just as much, if not more, care in experimental conception, design, and analysis than do traditional assays. As we suggested throughout this review, planning of experimental approaches and computational strategies should be integrated, paired with an understanding of the benefits and weaknesses of the approaches, and considered before any data are collected to ensure that the right types of data are collected in the right formats. An explosion of OMIC research has occurred in all fields of biological study, including in signal transduction. This explosion has been driven in part by successes in the OMIC study of genomics and transcriptomics. As we highlighted herein, unique challenges exist for applying

OMIC approaches to understanding signal transduction. Importantly, however, common challenges exist that all OMIC researchers face in producing reliable, meaningful, and practical insights. We expect that many of these challenges will be resolved by virtue of an expanding new generation of scientists who are fluent 184

in both experimental and computational matters, and indeed, the recent additions to the systems literature show demonstrable progress in improving our understanding and treatment of human disease243. OMIC approaches or computational tools are not a replacement for carefully thought out questions or hypotheses and well-designed experiments. The idea that one should “measure more for the sake of measuring more” or “model more for the sake of modeling more” has not worked well in the past and likely will not work well in the future. In many (well lit) corners of biology, a frustration has been expressed with the relative lack of benefit brought by resource-intensive systems and OMIC approaches. A clear task exists for those of us in the version 2.0 of systems biology: prove its value. We will be judged moving forward by our ability to make insightful predictions, particularly ones that are either counterintuitive or could not have been made in the absence of OMIC or systems tools. To meet this mandate, we will have to be better at developing appropriate OMIC-scale questions or hypotheses. Each one of us must have more dexterity in the use and integration of various OMIC-scale experimental approaches. We will have to be more fluent in the proper use of computational tools. And importantly, we will have to be more careful in marrying these experimental and computational approaches appropriately.

185

Acknowledgements:

M.J.L. and D.C.C. would like to thank the personnel and leadership of the

Massachusetts Institute of Technology Center for Cell Decision Processes

(GM68762) and the Massachusetts Institute of Technology Integrative Cancer

Biology Program (CA112967) for funding and intellectual support. In particular, these groups provided a forum for critical discussion and open sharing of ideas that was instrumental in advancing the ideas presented in this review.

Additionally, B.D.L. is funded by a training grant from the National Cancer

Institute (T32 CA130807). D.C.C. receives funding support from the Natural

Sciences and Engineering Research Council of Canada. M.J.L. receives funding support from the Richard and Susan Smith Family Foundation and from the

Breast Cancer Alliance.

186

REFERENCES

1. Cotlar AM, Dubose JJ, Rose DM. History of surgery for breast cancer: radical to the sublime. Curr Surg. 2003;60(3):329-337. doi:10.1016/S0149- 7944(02)00777-8 2. Matray-Devoti J. Cancer Drugs. New York: Chelsea House Publishers; 2007. 3. James TGI. An Introduction to the History of General Surgery by R. H. Meade, Philadelphia, London and Toronto, W. B. Saunders, 1970, pp. xi, 403, illus., £7.20. Med Hist. 1971;15(03):313. doi:10.1017/S0025727300016811 4. Fisher B. Biological Research in the Evolution of Cancer Surgery: A Personal Perspective. Cancer Res. 2008;68(24):10007 LP-10020. http://cancerres.aacrjournals.org/content/68/24/10007.abstract. 5. Fisher B, Fisher ER. The organ distribution of disseminated 51 Cr-labeled tumor cells. Cancer Res. 1967;27(2):412-420. http://www.ncbi.nlm.nih.gov/pubmed/6018568. 6. Fisher B, Slack NH. Number of lymph nodes examined and the prognosis of breast carcinoma. Surg Gynecol Obstet. 1970;131(1):79-88. http://www.ncbi.nlm.nih.gov/pubmed/5419966. 7. Fisher B, Hanlon JT, Linta J, Saffer E. Tumor specificity, serum inhibition, and influence of regional lymph nodes on cytotoxic macrophages from cultured bone marrow. Cancer Res. 1977;37(10):3628-3633. http://www.ncbi.nlm.nih.gov/pubmed/908011. 8. Fisher B, Jeong J-H, Anderson S, Bryant J, Fisher ER, Wolmark N. Twenty-Five-Year Follow-up of a Randomized Trial Comparing Radical Mastectomy, Total Mastectomy, and Total Mastectomy Followed by Irradiation. N Engl J Med. 2002;347(8):567-575. doi:10.1056/NEJMoa020128 9. Connell PP, Hellman S. Advances in Radiotherapy and Implications for the Next Century: A Historical Perspective. Cancer Res. 2009;69(2):383 LP- 392. http://cancerres.aacrjournals.org/content/69/2/383.abstract. 10. Curie E, Sheean V. Madame Curie : A Biography. Garden City, N.Y. : Doubleday, Doran & Co; 1937. http://umassmed.worldcat.org/title/madame-curie-a- biography/oclc/609355556. 11. Regaud C. Sterilization rontgenienne totale et definitive, sans radiodermite, 187

des testicules du Belier adulte: conditions de sa realisation. Compt Rend Soc Biol. 1911;70:202-203. 12. Regaud C. Sur les principes radiophysiologiques de la radiotherapie des cancers. Acta Radiol. 1930;11:456-486. 13. Regaud C. Actions des rayons X sur les diverses generations de la lignee spermatique: extreme sensibilite des spermatogonies a ces rayons. Compt Rend Soc Biol. 1906;61:163-165. 14. Bergonie’, J; Tribandeau L. “Interpretation de quelques resultats de la radiotherapie et essai de fixation d’une technique rationelle” note presented by Dr d’Arsonval. Compt Rend Acad Sci Paris. 1906;142:983- 985. 15. Muller HJ. Artificial transmutation of the gene. Science (80- ). 1927;66(1699):84-87. doi:10.1126/science.66.1699.84 16. Barnett GC, West CML, Dunning AM, et al. Normal tissue reactions to radiotherapy: towards tailoring treatment dose by genotype. Nat Rev Cancer. 2009;9(2):134-142. doi:10.1038/nrc2587 17. GILMAN A. Therapeutic applications of chemical warfare agents. Fed Proc. 1946;5:285-292. http://www.ncbi.nlm.nih.gov/pubmed/20990888. 18. Gilman A, Philips FS. The Biological Actions and Therapeutic Applications of the B-Chloroethyl Amines and Sulfides. Science (80- ). 1946;103(2675):409-436. doi:10.1126/science.103.2675.409 19. Goodman LS, Wintrobe MM, Dameshek W, Goodman MJ, Gilman A, Mclennan MT. Nitrogen mustard therapy: Use of Methyl-Bis(Beta- Chloroethyl)amine Hydrochloride and Tris(Beta-Chloroethyl)amine Hydrochloride for Hodgkin’s Disease, Lymphosarcoma, Leukemia and Certain Allied and Miscellaneous Disorders. J Am Med Assoc. 1946;132(3):126-132. doi:10.1001/jama.1946.02870380008004 20. Polavarapu A, Stillabower JA, Stubblefield SGW, Taylor WM, Baik M-H. The Mechanism of Guanine Alkylation by Nitrogen Mustards: A Computational Study. J Org Chem. 2012;77(14):5914-5921. doi:10.1021/jo300351g 21. Jackson SP, Bartek J. The DNA-damage response in human biology and disease. Nature. 2009;461(7267):1071-1078. http://dx.doi.org/10.1038/nature08467. 22. Ciccia A, Elledge SJ. The DNA Damage Response: Making It Safe to Play with Knives. Mol Cell. 2010;40(2):179-204. doi:https://doi.org/10.1016/j.molcel.2010.09.019 188

23. Goldstein M, Kastan MB. The DNA Damage Response: Implications for Tumor Responses to Radiation and Chemotherapy. Annu Rev Med. 2015;66(1):129-143. doi:10.1146/annurev-med-081313-121208 24. Hosoya N, Miyagawa K. Targeting DNA damage response in cancer therapy. Cancer Sci. 2014;105(4):370-388. doi:10.1111/cas.12366 25. Chonghaile TN, Sarosiek KA, Vo T-T, et al. Pretreatment Mitochondrial Priming Correlates with Clinical Response to Cytotoxic Chemotherapy. Science (80- ). 2011;334(6059):1129-1133. doi:10.1126/science.1206727 26. Bertino JR. The Mechanism of Action of the Folate Antagonists in Man. Cancer Res. 1963;23(8 Part 1):1286 LP-1306. http://cancerres.aacrjournals.org/content/23/8_Part_1/1286.abstract. 27. Hori T, Ayusawa D, Shimizu K, Koyama H, Seno T. Chromosome Breakage Induced by Thymidylate Stress in Thymidylate Synthase- negative Mutants of Mouse FM3A Cells. Cancer Res. 1984;44(2):703 LP- 709. http://cancerres.aacrjournals.org/content/44/2/703.abstract. 28. Houghton JA, Harwood FG, Tillman DM. Thymineless death in colon carcinoma cells is mediated via fas signaling. Proc Natl Acad Sci U S A. 1997;94(15):8144-8149. http://www.ncbi.nlm.nih.gov/pubmed/9223329. 29. García-González AP, Ritter AD, Shrestha S, Andersen EC, Yilmaz LS, Walhout AJM. Bacterial Metabolism Affects the C. elegans Response to Cancer Chemotherapeutics. Cell. 2017;169(3):431-441.e8. doi:10.1016/j.cell.2017.03.046 30. Wani MC, Taylor HL, Wall ME, Coggon P, McPhail AT. Plant antitumor agents. VI. The isolation and structure of taxol, a novel antileukemic and antitumor agent from Taxus brevifolia. J Am Chem Soc. 1971;93(9):2325- 2327. http://www.ncbi.nlm.nih.gov/pubmed/5553076. 31. Rowinsky EK, Donehower RC, Jones RJ, Tucker RW. Microtubule Changes and Cytotoxicity in Leukemic Cell Lines Treated with Taxol. Cancer Res. 1988;48(14):4093 LP-4100. http://cancerres.aacrjournals.org/content/48/14/4093.abstract. 32. Jordan MA, Wilson L. Microtubules as a target for anticancer drugs. Nat Rev Cancer. 2004;4(4):253-265. doi:10.1038/nrc1317 33. Bharadwaj R, Yu H. The spindle checkpoint, aneuploidy, and cancer. Oncogene. 2004;23(11):2016-2027. doi:10.1038/sj.onc.1207374 34. Ashkenazi A. Targeting the extrinsic apoptotic pathway in cancer: lessons learned and future directions. J Clin Invest. 2015;125(2):487-489. doi:10.1172/JCI80420 189

35. Delbridge ARD, Strasser A. The BCL-2 protein family, BH3-mimetics and cancer therapy. Cell Death Differ. 2015;22(7):1071-1080. doi:10.1038/cdd.2015.50 36. Potter DS, Letai A. To Prime, or Not to Prime: That Is the Question. Cold Spring Harb Symp Quant Biol. 2016;81:131-140. doi:10.1101/sqb.2016.81.030841 37. Bhola PD, Letai A. Mitochondria-Judges and Executioners of Cell Death Sentences. Mol Cell. 2016;61(5):695-704. doi:10.1016/j.molcel.2016.02.019 38. Del Gaizo Moore V, Letai A. BH3 profiling – Measuring integrated function of the mitochondrial apoptotic pathway to predict cell fate decisions. Cancer Lett. 2013;332(2):202-205. doi:10.1016/j.canlet.2011.12.021 39. Marusyk A, Tabassum DP, Janiszewska M, et al. Spatial Proximity to Fibroblasts Impacts Molecular Features and Therapeutic Sensitivity of Breast Cancer Cells Influencing Clinical Outcomes. Cancer Res. 2016;76(22):6495-6506. doi:10.1158/0008-5472.CAN-16-1457 40. Ferlay J, Steliarova-Foucher E, Lortet-Tieulent J, et al. Cancer incidence and mortality patterns in Europe: Estimates for 40 countries in 2012. Eur J Cancer. 2013;49(6):1374-1403. doi:10.1016/j.ejca.2012.12.027 41. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2017. CA Cancer J Clin. 2017;67(1):7-30. doi:10.3322/caac.21387 42. Neophytou C, Boutsikos P, Papageorgis P. Molecular Mechanisms and Emerging Therapeutic Targets of Triple-Negative Breast Cancer Metastasis. Front Oncol. 2018;8:31. doi:10.3389/fonc.2018.00031 43. Shiga K, Hara M, Nagasaki T, Sato T, Takahashi H, Takeyama H. Cancer- Associated Fibroblasts: Their Characteristics and Their Roles in Tumor Growth. Cancers (Basel). 2015;7(4):2443-2458. doi:10.3390/cancers7040902 44. Wen S, Niu Y, Yeh S, Chang C. BM-MSCs promote prostate cancer progression via the conversion of normal fibroblasts to cancer-associated fibroblasts. Int J Oncol. 2015;47(2):719-727. doi:10.3892/ijo.2015.3060 45. Mitra AK, Zillhardt M, Hua Y, et al. MicroRNAs Reprogram Normal Fibroblasts into Cancer-Associated Fibroblasts in Ovarian Cancer. Cancer Discov. 2012;2(12):1100-1108. doi:10.1158/2159-8290.CD-12-0206 46. Balkwill FR, Capasso M, Hagemann T. The tumor microenvironment at a glance. J Cell Sci. 2012;125(23):5591-5596. doi:10.1242/jcs.116392 47. Spaeth EL, Dembinski JL, Sasser AK, et al. Mesenchymal Stem Cell 190

Transition to Tumor-Associated Fibroblasts Contributes to Fibrovascular Network Expansion and Tumor Progression. Blagosklonny M V., ed. PLoS One. 2009;4(4):e4992. doi:10.1371/journal.pone.0004992 48. Tomasek JJ, Gabbiani G, Hinz B, Chaponnier C, Brown RA. Myofibroblasts and mechano-regulation of connective tissue remodelling. Nat Rev Mol Cell Biol. 2002;3(5):349-363. doi:10.1038/nrm809 49. Straussman R, Morikawa T, Shee K, et al. Tumour micro-environment elicits innate resistance to RAF inhibitors through HGF secretion. Nature. 2012;487(7408):500-504. doi:10.1038/nature11183 50. Meads MB, Gatenby RA, Dalton WS. Environment-mediated drug resistance: a major contributor to minimal residual disease. Nat Rev Cancer. 2009;9(9):665-674. doi:10.1038/nrc2714 51. Faurobert E, Bouin A-P, Albiges-Rizo C. Microenvironment, tumor cell plasticity, and cancer. Curr Opin Oncol. 2015;27(1):64-70. doi:10.1097/CCO.0000000000000154 52. Orimo A, Gupta PB, Sgroi DC, et al. Stromal fibroblasts present in invasive human breast carcinomas promote tumor growth and angiogenesis through elevated SDF-1/CXCL12 secretion. Cell. 2005;121(3):335-348. doi:10.1016/j.cell.2005.02.034 53. Shimoda M, Mellody KT, Orimo A. Carcinoma-associated fibroblasts are a rate-limiting determinant for tumour progression. Semin Cell Dev Biol. 2010;21(1):19-25. doi:10.1016/j.semcdb.2009.10.002 54. Lee MJJ, Ye ASS, Gardino AKK, et al. Sequential application of anticancer drugs enhances cell death by rewiring apoptotic signaling networks. Cell. 2012;149(4):780-794. doi:10.1016/j.cell.2012.03.031 55. Lamb J, Crawford ED, Peck D, et al. The Connectivity Map: Using Gene- Expression Signatures to Connect Small Molecules, Genes, and Disease. Science (80- ). 2006;313(5795):1929-1935. doi:10.1126/science.1132939 56. Cohen AL, Soldi R, Zhang H, et al. A pharmacogenomic method for individualized prediction of drug sensitivity. Mol Syst Biol. 2011;7(1):513. doi:10.1038/msb.2011.47 57. Barretina J, Caponigro G, Stransky N, et al. The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. J KG V, D S, A R, et al., eds. Nature. 2012;483(7391):603-607. doi:10.1038/nature11003 58. Cancer Genome Atlas Network. Comprehensive molecular characterization of human colon and rectal cancer. Nature. 2012;487(7407):330-337. 191

doi:10.1038/nature11252 59. Koboldt DC, Fulton RS, McLellan MD, et al. Comprehensive molecular portraits of human breast tumours. Nature. 2012;490(7418):61-70. doi:10.1038/nature11412 60. Shah SP, Roth A, Goya R, et al. The clonal and mutational evolution spectrum of primary triple-negative breast cancers. Nature. 2012;486(7403):395-399. doi:10.1038/nature10933 61. Li J, Zhao W, Akbani R, et al. Characterization of Human Cancer Cell Lines by Reverse-phase Protein Arrays. Cancer Cell. 2017;31(2):225-239. doi:10.1016/j.ccell.2017.01.005 62. Innocenti F, Cox NJ, Dolan ME. The Use of Genomic Information to Optimize Cancer Chemotherapy. Semin Oncol. 2011;38(2):186-195. doi:10.1053/j.seminoncol.2011.01.005 63. Jiang T, Shi W, Wali VB, et al. Predictors of Chemosensitivity in Triple Negative Breast Cancer: An Integrated Genomic Analysis. Ladanyi M, ed. PLoS Med. 2016;13(12):e1002193. doi:10.1371/journal.pmed.1002193 64. Kalluri R, Zeisberg M. Fibroblasts in cancer. Nat Rev Cancer. 2006;6(5):392-401. doi:10.1038/nrc1877 65. Pallasch CP, Leskov I, Braun CJ, et al. Sensitizing protective tumor microenvironments to antibody-mediated therapy. Cell. 2014;156(3):590- 602. doi:10.1016/j.cell.2013.12.041 66. McMillin DW, Delmore J, Weisberg E, et al. Tumor cell-specific bioluminescence platform to identify stroma-induced changes to anticancer drug activity. Nat Med. 2010;16(4):483-489. doi:10.1038/nm.2112 67. Carey LA, Dees EC, Sawyer L, et al. The triple negative paradox: primary tumor chemosensitivity of breast cancer subtypes. Clin Cancer Res. 2007;13(8):2329-2334. doi:10.1158/1078-0432.CCR-06-1109 68. Anders C, Carey LA. Understanding and treating triple-negative breast cancer. Oncology (Williston Park). 2008;22(11):1233-9; discussion 1239- 40, 1243. http://www.ncbi.nlm.nih.gov/pubmed/18980022. 69. Lehmann BD, Bauer JA, Chen X, et al. Identification of human triple- negative breast cancer subtypes and preclinical models for selection of targeted therapies. J Clin Invest. 2011;121(7):2750-2767. doi:10.1172/JCI45014 70. Perou CM, Sørlie T, Eisen MB, et al. Molecular portraits of human breast tumours. Nature. 2000;406(6797):747-752. doi:10.1038/35021093 192

71. Heiser LM, Wang NJ, Talcott CL, et al. Integrated analysis of breast cancer cell lines reveals unique signaling pathways. Genome Biol. 2009;10(3):R31. doi:10.1186/gb-2009-10-3-r31 72. Prat A, Parker JS, Karginova O, et al. Phenotypic and molecular characterization of the claudin-low intrinsic subtype of breast cancer. Breast Cancer Res. 2010;12(5):R68. doi:10.1186/bcr2635 73. Ahn SG, Kim SJ, Kim C, Jeong J. Molecular Classification of Triple- Negative Breast Cancer. J Breast Cancer. 2016;19(3):223-230. doi:10.4048/jbc.2016.19.3.223 74. Fallahi-Sichani M, Honarnejad S, Heiser LM, Gray JW, Sorger PK. Metrics other than potency reveal systematic variation in responses to cancer drugs. Nat Chem Biol. 2013;9(11):708-714. doi:10.1038/nchembio.1337 75. Yamada KM, Cukierman E. Modeling Tissue Morphogenesis and Cancer in 3D. Cell. 2007;130(4):601-610. doi:10.1016/j.cell.2007.08.006 76. Fang Y, Eglen RM. Three-Dimensional Cell Cultures in Drug Discovery and Development. SLAS Discov Adv Life Sci R&D. 2017;22(5):456-472. doi:10.1177/1087057117696795 77. Weaver VM, Leliè Vre S, Lakins JN, et al. Beta4 integrin-dependent formation of polarized three- dimensional architecture confers resistance to apoptosis in normal and malignant mammary epithelium. Cancer Cell. 2002;2(September):205-216. doi:10.1016/S1535-6108(02)00125-3 78. Nguyen T V, Sleiman M, Moriarty T, Herrick WG, Peyton SR. Sorafenib resistance and JNK signaling in carcinoma during extracellular matrix stiffening. Biomaterials. 2014;35(22):5749-5759. doi:10.1016/j.biomaterials.2014.03.058 79. Buchsbaum RJ, Oh SY. Breast Cancer-Associated Fibroblasts: Where We Are and Where We Need to Go. Cancers (Basel). 2016;8(2):19. doi:10.3390/cancers8020019 80. Montero J, Sarosiek KA, DeAngelo JD, et al. Drug-induced death signaling strategy rapidly predicts cancer response to chemotherapy. Cell. 2015;160(5):977-989. doi:10.1016/j.cell.2015.01.042 81. Huang X, Yang N, Fiore VF, et al. Matrix stiffness-induced myofibroblast differentiation is mediated by intrinsic mechanotransduction. Am J Respir Cell Mol Biol. 2012;47(3):340-348. doi:10.1165/rcmb.2012-0050OC 82. Janes K a, Yaffe MB. Data-driven modelling of signal-transduction networks. Nat Rev Mol Cell Biol. 2006;7(11):820-828. doi:10.1038/nrm2041 193

83. Lamprecht MR, Sabatini DM, Carpenter AE. CellProfiler: free, versatile software for automated biological image analysis. Biotechniques. 2007;42(1):71-75. doi:10.2144/000112257 84. Ni Chonghaile T, Sarosiek KA, Vo T-T, et al. Pretreatment mitochondrial priming correlates with clinical response to cytotoxic chemotherapy. Science. 2011;334(6059):1129-1133. doi:10.1126/science.1206727 85. Ryan JA, Brunelle JK, Letai A. Heightened mitochondrial priming is the basis for apoptotic hypersensitivity of CD4+ CD8+ thymocytes. Proc Natl Acad Sci U S A. 2010;107(29):12895-12900. doi:10.1073/pnas.0914878107 86. Ryan J, Letai A. BH3 profiling in whole cells by fluorimeter or FACS. Methods. 2013;61(2):156-164. doi:10.1016/j.ymeth.2013.04.006 87. Adams JM, Cory S. The BCL-2 arbiters of apoptosis and their growing role as cancer targets. Cell Death Differ. 2018;25(1):27-36. doi:10.1038/cdd.2017.161 88. Lee GY, Kenny PA, Lee EH, Bissell MJ. Three-dimensional culture models of normal and malignant breast epithelial cells. Nat Methods. 2007;4(4):359-365. doi:10.1038/nmeth1015 89. Ryan J, Montero J, Rocco J, Letai A. iBH3: simple, fixable BH3 profiling to determine apoptotic priming in primary tissue by flow cytometry. Biol Chem. 2016;397(7):671-678. doi:10.1515/hsz-2016-0107 90. Dethlefsen C, Højfeldt G, Hojman P. The role of intratumoral and systemic IL-6 in breast cancer. Breast Cancer Res Treat. 2013;138(3):657-664. doi:10.1007/s10549-013-2488-z 91. Landskron G, De la Fuente M, Thuwajit P, Thuwajit C, Hermoso MA. Chronic inflammation and cytokines in the tumor microenvironment. J Immunol Res. 2014;2014:149185. doi:10.1155/2014/149185 92. Xia Y, Shen S, Verma IM. NF-κB, an active player in human cancers. Cancer Immunol Res. 2014;2(9):823-830. doi:10.1158/2326-6066.CIR-14- 0112 93. Yoon S, Woo SU, Kang JH, et al. NF-κB and STAT3 cooperatively induce IL6 in starved cancer cells. Oncogene. 2012;31(29):3467-3481. doi:10.1038/onc.2011.517 94. Gilbert LA, Hemann MT. DNA Damage-Mediated Induction of a Chemoresistant Niche. Cell. 2010;143(3):355-366. doi:10.1016/j.cell.2010.09.043 95. Ning Y, Manegold PC, Hong YK, et al. Interleukin-8 is associated with 194

proliferation, migration, angiogenesis and chemosensitivity in vitro and in vivo in colon cancer cell line models. Int J Cancer. 2011;128(9):2038-2049. doi:10.1002/ijc.25562 96. Yokoyama C, Yajima C, Machida T, Kawahito Y, Uchida M, Hisatomi H. Interleukin-8 enhances the effect of colchicine on cell death. Biochem Biophys Res Commun. 2017;485(1):89-94. doi:10.1016/j.bbrc.2017.02.025 97. Kelland EE, Gilmore W, Weiner LP, Lund BT. The dual role of CXCL8 in human CNS stem cell function: Multipotent neural stem cell death and oligodendrocyte progenitor cell chemotaxis. Glia. 2011;59(12):1864-1878. doi:10.1002/glia.21230 98. Lee REC, Walker SR, Savery K, Frank DA, Gaudet S. Fold Change of Nuclear NF-κB Determines TNF-Induced Transcription in Single Cells. Mol Cell. 2014;53(6):867-879. doi:10.1016/j.molcel.2014.01.026 99. Forcina GC, Conlon M, Wells A, Cao JY, Dixon SJ. Systematic Quantification of Population Cell Death Kinetics in Mammalian Cells. Cell Syst. 2017;4(6):600-610.e6. doi:10.1016/j.cels.2017.05.002 100. Downward J. The ins and outs of signalling. Nature. 2001;411(6839):759- 762. doi:10.1038/35081138 101. Duncan JS, Whittle MC, Nakamura K, et al. Dynamic Reprogramming of the Kinome in Response to Targeted MEK Inhibition in Triple-Negative Breast Cancer. Cell. 2012;149(2):307-321. doi:10.1016/j.cell.2012.02.053 102. Yao Z, Petschnigg J, Ketteler R, Stagljar I. Application guide for omics approaches to cell signaling. Nat Chem Biol. 2015;11(6):387-397. doi:10.1038/nchembio.1809 103. Lee MJ, Yaffe MB, Cantley LC, Hunter T, Sever R TJ. Protein Regulation in Signal Transduction. Signal Transduction Principles, Pathways, and Processes. 2014:31–50. doi:10.1101/cshperspect.a005918 104. Peng J, Schwartz D, Elias JE, et al. A proteomics approach to understanding protein ubiquitination. Nat Biotechnol. 2003;21(8):921-926. doi:10.1038/nbt849 105. Huang PH, Mukasa A, Bonavia R, et al. Quantitative analysis of EGFRvIII cellular signaling networks reveals a combinatorial therapeutic strategy for glioblastoma. Proc Natl Acad Sci. 2007;104(31):12867-12872. doi:10.1073/pnas.0705158104 106. Behrends C, Sowa ME, Gygi SP, Harper JW. Network organization of the human autophagy system. Nature. 2010;466(7302):68-76. doi:10.1038/nature09204 195

107. Bantscheff M, Eberhard D, Abraham Y, et al. Quantitative chemical proteomics reveals mechanisms of action of clinical ABL kinase inhibitors. Nat Biotechnol. 2007;25(9):1035-1044. doi:10.1038/nbt1328 108. Lanning BR, Whitby LR, Dix MM, et al. A road map to evaluate the proteome-wide selectivity of covalent kinase inhibitors. Nat Chem Biol. 2014;10(9):760-767. doi:10.1038/nchembio.1582 109. White FM. The Potential Cost of High-Throughput Proteomics. Sci Signal. 2011;4(160):pe8-pe8. doi:10.1126/scisignal.2001813 110. Noble WS. Mass spectrometrists should search only for peptides they care about. Nat Methods. 2015;12(7):605-608. doi:10.1038/nmeth.3450 111. Nagaraj N, Wisniewski JR, Geiger T, et al. Deep proteome and transcriptome mapping of a human cancer cell line. Mol Syst Biol. 2014;7(1):548-548. doi:10.1038/msb.2011.81 112. Ong S-E, Foster LJ, Mann M. Mass spectrometric-based approaches in quantitative proteomics. Methods. 2003;29(2):124-130. http://www.ncbi.nlm.nih.gov/pubmed/12606218. 113. Evans C, Noirel J, Ow SY, et al. An insight into iTRAQ: where do we stand now? Anal Bioanal Chem. 2012;404(4):1011-1027. doi:10.1007/s00216- 012-5918-6 114. Ludwig C, Claassen M, Schmidt A, Aebersold R. Estimation of Absolute Protein Quantities of Unlabeled Samples by Selected Reaction Monitoring Mass Spectrometry. Mol Cell Proteomics. 2012;11(3):M111.013987. doi:10.1074/mcp.M111.013987 115. Janes KA. An analysis of critical factors for quantitative immunoblotting. Sci Signal. 2015;8(371):rs2-rs2. doi:10.1126/scisignal.2005966 116. Wilson B, Liotta LA, Petricoin E. Monitoring proteins and protein networks using reverse phase protein arrays. Dis Markers. 2010;28(4):225-232. doi:10.3233/DMA-2010-0705 117. Sevecka M, Wolf-Yadlin A, MacBeath G. Lysate Microarrays Enable High- throughput, Quantitative Investigations of Cellular Signaling. Mol Cell Proteomics. 2011;10(4):M110.005363. doi:10.1074/mcp.M110.005363 118. Wolf-Yadlin A, Sevecka M, MacBeath G. Dissecting protein function and signaling using protein microarrays. Curr Opin Chem Biol. 2009;13(4):398- 405. doi:10.1016/j.cbpa.2009.06.027 119. Ciaccio MF, Wagner JP, Chuu C-P, Lauffenburger DA, Jones RB. Systems analysis of EGF receptor signaling dynamics with microwestern arrays. Nat Methods. 2010;7(2):148-155. doi:10.1038/nmeth.1418 196

120. Hughes AJ, Spelke DP, Xu Z, Kang C-C, Schaffer D V, Herr AE. Single-cell western blotting. Nat Methods. 2014;11(7):749-755. doi:10.1038/nmeth.2992 121. Tibes R, Qiu Y, Lu Y, et al. Reverse phase protein array: validation of a novel proteomic technology and utility for analysis of primary leukemia specimens and hematopoietic stem cells. Mol Cancer Ther. 2006;5(10):2512-2521. doi:10.1158/1535-7163.MCT-06-0334 122. Akbani R, Ng PKS, Werner HMJ, et al. A pan-cancer proteomic perspective on The Cancer Genome Atlas. Nat Commun. 2014;5(1):3887. doi:10.1038/ncomms4887 123. Federici G, Gao X, Slawek J, et al. Systems Analysis of the NCI-60 Cancer Cell Lines by Alignment of Protein Pathway Activation Modules with “- OMIC” Data Fields and Therapeutic Response Signatures. Mol Cancer Res. 2013;11(6):676-685. doi:10.1158/1541-7786.MCR-12-0690 124. Albeck JG, Burke JM, Aldridge BB, Zhang M, Lauffenburger DA, Sorger PK. Quantitative Analysis of Pathways Controlling Extrinsic Apoptosis in Single Cells. Mol Cell. 2008;30(1):11-25. doi:10.1016/j.molcel.2008.02.012 125. Spencer SL, Gaudet S, Albeck JG, Burke JM, Sorger PK. Non-genetic origins of cell-to-cell variability in TRAIL-induced apoptosis. Nature. 2009;459(7245):428-432. doi:10.1038/nature08012 126. Cotari JW, Voisinne G, Dar OE, Karabacak V, Altan-Bonnet G. Cell-to-Cell Variability Analysis Dissects the Plasticity of Signaling of Common Chain Cytokines in T Cells. Sci Signal. 2013;6(266):ra17-ra17. doi:10.1126/scisignal.2003240 127. Perfetto SP, Chattopadhyay PK, Roederer M. Seventeen-colour flow cytometry: unravelling the immune system. Nat Rev Immunol. 2004;4(8):648-655. doi:10.1038/nri1416 128. Bandura DR, Baranov VI, Ornatsky OI, et al. Mass Cytometry: Technique for Real Time Single Cell Multitarget Immunoassay Based on Inductively Coupled Plasma Time-of-Flight Mass Spectrometry. Anal Chem. 2009;81(16):6813-6822. doi:10.1021/ac901049w 129. Bodenmiller B, Zunder ER, Finck R, et al. Multiplexed mass cytometry profiling of cellular states perturbed by small-molecule regulators. Nat Biotechnol. 2012;30(9):858-867. doi:10.1038/nbt.2317 130. Bendall SC, Simonds EF, Qiu P, et al. Single-Cell Mass Cytometry of Differential Immune and Drug Responses Across a Human Hematopoietic Continuum. Science (80- ). 2011;332(6030):687-696. doi:10.1126/science.1198704 197

131. Du J, Bernasconi P, Clauser KR, et al. Bead-based profiling of tyrosine kinase phosphorylation identifies SRC as a potential target for glioblastoma therapy. Nat Biotechnol. 2009;27(1):77-83. doi:10.1038/nbt.1513 132. Peck D, Crawford ED, Ross KN, Stegmaier K, Golub TR, Lamb J. A method for high-throughput gene expression signature analysis. Genome Biol. 2006;7(7):R61. doi:10.1186/gb-2006-7-7-r61 133. Bradbury A, Plückthun A. Reproducibility: Standardize antibodies used in research. Nature. 2015;518(7537):27-29. doi:10.1038/518027a 134. Fields S, Song O. A novel genetic system to detect protein–protein interactions. Nature. 1989;340(6230):245-246. doi:10.1038/340245a0 135. Fields S. High-throughput two-hybrid analysis. The promise and the peril. FEBS J. 2005;272(21):5391-5399. doi:10.1111/j.1742-4658.2005.04973.x 136. Rolland T, Taşan M, Charloteaux B, et al. A Proteome-Scale Map of the Human Interactome Network. Cell. 2014;159(5):1212-1226. doi:10.1016/j.cell.2014.10.050 137. Lam MHY, Stagljar I. Strategies for membrane interaction proteomics: No mass spectrometry required. Proteomics. 2012;12(10):1519-1526. doi:10.1002/pmic.201100471 138. Petschnigg J, Groisman B, Kotlyar M, et al. The mammalian-membrane two-hybrid assay (MaMTH) for probing membrane-protein interactions in human cells. Nat Methods. 2014;11(5):585-592. doi:10.1038/nmeth.2895 139. Hu C-D, Chinenov Y, Kerppola TK. Visualization of interactions among bZIP and Rel family proteins in living cells using bimolecular fluorescence complementation. Mol Cell. 2002;9(4):789-798. http://www.ncbi.nlm.nih.gov/pubmed/11983170. 140. Wehr MC, Laage R, Bolz U, et al. Monitoring regulated protein-protein interactions using split TEV. Nat Methods. 2006;3(12):985-993. doi:10.1038/nmeth967 141. Tarassov K, Messier V, Landry CR, et al. An in Vivo Map of the Yeast Protein Interactome. Science (80- ). 2008;320(5882):1465-1470. doi:10.1126/science.1153878 142. Remy I, Ghaddar G, Michnick SW. Using the β-lactamase protein-fragment complementation assay to probe dynamic protein–protein interactions. Nat Protoc. 2007;2(9):2302-2306. doi:10.1038/nprot.2007.356 143. Li Y-C, Rodewald LW, Hoppmann C, et al. A Versatile Platform to Analyze Low-Affinity and Transient Protein-Protein Interactions in Living Cells in Real Time. Cell Rep. 2014;9(5):1946-1958. 198

doi:10.1016/j.celrep.2014.10.058 144. Rossi F, Charlton CA, Blau HM. Monitoring protein-protein interactions in intact eukaryotic cells by beta-galactosidase complementation. Proc Natl Acad Sci U S A. 1997;94(16):8405-8410. http://www.ncbi.nlm.nih.gov/pubmed/9237989. 145. Barrios-Rodiles M, Brown KR, Ozdamar B, et al. High-Throughput Mapping of a Dynamic Signaling Network in Mammalian Cells. Science (80- ). 2005;307(5715):1621-1625. doi:10.1126/science.1105776 146. Taipale M, Tucker G, Peng J, et al. A Quantitative Chaperone Interaction Network Reveals the Architecture of Cellular Protein Homeostasis Pathways. Cell. 2014;158(2):434-448. doi:10.1016/j.cell.2014.05.039 147. Taipale M, Krykbaeva I, Whitesell L, et al. Chaperones as thermodynamic sensors of drug-target interactions reveal kinase inhibitor specificities in living cells. Nat Biotechnol. 2013;31(7):630-637. doi:10.1038/nbt.2620 148. Zhu H, Snyder M. Protein chip technology. Curr Opin Chem Biol. 2003;7(1):55-63. http://www.ncbi.nlm.nih.gov/pubmed/12547427. 149. Fan B, Lu K-Y, Reymond Sutandy FX, et al. A Human Proteome Microarray Identifies that the Heterogeneous Nuclear Ribonucleoprotein K (hnRNP K) Recognizes the 5′ Terminal Sequence of the Hepatitis C Virus RNA. Mol Cell Proteomics. 2014;13(1):84-92. doi:10.1074/mcp.M113.031682 150. Huang J, Zhu H, Haggarty SJ, et al. Finding new components of the target of rapamycin (TOR) signaling network through chemical genetics and proteome chips. Proc Natl Acad Sci. 2004;101(47):16594-16599. doi:10.1073/pnas.0407117101 151. Jones RB, Gordus A, Krall JA, MacBeath G. A quantitative protein interaction network for the ErbB receptors using protein microarrays. Nature. 2006;439(7073):168-174. doi:10.1038/nature04177 152. Tay S, Hughey JJ, Lee TK, Lipniacki T, Quake SR, Covert MW. Single-cell NF-κB dynamics reveal digital activation and analogue information processing. Nature. 2010;466(7303):267-271. doi:10.1038/nature09145 153. Batchelor E, Mock CS, Bhan I, Loewer A, Lahav G. Recurrent Initiation: A Mechanism for Triggering p53 Pulses in Response to DNA Damage. Mol Cell. 2008;30(3):277-289. doi:10.1016/j.molcel.2008.03.016 154. Purvis JE, Karhohs KW, Mock C, Batchelor E, Loewer A, Lahav G. p53 Dynamics Control Cell Fate. Science (80- ). 2012;336(6087):1440-1444. doi:10.1126/science.1218351 199

155. Montero J, Sarosiek KAA, Drapkin R, et al. Drug-Induced Death Signaling Strategy Rapidly Predicts Cancer Response to Chemotherapy Article Drug- Induced Death Signaling Strategy Rapidly Predicts Cancer Response to Chemotherapy. Cell. 2015;160(5):977-989. doi:10.1016/j.cell.2015.01.042 156. Albeck JG, Burke JM, Spencer SL, Lauffenburger DA, Sorger PK. Modeling a Snap-Action, Variable-Delay Switch Controlling Extrinsic Cell Death. Levchenko A, ed. PLoS Biol. 2008;6(12):e299. doi:10.1371/journal.pbio.0060299 157. Carpenter AE, Jones TR, Lamprecht MR, et al. CellProfiler: image analysis software for identifying and quantifying cell phenotypes. Genome Biol. 2006;7(10):R100. doi:10.1186/gb-2006-7-10-r100 158. Rajaram S, Pavie B, Wu LF, Altschuler SJ. PhenoRipper: software for rapidly profiling microscopy images. Nat Methods. 2012;9(7):635-637. doi:10.1038/nmeth.2097 159. Millard BL, Niepel M, Menden MP, Muhlich JL, Sorger PK. Adaptive informatics for multifactorial and high-content biological data. Nat Methods. 2011;8(6):487-492. doi:10.1038/nmeth.1600 160. Collins T. ImageJ for microscopy. Biotech. 2007;43:S25–S30. doi:10.2144/000112517 161. Janes KA, Albeck JG, Peng LX, Sorger PK, Lauffenburger DA, Yaffe MB. A High-throughput Quantitative Multiplex Kinase Assay for Monitoring Information Flow in Signaling Networks. Mol Cell Proteomics. 2003;2(7):463-473. doi:10.1074/mcp.M300045-MCP200 162. Janes KA, Albeck JG, Gaudet S, Sorger PK, Lauffenburger DA, Yaffe MB. A Systems Model of Signaling Identifies a Molecular Basis Set for Cytokine-Induced Apoptosis. Science (80- ). 2005;310(5754):1646-1653. doi:10.1126/science.1116598 163. Kubota K, Anjum R, Yu Y, et al. Sensitive multiplexed analysis of kinase activities and activity-based kinase identification. Nat Biotechnol. 2009;27(10):933-940. doi:10.1038/nbt.1566 164. Bose AK, Janes KA. A High-throughput Assay for Phosphoprotein-specific Phosphatase Activity in Cellular Extracts. Mol Cell Proteomics. 2013;12(3):797-806. doi:10.1074/mcp.O112.024059 165. Miller MA, Barkal L, Jeng K, et al. Proteolytic Activity Matrix Analysis (PrAMA) for simultaneous determination of multiple protease activities. Integr Biol (Camb). 2011;3(4):422-438. doi:10.1039/c0ib00083c 166. Shults MD, Imperiali B. Versatile Fluorescence Probes of Protein Kinase 200

Activity. J Am Chem Soc. 2003;125(47):14248-14249. doi:10.1021/ja0380502 167. Han W-G, Liu T, Himo F, et al. A Theoretical Study of the UV/Visible Absorption and Emission Solvatochromic Properties of Solvent-Sensitive Dyes. ChemPhysChem. 2003;4(10):1084-1094. doi:10.1002/cphc.200300801 168. Goguen BN, Loving GS, Imperiali B. Development of a fluorogenic sensor for activated Cdc42. Bioorg Med Chem Lett. 2011;21(17):5058-5061. doi:10.1016/j.bmcl.2011.04.051 169. Pertz O, Hahn KM. Designing biosensors for Rho family proteins -- deciphering the dynamics of Rho family GTPase activation in living cells. J Cell Sci. 2004;117(8):1313-1318. doi:10.1242/jcs.01117 170. Subramanian A, Tamayo P, Mootha VK, et al. Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci. 2005;102(43):15545-15550. doi:10.1073/pnas.0506580102 171. Margolin AA, Nemenman I, Basso K, et al. ARACNE: An Algorithm for the Reconstruction of Gene Regulatory Networks in a Mammalian Cellular Context. BMC Bioinformatics. 2006;7(Suppl 1):S7. doi:10.1186/1471-2105- 7-S1-S7 172. Lefebvre C, Rajbhandari P, Alvarez MJ, et al. A human B-cell interactome identifies MYB and FOXM1 as master regulators of proliferation in germinal centers. Mol Syst Biol. 2010;6:377. doi:10.1038/msb.2010.31 173. Wang K, Saito M, Bisikirska BC, et al. Genome-wide identification of post- translational modulators of transcription factor activity in human B cells. Nat Biotechnol. 2009;27(9):829-837. doi:10.1038/nbt.1563 174. Linding R, Jensen LJ, Ostheimer GJ, et al. Systematic Discovery of In Vivo Phosphorylation Networks. Cell. 2007;129(7):1415-1426. doi:10.1016/j.cell.2007.05.052 175. Gilbert LA, Horlbeck MA, Adamson B, et al. Genome-Scale CRISPR- Mediated Control of Gene Repression and Activation. Cell. 2014;159(3):647-661. doi:10.1016/j.cell.2014.09.029 176. Silva JM, Li MZ, Chang K, et al. Second-generation shRNA libraries covering the mouse and human genomes. Nat Genet. 2005;37(11):1281- 1288. doi:10.1038/ng1650 177. Leek JT, Scharpf RB, Bravo HC, et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet. 201

2010;11(10):733-739. doi:10.1038/nrg2825 178. Collins FS, Tabak LA. Policy: NIH plans to enhance reproducibility. Nature. 2014;505(7485):612-613. http://www.ncbi.nlm.nih.gov/pubmed/24482835. 179. Chervitz SA, Deutsch EW, Field D, et al. Data Standards for Omics Data: The Basis of Data Sharing and Reuse. In: Methods in Molecular Biology (Clifton, N.J.). Vol 719. ; 2011:31-69. doi:10.1007/978-1-61779-027-0_2 180. Mann M. Comparative analysis to guide quality improvements in proteomics. Nat Methods. 2009;6(10):717-719. http://www.ncbi.nlm.nih.gov/pubmed/19953682. 181. Cairns DA. Statistical issues in quality control of proteomic analyses: Good experimental design and planning. Proteomics. 2011;11(6):1037-1048. doi:10.1002/pmic.201000579 182. Kreutz C, Timmer J. Systems biology: experimental design. FEBS J. 2009;276(4):923-942. doi:10.1111/j.1742-4658.2008.06843.x 183. Oberg AL, Vitek O. Statistical Design of Quantitative Mass Spectrometry- Based Proteomic Experiments. J Proteome Res. 2009;8(5):2144-2156. doi:10.1021/pr8010099 184. Rocke DM. Design and analysis of experiments with high throughput biological assay data. Semin Cell Dev Biol. 2004;15(6):703-713. doi:10.1016/j.semcdb.2004.09.007 185. Glass DJ. Experimental Design for Biologists. Cold Spring Harbor, New York: Cold Spring Harbor Laboratory Press; 2007. 186. Naegle K, Gough NR, Yaffe MB. Criteria for biological reproducibility: What does “ n ” mean? Sci Signal. 2015;8(371):fs7-fs7. doi:10.1126/scisignal.aab1125 187. Lazic SE. The problem of pseudoreplication in neuroscientific studies: is it affecting your analysis? BMC Neurosci. 2010;11(1):5. doi:10.1186/1471- 2202-11-5 188. Aarts E, Verhage M, Veenvliet J V, Dolan C V, van der Sluis S. A solution to dependency: using multilevel analysis to accommodate nested data. Nat Neurosci. 2014;17(4):491-496. doi:10.1038/nn.3648 189. Altman N, Krzywinski M. Points of significance: Sources of variation. Nat Methods. 2015;12(1):5-6. http://www.ncbi.nlm.nih.gov/pubmed/25699313. 190. Ioannidis JPA, Hozo I, Djulbegovic B. Optimal type I and type II error pairs when the available sample size is fixed. J Clin Epidemiol. 2013;66(8):903- 910.e2. doi:10.1016/j.jclinepi.2013.03.002 202

191. Glass DJ. A Critique of the Hypothesis, and a Defense of the Question, as a Framework for Experimentation. Clin Chem. 2010;56(7):1080-1085. doi:10.1373/clinchem.2010.144477 192. Kell DB, Oliver SG. Here is the evidence, now what is the hypothesis? The complementary roles of inductive and hypothesis-driven science in the post-genomic era. BioEssays. 2004;26(1):99-105. doi:10.1002/bies.10385 193. Platt JR. Strong Inference: Certain systematic methods of scientific thinking may produce much more rapid progress than others. Science (80- ). 1964;146(3642):347-353. doi:10.1126/science.146.3642.347 194. Beard DA, Kushmerick MJ. Strong Inference for Systems Biology. McCulloch AD, ed. PLoS Comput Biol. 2009;5(8):e1000459. doi:10.1371/journal.pcbi.1000459 195. Glass DJ, Hall N. A Brief History of the Hypothesis. Cell. 2008;134(3):378- 381. doi:10.1016/j.cell.2008.07.033 196. Ioannidis JPA, Khoury MJ. Improving Validation Practices in “Omics”; Research. Science (80- ). 2011;334(6060):1230-1232. doi:10.1126/science.1211811 197. Ioannidis JPA. Why Most Published Research Findings Are False. PLoS Med. 2005;2(8):e124. doi:10.1371/journal.pmed.0020124 198. Begley CG, Ellis LM. Raise standards for preclinical cancer research. Nature. 2012;483(7391):531-533. doi:10.1038/483531a 199. Ioannidis JPA. Expectations, validity, and reality in omics. J Clin Epidemiol. 2010;63(9):945-949. doi:10.1016/j.jclinepi.2010.04.002 200. Chuang H-Y, Hofree M, Ideker T. A Decade of Systems Biology. Annu Rev Cell Dev Biol. 2010;26(1):721-744. doi:10.1146/annurev-cellbio-100109- 104122 201. Ideker T, Galitski T, Hood L. A New Approach to Decoding Life: Systems Biology. Annu Rev Genomics Hum Genet. 2001;2(1):343-372. doi:10.1146/annurev.genom.2.1.343 202. Aldridge BB, Burke JM, Lauffenburger DA, Sorger PK. Physicochemical modelling of cell signalling pathways. Nat Cell Biol. 2006;8(11):1195-1203. doi:10.1038/ncb1497 203. Saez-Rodriguez J, Alexopoulos LG, Epperlein J, et al. Discrete logic modelling as a means to link protein signalling networks with functional analysis of mammalian signal transduction. Mol Syst Biol. 2009;5:331. doi:10.1038/msb.2009.87 203

204. Sun J, Garibaldi JM, Hodgman C. Parameter Estimation Using Metaheuristics in Systems Biology: A Comprehensive Review. IEEE/ACM Trans Comput Biol Bioinforma. 2012;9(1):185-202. doi:10.1109/TCBB.2011.63 205. Santos SDM, Wollman R, Meyer T, Ferrell JE. Spatial Positive Feedback at the Onset of Mitosis. Cell. 2012;149(7):1500-1513. doi:10.1016/j.cell.2012.05.028 206. Litvin O, Schwartz S, Wan Z, et al. Interferon α/β Enhances the Cytotoxic Response of MEK Inhibition in Melanoma. Mol Cell. 2015;57(5):784-796. doi:10.1016/j.molcel.2014.12.030 207. Lee MJ. Let a Hundred Flowers Bloom: The Role of Context Dependence in Creating Phenotypic Diversity following Targeted Therapy. Mol Cell. 2015;57(5):763-764. doi:10.1016/j.molcel.2015.02.030 208. Morris MK, Saez-Rodriguez J, Clarke DC, Sorger PK, Lauffenburger DA. Training Signaling Pathway Maps to Biochemical Data with Constrained Fuzzy Logic: Quantitative Analysis of Liver Cell Responses to Inflammatory Stimuli. Beard DA, ed. PLoS Comput Biol. 2011;7(3):e1001099. doi:10.1371/journal.pcbi.1001099 209. MacNamara A, Terfve C, Henriques D, Bernabé BP, Saez-Rodriguez J. State–time spectrum of signal transduction logic models. Phys Biol. 2012;9(4):045003. doi:10.1088/1478-3975/9/4/045003 210. Pe’er D, Hacohen N. Principles and Strategies for Developing Network Models in Cancer. Cell. 2011;144(6):864-873. doi:10.1016/j.cell.2011.03.001 211. Ideker T, Krogan NJ. Differential network biology. Mol Syst Biol. 2012;8:565. doi:10.1038/msb.2011.99 212. Vidal M, Cusick ME, Barabási A-L. Interactome Networks and Human Disease. Cell. 2011;144(6):986-998. doi:10.1016/j.cell.2011.02.016 213. Goh K-I, Cusick ME, Valle D, Childs B, Vidal M, Barabasi A-L. The human disease network. Proc Natl Acad Sci. 2007;104(21):8685-8690. doi:10.1073/pnas.0701361104 214. Amit I, Citri A, Shay T, et al. A module of negative feedback regulators defines growth factor signaling. Nat Genet. 2007;39(4):503-512. doi:10.1038/ng1987 215. Naegle KM, Welsch RE, Yaffe MB, White FM, Lauffenburger DA. MCAM: Multiple Clustering Analysis Methodology for Deriving Hypotheses and Insights from High-Throughput Proteomic Datasets. Papin JA, ed. PLoS 204

Comput Biol. 2011;7(7):e1002119. doi:10.1371/journal.pcbi.1002119 216. Tentner AR, Lee MJ, Ostheimer GJ, Samson LD, Lauffenburger DA, Yaffe MB. Combined experimental and computational analysis of DNA damage signaling reveals context-dependent roles for Erk in apoptosis and G1/S arrest after genotoxic stress. Mol Syst Biol. 2012;8:568. doi:10.1038/msb.2012.1 217. Brown KS, Sethna JP. Statistical mechanical approaches to models with many poorly known parameters. Phys Rev E. 2003;68(2):021904. doi:10.1103/PhysRevE.68.021904 218. Raue A, Kreutz C, Maiwald T, et al. Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics. 2009;25(15):1923-1929. doi:10.1093/bioinformatics/btp358 219. Maiwald T, Timmer J, Kreutz C, Klingmüller U, Raue A. Addressing parameter identifiability by model-based experimentation. IET Syst Biol. 2011;5(2):120-130. doi:10.1049/iet-syb.2010.0061 220. Gutenkunst RN, Waterfall JJ, Casey FP, Brown KS, Myers CR, Sethna JP. Universally Sloppy Parameter Sensitivities in Systems Biology Models. PLoS Comput Biol. 2007;3(10):e189. doi:10.1371/journal.pcbi.0030189 221. Clarke DC, Morris MK, Lauffenburger DA. Normalization and Statistical Analysis of Multiplexed Bead-based Immunoassay Data Using Mixed- effects Modeling. Mol Cell Proteomics. 2013;12(1):245-262. doi:10.1074/mcp.M112.018655 222. Natarajan M, Lin K-M, Hsueh RC, Sternweis PC, Ranganathan R. A global analysis of cross-talk in a mammalian cellular signalling network. Nat Cell Biol. 2006;8(6):571-580. doi:10.1038/ncb1418 223. Cho K-H, Shin S-Y, Kolch W, Wolkenhauer O. Experimental Design in Systems Biology, Based on Parameter Sensitivity Analysis Using a Monte Carlo Method: A Case Study for the TNFα-Mediated NF-κ B Signal Transduction Pathway. Simulation. 2003;79(12):726-739. doi:10.1177/0037549703040943 224. Casey FP, Baird D, Feng Q, et al. Optimal experimental design in an epidermal growth factor receptor signalling and down-regulation model. IET Syst Biol. 2007;1(3):190-202. http://www.ncbi.nlm.nih.gov/pubmed/17591178. 225. Hagen DR, White JK, Tidor B. Convergence in parameters and predictions using computational experimental design. Interface Focus. 2013;3(4):20130008-20130008. doi:10.1098/rsfs.2013.0008 205

226. Apgar JF, Witmer DK, White FM, Tidor B. Sloppy models, parameter uncertainty, and the role of experimental design. Mol Biosyst. 2010;6(10):1890. doi:10.1039/b918098b 227. Schelker M, Raue A, Timmer J, Kreutz C. Comprehensive estimation of input signals and dynamics in biochemical reaction networks. Bioinformatics. 2012;28(18):i529-i534. doi:10.1093/bioinformatics/bts393 228. Vilar JMG, Jansen R, Sander C. Signal Processing in the TGF-β Superfamily Ligand-Receptor Network. PLoS Comput Biol. 2006;2(1):e3. doi:10.1371/journal.pcbi.0020003 229. Clarke DC, Brown ML, Erickson RA, Shi Y, Liu X. Transforming Growth Factor Depletion Is the Primary Determinant of Smad Signaling Kinetics. Mol Cell Biol. 2009;29(9):2443-2455. doi:10.1128/MCB.01443-08 230. Mecham BH, Nelson PS, Storey JD. Supervised normalization of microarrays. Bioinformatics. 2010;26(10):1308-1315. doi:10.1093/bioinformatics/btq118 231. Callister SJ, Barry RC, Adkins JN, et al. Normalization Approaches for Removing Systematic Biases Associated with Mass Spectrometry and Label-Free Proteomics. J Proteome Res. 2006;5(2):277-286. doi:10.1021/pr050300l 232. Liu W, Ju Z, Lu Y, Mills GB, Akbani R. A Comprehensive Comparison of Normalization Methods for Loading Control and Variance Stabilization of Reverse-Phase Protein Array Data. Cancer Inform. 2014;13:CIN.S13329. doi:10.4137/CIN.S13329 233. Neeley ES, Kornblau SM, Coombes KR, Baggerly KA. Variable slope normalization of reverse phase protein arrays. Bioinformatics. 2009;25(11):1384-1389. doi:10.1093/bioinformatics/btp174 234. Artigaud S, Gauthier O, Pichereau V. Identifying differentially expressed proteins in two-dimensional electrophoresis experiments: inputs from transcriptomics statistical tools. Bioinformatics. 2013;29(21):2729-2734. doi:10.1093/bioinformatics/btt464 235. Kreutz C, Rodriguez MMB, Maiwald T, et al. An error model for protein quantification. Bioinformatics. 2007;23(20):2747-2753. doi:10.1093/bioinformatics/btm397 236. Hahne F, Khodabakhshi AH, Bashashati A, et al. Per-channel basis normalization methods for flow cytometry data. Cytom Part A. 2009;9999A(2):NA-NA. doi:10.1002/cyto.a.20823 237. O’Neill K, Aghaeepour N, Špidlen J, Brinkman R. Flow Cytometry 206

Bioinformatics. Wodak S, ed. PLoS Comput Biol. 2013;9(12):e1003365. doi:10.1371/journal.pcbi.1003365 238. Gaudet S, Janes KA, Albeck JG, Pace EA, Lauffenburger DA, Sorger PK. A Compendium of Signals and Responses Triggered by Prodeath and Prosurvival Cytokines. Mol Cell Proteomics. 2005;4(10):1569-1590. doi:10.1074/mcp.M500158-MCP200 239. Hamid JS, Hu P, Roslin NM, Ling V, Greenwood CMT, Beyene J. Data Integration in Genetics and Genomics: Methods and Challenges. Hum Genomics Proteomics. 2009;2009:1-13. doi:10.4061/2009/869093 240. Braun P, Tasan M, Dreze M, et al. An experimentally derived confidence score for binary protein-protein interactions. Nat Methods. 2009;6(1):91-97. doi:10.1038/nmeth.1281 241. Tian Q, Price ND, Hood L. Systems cancer medicine: towards realization of predictive, preventive, personalized and participatory (P4) medicine. J Intern Med. 2012;271(2):111-121. doi:10.1111/j.1365-2796.2011.02498.x 242. Cornish-Bowden A. Systems biology. How far has it come? Biochem (Lond). 2011;33:16-18. 243. Huber KVM, Salah E, Radic B, et al. Stereospecific targeting of MTH1 by (S)-crizotinib as an anticancer strategy. Nature. 2014;508(7495):222-227. doi:10.1038/nature13194