Reconstructing Blood Stem Cell Regulatory Network Models From

Downloaded by guest on October 1, 2021 omdi niiulclst bansnl-elgn expression gene single-cell obtain to (4). profiles cells per- be individual can in (RNA-Seq) real-time formed sequencing quantitative RNA as and of (qRT-PCR) such PCR basis Methods molecular heterogeneity. a measure the type within resolve which cell homogeneity the can assume studies, assays in and single-cell population choices population, expression fate bulk of to states Unlike average related 3). is (2, progen- of populations and blood development (HSPC) stem recent cell hematopoietic the within itor with heterogeneity how only understand- are ing is we that it technologies single-cell yet high-throughput and (1), system development. cell blood as during understand such controlled to are disorders important decisions blood therefore fate how severe is production It in biased leukemia. result to myeloid can acute leads which choices types, fate cell population of of the imbalance at regulated An and balanced level. differ- is toward fate types output cell alternative the mature cells, Although ent individual pool. by made cell be can stem choices blood self-renew as the well maintain as types, to cell blood mature all into differentiate T network networks regulatory gene applica- regulatory uncover widely to hypotheses is systems relationships. biological and provides hierarchical differentiation, other to hematopoiesis, HSC ble of of homolog con- regulation approach aspects 3 about Our to, Gata2. known translocated factor firms transcription 2, the fac- subunit core-binding by alpha and (Cbfa2t3h) domain, (Nfe2) and 2 runt regu- identified erythroid tor, factor, the we nuclear in of difference models, lation a two validated these experimentally subsequently comparing pro- megakaryocyte– multipotent By toward lymphoid-primed genitors. trajectories and on progeni- progenitors focusing into erythrocyte types, HSCs This of cell cells. differentiation network tor progenitor regulatory recapitulated and transcriptional that stem infer models blood pro- to 2,167 us expression in allowed on approach genes based 48 of data inferencefiles from expression information network dynamic snapshot a infer cell-by- applied single-cell to a ability and the on developed important exploiting controlled method, we is are Here it decisions basis. leukemia, these cell as how such decisions understand malignancies fate to fatal HSC to single linked of reg- is dysregulation transcriptional as As organized modeled networks. within be ulatory operate can that and programs decisions regulatory these key regulating a play in fate factors the role Transcription by cells. level individual population by the made at decisions alterna- balanced toward be HSCs must the of lineages for Differentiation tive types organism. cell the mature with of all lifetime each generate to types, shown have cell functionally (HSCs) been cells mature stem 2017 of hematopoietic 18, Single January mixture functions. Shubin a specialized H. Neil contains Member Board blood Editorial Adult by accepted and CA, 2016) Pasadena, 25, Technology, July of review Institute for California (received Rothenberg, V. Ellen by Edited Kingdom United 0XY, CB2 Cambridge Research, Medical a G Berthold and www.pnas.org/cgi/doi/10.1073/pnas.1610609114 Hamey K. profiles Fiona molecular single-cell from network models regulatory cell stem blood Reconstructing eateto amtlg,Wlcm rs–eia eerhCuclCmrdeSe elIsiue nvriyo abig,CmrdeIsiuefor Institute Cambridge Cambridge, of University Institute, Cell Stem Cambridge Council Research Trust–Medical Wellcome Haematology, of Department eaooei sa xesvl tde n well-characterized and studied extensively an is Hematopoiesis andb eaooei tmcls(Ss.HC r beto able are main- HSCs is (HSCs). cells system stem blood hematopoietic by mammalian tained the life, adult hroughout | tmpoeio cells progenitor stem ottgens ¨ a,1 oi Nestorowa Sonia , | hematopoiesis a,2 | igecell single a,1 aa .Kinston J. Sarah , | Boolean a ai .Kent G. David , ulse 2 eeetne ooti opeesv coverage comprehensive obtain previously to we extended data were qRT-PCR (2) investigation, published this large for a cells provide of To pool relationships. infer regulatory to factor profiling expression transcription gene single-cell used have we trolled, perturbations. making reca- of simulating capable by that was predictions model and behavior, network patterning computational known transcriptional pitulated a of create evidence to experimental regulation et extensive Peter gap used embryo, urchin the (14) sea the in al. developing the patterning present In (13). interactions in network the gene genes of gap understanding of improved mathemati- role the a develop- study used Drosophila to particular, have approach studies in cal Several systems, gene about processes. of understanding regulatory new variety providing networks, a gene has mental to networks expres- regulatory applied gene gene of been of simulation modeling silico Computational in allow sion. also can methods ence (8–11). data previously single-cell emerged on have has reconstruction approaches network recently, relationships data basing More 7). regulatory single-cell (6, simple systems popu- of blood discovering in whole power in on recognized The has measured been cells. methods data of reconstruction expression lations to network restricted most However, been of relationships. functional application become these the predict therefore to have used methods widely Com- inference approach. system-wide network a to putational scale readily tar- not and does genes regulator get between experimental relationships the functional because of part remains in validation least interactions at challenge, net- transcriptional enormous regulatory true an transcriptional Identifying of (5). components works as acting factors 1073/pnas.1610609114/-/DCSupplemental at online information supporting contains article This identifiers the assigned 2 and the BioModels in in 1 deposited deposited been respectively. no. have were MODEL1610060001, and (accession paper MODEL1610060000 models database this network in (GEO) LMPP reported Omnibus data Editorial Expression the ChIP by Gene The invited editor deposition: guest Data a is E.V.R. Submission. Direct PNAS Board. a is article This interest. paper. of the conflict wrote no B.G. declare and authors N.K.W., The D.G.K., S.N., S.N., F.K.H., F.K.H., and tools; data; reagents/analytic analyzed and N.K.W. new D.G.K., and S.J.K., contributed F.K.H. S.N., F.K.H., research; research; performed designed N.K.W. B.G. and N.K.W. contributions: Author cdme fSine n niern nIvn,C.Tecmlt rga n video and at program website complete NAS National the Gene The on the available CA. of are Irvine, presentations Center most in Beckman of recordings Engineering Mabel and and Arnold Evolu- Sciences the and of at Academies Development 2016, in 12–14, Models April Network held tion,” and Networks Regulatory ”Gene Sciences, hspprrslsfo h rhrM ake olqimo h ainlAaeyof Academy National the of Colloquium Sackler M. Arthur the from results paper This cam.ac.uk. owo orsodnemyb drse.Eal [email protected] rnkw22@ or [email protected] Email: addressed. be may correspondence whom To work. this to equally contributed S.N. and F.K.H. oadesteqeto fhwHP aedcsosaecon- are decisions fate HSPC how of question the address To infer- network relationships, regulatory identifying as well As transcription by influenced heavily is decision-making Cellular Regulatory a mro(2 hr osrciggn ici models circuit gene constructing where (12) embryo ioaK Wilson K. Nicola , Networks . . a,2 , www.pnas.org/lookup/suppl/doi:10. NSEryEdition Early PNAS GSE84328 www.nasonline.org/ .TeMPand MEP The ). | f8 of 1

BIOPHYSICS AND COLLOQUIUM COMPUTATIONAL BIOLOGY PAPER of the murine bone marrow HSPC compartment. Using these can capture a variety of more complex structures. The diffusion data, differentiation trajectories from HSCs to progenitor cells map method has been specifically adapted for use with single- were constructed. These were used to infer and validate regu- cell expression data (17) and has proved to be a powerful tool for latory network models, thereby gaining greater insight into the representing spatial heterogeneity in single-cell data from mouse transcriptional programs governing HSC differentiation. embryos (18), and branching differentiation dynamics for both single-cell qRT-PCR data describing embryonic blood develop- Results ment (10) and single-cell RNA sequencing (scRNA-Seq) data for Single-Cell Snapshot Measurements Capture Progression Through adult HSPCs (19). HSPC Differentiation. To study the transcriptional control of When applied to our data, diffusion map analysis using all of HSPC differentiation, we previously collected single-cell qRT- the genes analyzed by single-cell qRT-PCR demonstrated that PCR data for HSCs and progenitor cells, in which we quan- the new and old data sets integrated well (SI Appendix, Fig. S1). tified the expression levels of 48 genes in 1,626 HSPCs The location of specific HSPC populations in the diffusion using the Fluidigm Biomark system (2). This study profiled map was consistent with known lineage relationships between megakaryocyte–erythroid progenitors (MEPs), granulocyte– mature cell types and their respective precursor populations. monocyte progenitors (GMPs), lymphoid-primed multipotent Fig. 1B highlights two progenitor cell populations, MEPs and progenitors (LMPPs), common myeloid progenitors (CMPs), LMPPs, along with the so-called molecular overlap, or “MolO” HSCs with finite self-renewal (FSR-HSCs), and long-term HSCs HSCs, as identified by Wilson et al. (2). MolO cells are HSCs (LT-HSCs). However, the primary focus was to resolve hetero- with a shared transcriptional profile and increased probability geneity within four different LT-HSC populations isolated by of long-term multilineage reconstitution upon single-cell trans- fluorescence-activated cell sorting. Furthermore, the study pro- plantation. Cells belonging to intermediate populations, such as filed a limited number of progenitor populations. As we were MPPs and preMegEs, were present in regions of the diffusion interested in understanding progression through differentiation, map between the highlighted cell types. Taken together, dif- we generated equivalent expression profiles for over 500 single fusion map analysis of this comprehensive single-cell data set cells from three additional populations to increase the coverage reveals a transcriptional landscape of expression states charac- of intermediate cell stages and therefore improve our resolution teristic for early HSPC differentiation (Fig. 1C). In addition, of the hematopoietic hierarchy (Fig. 1A). FSR-HSC2, multipo- the coordinates of the data in the diffusion map provide more tent progenitor (MPP), and pre-megakaryocyte-erythroid pro- than a visualization, as distances in diffusion space represent genitor (preMegE) (15) populations were profiled using the a measure of similarity between cells that avoids some of the same single-cell qRT-PCR assays as before. Combined with the effects of noise present in single-cell expression measurements earlier profiles, these data provide extensive coverage of murine (11, 20). HSPC populations (Fig. 1A). The gene set used included 33 transcription factors known to play a role in HSC or myeloid differ- Single-Cell Expression Profiles Can Be Used to Construct Differen- entiation, 12 nontranscription factor genes implicated in HSPC tiation Trajectories. Motivated by the consistency between the biology, and three housekeeping genes. location of HSPC populations in the diffusion map and the To visualize the broader expression landscape captured by hematopoietic hierarchy, we aimed to use the underlying coordi- these 2,167 single-cell transcriptional profiles, we used diffu- nate space to better understand transcriptional changes through- sion maps (16). Diffusion maps use properties of random walks out differentiation. Recent work introduced the concept of between cells to describe the underlying structure of the data. inferring “pseudotime” trajectories from single-cell expression This method offers an advantage over linear dimensionality data, where a sample of cells is ordered by progress through reduction techniques, such as principal component analysis, as it differentiation based on the strength of similarities between

Fig. 1. Single-cell gene expression profiling captures the transcriptional landscape of HSC differentiation. (A) The hematopoietic hierarchy, with populations profiled by qRT-PCR highlighted in boxes. The sorting strategies used to isolate each population are displayed to the right of the lineage tree. The three cell types with starred sorting strategies were collected and profiled specifically for this study; unstarred populations were profiled in our previous study, and the lineage tree diagram is also adapted from this paper (2). (B) Diffusion map dimensionality reduction of the populations highlighted in A based on gene expression as quantified by qRT-PCR. MolO stem cells (a subset of the LT-HSC sorting strategies enriched for functional LT-HSCs) are shown in purple, MEPs in red, and LMPPs in blue. All other cell types are in gray. For diffusion map, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) plots showing all cell types see SI Appendix, Fig. S1.(C) Diagram highlighting how these single-cell data capture HSC fate choice. HSCs can self-renew, or differentiate toward alternative lineages. Single-cell expression data are sampled from a transcriptional landscape that contains cells at different stages along differentiation trajectories toward MEP or LMPP progenitor cells.

2 of 8 | www.pnas.org/cgi/doi/10.1073/pnas.1610609114 Hamey et al. Downloaded by guest on October 1, 2021 Downloaded by guest on October 1, 2021 nMPo MPtaetre.Clsaeclrdb hi suoievalue, pseudotime their by colored are cells Cells showing trajectories. map LMPP Diffusion or (A) MEP dynamics. on expression gene captures tories 2. Fig. ae tal. et Hamey B A opttoal reigsnl el ln ifrnito trajec- differentiation along cells single ordering Computationally .As 3C). (Fig. cell input the to applied when partial cell agreed the output it frequently from the how with function on based relevant Boolean scored most was each analysis the correlation end, and identify this function, to To be Boolean opportunity functions. can a an trajectory for provide ordered states therefore an input–output from gene as cells binary considered of the that pairs propose of for we orderings expression Here dynamic pseudotime branches. a lineage finding two is by the differentiation captured we above, func- transi- which discussed Boolean that process, As of identify to established 25). state used (10, been be allowed tions can has an states it these as between poten- and tions considered of network, be for Boolean set data can the expression this cell binary from the individual data, gene each single-cell each In functions. for tial function Boolean able regulation 3B). the (Fig. genes model factor to transcription functions the Boolean of each potential of abstracted of then set were a the to correlations with pairs Partial gene the correlation. potential from strongest formed of was network gene a each and for regulators calculated, were correla- pairwise genes relationships data, between expression regulatory tions single-cell gene the potential Using continuous identify 3A). about (Fig. that to information reasoned use levels therefore to expression valuable We be information. would this it lose binary only would considering data differentiation; throughout expressed is For expres- 2B). (Fig. the pseudotime in throughout example, factors visible changes genes transcription of with some levels behavior, sion that clear complex was more val- it exhibited ON/OFF data, binary our to From converted ues. be must levels expression that develop- 24). blood (8, cells embryonic stem embryonic (23), and HSCs (10), transcriptional ment in model networks to have abstraction studies regulatory Boolean Previous fates. used blood successfully differ- trajectories alternative controlling toward programs differentiation understand entiation regulatory help to the two used between be differences could the orderings these between that Pseu- suggested the dynamics Data. from Single-Cell Inferred sion of Be Can Dynamics Models dotime Network Regulatory and Gene CMPs to related S1). closely Fig. appear Appendix, of (SI representation LMPPs nature map GMPs types. where diffusion the the cell data, to in the of part, seen also of in mixture is least a and set, at of gene attributed, our composed be is may the path finding contrast, In This LMPP trajectory. the its of of end end the near passes cells mostly MEP MEPs through toward path The transcrip- MEPs. toward the ferentiation of Expression factor trajectory. tion MEP regula- the developmental in the undetected of visi- expression trajectories the tor two example, the For between ble. differences with lineages, the differentiation. HSPC investigation during allow dynamics to expression pseudo- gene 2A) respective of (Fig. two the trajectories into differentiation each ordered time to were belonging branches Cells these cells. terminal of as LMPPs with or identified were MEPs map, HSCs either diffusion from our originating branches From lineage 22). two (21, profiles expression individual rdbr ttetpo h etasidct h ye fclsaogthe along cells of types the indicate heatmaps the Col- of genes. ordering. top pseudotime on the clustering hierarchical at levels of left bars results later. the expression ored the on cells indicate Dendrograms factor heatmap trajectories. red the LMPP transcription of and and in MEP trajectory for changes differentiation pseudotime along the showing in Heatmaps early (B) cells blue with ehdwsnee,hwvr oietf h otsuit- most the identify to however, needed, was method A is however, modeling, network Boolean of limitation clear A of both or one in dynamics strong displayed factors Several Notch Myb nrae ln h MPtaetr e a largely was yet trajectory LMPP the along increased oee,wsseiclyatvtddrn dif- during activated specifically was however, Gata1, xrsinicessaogbt rjcoisbut trajectories both along increases expression aito ngn expres- gene in Variation NSEryEdition Early PNAS | f8 of 3

BIOPHYSICS AND COLLOQUIUM COMPUTATIONAL BIOLOGY PAPER A Correlation network Single-cell Boolean qRT-PCR network data inference Order cells in Fig. 3. Single-cell molecular profiles allow inference of regulatory network models. ( ) pseudotime A Schematic showing the network inference steps starting from gene expression profiling using B C single-cell qRT-PCR data. (B) Potential regula- Gene correlation network gives Score functions using pseudotime ordering tors of each gene are identified by calculat- potential Boolean functions ing a pairwise gene–gene correlation network. The highest correlating gene pairs are linked G1 in the gene network. Activating (red edge) or G8 G2 repressing (blue edge) relationships correspond to positive or negative correlations, respectively. G7 G3 The regulators of each gene then define a set of potential Boolean functions governing the (I0,O0) (I1,O1) (I2,O2) G6 G4 expression of that gene. Three of the possible functions for G1 are shown here. (C) The pseudo- G5 time trajectory is then used to identify the most suitable Boolean functions. Cells are ordered in pseudotime (based on continuous expression data) and then converted to binary expression. ? Pairs of cells a fixed distance apart then repre- Fa G2 → G1 F(I ) sent input–output pairs to the Boolean function. Ik F k O G2 Λ G3 → G1 k These pairs are used to score a Boolean function Fb F by comparing F(Ik) to Ok for a pair (Ik, Ok). The G2 V G3 → G1 highest scoring function is the one where these Fc Input Boolean Predicted Observed

... Function Output Output values agree for the greatest number of pairs.

well as providing a score for the Boolean functions, this method MEP stable states having expression close to cells on the MEP can also enable a direction of regulation to be inferred from trajectory. Similarly, for the LMPP network model, stable states the undirected correlation network. Using this method, we iden- were found that either closely or exactly matched the expression tified potential transcriptional regulatory network models for profiles of LMPP cells from the primary bone marrow data (SI differentiation from HSCs to MEPs and LMPPs, with regula- Appendix, Fig. S3). To visualize how closely these stable states tory rules for each gene given by the highest scoring Boolean matched the location of cells sorted on LMPP and MEP surface functions (see SI Appendix, Table S1 for full set of results). markers, we also projected the stable states onto the diffusion Examples of the dynamic expression patterns seen in Fig. 2B can map (Fig. 4B). Close matches between the sorted qRT-PCR data be readily explained by the Boolean rules, such as differences and the stable states were seen along the relevant lineage trajec- in Notch expression between the two trajectories. In the LMPP tories for both network models. trajectory, the expression increases throughout differentiation, Stable state analysis identifies all stable states of the network whereas the majority of cells on the MEP differentiation trajec- model, regardless of whether they can be reached from a biolog- tory do not express Notch. Investigating the Boolean rules for ically meaningful starting condition. We therefore simulated the Notch shows that it is predicted within the LMPP trajectory to be network with initial conditions corresponding to binary expres- regulated via Lmo2 AND NOT (Gata2 AND Gfi1b). A similar sion in MolO cells (see Materials and Methods for details). Sim- rule was found as one of the alternatives for the MEP trajec- ulations starting from several of the MolO binary states could tory [Lmo2 AND NOT (Gata2 OR Gfi1b) activates Notch]. The stabilize on both the MEP and LMPP binary states when simu- different behavior of Gata2 and Gfi1b along both trajectories lated with the relevant networks, demonstrating that the two net- can account for the different dynamics of Notch expression, as work models could recapitulate differentiation trajectories from Gata2 and Gfi1b are downregulated toward LMPPs but remain HSCs to MEPs and LMPPs, respectively. expressed in MEPs. Differences in Network Model Connectivity Are Supported by Tran- Stable State Analysis of Network Models Identifies States Corre- scription Factor Binding. Given the differences in dynamic expres- sponding to in Vivo Cell Types. The Boolean network models sion of genes such as Notch and Gata1 between the two reconstructed from pseudotime ordering of LMPP and MEP dif- differentiation trajectories, it was not unexpected that the ferentiation trajectories were found to have complex structures, inferred Boolean networks for the two trajectories show differ- with each gene receiving inputs from an average of four upstream ences in the regulatory rules for some genes. Comparing rules regulators, often as part of composite Boolean functions such as in the two network models highlighted a trio of genes with “(Notch AND Tcf7) AND NOT Etv6 activates Ets1.” Simplified regulation unique to the MEP network model (Fig. 5A). In graphical representation, depicting regulation as only activation the MEP network model, Gata2 positively regulates Cbfa2t3h or repression rather than the Boolean AND/OR relations form- and Nfe2, with this regulation not present in the LMPP net- ing the regulatory rules, illustrates the highly connected nature work model. Classical assays for the functional validation of the of both networks (Fig. 4A and SI Appendix, Fig. S2). To assess specificity of regulatory relationships require the use of model whether the reconstructed LMPP and MEP network models cell lines. We therefore considered previously published single- were able to recapitulate HSPC differentiation, we identified the cell expression profiles of the 416B myeloid progenitor cell line stable states of both models. Importantly, this analysis demon- (26, 27), which can be induced toward megakaryocyte differ- strated that, within the set of stable states for the MEP network entiation (28). Projection of the 416B expression profiles onto model, there were several states exactly matching binary gene our bone marrow HSPC diffusion map indeed demonstrated expression profiles of MEP but not LMPP cells, with the other that 416B cells occupy a territory that forms part of the MEP

4 of 8 | www.pnas.org/cgi/doi/10.1073/pnas.1610609114 Hamey et al. Downloaded by guest on October 1, 2021 Downloaded by guest on October 1, 2021 ae tal. et Hamey factors transcription the via inde- net- regulated factor predicted being [growth (Gfi1b) MEP reported 1B the previously pendent for been rules have model Boolean work the of proposed Several and Nfe2. relationships Gata2 regulatory between the silico in validate Luciferase to and served pro- ChIP-Seq assays both the Therefore, with both differentiation. of MEP consistent activation Gata2 therefore of role was posed reduced and significantly in activity, resulted luciferase sites binding Gata2 Moreover, of 5D). mutation over (Fig. activation constructs fold control significant promoter/enhancer-less demonstrated showed cells sites constructs 416B wild-type binding in that Gata2 assays reporter relevant Luciferase with mutated. constructs corresponding the by as well gener- as we the moter, for model, constructs our reporter by ated predicted as activation transcriptional level. lower binding much a Gata2 at again, was and, cells cells, HoxB8-FL 416B in in region our enhancer −7-kb of specificity the At the model. with the consistent at observed cells, Gata2-activates-Cbfa2t3h was HoxB8-FL binding in LMPP Gata2 locus marrow limited bone express very primary cells of only in minority small like HoxB8- a just only of cells, that, profiling showed single-cell cells Our repre- FL region. being peaks promoter conserved minimal two most the The promoter, the cells. full and 416B minimal in the region sent promoter the at fied to cells HoxB8-FL and to the in (27) Gata2 Gata2 of cells for binding 416B investigate data in ChIP-Seq Gata2 new for generated data (ChIP-Seq) LMPP Sequencing the from cells the marrow state bone expression onto primary HoxB8-FL trajectory. projected of the when that for that resembles which, profiles confirmed cells, expression plot, poten- generated diffusion single lymphoid also HoxB8-FL and therefore myeloid 107 We both (29). have was tial line to cell reported HoxB8-FL The recently 5B). (Fig. trajectory differentiation ovldt hte h idn fGt2i 1Bclscauses cells 416B in Gata2 of binding the whether validate To Immunoprecipitation Chromatin existing interrogated We B A Cbfa2t3h ou,topoietbnigpaswr identi- were peaks binding prominent two locus, Nfe2 ou,apoietpa a dnie tthe at identified was peak prominent a locus, Nfe2 uefudol nteMPnetwork MEP the in only found rule nacr hc eecomplemented were which enhancer, n ewe aa and Gata2 between and Cbfa2t3h, Cbfa2t3h Cbfa2t3h Cbfa2t3h and naccordance, In Gata2. iia n ulpro- full and minimal Nfe2 and .At 5C). (Fig. Nfe2 Cbfa2t3h during aat icvrrgltr eainhp,btol oue on focused only but relationships, regulatory single-cell discover used bulk have to to studies limited data Several was data. and single-cell feasible, than always rather not is which experimental multiple conditions, in of profiling relied results expression study gene the performing this predict on However, to perturbations. able network was in and experimental network (24), pluripotency cells the stem study model recent embryonic A to limitation. abstraction this from Boolean suffer used not does this net- in it used the that as is in modeling, work, network factors Boolean of transcription advantage can- An therefore between work. and feedback graphs capture acyclic to not limited is Bayesian topology However, simulated. network be and to efficient perturbations network computationally allow are as which such (27), methods networks using Bayesian modeled successfully factor been Transcription have data. networks expression gene from models disorders. network these in play decisions fate cell into subverted insights that provide role can the cells regu- stem mechanisms blood the of differentiation discovering lating con- (31), blood leukemia serious as As to such blood. linked ditions the is in under- programs regulated regulatory our are of improve decisions disruption fate will cell hypotheses perturbations how test of network standing and of develop effect to the models on Using laboratory. the feasible approaches in experimental per- to network relate simulate readily to single-cell and used turbations by be easily captured can information models Boolean dynamic data. the mod- exploit network to Boolean eling with trajectories pseudotime about ideas evidence. experimental of by supported control be Gata2 identified contrast- we Cbfa2t3h rules, By network to lineages. the profiling blood ing alternative expression capturing models toward gene network differentiation single-cell regulatory used transcriptional two we define study, this In Discussion 30). (27, study net- this Boolean in proposed fac- the model Ets of work via utility the regulated reiterating being thereby (Fli1) or tors], 1 (Tal1)/Ets/Gata integration 1 leukemia protein friend leukemia lymphocytic acute T-cell aymtoseitwt h i fcntutn regulatory constructing of aim the with exist methods Many developed recently combines method inference network Our nqet h E ewr oe,wihw on to found we which model, network MEP the to unique uldsrpino ola ue o both for rules in Boolean available of is repres- networks description arrow. and flat-headed full blue arrow, indi- a A with pointed is indicated red Activation is sion LMPPs. a to with HSCs cated HSCs or HSC from MEPs for differentiation to for regulatory networks models Transcriptional network the (A) of differentiation. relevance logical 4. Fig. h rmr oemro aaecp o sin- a for gene. except matched gle data state marrow stable bone the primary that the a means and 1 data, exact of marrow an value bone primary indicates the zero to of match closely value values; a how expression example, indicates measured for binary state matches each it of of intensity color The the circles). red/blue high- (large and map lighted was diffusion the data) into nonbinary projected then the gene (in near- binary neighbors nearest these the its of expression in average state, The found data. expression stable were each neighbors est For (small data points). qRT-PCR of gray marrow map bone (blue) diffusion primary LMPP the the and onto (red) projected MEP networks of states Stable (B) tbesaeaayi eosrtsbio- demonstrates analysis state Stable NSEryEdition Early PNAS al S1 Table Appendix, SI Nfe2 | f8 of 5 and .

BIOPHYSICS AND COLLOQUIUM COMPUTATIONAL BIOLOGY PAPER A B C MEP Network only: Gata2 binding to Cbfa2t3h mm10 20 kb

Gata2 _ 246

416B 416B WT DC1 HoxB8-FL _ 0 Nfe2 Cbfa2t3h In vivo _ 246 Min HoxB8-FL

DC2 _ 0 Cbfa2t3h Promoter Gata2 binding to Nfe2 D Fold change in luciferase acvity at Fold change in luciferase acvity at mm10 5 kb Cbfa2t3hpromoter Nfe2 enhancer _ 399 1.2 1.2 416B 1 1 0.8 0.8 _ 0 *** 0.6 *** 0.6 * _ 399

Fold Change 0.4

Fold Change 0.4 HoxB8-FL 0.2 0.2

0 0 _ 0 WT Gata2 mutant WT Gata2 mutant WT Gata2 mutant Nfe2 Cbfa2t3h promoter Cbfa2t3hmin promoter Nfe2 enhancer -7kb enhancer

Fig. 5. Regulatory relationships unique to the MEP network model are supported by transcription factor binding. (A) Diagram of the trio of genes with a regulatory pattern identiﬁed as unique to the MEP network model. Red arrows indicate binding and positive regulation of genes by Gata2. (B) Diffusion map with projected qRT-PCR data for 416B and HoxB8-FL cells, showing gene expression similarities between the cell lines and in vivo data. (C) ChIP-Seq analysis of Gata2 in 416B and HoxB8-FL cell lines, showing Gata2 binds the Cbfa2t3h promoter in 416B cells only, and binds the Nfe2 enhancer in both cell lines but with greater binding in 416B cells. (D) Fold change in luciferase activity at the Cbfa2t3h promoter and Nfe2 enhancer, comparing the wild-type and Gata2 mutant regulatory regions. (*P < 0.05, **P < 0.01, ***P < 0.001; two-tailed unpaired t test, n = 3 ± SD.)

simple correlation analyses (6, 7), which cannot infer the direc- involved in HSC cell fate decisions made by single cells. Future tion of regulation without additional experimental data. work may focus on expanding the set of profiled genes by using More recently, single-cell gene expression data have been used other high-throughput single-cell approaches, such as RNA-seq, to construct Boolean network models, but the models either which may also resolve heterogeneities within HSPC populations relied on the assumption of cells being in a steady-state (8), which linked to fate choices (3, 32, 33). However, single-cell RNA-seq is is not applicable to differentiating systems, or only used binary currently less sensitive than qRT-PCR, which presents its own set gene expression data, thereby losing information present in the of challenges for network inference methods. level of gene expression (10). Regulatory factors in Boolean The two MEP trajectory-specific network rules we identified, rules with many OR logic inputs could play different roles in the namely, the positive regulation of Cbfa2t3h and Nfe2 by Gata2, regulation of expression levels, which would not be captured by are both consistent with the known biological functions of the the Boolean model. For example, when Gata2 is predicted by genes involved. Cbfa2t3h functions as a key component of multi- our model to act in OR logic control of Nfe2, this would lead to meric transcription factor complexes that regulate both erythroid the prediction that the loss of Gata2 is as important as any of and megakaryocytic expression programs (34–36), whereas Nfe2 the other factors also involved in the OR rules. This may not be was originally discovered as an upstream regulator of globin gene true in vivo, as the relative expression levels of genes will vary, expression (37) and is required for megakaryocyte maturation and loss of a transcription factor that is very highly expressed (38). Gata2 is primarily recognized as a regulator of HSPC func- may have different consequences than that of the loss of a factor tion (39, 40). It is involved in HSC maintenance and expansion, that is lowly expressed. An alternative approach, using the pseu- playing a role in early hematopoietic cell formation (41), where dotime ordering of single-cell expression profiles to construct Gata2 knockout mice display defects in primary hematopoiesis an ordinary differential equation network model, was recently (42). Cbfa2t3h encodes the transcription factor Eto2, a core- described (11). This approach can model more sensitive changes pressor in complex with Scl/Tal1 (43). Gata2 binds and activates in gene expression levels, but is limited to smaller networks. Cbfa2t3h and, during differentiation, Eto2 represses its own pro- We believe that the ability of our method to simulate and infer moter, leading to erythroid maturation and a Gata1-driven tran- larger networks is a reasonable trade-off for modeling binary scriptional program (44). Directly linking Gata2 to Cbfa2t3h and gene expression states. Nfe2 in the MEP regulatory network model but not the LMPP Our method is particularly useful for studying regulation of network model therefore provides an illustration of how differ- differentiation processes, as it uses the dynamic pseudotime ences in network topology guide the interaction between HSPC ordering to identify regulatory rules. A limitation of using qRT- regulators such as Gata2 and more lineage-restricted regulators PCR profiling is that it can only measure the expression of a lim- such as Cbfa2t3h and Nfe2. Interestingly, although Cbfa2t3h is ited number of genes. This restriction will affect the accuracy of traditionally reported to be a corepressor, our model predicts the pseudotime ordering and means some important regulatory that it would activate several genes in the network. This activity relationships cannot be described in the network, as the relevant could be directly a result of Cbfa2t3h (depending on its cofactors) genes were not included. Nevertheless, this study demonstrates an or a double repressive link (involving a gene not included in our advantage to performing single-cell rather than bulk expression dataset). An important area of future research will be the iden- analysis, as it allows the construction of differentiation trajectories tification of the mechanisms that direct stem cells into entering (11, 21, 22) and the reconstruction of transcriptional relationships specific differentiation trajectories. By identifying and validating

6 of 8 | www.pnas.org/cgi/doi/10.1073/pnas.1610609114 Hamey et al. Downloaded by guest on October 1, 2021 Downloaded by guest on October 1, 2021 (O formedfrom,atmost, and function f activators possible (I taking by generated then tr e oadfutof default a to set eters functions The 2. in-degree of nodes OR and nodes AND to restricted (c cells ordered ae tal. et Hamey vector a by represented is cell (ON/OFF) each ( binary in to where expression converted After Binary was order. expression pseudotime expression. the gene from ordering, pairs input–output val- pseudotime the correlation generate of to (distribution in used for edges available self-activation potential plus is as pairs, ues correlating taken 100 were top gene, The each zero. significance to with func- set coefficients were using Correlation 0.01 package. factors R transcription ppcor of the pairs from tions all on calculated were coefficients Construction. Network algorithm Wanderlust the (21). using branch parameters pseudotime default the in with and ordered found, then was were cell path this end the to the start algorithm, from formed the Dijkstra’s from Using path components. shortest diffusion the four first then the were on end Branches distance and populations. a start on these constructing a within by highlighted select from identified were to trajectory of used cells each was method LMPP for visualization the and cell this following MEP and identified map, MolO, diffusion Briefly, were the (11). LMPP) al. to et HSC Ocone and MEP to (HSC from downloaded Inference. function be Pseudotime can dm.predict data the qRT-PCR All using package. map R blood.stemcells.cam.ac.uk/single diffusion Destiny the the and Destiny from onto distance the cosine projected using centered (see were using calculated 45) (46) were package 6, reductions R (2, dimensionality described map previously fusion as performed Analysis. was Expression Gene Single-Cell plate. PCR 96-well a of wells individual into mice C57BL/6 12-wk-old to EasySep 8- the of using lineage-depleted Technologies). crest were (STEMCELL Cells lysis iliac chloride ammonium and by red-cell-depleted tibiae, and femurs, Cells. the Progenitor from and Stem of Purification Methods and Materials systems. be biological will differentiating approach other inference to network applicable our widely and differen- regulated, HSC capture is how about tiation models hypotheses network provide Our and biology differences models. known of network validation two provide the blood and between in cells, models progenitor network and regulatory describe stem differentia- to this cell use We governing tion. programs regulatory transcriptional mutation. by cell affected specific be as well may as alone, that interact, types knockouts that genes gene of from combinations as available such can not in we information and investigations, gain vitro silico in in also these guide doing to By inference investigations. network vivo the silico show in we using models, network of LMPP value and MEP the in rules simple slto fFRHC,MP,adPeeE r itdin listed are for PreMegEs and S2 used MPPs, Antibodies FSR-HSC2, Technologies). of (STEMCELL isolation Kit Enrichment Cell genitor .Wlo K ta.(05 obndsnl-elfntoa n eeexpression myeloid in commitment gene lineage and heterogeneity and Transcriptional (2015) al. functional et F, Paul single-cell 3. Combined (2015) al. paradigmatic et The cells: NK, stem Wilson Hematopoietic (2006) 2. IL Weissman DJ, Rossi D, Bryder 1. = igeclswr otd sn v-ae etnDcisnIflxsorter, Influx Dickinson Becton five-laser a using sorted, were cells Single . t o ee h de fteprilcreainntokdfieisst of sets its define network correlation partial the of edges the gene, a For ncnlso,w rsn nagrtmfrdsoeigthe discovering for algorithm an present we conclusion, In progenitors. populations. cell stem within 712–724. heterogeneity resolves analysis cell. stem tissue-specific ) f j 1 = ¬f ∧ (c c i t ) ) j 2 j . seult h xrsinlvlo gene of level expression the to equal is eecniee,where considered, were f Cell 2 0 ersnsterpesn at Here part. repressing the represents c , n 163(7):1663–1677. = {a 1 .Ase-ieparameter step-size A ). S4 Fig. Appendix, SI , 0 ers egbr fec elo hspt.Clson Cells path. this on cell each of neighbors nearest 100 c ..., i d } 1 eoepedtm reig w ieg branches lineage two ordering, pseudotime Before ocntuta nta ewr,prilcorrelation partial network, initial an construct To d mJPathol J Am n repressors and and 1 n = .Testo nu–uptpairs input–output of set The ). d k d 4, 2 t = ) regulatingTFs,respectively,withtheseparam- j 2 0naetniho rp sn Euclidean using graph nearest-neighbor 30 cell = = 169(2):338–346. o ahgene. each for 2 f mode[(c qpcr.html. 1 igecl eeepeso analysis expression gene Single-cell {r ersnsteatvtn ato the of part activating the represents j ola functions Boolean }. oemro el eeisolated were cells marrow Bone t −k TM −1 os eaooei Pro- Hematopoietic Mouse ) j f , σ j (c saBoenfunction Boolean a is = t j −k Table Appendix, SI IAppendix SI ncell in .Cl iedata line Cell 0.3. elSe Cell Stem Cell ) j , {(I (c f t t −k O , fteform the of i Consider . f k +1 1 t f , = )} ) 2 j .Dif- ). was 3 ] were were 16(6): and c > i , {(I genes activating For 4)(see (49) Assays. Luciferase Omnibus Expression Gene NCBI the GSE84328. to identifier submitted with been have data ChIP-Seq (see (48) described previously Sequencing. Immunoprecipitation Chromatin function dm.predict expression the average using package. map R the Destiny diffusion and the the from neighbors, onto expres- gene projected across continuous was averaged the state match, were best If levels the map. state, diffusion was sion stable the neighbor in each one highlighted than For and more data. identified was expression neighbor marrow nearest its bone primary was binary simulation the with of cor- state states end expression stable binary the 237 and recorded. the cells, of MolO each simula- to from 1,000 responding starting networks, LMPP no run and and were and MEP tions stabilized R For in network expression. change encoded the could were until genes network updates from a asynchronous reachable for with states rules simulated the Boolean identify To points, (47). starting time) MolO a at randomly changing one gene of chosen (expression updates asynchronous using algorithm GenYsis Analysis. State Stable rules. two other the contains n uoenHmtlg soito o-lnclAvne Research Advanced Non-Clinical Association Hematology Fellowship. European (15008) Fellowship PhD a Bennett Council Bloodwise Research and a Medical by Stem supported of Medical is Cambridge recipients D.G.K. for are Studentships. F.K.H. Council Institute and Research Cambridge S.N. Institute. Trust–Medical the Cell sup- to Wellcome core Trust and Wellcome and Research the Center by Research Institute grants Biomedical National port Cambridge the Research and grants Health Society, by for Lymphoma supported Sciences Leukemia Biological net- is and Council, the Biotechnology laboratory Research UK, implement B.G.’s Research to Cancer in Bloodwise, code Work isolation, from writing method. cell technical with marrow inference for help bone work Hamilton for Tina with Woodhouse and assistance Steven Pask Core for and Dean Cytometry Shepherd sorting, Flow Mairi cell Research assistance, with Medical help for their Institute for Cambridge the at Maj ACKNOWLEDGMENTS. functions three reduce if to multiple example, simplified way, For Gata1 also contained rules. this was of rules and in number of simpler simplified total list the the be is possible, not If it retained. could as were rules rules chosen, When be latter. of would the set within former minimum the the functions to scores, simplified if equal were example, results For the functions. equally case, with simplest functions this several In gave scores. method high the genes, many For inference. can inference from network downloaded this performing be for code Python constraints. fiability activators pseudotime observed the with output. agrees input pseudotime the from calculated n ue o ahgn.Ti ehdwsipeetdi h yhnpro- Python the (z3.codeplex.com/ in solver implemented Z3 the was using method language This gramming gene. each for rules find score the maximize function(s) (i criteria: following the satisfying form above the .Minr ,e l 21)Caatrzto ftasrpinlntok nbodstem blood in networks transcriptional of Characterization (2013) al. et V, Moignard 6. (2015) EH Davidson I, Peter 5. stem haematopoietic Advancing (2016) B Gottgens NK, Wilson S, Nestorowa FK, Hamey 4. t opoetsal ttsot h ifso a,sae eecompared were states map, diffusion the onto states stable project To ofidterl o gene a for rule the find To function a of output predicted the times many how counts function This gene a for rules best the identify To h bv rtrawr noe saBoenstsaiiypolmto problem satisfiability Boolean a as encoded were criteria above The n rgntrclsuighg-hogptsnl-elgn xrsinanalysis. expression gene single-cell Biol Cell high-throughput using cells progenitor and Ed. 2nd York), New (Academic, profiling. cell single through biology cell progenitor and O , t ∨ )} Nfe2 IAppendix SI t m =1 15(4):363–372. {a ehave we , i S } eertre,te nyteO uewudb eand sit as retained, be would rule OR the only then returned, were (f n repressors and ) = uieaeasy eepromda rvosydescribed previously as performed were assays Luciferase X t =1 m o details). for ersiggenes repressing b, a, tbesae ftentokwr dnie sn the using identified were network the of states Stable https://github.com/fionahamey/Pseudotime-network- f etakRie cut,Cir osti n Michal and Cossetti, Chiara Schulte, Reiner thank We (I s t t (f ) = s ), eoi oto rcs:DvlpetadEvolution and Development Process: Control Genomic v f 1 h loih hnsace o ucin of functions for searches then algorithm the , IAppendix SI {r S [ (I t (f (f j t } ). ) ) a of , = (I t ( v ) b edfie h cr function score the defined we g, 1, 0, sdfie bv.(ii above. defined as ] ¬f ∧ o eal) a n processed and Raw details). for hPasy eepromdas performed were assays ChIP and s, r, otherwise if Gata1 2 [f [ (I NSEryEdition Early PNAS (I ) t t ESLett FEBS f ) )] r 1 , and g = (I m = t ) f s nu–uptpairs input–output 1 (O Gata1 ] (a Gata1, and oecd satis- encode to ) 590(22):4052–4067. t i ) ), g h allowed The ) . f 2 ∧ = and Gfi1b, Nfe2 f | 2 (r f8 of 7 j for ) had Nat S .

BIOPHYSICS AND COLLOQUIUM COMPUTATIONAL BIOLOGY PAPER 7. Pina C, et al. (2015) Single-cell network analysis identifies DDIT3 as a nodal lineage 29. Redecke V, et al. (2013) Hematopoietic progenitor cell lines with myeloid and lym- regulator in hematopoiesis. Cell Rep 11(10):1503–1510. phoid potential. Nat Methods 10(8):795–803. 8. Xu H, Ang YS, Sevilla A, Lemischka IR, Ma’ayan A (2014) Construction and validation 30. Pimanda JE, et al. (2007) Gata2, Fli1, and Scl form a recursively wired gene- of a regulatory network for pluripotency and self-renewal of mouse embryonic stem regulatory circuit during early hematopoietic development. Proc Natl Acad Sci USA cells. PLoS Comput Biol 10(8):e1003777. 104(45):17692–17697. 9. Chen H, et al. (2015) Single-cell transcriptional analysis to uncover regulatory cir- 31. Wilkinson AC, Gottgens¨ B (2013) Transcriptional Regulation of Haematopoietic Stem cuits driving cell fate decisions in early mouse development. Bioinformatics 31(7): Cells, eds Hime G , Abud H (Springer, Dordrecht, The Netherlands), pp 187–212. 1060–1066. 32. Moignard V, Gottgens¨ B (2014) Transcriptional mechanisms of cell fate decisions 10. Moignard V, et al. (2015) Decoding the regulatory network of early blood develop- revealed by single cell expression profiling. Bioessays 36(4):419–426. ment from single-cell gene expression measurements. Nat Biotechnol 33(3):269–276. 33. Buettner F, et al. (2015) Computational analysis of cell-to-cell heterogeneity in single- 11. Ocone A, Haghverdi L, Mueller NS, Theis FJ (2015) Reconstructing gene regulatory cell RNA-sequencing data reveals hidden subpopulations of cells. Nat Biotechnol dynamics from high-dimensional single-cell snapshot data. Bioinformatics 31(12): 33(2):155–160. i89–i96. 34. Goardon N, et al. (2006) ETO2 coordinates cellular proliferation and differentiation 12. Jaeger J (2009) Modelling the Drosophila embryo. Mol BioSyst 5:1549–1568. during erythropoiesis. EMBO J 25(2):357–366. 13. Ashyraliyev M, et al. (2009) Gene circuit analysis of the terminal gap gene huckebein. 35. Fujiwara T, Lee HY, Sanalkumar R, Bresnick EH (2010) Building multifunctionality into PLoS Comput Biol 5(10):1–16. a complex containing master regulators of hematopoiesis. Proc Natl Acad Sci USA 14. Peter IS, Faure E, Davidson EH (2012) Predictive computation of genomic logic 107(47):20429–20434. processing functions in embryonic development. Proc Natl Acad Sci USA 109(41): 36. Hamlett I, et al. (2008) Characterization of megakaryocyte GATA1-interacting pro- 16434–16442. teins: The corepressor ETO2 and GATA1 interact to regulate terminal megakaryocyte 15. Pronk CJH, et al. (2007) Elucidation of the phenotypic, functional, and molecu- maturation. Blood 112(7):2738–2749. lar topography of a myeloerythroid progenitor cell hierarchy. Cell Stem Cell 1(4): 37. Ney PA, et al. (1993) Purification of the human NF-E2 complex: cDNA cloning of the 428–442. hematopoietic cell-specific subunit and evidence for an associated partner. Mol Cell 16. Coifman RR, et al. (2005) Geometric diffusions as a tool for harmonic analysis Biol 13(9):5604–5612. and structure definition of data: Diffusion maps. Proc Natl Acad Sci USA 102(21): 38. Shivdasani RA, et al. (1995) Transcription factor NF-E2 is required for platelet forma- 7426–7431. tion independent of the actions of thrombopoietin/MGDF in megakaryocyte devel- 17. Haghverdi L, Buettner F, Theis FJ (2015) Diffusion maps for high-dimensional single- opment. Cell 81(5):695–704. cell analysis of differentiation data. Bioinformatics 31(18):2989–2998. 39. Rodrigues NP, et al. (2005) Haploinsufficiency of GATA-2 perturbs adult hematopoietic 18. Scialdone A, et al. (2016) Resolving early mesoderm diversification through single-cell stem-cell homeostasis. Blood 106(2):477–484. expression profiling. Nature 535(7611):289–293. 40. Lim KC, et al. (2012) Conditional Gata2 inactivation results in HSC loss and lymphatic 19. Nestorowa S, et al. (2016) A single cell resolution map of mouse haematopoietic stem mispatterning. J Clin Invest 122(10):3705–3717. and progenitor cell differentiation. Blood 128(8):e20–e31. 41. Tsai FY, Orkin SH (1997) Transcription factor GATA-2 is required for prolifera- 20. Setty M, et al. (2016) Wishbone identifies bifurcating developmental trajectories from tion/survival of early hematopoietic cells and mast cell formation, but not for ery- single-cell data. Nat Biotechnol 34(6):637–645. throid and myeloid terminal differentiation. Blood 89(10):3636–3643. 21. Bendall SC, et al. (2014) Single-cell trajectory detection uncovers progression and reg- 42. Tsai FY, et al. (1994) An early haematopoietic defect in mice lacking the transcription ulatory coordination in human B cell development. Cell 157(3):714–725. factor GATA-2. Nature 371(6494):221–226. 22. Trapnell C, et al. (2014) The dynamics and regulators of cell fate decisions are revealed 43. Schuh AH, et al. (2005) ETO-2 associates with SCL in erythroid cells and megakaryo- by pseudotemporal ordering of single cells. Nat Biotechnol 32(4):381–386. cytes and provides repressor functions in erythropoiesis. Mol Cell Biol 25(23): 23. Bonzanni N, et al. (2013) Hard-wired heterogeneity in blood stem cells revealed using 10235–10250. a dynamic regulatory network model. Bioinformatics 29:i80–i88. 44. Fujiwara T, et al. (2009) Discovering hematopoietic mechanisms through genome- 24. Dunn SJ, Martello G, Yordanov B, Emmott S, Smith AG (2014) Defining an essential wide analysis of GATA factor chromatin occupancy. Mol Cell 36(4):667–681. transcription factor program for na¨ıve pluripotency. Science 344(6188):1156–1160. 45. Guo G, et al. (2010) Resolution of cell fate decisions revealed by single-cell gene 25. Woodhouse S, Moignard V, Gottgens B, Fisher J (2016) Processing, visualising and expression analysis from zygote to blastocyst. Dev Cell 18(4):675–685. reconstructing network models from single-cell data. Immunol Cell Biol 94(3): 46. Angerer P, et al. (2016) Destiny: Diffusion maps for large-scale single-cell data in R. 256–265. Bioinformatics 32(8):1241–1243. 26. Dexter TM, Allen TD, Scott D, Teich NM (1979) Isolation and characterisation of a 47. Garg A, Cara AD, Xenarios I, Mendoza L, Micheli GD (2008) Synchronous versus asyn- bipotential haematopoietic cell line. Nature 277(5696):471–474. chronous modeling of gene regulatory networks. Bioinformatics 24(17):1917–1925. 27. Schutte J, et al. (2016) An experimentally validated network of nine haematopoietic 48. Wilson NK, et al. (2010) Combinatorial transcriptional control in blood stem/ transcription factors reveals mechanisms of cell state stability. Elife 5:e11469. progenitor cells: Genome-wide analysis of ten major transcriptional regulators. Cell 28. Visvader J, Adams J (1993) Megakaryocytic differentiation induced in 416B myeloid Stem Cell 7(4):532–544. cells by GATA-2 and GATA-3 transgenes or 5-azacytidine is tightly coupled to GATA-1 49. Bockamp E, et al. (1995) Lineage-restricted regulation of the murine SCL/TAL-1 pro- expression. Blood 82(5):1493–1501. moter. Blood 86(4):1502–1514.

8 of 8 | www.pnas.org/cgi/doi/10.1073/pnas.1610609114 Hamey et al. Downloaded by guest on October 1, 2021