©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION �

r e s o u r c e

The differentiation of hematopoietic stem cells into cells of the immune system has been studied extensively in mammals, but the transcriptional circuitry that controls it is still only partially understood. Here, the Immunological Genome Project gene-expression profiles across mouse immune lineages allowed us to systematically analyze these circuits. To analyze this data set we developed Ontogenet, an algorithm for reconstructing lineage-specific regulation from gene-expression profiles across lineages. Using Ontogenet, we found differentiation stage–specific regulators of mouse hematopoiesis and identified many known hematopoietic regulators and 175 previously unknown candidate regulators, as well as their target genes and the cell types in which they act. Among the previously unknown regulators, we emphasize the role of ETV5 in the differentiation of gd T cells. As the transcriptional programs of human and mouse cells are highly conserved, it is likely that many lessons learned from the mouse model apply to humans.

­programs­of­human­and­mouse­cells­are­highly­conserved4,­many­lessons­learned­from­the­mouse­model­will­probably­be­applicable­to­humans.­Two­key­approaches­for­the­identification­of­regulatory­networks5­are­physical­models­based­on­the­association­of­a­transcrip-tion­factor­or­a­cis-regulatory­element­with­a­target’s­promoter­(for­example,­from­chromatin­immunoprecipitation­(ChIP)­followed­by­deep­sequencing)­and­observational­models­with­which­regulation­can­be­inferred­from­a­significant­correlation­between­the­abundance­or­activity­of­a­transcription­factor­(as­protein­or­mRNA)­and­that­of­its­presumed­target.­In­both­cases,­analysis­of­the­relationship­between­a­putative­regulator­and­a­module­of­coregulated­targets­enhances­robustness­and­biological­ interpretability6,7.­Physical­data­provide­direct­evidence­of­biochemical­ interactions­but­do­not­necessarily­indicate­function8­and­are­challenging­to­collect9,­whereas­mRNA­profiles­are­highly­accessible­but­provide­only­correlative­evidence.­As­physical­and­observational­models­are­complementary,­using­both3­can­enhance­confidence5–7­and­expand­the­scope­of­discovery.

Analysis­of­cells­organized­in­a­known­lineage,­as­in­hematopoiesis,­offers­unique­opportunities­that­have­not­been­leveraged­before.­In­particular,­published­models3­have­not­explicitly­considered­the­fact­that­cells­that­are­more­closely­related­(according­to­the­known­line-age­tree)­probably­share­many­of­their­regulatory­mechanisms­and­that­regulatory­relationships­that­exist­in­one­sublineage­may­not­be­active­in­another.­Incorporating­such­information­may­help­in­the­identification­of­true­regulators­of­hematopoiesis.

Here­we­used­ those­ insights­ to­develop­a­new­computational­method,­Ontogenet,­and­to­apply­it­to­the­ImmGen­compendium­

1Computer Science Department, Stanford University, Stanford, California, USA. 2Broad Institute of MIT and Harvard, Cambridge, Massachusetts, USA. 3Department of Pathology, University of Massachusetts Medical School, Worcester, Massachusetts USA. 4Laboratory of Genetics, University of Wisconsin-Madison, Madison, Wisconsin, USA. 5Howard Hughes Medical Institute, Department of Biology, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. 6A full list of members and affiliations appears at the end of the paper. 7These authors contributed equally to this work. 8These authors jointly directed this work. Correspondence should be addressed to A.R. (aregev@broadinstitute.org) or D.K. (koller@cs.stanford.edu).

Received 30 January; accepted 13 March; published online 28 April 2013; doi:10.1038/ni.2587

The­Immunological­Genome­Project­(ImmGen)­is­a­consortium­of­immunologists­and­computational­biologists­who­aim,­through­the­use­of­shared­and­rigorously­controlled­data-generation­pipelines,­to­exhaustively­chart­gene-expression­profiles­and­their­underlying­regulatory­networks­in­the­mouse­immune­system1.­In­this­context,­we­provide­the­ first­comprehensive­analysis­of­ the­ImmGen­com-pendium­and­use­a­new­computational­algorithm­to­reconstruct­a­modular­model­of­the­regulatory­program­of­mouse­hematopoiesis.­Understanding­ the­ regulatory­ mechanisms­ that­ underlie­ the­ dif-ferentiation­of­cells­of­the­immune­system­has­important­implica-tions­for­the­study­of­development­and­for­understanding­the­basis­of­ human­ immunological­ disorders­ and­ hematological­ malignan-cies.­Most­studies­of­hematopoiesis­view­differentiation­as­a­process­controlled­ by­ relatively­ few­ ‘master’­ transcription­ factors­ that­ are­expressed­ in­specific­ lineages­and­act­ to­set­and­reinforce­distinct­cell­states2.­However,­analysis­of­gene­expression­in­38­cell­types­in­human­hematopoiesis3­has­suggested­a­more­complex­organization­that­ involves­a­ larger­number­of­ transcription­factors­ that­control­combinations­of­modules­of­coexpressed­genes­and­are­arranged­in­densely­ interconnected­ circuits.­ However,­ that­ human­ study­ was­restricted­to­human­cells­that­could­be­obtained­in­sufficient­quanti-ties­from­peripheral­or­cord­blood­and­thus­could­not­access­many­cell­populations­of­the­immune­system.

The­246­cell­types­of­the­mouse­immune­system­in­the­816­arrays­of­the­ImmGen­compendium­offer­an­unprecedented­opportunity­for­ studying­ the­ regulatory­ organization­ of­ hematopoiesis­ in­ the­­context­of­a­rich­and­diverse­lineage­tree.­Because­the­transcriptional­

Identification of transcriptional regulators in the mouse immune systemVladimir Jojic1,7, Tal Shay2,7, Katelyn Sylvia3, Or Zuk2, Xin Sun4, Joonsoo Kang3, Aviv Regev2,5,8, Daphne Koller1,8 & the Immunological Genome Project Consortium6

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

�  aDVaNCE ONLINE PUBLICaTION nature immunology

r e s o u r c e

to­build­an­observational­model­associating­578­candidate­regula-tors­with­modules­of­coexpressed­genes.­We­defined­modules­at­two­different­granularities,­with­81­larger­coarse-grained­modules,­some­of­which­we­further­refined­into­smaller­modules­with­more­coherent­expression;­ this­resulted­ in­334­ fine-grained­modules.­The­model­identified­many­of­the­already­known­hematopoietic­regulators,­ was­ supported­ through­ the­ use­ of­ a­ complementary­physical­model­and­proposed­dozens­of­previously­unknown­can-didate­regulators.­Our­model­provides­a­rich­resource­of­testable­hypotheses­ for­ experimental­ studies,­ and­ the­ Ontogenet­ algo-rithm­can­be­used­ to­delineate­ regulation­ in­ the­context­of­any­­cell­lineage.

RESULTSTranscriptional compendium of the mouse immune systemThe­ ImmGen­ consortium­ data­ set1­ (April­ 2012­ release)­ consists­­of­816­expression­profiles­from­246­cell­types­of­the­mouse­immune­­system­ (Fig. 1­ and­ Supplementary Table 1).­ The­ cell­ types­ span­­all­ major­ hematopoietic­ lineages,­ including­ stem­ and­ progenitor­­cells,­granulocytes,­monocytes,­macrophages,­dendritic­cells­(DCs),­­natural­ killer­ (NK)­ cells,­ B­ cells­ and­ T­ cells.­ The­ T­ cells­ include­­many­ types­ of­ αβ­ T­ cells,­ regulatory­ T­ cells­ (Treg­ cells),­ natural­­killer­ T­ cells­ (NKT­ cells)­ and­γδ­ T­ cells.­ The­ ‘same’­ cell­ type­ was­often­sampled­ from­several­ tissues,­ such­as­bone­marrow,­ thymus­and­spleen.

Similarities in global profiles trace the cell ontogenyCorrelations­in­global­profiles­between­samples­were­largely­consist-ent­with­the­known­lineage­tree­(Fig. 2).­In­general,­the­closer­two­cell­populations­were­in­the­lineage­tree,­the­more­similar­their­expression­profiles­were­(Pearson­r­=­–0.71;­Supplementary Fig. 1).­For­myeloid­cells,­profiles­were­similar­overall,­with­granulocytes­being­the­least­variable,­DCs­being­the­most­variable­(consistent­with­DC­samples’­being­obtained­from­diverse­tissues­and­their­known­inherent­diver-sity10)­and­all­myeloid­cells­being­weakly­ similar­ to­ stromal­cells.­Conversely,­lymphocytes­had­larger­differences­between­lineages.­NK­cells,­although­tightly­correlated,­did­show­weaker­similarity­to­T­cells,­especially­CD8+­T­cells­and­NKT­cells.­T­cells­were­very­heterogene-ous,­which­partly­reflected­the­finer­sampling­for­this­lineage.­Stem­cells­were­most­similar­to­early­myeloid­and­lymphoid­progenitors­(S&P­group,­Fig. 2),­followed­by­pre-B­cells­and­pre-T­cells,­consist-ent­with­a­gradual­loss­of­differentiation­potential.­As­a­resource­for­studying­each­lineage,­we­used­one-way­analysis­of­variance­to­define­characteristic­signatures­of­over-­and­underexpressed­genes­for­each­of­the­main­eleven­lineages­compared­with­the­expression­of­those­genes­in­all­other­lineages­(Supplementary Table 2).

Coarse- and fine-grained expression modules in hematopoiesisTo­characterize­the­key­patterns­of­gene­regulation,­we­next­defined­modules­of­coexpressed­genes­at­two­granularities­(Supplementary Fig. 2a,b).­We­first­constructed­81­coarse-grained­modules­(C1–C81;­

SC.LTSL.BM

CLP.BM

preT.ETP.Th

SC.STSL.BM

NK.49CI–.Sp NK.49CI+.Sp NK.49H+.Sp NK.49H-.Sp

NK.MCMV1.Sp

proB.FrA.BM

proB.FrBC.BM

preB.FrC.BM

preB.FrD.BM

B.FrE.BM

B.T1.Sp

B.T2.Sp

B.T3.Sp

B.Fo.Sp

B.GC.SpB.FO.PC

preT.DN2.Th

T.DN4.Th

T.4int8+.Th

T.8SP69+.Th

T.4SP24int.Th

Tgd.Th

T.DP69+.Th

T.4+8int.Th

T.4SP69+.Th T.4FP3+25+.AA

T.DPbl.Th

T.DPsm.Th

T.DP.Th

Lu

NK.SpT.8SP24int.Th

T.8SP24-.Th

T.8Nve.SpT.8Nve.LN T.8Nve.MLN T.8Nve.PP

T.8Mem.LN T.8Mem.Sp T.8Nve.Sp.OT1

T.8Mem.Sp.d45

T.8Mem.Sp.d106

T.4SP24-.Th

T.4Nve.Sp T.4Nve.PPT.4Nve.MLNT.4Nve.LN

T.4Mem.LN

T.4.Pa.BDC

T.4.PLN.BDC

T.4.LN.BDC

T.4Mem.Sp

T.4Mem44h62l.Sp T.4Mem44h62l.LN

NKT.44–NK1.1–.Th

NKT.44+NK1.1+.Th

NKT.44+NK1.1–.Th

NKT.4+.Sp NKT.4+Lv

NKT.4–.Sp

NKT.4+.Lu

NKT.4–.Lv

NK.H–.MCMV1.SpNK.H+.MCMV1.Sp

NK.H+.MCMV7.SpNK.MCMV7.Sp

B CTRL

T.8Eff.Sp.d5T.8Eff.Sp.48hr

T.8Eff.Sp.d8

T.8Eff.Sp.d10

T.8Eff.Sp.d6

T.8Eff.Sp.d8

T.8Eff.Sp.d15 T.8Eff.Sp.d15

MLP.BM

IIhilang+103+11bLO.SLN

IIhilang+103–11b+.SLN

IIhilang–103–11b+.SLN

IIhilang–103–11bLO.SLN

LC.Sk

SC.LTSL.FL

SC.STSL.FL

MLP.FL

proB.FrA.FL

proB.FrBC.FL

preB.FrD.FL

B.FrE.FL

B.1a.PC

Fetal Adult

T CTRL

RP.Sp

Thio5.II+480lo.PC

Thio5.II+480int.PC

Thio5.II–480int.PC

Thio5.II-480hi.PC

II–480hi.PC

4+

8–

8+

8–4–11b–

MLN Th Sp

6C+II+

6C–II–

6C+II–

6C–II+

BI LN

6C–IIint

B.MZ.Sp B1a.SpB1b.PC

8–4–11b+

SLN

BM

Microglia.CNS

SC.CMP.BM

SC.GMP.BM

BMBM

T.8Mem.Sp.d100

ARTH.BM

ARTH.SYNF

Bl

THIO.PC

URAC.PC

B.FO.LN B.FO.MLN

LvLu.LN IC

SC.MPP34F.BM

SC.ST34F.BM

SC.MEP.BM

preT.DN2-3.Th

T.ISP.Th

PreT.DN3-4.Th

T.4FP3+25+.LN T.4FP3+25+.Sp

T.8Eff.Sp.12hr

T.8Eff.Sp.24hr

T.8Eff.Sp.d6

LisOva.OT1 VSVOva.OT1

103+.BR103-.BR 103-.Sp

T.8Mem.d20

Granulocytes

Monocytes

Macrophages

103+.11b+

103+.11b–

SI S.SI

Vg2+.Sp

Vg2+.act.Sp

Vg1+Vd6–24–.Th

Vg1+Vd6–.Th

Vg1+Vd6–24+.Th

Tgd.Vg1+Vd6+24–.Th

Vg1+Vd6+.Th

Vg1+Vd6+24+.Th

Vg2+24–.Th

Vg2+24+.Th

Tgd.Sp

Vg2–.Sp

Vg2–.act.Sp

Vg5–.IEL

Vg5–.act.IEL

Vg5+.IEL

Vg5+.act.IEL

T.8Mem.Sp.d45

T.4FP3–.Sp

Putative edges

Nonintersecting edges

B.FrF.BM

CLP.FL

CD8+ Tcells

CD4+ Tcells

Treg cells

NKT cells

γδ T cells

DCsEp.MEChi

Fi

SLN

FRC

LEC

BEC

MLNSk

ST.31-38-44-

Th

Stromal cells

NK cellsB cellsDC MF-MO T cellsGN

SC.LT34F

SC.MDP.BM

SC.CDP.BM

103–.11b+24+.Lu

103-.11b+.Lu

103-.11b+.SI

103-.11b+.Salm.SI

11cloSer.SI

11cloSer.Salm.SI

II+480lo.PC

DC.103-11b+F4/80lo.Kd

103-.11b+

pDC.8–

pDC.8+

Figure 1 Mouse cell populations in the ImmGen compendium. Lineage tree of the hematopoietic mouse cell types profiled by the ImmGen Consortium (nomenclature and markers for sorting, Supplementary Table 1). Some samples of stem cells, progenitor cells and B cells were obtained from adult and fetal liver. Stromal cells (box, bottom right) were also measured as part of ImmGen but are not part of the lineage tree. Color in bar beneath matches color of branches in tree above. GN, granulocyte; MF-MO, macrophages-monocyte. Adapted from ref. 4.

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION �

r e s o u r c e

Supplementary Fig. 2c–h­ and­ Supplementary Table 3)­ and­ then­further­identified­for­each­coarse-grained­module­a­set­of­nested­fine­modules­(Supplementary Fig. 2a),­which­resulted­in­334­fine­mod-ules­spanning­7,965­genes­(F1–F334;­Supplementary Table 4).­Coarse­modules­helped­us­capture­the­mechanisms­that­coregulate­a­larger­set­of­genes­in­one­lineage,­whereas­fine­modules­may­help­in­the­identi-fication­of­distinct­regulatory­mechanisms­that­control­only­a­smaller­subset­of­these­genes­in­the­other­lineage(s).­Many­of­the­modules­showed­enrichment­for­coherent­functional­annotations,­cis-regulatory­­elements­(Supplementary Table 5)­and­binding­of­transcription­fac-tors­(Supplementary Table 6­and­Supplementary Note 1),­including­binding­sites­for­factors­known­to­act­as­regulators­in­the­lineage(s)­in­which­the­module’s­genes­are­expressed­(Supplementary Note 2).­All­modules­and­their­associated­enrichments­can­be­searched,­browsed­and­ downloaded­ at­ the­ ImmGen­ portal­ (http://www.immgen.org/ModsRegs/modules.html).

Most­ coarse-grained­ modules­ (48­ of­ 81­ modules;­ 4,478­ of­ 7,965­genes)­ showed­ either­ lineage-specific­ induction­ (Supplementary Figs. 2c and 3)­ or­ ‘pan-differentiation’­ regulation­ (Supplementary Figs. 2de, 4 and 5).­ In­addition,­6­modules­were­ ‘mixed-use’­across­lineages­ (Supplementary Figs. 2f and 6),­ 8­ were­ stromal­ specific­(Supplementary Fig. 2g)­and­19­had­expression­patterns­that­did­not­fall­into­those­categories­(Supplementary Figs. 2h and 7).­Lineage-specific­repression­was­rare­(only­in­C53­(B­cells)­and­C17­(stromal­cells)).

Ontogenet: reconstructing lineage-sensitive regulationWe­next­developed­a­new­algorithm,­Ontogenet,­to­delineate­the­regu-latory­circuits­that­drive­hematopoietic­cell­differentiation.­Ontogenet­aims­to­fulfill­the­following­biological­considerations:­criterion­1,­the­expression­of­each­module­of­genes­is­determined­by­a­combination­of­activating­and­repressing­transcription­factors;­criterion­2,­the­activity­of­those­factors­may­change­in­different­cell­types­(for­example,­factor­A­may­activate­a­module­in­one­lineage­but­not­in­another,­even­if­A­is­expressed­in­both­lineages);­criterion­3,­the­identity­and­activity­of­the­factors­that­regulate­a­module­are­more­similar­in­cells­that­are­close­to­each­other­in­the­lineage­tree­(for­example,­from­the­same­sublineage)­than­in­‘distant’­cells­(for­example,­from­two­different­sublineages),­in­accordance­with­the­greater­similarity­in­expression­profiles­of­closer­cell­types­(Supplementary Fig. 1);­and­criterion­4,­master­regulators­of­a­lineage­(for­example,­GATA-3­for­T­cells)­are­active­across­the­sublineages,­but­the­subtypes­can­also­have­additional,­more­specific­regulators­(for­example,­Foxp3­for­Treg­cells).­The­former­should­be­captured­as­shared­regulators­of­a­coarse­module­and­its­nested­fine­modules,­whereas­the­latter­regulate­only­particular­fine­modules.

Ontogenet­receives­as­input­the­gene-expression­module,­the­lineage­tree­and­the­expression­profiles­of­a­predesignated­set­of­‘candidate­regulators’­(transcription­factors,­chromatin­regulators­and­so­on).­­It­ then­ associates­ each­ module­ with­ a­ combination­ of­ regulators­(criterion­1­above),­whereby­each­regulator­is­assigned­an­‘activity­weight’­for­each­cell­type­that­indicates­its­activity­as­a­regulator­for­that­module­in­that­cell­(criterion­2­above).­The­regulator­activity­is­

at­the­protein­level­but­is­inferred­solely­from­transcript­abundance.­Following­the­approach­in­the­Lirnet­method­for­regulatory-network­reconstruction6,­the­activity-weighted­expression­of­the­regulators­is­combined­in­a­linear­model­to­generate­a­prediction­of­a­module’s­gene­expression­in­each­cell­type­(Fig. 3).­In­this­model,­the­expres-sion­of­the­module’s­genes­in­a­given­cell­type­is­approximated­by­the­linear­sum­of­the­regulators’­expression­in­that­cell­type­multiplied­by­each­regulator’s­activity­weight­in­that­cell­type.­As­a­result,­the­model­makes­predictions­such­as­“in­pre­B­cells,­Module­1­is­activated­by­transcription­factors­A­and­B­and­is­repressed­by­factor­C,­whereas­in­B­cells,­factors­A­and­C­are­no­longer­active­(even­if­the­factors­are­expressed),­and­Module­1­is­activated­by­B­and­D.”­Our­model­assumes­that­all­genes­in­the­same­module­are­regulated­in­the­same­way.­This­is­essential­for­statistical­robustness,­although­it­comes­at­the­cost­of­missing­some­gene-specific­expression­patterns.­The­fine­modules­let­us­examine­subtler­expression­patterns­shared­by­fewer­genes­but­are­more­susceptible­to­noise.

Although­Ontogenet­reconstructs­a­potentially­different­regulatory­program­for­each­cell­type,­as­reflected­by­the­cell-specific­activity­weights­of­each­regulator,­it­is­geared­toward­maintaining­the­same­activity­across­consecutive­stages­in­differentiation­(criterion­3).­This­is­achieved­by­penalizing­changes­in­the­activity­weights­of­the­regula-tory­program­between­a­cell­type­and­its­progenitor.­The­fine-grained­modules­derived­ from­a­coarse-grained­module­ ‘inherit’­ the­same­regulators­and­activity­weights­that­were­inferred­for­their­coarse-grained­module­(while­possibly­gaining­additional­regulators;­crite-rion­4).­Collectively,­we­use­an­optimization­approach­that­constructs­an­ensemble­of­regulatory­programs­that­try­to­achieve­the­following­goals:­each­regulatory­program­explains­as­much­of­the­gene-expression­­variance­in­the­module­as­possible;­the­regulatory­programs­remain­as­simple­as­possible;­regulatory­programs­are­consistent­across­related­cell­types­in­the­ontogeny;­and­fine­modules­have­regulators­similar­to­those­of­the­coarse­modules­to­which­they­belong.

Notably,­the­approach­used­before­to­identify­combinations­of­regu-lators­(for­example,­linear­regression­regularized­with­the­Elastic­Net­

GN

MF

MO

DC

S&P

PROB

B

NK

PRET

T4

T8

ACTT8

NKT

gdT

Stromal

NKBDC T8GN MF MO S&P

PROB

PRET T4 NKTACTT8 gdT

Stromal

−1.0

−0.8

−0.6

−0.4

−0.2 0

0.2

0.4

0.6

0.8

1.0

Correlation

Figure 2 Related cells have very similar expression profiles. Pearson correlation coefficients (purple, positive; yellow, negative; white, none) for each pair of profiled cell types, calculated for the 1,000 genes (of the 8,431 unique expressed genes) with the highest s.d. value of all samples. Samples are sorted by breadth-first search on the tree (Fig. 1), with stromal cells at the lower or right end. Black vertical and horizontal lines delineate major lineages according to labels along left margin and beneath (color in bar beneath matches colors in Fig. 1). S&P, stem and progenitor cells; PROB, pre-B cells and pro-B cells; T4, CD4+ T cells; T8, CD8+ T cells; ACTT8, activated CD8+ T cells; gdT, γδ T cells.

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

�  aDVaNCE ONLINE PUBLICaTION nature immunology

r e s o u r c e

penalty6,11)­ assumed­ that­ regulatory­ activ-ity­ (and­hence­activity­weight)­ is­ the­ same­across­all­cell­types.­Thus,­if­a­regulator­was­expressed­similarly­in­two­different­cells,­it­was­deemed­to­be­active­to­the­same­extent.­This­violates­the­known­context­specificity­of­regulation­in­complex­lineages.­Conversely,­allowing­the­algorithm­to­construct­a­sepa-rate­ regulatory­ program­ for­ each­ cell­ type­independently­is­impractical­and­also­ignores­the­expected­similarity­between­related­cell­types­in­the­lineage­in­terms­of­gene­regulation.­Ontogenet­solves­this­problem­by­leveraging­the­lineage­tree­when­inferring­the­regulatory­connections­and­their­activity,­such­that­the­module’s­genes­are­more­likely­to­be­regulated­in­a­similar­way­in­related­cell­types.

Ontogenet regulatory model for mouse hematopoiesisWe­applied­Ontogenet­to­the­81­coarse-grained­modules­and­334­fine­modules,­a­lineage­tree­consisting­of­195­cell­types­and­580­candidate­regulators.­The­Ontogenet­model­identified­1,417­regulatory­relations­(1,091­activating,­317­repressing­and­9­mixed)­between­81­coarse-grained­modules­and­480­unique­regulators­(Fig. 4,­Supplementary Fig. 8­and­Supplementary Table 5).­On­average,­there­were­17­regula-tors­per­coarse-grained­module,­and­three­coarse-grained­modules­per­regulator.­As­determined­by­cross-validation,­Ontogenet­constructs­regulatory­programs­ that­are­ strictly­better­at­predicting­new­and­previously­unknown­expression­data­than­those­obtained­by­Elastic­Net6,­a­method­that­does­not­use­the­tree­and­has­fixed­activity­weights­(Supplementary Fig. 9­and­Supplementary Note 3).

In­most­cases­(59%),­a­regulator’s­activity­weights­varied­in­differ-ent­cell­types­(‘frequently­changing’),­reflective­of­context-specific­regulation­(Supplementary Fig. 10).­When­we­pruned­regulatory­interactions­whose­maximal­effect­(defined­as­the­product­of­activ-ity­weight­and­expression)­was­low,­we­obtained­a­sparser­network,­in­ which­ ‘pan-differentiation’­ and­ lineage-specific­ modules­ were­controlled­mostly­by­distinct­regulators­(Fig. 5),­whereas­mixed-use­modules­shared­regulators­with­modules­ in­the­other­classes.­The­regulatory­model­associating­334­fine­modules­and­554­regulators­in­6,151­interactions­had­qualitatively­similar­patterns,­except­for­hav-ing­more­regulators­with­mixed­activity­(that­is,­a­regulator’s­activity­weights­frequently­changed­in­some­modules­and­remained­constant­in­others),­probably­reflective­of­both­the­greater­number­of­interac-tions­and­the­finer­regulatory­program­(Supplementary Fig. 10­and­Supplementary Table 7).­This­rich­regulatory­model­for­differentia-tion­of­the­mouse­immune­system­identified­many­known­regulatory­interactions­and­suggested­new­regulatory­ interactions­ in­specific­immunological­contexts.

Ontogenet prediction of known regulatory interactionsMany­of­ the­regulatory­ interactions­ identified­by­Ontogenet­were­already­ known,­ which­ supported­ the­ accuracy­ of­ our­ model.­ For­example,­among­individual­regulators,­PU.1­(encoded­by­Sfpi1)­was­selected­as­a­regulator­of­the­myeloid­and­B­cell­module­C25­(and­13­of­its­15­fine­modules);­C/EBPα­(encoded­by­Cebpa)­regulates­the­myeloid­modules­C24,­C30­and­C74,­the­macrophage­module­C29,­and­many­myeloid­fine­modules;­C/EBPβ­(encoded­by­Cebpb)­regu-lates­the­myeloid-specific­modules­C25­and­C30­and­many­myeloid­fine­modules;­MafB­(encoded­by­Mafb)­regulates­the­macrophage-specific­modules­C29,­F128­and­F131;­STAT1­regulates­the­interferon-response­module­C52;­T-bet­(encoded­by­Tbx21)­regulates­the­NK­cell­module­C19­and­NKT­cell­module­F288;­and­CIITA­(encoded­by­Ciita)­regulates­the­antigen-presenting­cell­module­F136.

Furthermore,­ the­ combination­ of­ regulators­ associated­ with­ a­­single­module­was­also­consistent­with­known­regulatory­relations.­­For­ example,­ the­ B­ cell­ module­ C33­ is­ regulated­ by­ the­ known­­B­ cell­ regulators­ Pax5,­ EBF1,­ POU2AF1­ and­ Spi-B­ (Fig. 4);­ the­­T­ cell­ module­ C18­ (Supplementary Fig. 8)­ is­ regulated­ by­ the­­known­ T­ cell­ regulators­ Bcl-11B,­ GATA-3,­ Lef1,­ TOX­ and­ TCF7;­­the­γδ­T­cell­module­C56­is­regulated­by­the­known­γδ­T­cell­regula-tors­PLZF­(ZBTB16),­Sox13­and­Id3,­all­also­involved­in­NKT­cell­development­and­function;­ the­NKT­module­F188­ is­regulated­by­­GATA-3,­T-bet­and­PLZF;­and­fine­modules­F150­and­F152,­in­which­the­expression­of­their­member­genes­by­CD8+­DCs­is­higher­than­that­of­CD4+­DCs,­are­regulated­by­IRF8­(but­not­IRF4),­consistent­with­the­known­role­of­subset-selective­expression­IRF4­and­IRF8­in­DC­commitment12.

Ontogenet’s­predictions­were­also­supported­by­their­significant­overlap­with­those­based­on­enrichment­of­cis-regulatory­motifs­and­ChIP-based­binding­profiles­in­the­modules­(Supplementary Tables 5 and 6),­which­supported­the­idea­of­direct­physical­interaction­between­a­regulator­and­the­genes­in­the­module­with­which­it­was­associated­by­Ontogenet­(Supplementary Table 8).­For­example,­27­of­the­asso-ciations­between­a­regulator­and­a­coarse­module­were­supported­by­enrichment­for­cis-regulatory­motifs­(P­=­2.6­×­10−5­(hypergeometric­test­ for­two­groups)­and­P­<­1­×­10−5­(permutation­test)),­such­as­

ba

Target module mean

Global activatorExpression

Activity

Context-specific regulator

Global repressor

Target module expression

Expression

Activity

Expression

Activity

Samples

−2 0 2

Expression (fold)

Repression None Activation

Activity

= Regulator expressionRegulator activity

Irf5 expression

Gata3 expression

Tox expression

Trim14 expression

IRF5 activity

GATA-3 activity

TOX activity

TRIM14 activity

+

+

+

Mean module expression

= ×

×

×

×

×

Σ

Figure 3 Overview of Ontogenet. (a) Use of the regulator expression profile (blue-red; key below) and activity profile (orange-purple; key below) to demonstrate how a type of regulator (bottom) can ‘explain’ the expression of a module (top): a regulator may have a uniformly positive activity weight across the lineage (constitutive activator; top), a uniformly negative activity weight (constitutive repressor; middle) or variable activity weights (context-specific regulator; bottom). (b) Mean expression of a module (top) calculated as a linear combination of regulator expression (blue-red; left) and activity (orange-purple; right).

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION �

r e s o u r c e

the­GATA-2­motif­in­the­hematopoietic­stem­cell­(HSC)­module­C40,­and­the­PU.1­(SFPI1)­­motif­ in­ myeloid­ cell­ module­ C25.­ The­­ChIP­ profiles­ supported­ the­ prediction­ of­­21­ regulator–coarse­ module­ associations­­(P­=­2.2­×­10−5­(hypergeometric­test­for­two­groups)­and­P­<­1­×­10−5­(permutation­test)),­such­as­the­binding­of­C/EBPα­and­C/EBPβ­in­the­myeloid­cell­module­C24­and­the­bind-ing­of­EBF1­in­the­B­cell­module­C33.

Although­those­overlaps­were­statistically­significant,­they­nevertheless­also­indicated­that­the­predictions­of­most­regulatory­inter-actions­were­not­supported­by­enrichment­for­known­cis-regulatory­motifs­or­available­transcription­factor–binding­data,­and­vice­versa.­There­are­three­reasons­for­this.­First,­assigning­scores­for­binding­sites­and­their­enrichment­is­a­process­that­is­highly­prone­to­false-negative­results;­this­is­particularly­likely­to­occur­for­much­smaller­fine­modules.­Second,­the­majority­of­regulators­chosen­by­Ontogenet­do­not­have­a­characterized­binding­motif­(60%­of­regulators;­334­of­554)­or­ChIP­binding­data­ in­any­cell­ type­ (90%­of­ regulators;­497­of­554).­Such­regulators­can­be­nominated­only­by­an­expres-sion-based­method,­such­as­Ontogenet,­and­should­not­be­considered­false-positive­results­of­our­method.­Third,­in­many­cases­in­which­we­do­find­enrichment­for­a­cis-regulatory­element­or­binding­profile­for­­(for­example)­ transcription­ factor­A­ in­module­B­(300­of­551­cis-­regulatory­ interactions­ (54%);­ 52­ of­ 90­ ChIP-based­ interactions­(57%)),­the­transcription­factor­(A)­and­its­target­module­(B)­show­little­or­no­correlation­in­expression­(absolute­Pearson­r­<­0.5).­In­some­cases,­this­is­due­to­a­factor­that­is­not­itself­transcriptionally­regulated­(a­real­ ‘false-negative’­result­of­Ontogenet),­but­in­many­other­cases­the­factor­probably­controls­these­targets­in­another­cell­type­ not­ measured­ in­ our­ study­ (and­ hence­ is­ not­ in­ fact­ a­ false-­negative­result­of­Ontogenet).

A­few­known­regulators­of­differentiation­of­the­immune­system13­were­not­identified­by­the­model­for­various­reasons.­Tal-1­and­BMI1­did­not­meet­the­initial­filtering­criteria,­as­they­were­expressed­only­in­HSCs,­and­hence­were­not­provided­as­input.­GFI1­was­not­assigned­

as­a­regulator­in­stem­and­progenitor­cells­or­granulocytes­because­its­expression­was­highest­in­pre-T­cells­and­was­only­sparse­and­inter-mediate­in­stem­and­progenitor­cells­and­granulocytes.­E2A­(encoded­by­Tcf3)­was­not­identified­as­a­T­cell­regulator,­perhaps­because­it­was­not­specifically­expressed­in­T­cells­and­had­low­expression­in­general,­possibly­because­of­a­bad­probe­set.­XBP1­was­not­identified­as­a­B­cell­regulator­because­it­had­relatively­low­expression­in­B­cells­in­our­arrays­and­had­higher­expression­in­myeloid­cells.

The­reidentification­of­known­regulators­lends­support­to­the­many­previously­unknown­regulatory­interactions­in­the­model.­Of­the­475­regulators­that­Ontogenet­associated­with­lineage-specific­modules­or­‘pan-differentiation’­modules,­at­least­175­(37%)­were­completely­unknown­in­this­context.­Among­those,­for­example,­KLF12­was­pre-dicted­to­be­a­regulator­of­the­NK­cell­module­C19­but­was­not­associ-ated­before­with­the­regulation­of­NK­cells.­GATA-6­was­predicted­to­be­a­regulator­of­the­macrophage-specific­modules­C31,­C50­and­C58­but­was­not­associated­before­with­macrophages.­That­is­in­agreement­with­the­much­lower­number­of­granulocyte-macrophage­colonies­gener-ated­by­embryoid­bodies­of­GATA-6-deficient­mice14.­Finally,­ETV5­was­predicted­by­the­model­to­be­a­regulator­of­the­γδ­T­cell­modules­F287­and­F289,­a­previously­unknown­role­discussed­below.

Context-specific regulation underlies mixed-use modulesContext-specific­regulation,­in­which­the­same­set­of­genes­is­regu-lated­by­one­set­of­regulators­in­the­context­of­one­lineage­and­by­

POU2AF1

Pax5

EBF1

Spi-B

RFX5

PRDM9

HDAC10

c

b

aF175

F176

F177

F178

F179

F180

F181

Bcl7a

RestBlk

Cd19Cd79aCd79bEbf1Pax5

Pou2af1Vpreb3Cd22Cd55

Fcer2aFcrl1Fcrl5Il9r

Igll1Vpreb1

BlnkIl12a

Tcf4

Siglecg

Spib

NKBDC T8GN MF MO S&P PREB PRET T4 NKT ACTT8 GDT

Regulator expression

Regulator activity weight

Pou2af1

Pax5

Ebf1

Spib

Rfx5

Prdm9

Hdac10

GN MF DCs NK T-reg NKTT4T8

gdT

MO

d

POU2AF1

−2 0 2

Repression None Activation

Expression (fold)

ActivityPax5EBF1POU2AF1Spi-BRFX5HDAC10

Figure 4 Ontogenet regulatory model for coarse–grained module C33. (a) Module C33: mean-centered expression (red-blue; key, bottom right) of the module’s genes (rows) in each cell (column); major lineages are delineated by dashed vertical lines (which correspond to color bar beneath c; matches colors in Fig. 1); fine modules F175–F181 (right margin) nested within C33 are delineated by thin horizontal lines; left margin, examples of genes in module. This module contains some typical B cell genes, including Cd19, Blnk, Ebf1 and Cd79a. (b) Regulator expression, presented as mean-centered expression (red-blue color bar, below c) of the regulators (rows) assigned by Ontogenet to module C33. (c) Regulator activity weight (orange-purple; key, bottom right) assigned by Ontogenet for each of the regulators from b in each cell type. (d) Projection of mean-centered mean expression (blue, low; red, high) of module C33 onto the hematopoietic tree (below, differentiated populations on that tree); arrowheads indicate ‘edges’ (differentiation steps) at which the activity weight of selected inferred regulators (labeled in diagram) changes.

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

�  aDVaNCE ONLINE PUBLICaTION nature immunology

r e s o u r c e

another­set­of­regulators­in­the­context­of­another­lineage,­has­been­reported­in­selected­cases,­such­as­the­regulation­of­Rag2­by­GATA-3­in­T­cells­and­by­Pax5­in­B­cells15.­The­ability­of­Ontogenet­to­identify­different­regulatory­programs­for­the­same­module­in­different­parts­of­the­lineage­tree­can­help­delineate­the­regulatory­mechanisms­that­underlie­‘mixed-use’­modules­expressed­in­more­than­one­lineage.­For­example,­module­C70­is­induced­both­in­Treg­cells­and­some­myeloid­populations.­Each­activation­event­is­associated­with­different­regu-lators­in­our­model:­Foxp3­in­CD4+­T­cells­(itself­a­member­of­the­module,­although­not­expressed­in­the­DC­subsets),­and­PIAS3,­HSF2­and­INSM1­in­DCs.­In­another­example,­ the­fine-grained­module­F300­is­independently­induced­in­both­mature­B­cells­and­T­cells.­Although­some­of­its­regulators­are­themselves­‘mixed-use’­in­both­lineages,­others­are­B­cell­specific­(ZFP318,­RFX5­and­CIITA)­or­T­cell­­specific­(EGR2).

Regulatory recruitment and ‘rewiring’ during differentiationMost­regulatory­relations­identified­by­Ontogenet­were­dynamic,­as­reflected­by­the­change­in­their­associated­activity­weights­during­differ-entiation.­This­change­provided­a­‘bird’s-eye’­view­of­the­‘recruitment’­­

and­‘disposal’­of­regulators­(Fig. 6a).­To­characterize­this,­for­each­cell­type,­we­identified­all­the­regulatory­interactions­whose­activ-ity­weight­changed­(increased­or­decreased)­between­that­cell­type­and­its­immediate­progenitor­(Supplementary Table 9),­as­well­as­the­unique­regulators­and­modules­involved­in­those­interactions.­In­this­way,­we­identified­modules­and­regulators­that­were­recruited­and­strengthened­(activity­weight­greater­than­that­of­its­progenitor)­or­were­disposed­of­and­weakened­(activity­weight­lower­than­that­of­its­progenitor)­at­each­differentiation­step.­Notably,­recruitment­(or­disposal)­of­regulators­does­not­necessarily­mean­that­the­regulators’­expression­changes­but­that­the­model­suggests­that­their­regulatory­activity­has­changed­for­this­set­of­targets.­For­example,­during­the­differentiation­of­CD8+­T­cells­from­common­lymphoid­progenitors,­61­regulatory­interactions­were­recruited,­involving­34­modules­and­49­regulators,­only­15­of­which­have­been­associated­before­with­T­cell­differentiation.­In­particular,­for­the­differentiation­step­from­double-negative­(CD4−CD8−)­stage­4­T­cell­to­immature­single-positive­(CD4+­or­CD8+)­T­cell,­Ontogenet­independently­identified­the­previously­reported­involvement­of­MXD4,­Batf­and­NFIL3­and­newly­identi-fied­the­involvement­of­RCBTB1,­PIAS3­and­ITGB3BP­(Fig. 6b,c).­­

TRPS1

SPIC

NFE2L2

JDP2

ZFP691

POU2AF1

SOX5

RFX5

SPIB

EBF1

PRDM9

PAX5

BCL11B

HLF

76

65

18

19

62

75

31

ETS1

AES

33

61

1628

27

41

74

24

TEAD2

PA2G4

ACTL6A

23

RUVBL1

PHF5A

HMGB2

ASF1B

FOXM1

11

THRA

47

42

HMGA2

PHB2RUVBL2

MECOM

PLAG1

CNBP

9

6

15

55

SOX7

7

3

4010

504

14

CBX5

DNMT1

GATA1

HSF2

MYBL2

43

UHRF1

12

ATAD2

RB

BP

7

MO

RF4

L2

NP

M1

BUD31

CHURC1

HOXA9

SU

V39

H2

HDAC277

443029

FOXN2KLF6

32

EPAS1

AHR

26

48

58

CEBPBHDAC5

IRF6

ZEB1

KLF12IRF8RCOR2

PPARGTRERF1

GATA6

RARB

CREG1NR1H3

CEBPAETV1 KLF9 MAFB M

AF

CH

D3

EY

A1

TCFE

BID

2N

FATC

2G

M99

07

AR

ID4A

CH

D7

FLI1

WH

SC

1L1

ME

IS1

8

54 67

BH

LHE

40

34 5259

STA

T2

DTX

3LIR

F7 AR

SO

X13

ZBTB

16

NFE

2L1

STA

T1

NP

AS

2

71

FHL2

SP4

RN

F2

ID3

70

56

BHLHE41

NCOA7

25

64

ETV3

HIC1

EGR2TRIM25MXD4PIAS3

HIF1AFOXP3INSM1IRF5

MITF

HIV

EP

1

EYA2RUNX3STAT4EZH1RXRAKCNIP3

EOMES

GATA2HMGN3

TCFECWWTR1

IRF3PLAGL1

IRF4RBPJMYCL1

TEAD1RUNX2VDRNR4A2

DACH1

KDM2B

CEBPE

BATF3

PADI4NFE2

FUBP1

PBX1

KLF8

RELB

TBX21

Regulator

Lineage-specific module (MF)

Lineage-specific module (DC)

Lineage-specific module (GN)

Lineage-specific module(myeloid cell)

Lineage-specific module (NK cell)

Lineage-specific module (T cell)

Lineage-specific module (B cell)

Module downregulated withdifferentiationModule upregulated withdifferentiationMixed-use modules

Regulatory interaction

PRDM1

AFF3

ZHX2

BACH2NR4A3

DTX1

MBD2

CREB3L1

NR3C2

ZFP263NCOA3

SREBF2ARID3B

SMAD3ZFP318

SAP18

Figure 5 Ontogenet regulatory model for coarse-grained modules. Lineage-specific modules (inner circle; colors as in Fig. 2, except myeloid induced modules (dark purple)), ‘pan-differentiation’ induced (red) and repressed (gray) modules and mixed-use modules (yellow) and their Ontogenet assigned regulators (outer circle; cream) with regulatory interactions with a maximal effect (absolute activity weight × expression) >1. Lines (regulatory interactions) connect each regulator to the module(s) it regulates.

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION �

r e s o u r c e

In­another­example,­during­the­differentiation­step­that­leads­to­NK­cells,­the­NK­cell­module­C19­was­assigned­the­known­NK­cell­regula-tors­Eomes­and­T-bet­as­activators.­Both­Eomes­and­T-bet­were­also­recruited­as­repressors­at­this­differentiation­step­in­other­modules.­The­differentiation­step­that­leads­to­Treg­cells­recruited­the­Treg­cell­module­C70­and­its­known­regulators­Foxp3­and­CREM­(which­has­been­proposed­as­a­Treg­cell­regulator16).­Notably,­because­HSCs­have­no­parent­in­our­model,­regulators­active­in­HSCs­will­be­noted­only­when­they­are­no­longer­used­at­later­points­(for­example,­HOXA7­and­HOXA9­were­no­longer­used­as­activators­at­the­multilymphoid­

­progenitor­stage).­The­first­differentiation­step­with­activator­recruit-ment­is­the­step­that­leads­to­multilymphoid­progenitors,­at­which­MEIS1­is­recruited­to­module­C42.­MEIS1­is­ later­no­ longer­used­by­C42­in­T­cells,­in­agreement­with­the­reported­methylation­and­silencing­of­the­gene­encoding­MEIS1­during­differentiation­toward­T­cells17.

Ranking of lineage activators and repressorsThe­activity­weights­assigned­for­each­regulator­at­each­differentiation­point­allowed­us­to­identify­and­rank­regulators­as­lineage­activators­­

100

200

300

400

500

600

700

800

NKBDC T8GN MF MO S&P PROB PRET T4 NKTACTT8 gdT

d

GN MOMF DC B abTNK NKT gdT

CLP.B

M

preT

.ETP.T

h

preT

.DN2.

Th

preT

.DN2−

3.Th

preT

.DN3−

4.Th

T.DN4.

Th

T.ISP.T

h

T.DP.T

h

T.DPbl.

Th

T.DPsm

.Th

T.DP69

+.Th

T.4int

8+.T

h

T.8SP69

+.Th

T.8SP24

int.T

h

T.8SP24

−.Th

T.8Nve

.PP

T.8Nve

.MLN

T.8Nve

.LN

T.8M

em.L

N

T.8Nve

.Sp

T.8M

em.S

p

Activa

ted

αβT C

ell

HES1TRP73

IRF9SCMH1

CBX7TCF15

RUNX2ETS2

KLF15PHF1

SMAD7PIAS3

TBX21RCBTB1ZBTB16

TAF11KLF15

ZBTB16KLF3

CLOCKZFP367

AFF1WHSC1ZFP367ZBTB38TRIM14

SP4MAF

TWIST1KLF15

POU4F2FOXE3AEBP1KLF15SOX4KLF15ATF6

MYNNKLF15

TRIM14ZHX2SOX9MXD4BATF

RCBTB1PIAS3

ITGB3BPMXD4NFIL3CUX1

HOXD8SMARCD2

ATF6STAT4BCL6BSMYD1

MORF4L2LMO4

POU6F1RORA

GFI1

CLP.BM

preT.ETP.Th

preT.DN2.Th

T.DN4.Th

T.4int8+.Th

T.8SP69+.Th

T.DP69+.Th

T.DPbl.Th

T.DPsm.Th

T.8SP24int.Th

T.8SP24-.Th

T.8Nve.SpT.8Nve.LN T.8Nve.MLN T.8Nve.PP

T.8Mem.LN T.8Mem.Sp T.8Nve.Sp.OT1

Activated CD8+ T

preT.DN2-3.Th

T.ISP.Th

PreT.DN3-4.Th

b

a

HES1, TRP73, IRF9, SCMH1, CBX7, TCF15, Runx2, ETS2,KLF15, PHF1, Smad7, PIAS3, 42; 69; 67; 77; 54; 38; 71; 57; 37; 45

GFI1, TBX21, RCBTB1, ZBTB16(2), TAF11, KLF15(2), KLF3, CLOCK, ZFP367(2), AFF1, WHSC1, ZBTB38, TRIM14, SP4, MAF, TWIST1, 68; 71; 52; 56; 35; 62; 6778; 14; 77; 23; 15; 65; 17; 66

ZHX2, Sox9, 18; 37

KLF15(3), POU4F2, Foxe3, AEBP1, Sox4; 39; 79; 38; 66; 77

ATF6, MYNN, KLF15, TRIM14, 71; 78; 73; 34

MXD4(2), BATF, RCBTB1, PIAS3, ITGB3BP, NFIL378; 34; 18; 71; 75; 81

CUX1, HOXD8, SMARCD2, 77; 69; 78

ATF6, 78

STAT4, BCL6B, SMYD1, 59; 73; 19

c

MORF4L2, 20

RORA, 19

LMO4, 44

POU6F1, 4

Reported T cell regulatorNot reported in T cells

SC

SC.COMP.BM CLP.BM

PHB2, CNBP, TRIM28, RUVBL2, HLF, LMO2, PA2G4, NPM1, RBBP7, PHF5A

ENO1PADI4C/EBPβ BACH1XBP1NCOA4STAT6NFE2KLF6TCFE3

CREG1XBP1C/EBPαLMO2ENO1C/EBPβ CNBPBACH1MAFNCOA4

C/EBPβ ENO1XBP1IRF5SFPI1BACH1TRPS1ATF6MEF2ALMO2

RELBETV3KLF6IKBαRBPJXBP1CIITAINSM1RELNFE2L1

POU2AF1PAX5HMGB2CBX5UHRF1EBF1FUBP1SPIBPA2G4ASF1B

EOMESTBX21ETS1ENO1SMAD3AESKLF2ID2SGM9907STAT6

AESTCF7ENO1LEF1BCL11BTRIM28NPM1ETS1FUBP1SP100

GATA3AESZBTB16ID2ENO1SP100NPM1TBX21STAT1LEF1

MBD2AESSTAT4DTX1ETS1ENO1HIC1GATA3TCF7SREBF2

Fre

quen

tly c

hang

ing

regu

lato

ry in

tera

ctio

ns

Repression None Activation

Activity

Figure 6 Changes in activity weights across the hematopoietic lineage tree. (a) Activity weights in each cell type (columns; correspond to color bar below, which matches colors in Fig. 1) for every ‘frequently changing’ regulatory interaction between a regulator and a coarse-grained module: orange, positive (activation) activity weight; purple, negative (repression) activity weight; white, 0 (no regulation). (b) Enlargement of boxed area in a of ‘frequently changing’ interactions for only those activating interactions recruited in the CD8+ T cell lineage. (c) Known and previously unknown regulators recruited in the CD8+ T cell lineage, including the CD8+ T cell lineage branch (squares, cell types; lines (‘edges’), differentiation steps); right, regulators recruited along each differentiation step and their associated modules (red, regulators reported to have a role in T cells; black, not reported in T cells). Numbers in parentheses indicate activity weight changes for the regulator on this edge, if >1. (d) Simplified ImmGen tree (cell type colors correspond to those in Fig. 1) with Ontogenet-inferred lineage regulators (red and black as in c).

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

�  aDVaNCE ONLINE PUBLICaTION nature immunology

r e s o u r c e

and­ repressors­ on­ the­ basis­ of­ the­ entire­ model­ (Fig. 6d­ and­Supplementary Table 10).­In­this­way­we­correctly­captured­many­known­regulators­of­each­lineage­among­the­top-ranked­activators.­For­example,­our­model­associated­c-Myc,­N-Myc,­GATA-2­and­MEIS1­with­stem­and­progenitor­cells;­Bcl-11B,­TCF7­and­GATA-3­with­αβ­T­cells;­POU2AF1,­Pax5,­EBF1­and­Spi-B­with­B­cells;­Eomes,­T-bet­and­Smad3­with­NK­cells;­and­GATA-3­and­PLZF­with­NKT­cells.­In­addition,­the­model­made­many­predictions­of­lineage­regulators­not­previously­associated­with­those­lineages,­such­as­the­following:­in­stem­and­progenitor­cells,­HLF;­in­granulocytes,­DACH1­(reported­to­regulate­cell-cycle­progression­in­myeloid­cells18),­Bach1­and­NFE2;­in­macrophages,­CREG1;­in­DCs,­ATF6,­ETV3,­SKIL,­NR4A2­and­NR4A3­(shown­to­be­induced­in­viral­infected­DCs19,20);­in­mono-cytes,­POU2F2­(Oct2;­reported­to­be­upregulated­during­macrophage­differentiation21)­and­KLF13­(a­regulator­of­B­cells­and­T­cells22­with­higher­expression­in­monocytes);­in­B­cells,­ZFP318;­and­in­NK­cells,­ELF4­(Gm9907;­shown­to­control­the­proliferation­and­homing­of­CD8+­T­cells23).­Notably,­although­this­‘pan-model’­analysis­is­useful,­it­can­deemphasize­the­contribution­of­important­regulators­captured­by­the­model­in­a­more­nuanced­way—for­example,­as­acting­only­dur-ing­a­limited­window­of­differentiation­but­not­present­in­the­mature­stage.­Those­are­captured­by­the­recruitment­and­disposal­analysis­presented­above­(Fig. 6).

Finally,­by­counting­the­changes­in­activity­weight­that­occur­(across­all­regulators­and­modules)­at­each­differentiation­step­(‘edge’),­we­can­identify­those­differentiation­points­at­which­regulation­is­‘rewired’­most­substantially­(Supplementary Fig. 11).­For­example,­19­regula-tors­were­recruited­to­coarse­modules­(that­is,­their­activity­weight­increased­ from­0)­at­ the­early­ thymocyte­progenitor­stage,­and­28­regulators­were­recruited­to­coarse­modules­at­ the­γδ­T­cell­ stage,­including­the­known­T­cell­regulator­GATA-3­and­the­known­γδ­T­cell­regulators­Id3­and­Sox13­(Supplementary Fig. 11a).­At­the­common­lymphoid­progenitor­stage,­four­regulators­were­disposed­of­(that­is,­their­activity­weight­diminished­to­0)­by­coarse­modules,­including­the­HSC­regulators­HOXA7,­HOXA9­and­HOXB3.­Eighteen­regula-tors­were­disposed­of­at­the­double-negative-2­T­cell­precursor­stage,­including­GATA-1,­c-Myc­and­N-Myc­(Supplementary Fig. 11b).­

Overall,­‘rewiring’­was­more­prominent­at­higher­levels­in­the­lineage­than­at­ lower­(more-differentiated)­ levels,­although­this­may­have­been­partly­due­to­the­diminished­power­to­detect­changes­in­cell­types­with­no­other­cells­differentiating­from­them­(terminally­differenti-ated;­also­called­‘leaves­in­the­tree’).­The­individual­differentiation­steps­with­the­largest­number­of­activity­weight­changes­were­those­in­small-intestine­DCs,­thymus­γδ­T­cells,­ liver­and­lung­DCs­and­­double-negative-2­T­cell­precursor­stage,­which­suggests­substantial­regulatory­ ‘rewiring’­ in­ these­ cells,­ possibly­ due­ to­ tissue-specific­effects.­ The­ regulatory­ model­ for­ fine­ modules­ identified­ a­ larger­number­of­regulatory­changes­(a­change­in­activity­weight­for­82%­of­the­differentiation­steps,­compared­with­65%­for­the­coarse-grained­module­model),­in­particular­in­differentiation­steps­leading­to­‘leaves’­(terminally­ differentiated­ cells;­ 67%­ versus­ 48%).­ Thus,­ the­ fine-grained­modules­help­to­identify­more­cell­type–specific­regulation.

ETV5 regulates gd T cell differentiationTo­test­one­of­the­model’s­predictions­in vivo,­we­centered­on­regu-latory­activators­of­lineage-specific­modules­with­no­known­func-tion­in­that­lineage.­A­practical­criterion­was­that­the­gene­could­be­manipulated­in vivo­in­a­cell­type–restricted­manner.­We­focused­on­the­Ets­ family­member­ETV5­and­its­predicted­role­as­a­regulator­of­the­differentiation­of­γδ­T­cells­in­modules­F287­and­F289,­as­its­expression­is­highly­restricted­to­the­γδ­T­cell­lineage.­Although­the­model­assigned­several­regulators­to­these­modules,­only­two,­Sox13­and­ETV5,­are­specific­to­the­γδ­T­cell­lineage.­Both­are­expressed­in­distinct­thymic­precursors,­which­raised­the­possibility­that­they­are­among­the­earliest­determinants­of­the­lineage.­Sox13­is­a­known­regulator­of­γδ­T­cells,­but­ETV5­has­not­been­ linked­to­γδ­T­cell­development­thus­far.

To­assess­the­regulatory­role­of­ETV5­in­γδ­T­cells,­we­analyzed­γδ­T­cell­development­and­function­in­mice­lacking­ETV5­specifically­in­T­cells­(CD2p-CreTg+Etv5fl/fl­mice).­As­thymocytes­that­express­the­γδ­T­cell­antigen­receptor­transit­from­immature­cells­with­high­expression­ of­ the­ cell­ surface­ marker­ CD24­ (CD24hi)­ to­ mature­CD24lo­cells,­they­acquire­effector­functions24.­ETV5­has­its­high-est­expression­in­γδ­thymocytes­expressing­γ-chain­variable­region­2­

ba

7920 962.8

30 30

2516

2.5 2.7

7421

6

3

0

10

5

0

Thymus

Spleen

CD24

CD62L

CD

44

ThymusCtrl CKO

Eve

nts

(% o

f max

)

100806040200

0

104

103

102

105

0 104103102 105

0 104103102 105

c

67

9.7

42

32

RORγt

Thymus

IL-17

8.849

100

80

60

40

20

0

100

80

60

40

20

0

18

18

9.8

45

42

43

14

74

CD27

IFN-γ

IL-1

7C

CR

6

LNCtrl CKO

0

104

103

102

105

0

104

103

102

105

0 104103102 105

CtrlCKO

Ctrl CKO

Eve

nts

(% o

f max

)E

vent

s(%

of m

ax)

γδ T

cel

ls (

×103 )

γδ T

cel

ls (

×105 )

Figure 7 ETV5 is a regulator of γδ T cells. (a) Total γδ T cells in the thymus and spleen of mice with T cell–specific ETV5 deficiency (CKO) and their CD2p-CreTg+Etv5+/+ littermates (control (Ctrl)). (b) Expression of CD24 (top) and of CD44 and CD62L (bottom) by Vγ2+ thymocytes from 7-day-old mice as in a. Numbers above bracketed lines (top) indicate percent CD24lo (mature) thymocytes (left) or CD24hi (immature) thymocytes; numbers in quadrants (bottom) indicate percent cells in each. Similar results were obtained for mice of different ages. (c) Expression of RORγt and production of IL-17 (left) by mature (CD24lo) Vγ2+ thymocytes of mice as in a; numbers above bracketed lines indicate percent RORγt− cells (top; left number) or RORγt+ cells (top; right number), or IL-17+ cells (bottom). Right, expression of CCR6 and CD27 (top) and of IL-17 and interferon-γ (IFN-γ; bottom) by Vγ2+ T cells from the lymph nodes (LN) of 4-week-old mice as in a; numbers in or adjacent to outlined areas indicate percent cells in each. Data are representative of three experiments with one to three independent litters with a minimum of two mice per genotype (a; error bars, s.e.m.), five experiments (b) or three experiments (c).

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION �

r e s o u r c e

(Vγ2)­of­the­T­cell­antigen­receptor,­which­constitute­nearly­half­of­all­γδ­T­cells­in­postnatal­mice.­Most­Vγ2+­cells­differentiate­into­inter-leukin­17­(IL-17)-producing­γδ­effector­cells­in­the­thymus24.­Thus,­one­prediction­of­the­model­was­that­the­intrathymic­development­of­­IL-17-producing­γδ­effector­cells­would­be­particularly­impaired­in­the­absence­of­ETV5.­In­mice­with­conditional­T­cell–specific­deficiency­in­ETV5,­the­overall­number­of­γδ­T­cells­generated­was­similar­to­that­of­control­mice­(their­CD2p-CreTg+Etv5+/+­littermates):­in­7-day-old­neonates,­total­number­of­thymocytes­in­mice­with­T­cell–specific­ETV5­deficiency­was­~50%­of­normal,­but­the­frequency­of­thymo-cytes­that­expressed­the­γδ­T­cell­antigen­receptor­was­about­twofold­higher,­which­resulted­in­an­abundance­of­γδ­T­cells­in­the­thymus­and­spleen­similar­to­that­in­control­mice­(Fig. 7a).­However,­there­was­specific­loss­of­mature­Vγ2+­thymocytes­in­mice­with­T­cell–specific­ETV5­deficiency­(Fig. 7b,­top).­This­may­have­been­due­to­inefficient­activation,­as­indicated­by­the­lower­expression­of­CD44­(the­nominal­marker­of­lymphocyte­activation)­on­Vγ2+­thymocytes­from­mice­with­T­cell–specific­ETV5­deficiency­and­the­correspondingly­higher­expres-sion­of­CD62L­(a­marker­of­the­naive­state)­on­those­cells­(Fig. 7b,­­bottom).­For­γδ­thymocytes­that­expressed­other­Vγ­chains,­the­pro-portion­of­mature­cells­or­activated­cells­in­mice­with­T­cell–specific­­ETV5­deficiency­was­not­different­from­that­of­controls.­Critically,­the­residual­mature­thymocytes­in­mice­with­T­cell–specific­ETV5­deficiency­were­impaired­in­the­generation­of­IL-17-producing­γδ­effector­cells­(Fig. 7c).­Mature­Vγ2+­thymocytes­from­ETV5-deficient­mice­had­lower­expression­of­the­transcription­factor­RORγt­(which­induces­Il17­transcription),­and­both­thymic­and­peripheral­γδ­T­cells­were­impaired­in­the­generation­of­CCR6+CD27−­IL-17-producing­γδ­effector­cells­(Fig. 7c).­These­results­supported­the­prediction­of­our­­model­and­demonstrated­that­ETV5­was­essential­for­proper­intra-thymic­maturation­of­the­IL-17-producing­γδ­effector­cell­subset.

Studying the Ontogenet model on the ImmGen portalTo­ facilitate­ exploration­ and­ testing­ of­ other­ predictions­ of­ our­model,­we­provide­the­full­set­of­modules­and­regulatory­model­as­part­of­the­ImmGen­portal,­with­relevant­tools­for­searching,­brows-ing­and­visually­inspecting­the­results.­Specifically,­the­‘Modules­and­Regulators’­data­browser­of­the­ImmGen­portal­(http://www.immgen.org/ModsRegs/modules.html)­is­the­gateway­to­the­Ontogenet­regu-latory­model­of­the­ImmGen.­It­allows­the­user­to­browse­coarse-grained­or­fine-grained­modules­by­their­number,­their­pattern­of­expression,­ a­ gene­ they­ contain,­ a­ regulator­ predicted­ to­ regulate­them­or­the­cell­type­in­which­they­are­induced.­For­each­module,­we­present­the­expression­of­its­genes­and­predicted­regulators­(each­as­a­heat­map),­the­activity­weights­of­each­regulator­in­each­cell,­and­the­module’s­mean­expression­projected­on­the­lineage­tree­(as­in­Fig. 4a).­The­module­page­also­links­to­a­list­of­the­genes­in­the­module,­the­regulators­that­are­members­of­the­module,­the­regulators­predicted­to­regulate­the­module,­the­regulators­suggested­by­enrichment­of­cis­motifs­and­binding­events­of­the­module­genes,­and­functional­enrichments­of­the­module.­Finally,­we­provide­links­for­downloading­a­table­with­the­assignment­of­all­genes­to­coarse­and­fine­modules,­the­regulatory­program­of­all­modules,­and­the­Ontogenet­code.

DISCUSSIONThe­ImmGen­data­set­provides­the­most­detailed­and­comprehensive­view­of­the­transcriptional­activity­of­any­mammalian­immune­system­and­(arguably)­of­any­developmental­cell–differentiation­process.­We­have­used­those­data­to­analyze­the­regulatory­circuits­underlying­such­processes,­from­global­profiles­to­modules­to­the­transcription­factors­ that­control­ them.­The­unique­ features­of­Ontogenet­have­

allowed­us­to­identify­regulatory­programs­active­at­specific­differen-tiation­stages­and­to­follow­them­as­they­‘unfold’­and­‘rewire’.

Our­analysis­has­automatically­ reidentified­many­of­ the­known­regulators­and­their­correct­function,­has­suggested­additional­roles­for­at­least­175­more­regulators­not­associated­before­with­hemato-poiesis­and­has­identified­points­in­the­lineage­at­which­regulators­are­recruited­to­control­a­specific­gene­program­or­lose­their­regula-tory­function.­Our­ability­to­automatically­reidentify­many­known­regulators­at­the­appropriate­developmental­stage­and­the­significant­correspondence­among­the­predicted­regulators,­known­functions,­enrichment­for­cis-motifs­and­enrichment­by­ChIP­followed­by­deep­sequencing­supports­the­probably­high­quality­of­our­new­predictions.­Among­those,­we­experimentally­tested­and­confirmed­a­previously­unknown­role­for­ETV5­in­the­differentiation­of­the­γδ­T­effector­cell­subset.­Additional­studies­should­determine­whether­ETV5­regulates­the­differentiation­of­IL-17-producing­γδ­effector­cells­by­selectively­controlling­the­expression­of­genes­in­γδ­lineage–specific­modules.

Ontogenet’s­rich­model­allows­us­to­predict­the­specific­biologi-cal­context­at­which­regulation­occurs,­to­generalize­broad­roles­for­regulators­and­to­identify­global­principles­of­the­regulatory­program.­The­ability­to­ identify­regulators­that­act­only­during­specific­dif-ferentiation­windows­helps­to­detect­‘early’­programming­transcrip-tion­ factors­whose­expression­ is­ shut­off­when­cells­ transit­ to­ the­mature­stage.­However,­integrating­across­the­model’s­predictions­in­an­entire­lineage­helps­to­identify­transcription­factors­important­for­the­maintenance­of­lineage­identity­or­function,­such­as­those­that­directly­regulate­the­expression­of­effector­molecules.­Finally,­general-izing­across­multiple­regulators,­we­can­identify­those­differentiation­steps­at­which­regulatory­control­‘rewires’­most­substantially­and­the­regulators­that­control­such­‘rewiring’.

As­with­all­expression-based­methods­used­to­predict­regulation,­Ontogenet­cannot­directly­distinguish­causal­directionality.­To­avoid­arbitrary­ resolution­ of­ this­ ambiguity,­ Ontogenet­ allows­ several­regulators­with­similar­expression­profiles­to­be­assigned­together­as­ regulators­ of­ a­ module.­ The­ dense­ interconnected­ circuits­ and­extensive­autoregulation­in­other­mammalian­circuits­that­controls­cell­states3,25­suggest­that­such­regulatory­interactions­are­probably­functional,­although­some­may­be­‘false-positive’­results.­Conversely,­the­activation­of­other­functional­regulators­may­not­be­reflected­by­their­expression,­and­some­may­have­been­filtered­by­our­stringent­criteria­(for­example,­Tal1,­which­encodes­a­known­HSC­regulator).­Those­may­be­captured­by­our­complementary­analysis­of­enrich-ment­of­modules­in­cis-regulatory­motifs­and­binding­of­regulators.­Another­challenge­is­posed­by­genes­with­unique­expression­profiles­that­are­assigned­to­modules­with­similar­but­distinct­expression­pro-files­(such­as­Rag1­and­Rag2­in­module­C5).­The­inferred­regulatory­program­is­unlikely­to­hold­true­for­those­genes.

A­similar­study­of­human­hematopoiesis3­has­suggested­substan-tial­mixed­use­of­modules­by­lineages,­whereas­the­mouse­compen-dium­suggests­that­most­modules­are­lineage­specific.­As­has­been­shown­before,­global­profiles,­lineage-specific­signatures­and­gene-­coexpression­ patterns­ are­ otherwise­ broadly­ conserved­ between­humans­and­mice4.­One­possible­reason­for­the­diminished­‘mixed­use’­in­the­mouse­program­is­that­whereas­the­mouse­data­set­con-tains­many­more­cell­types,­it­does­not­include­erythrocytes,­mega-karyocyte,­basophils­and­eosinophils,­the­cells­for­which­many­of­the­‘mixed-use’­patterns­have­been­observed­in­humans3.­Notably,­many­regulators­were­shared­across­lineages.­In­particular,­some­regulators­were­active­ in­only­one­ lineage­ in­some­modules­but­were­shared­by­lineages­in­other­modules.­For­example,­ATF6­was­an­activator­in­all­lineages­in­the­myeloid­modules­C25,­C45­and­C49­but­was­a­­

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

�0  aDVaNCE ONLINE PUBLICaTION nature immunology

r e s o u r c e

T­cell–specific­repressor­in­the­T­cell­precursor­module­C57­and­was­a­T­cell–specific­activator­in­the­B­cell­module­C71.

Ontogenet­is­applicable­to­other­differentiation­data­sets,­including­data­obtained­with­fetal­samples­or­for­cancer­studies,­when­other­pre-dictors­are­used­as­candidate­regulators­(for­example,­genetic­variants­as­in­Lirnet6),­when­cells­are­measured­in­both­the­resting­state­and­stimulated­state,­or­for­protein-expression­data­(for­example,­single-cell,­high-dimensional­phosphoproteomic­mass­cytometry­data26).­In­each­case,­the­ability­to­share­regulatory­programs­by­related­cell­types­or­conditions­can­both­enhance­the­power­and­help­with­biological­interpretation.­Notably,­Ontogenet­now­depends­on­a­preconstructed­ontogeny.­Although­much­is­known­about­the­hematopoietic­lineage,­some­parts­remain­unstructured­(for­example,­all­DCs­in­the­myeloid­lineage)­and­some­progenitors­are­not­known­(for­example,­ those­of­γδ­T­cells­or­other­innate-like­lymphocytes).­This­reflects­in­part­inherent­lineage­flexibility,­whereby­several­cell­types­can­differenti-ate­into­the­same­cell­type,­but­reflects­in­part­simply­the­present­lack­of­knowledge­of­the­particular­progenitor­of­a­given­cell­type.­New­methods­would­be­needed­to­construct­an­ontogeny­automatically­or­to­revise­an­existing­one.­In­other­cases,­Ontogenet’s­output­can­be­used­to­refine­the­topology­of­the­ontogeny­by­identifying­‘edges’­that­do­not­correspond­to­any­changes­in­regulatory­programs­and­can­be­removed­without­disconnecting­the­lineage.­The­ImmGen­compen-dium,­coarse-­and­ fine-grained­modules­and­ identified­regulators­and­regulatory­relations­are­all­available­for­interactive­searching­and­browsing­and­for­downloading­at­the­ImmGen­portal­and­will­provide­an­invaluable­resource­for­future­studies­of­the­role­of­gene­regulation­in­cell­differentiation­and­immunological­disease.

METHODSMethods­and­any­associated­references­are­available­ in­ the­online­version­of­the­paper.

Accession code.­GEO:­microarray­data,­GSE15907.

Note: Supplementary information is available in the online version of the paper.

ACKnOwleDGmenTSWe­thank­the­ImmGen­core­team­(including­M.­Painter­and­S.­Davis)­for­help­­with­data­generation­and­processing;­eBioscience,­Affymetrix­and­Expression­Analysis­for­support­of­ImmGen;­L.­Gaffney­for­help­with­figure­preparation­­and­layout­of­the­lineage­tree;­S.­Hart­for­initial­layout­of­the­lineage­tree;­and­­A.­Liberzon­(Molecular­Signatures­Database)­for­the­positional­gene­sets­for­mouse.­Supported­by­National­Institute­of­Allergy­and­Infectious­Diseases­(R24­AI072073­to­the­ImmGen­Consortium),­the­US­National­Institutes­of­Health­(A.R.;­and­U54-CA149145­and­149644.0103­to­V.J.­and­D.K.),­the­Burroughs­Wellcome­Fund­(A.R.),­the­Klarman­Cell­Observatory­(A.R.),­the­Howard­Hughes­Medical­Institute­(A.R.),­the­Merkin­Foundation­for­Stem­Cell­Research­at­the­Broad­Institute­(A.R.)­and­the­National­Science­Foundation­(DBI-0345474­to­V.J.­and­D.K.).

AUTHOR COnTRIBUTIOnSV.J.,­T.S.,­A.R.­and­D.K.­designed­the­study;­V.J.­developed­the­algorithm,­with­input­from­T.S.,­A.R.­and­D.K.;­T.S.­analyzed­the­data;­K.S.­did­experiments;­O.Z.­provided­the­motif-related­code­and­participated­in­writing;­X.S.­generated­

a­mouse­model;­J.K.­designed­the­experiments­and­participated­in­writing­the­manuscript;­and­V.J.,­T.S.,­A.R.­and­D.K.­wrote­the­manuscript­with­input­from­­all­authors.­

COmPeTInG FInAnCIAl InTeReSTSThe­authors­declare­no­competing­financial­interests.

reprints and permissions information is available online at http://www.nature.com/reprints/index.html.

1. Heng, T.S.P. et al. The Immunological Genome Project: networks of gene expression in immune cells. Nat. Immunol. 9, 1091–1094 (2008).

2. Iwasaki, H. & Akashi, K. Myeloid lineage commitment from the hematopoietic stem cell. Immunity 26, 726–740 (2007).

3. Novershtern, N. et al. Densely interconnected transcriptional circuits control cell states in human hematopoiesis. Cell 144, 296–309 (2011).

4. Shay, T. et al. Conservation and divergence in the transcriptional programs of the human and mouse immune systems. Proc. Natl. Acad. Sci. USA 110, 2946–2951 (2013).

5. Kim, H.D., Shay, T., O’Shea, E.K. & Regev, A. Transcriptional regulatory circuits: predicting numbers from alphabets. Science 325, 429–432 (2009).

6. Lee, S.-I. et al. Learning a prior on regulatory potential from eQTL data. PLoS Genet. 5, e1000358 (2009).

7. Segal, E. et al. Module networks: identifying regulatory modules and their condition-specific regulators from gene expression data. Nat. Genet. 34, 166–176 (2003).

8. Capaldi, A.P. et al. Structure and function of a transcriptional network activated by the MAPK Hog1. Nat. Genet. 40, 1300–1306 (2008).

9. Ram, O. et al. Combinatorial patterning of chromatin regulators uncovered by genome-wide location analysis in human cells. Cell 147, 1628–1639 (2011).

10. Miller, J.C. et al. Deciphering the transcriptional network of the dendritic cell lineage. Nat. Immunol. 13, 888–899 (2012).

11. Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. R. Stat. Soc. Series B Stat. Methodol. 67, 301–320 (2005).

12. Tamura, T. et al. IFN regulatory factor-4 and -8 govern dendritic cell subset development and their functional diversity. J. Immunol. 174, 2573–2581 (2005).

13. Orkin, S.H. & Zon, L.I. SnapShot: hematopoiesis. Cell 132, 712.e711–712.e712 (2008).

14. Pierre, M., Yoshimoto, M., Huang, L., Richardson, M. & Yoder, M.C. VEGF and IHH rescue definitive hematopoiesis in Gata-4 and Gata-6–deficient murine embryoid bodies. Exp. Hematol. 37, 1038–1053 (2009).

15. Kishi, H. et al. Lineage-specific regulation of the murine RAG-2 promoter: GATA-3 in T cells and Pax-5 in B cells. Blood 95, 3845–3852 (2000).

16. Bodor, J., Fehervari, Z., Diamond, B. & Sakaguchi, S. ICER/CREM-mediated transcriptional attenuation of IL-2 and its role in suppression by regulatory T cells. Eur. J. Immunol. 37, 884–895 (2007).

17. Ji, H. et al. Comprehensive methylome map of lineage commitment from haematopoietic progenitors. Nature 467, 338–342 (2010).

18. Lee, J.-W. et al. DACH1 regulates cell cycle progression of myeloid cells through the control of cyclin D, Cdk 4/6 and p21Cip1. Biochem. Biophys. Res. Commun. 420, 91–95 (2012).

19. Ng, S.S.M., Chang, T.-H., Tailor, P., Ozato, K. & Kino, T. Virus-induced differential expression of nuclear receptors and coregulators in dendritic cells: Implication to interferon production. FEBS Lett. 585, 1331–1337 (2011).

20. Wang, T. et al. Inhibition of activation-induced death of dendritic cells and enhancement of vaccine efficacy via blockade of MINOR. Blood 113, 2906–2913 (2009).

21. Neumann, M. et al. Differential expression of Rel/NF-κB and octamer factors is a hallmark of the generation and maturation of dendritic cells. Blood 95, 277–285 (2000).

22. Outram, S.V. et al. KLF13 influences multiple stages of both B and T cell development. Cell Cycle 7, 2047–2055 (2008).

23. Yamada, T., Park, C.S., Mamonkin, M. & Lacorazza, H.D. Transcription factor ELF4 controls the proliferation and homing of CD8+ T cells via the Kruppel-like factors KLF4 and KLF2. Nat. Immunol. 10, 618–626 (2009).

24. Narayan, K. et al. Intrathymic programming of effector fates in three molecularly distinct γδ T cell subtypes. Nat. Immunol. 13, 511–518 (2012).

25. Yosef, N. & Regev, A. Impulse control: temporal dynamics in gene transcription. Cell 144, 886–896 (2011).

26. Bendall, S.C. et al. Single-cell Mass cytometry of differential immune and drug responses across a human hematopoietic continuum. Science 332, 687–696 (2011).

The complete list of authors is as follows: Adam J Best9, Jamie Knell9, Ananda Goldrath9, Vladimir Jojic1, Daphne Koller1, Tal Shay2, Aviv Regev2,5, nadia Cohen11, Patrick Brennan11, michael Brenner11, Francis Kim12, Tata nageswara Rao12, Amy wagers12, Tracy Heng13, Jeffrey ericson13, Katherine Rothamel13, Adriana Ortiz-lopez13, Diane mathis13, Christophe Benoist13, natalie A Bezman14, Joseph C Sun14, Gundula min-Oo14, Charlie C Kim14, lewis l lanier14, Jennifer miller15, Brian Brown15, miriam merad15, emmanuel l Gautier15,16, Claudia Jakubzick15, Gwendalyn J Randolph15,16, Paul monach17, David A Blair18,

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology  aDVaNCE ONLINE PUBLICaTION ��

r e s o u r c e

9Division of Biological Sciences, University of California San Diego, La Jolla, California, USA. 10Broad Institute and Department of Biology, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. 11Division of Rheumatology, Immunology and Allergy, Brigham and Women’s Hospital, Boston, Massachusetts, USA. 12Joslin Diabetes Center, Boston, Massachusetts, USA. 13Division of Immunology, Department of Microbiology & Immunobiology, Harvard Medical School, Boston, Massachusetts, USA. 14Department of Microbiology & Immunology, University of California San Francisco, San Francisco, California, USA. 15Icahn Medical Institute, Mount Sinai Hospital, New York, New York, USA. 16Department of Pathology & Immunology, Washington University, St. Louis, Missouri, USA. 17Department of Medicine, Boston University, Boston, Massachusetts, USA. 18Skirball Institute of Biomolecular Medicine, New York University School of Medicine, New York, New York, USA. 19Fox Chase Cancer Center, Philadelphia, Pennsylvania, USA. 20Computer Science Department, Brown University, Providence, Rhode Island, USA. 21Department of Biomedical Engineering, Howard Hughes Medical Institute, Boston University, Boston, Massachusetts, USA. 22Program in Molecular Medicine, Children’s Hospital, Boston, Massachusetts, USA. 23Dana-Farber Cancer Institute and Harvard Medical School, Boston, Massachusetts, USA.

michael l Dustin18, Susan A Shinton19, Richard R Hardy19, David laidlaw20, Jim Collins21, Roi Gazit22, Derrick J Rossi22, nidhi malhotra3, Katelyn Sylvia3, Joonsoo Kang3, Taras Kreslavsky23, Anne Fletcher23, Kutlu elpek23, Angelique Bellemare-Pelletier23, Deepali malhotra23 & Shannon Turley23

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology doi:10.1038/ni.2587

ONLINE METHODSData set.­Expression­of­mouse­genes­was­measured­on­Affymetrix­Mogen1­arrays­(Affymetrix­annotation­version­31).­Sorting­strategies­for­the­ImmGen­populations­are­available­on­the­ImmGen­website­(http://www.immgen.org/Protocols/ImmGen­Cell­prep­and­sorting­SOP.pdf).

As­the­data­set­of­the­ImmGen­gradually­grew­from­2010­to­2012,­cluster-ing,­regulatory­program­reconstruction­and­final­presentation­were­done­on­three­different­ImmGen­releases­(September­2010,­March­2011­and­April­2012)­with­attempts­to­maximize­backward­compatibility­as­much­as­possible.­The­clusters­and­the­regulatory­program­are­from­the­September­2010­and­March­2011­releases,­chosen­to­ensure­consistency­with­the­other­ImmGen­Report­papers­that­refer­to­them.­Clustering­was­done­on­the­ImmGen­release­of­September­2010,­with­744­samples,­647­of­which­remained­in­the­April­2012­release.­Ontogenet­was­applied­to­the­ImmGen­release­of­March­2011,­only­to­the­data­of­the­676­samples­(195­hematopoietic­cell­types)­that­were­connected­to­the­hematopoietic­tree.­Thus,­we­maintained­membership­in­clusters­from­the­earlier­analysis­but­used­only­some­of­the­samples­to­learn­the­regulatory­program.­The­heat­maps­presented­here­include­755­samples­(244­cell­ types),­excluding­control­samples.­For­simplicity,­only­720­sam-ples­are­presented­on­the­full­tree­(210­cell­types).­Supplementary Table 1­lists­all­the­samples­in­the­last­ImmGen­release­(April­2012)­and­states­for­each­sample­if­it­was­used­in­generating­the­modules,­regulatory­program­reconstruction,­the­presented­heat­maps­and­tree.­The­ImmGen­website­is­continuously­updated.

Data preprocessing.­Expression­data­were­normalized­as­part­of­the­ImmGen­pipeline­ by­ the­ robust­ multiarray­ average­ method.­ Data­ were­ log2­ trans-formed.­For­genes­with­more­than­one­probe­set­on­the­array,­only­the­probe­set­with­the­highest­mean­expression­was­retained.­Of­those,­only­probe­sets­with­a­s.d.­value­above­0.5­for­the­entire­data­set­were­used­for­the­clustering,­which­resulted­with­7,965­unique­genes­with­a­difference­in­expression­in­the­September­2011­release­and­8,431­in­the­April­2012­release.

Lineage-specific signatures.­We­calculated­signatures­for­11­lineages:­granu-locyte,­macrophage,­monocyte,­DC,­B­cell,­NK­cell,­CD4+­T­cell,­CD8+­T­cell,­NKT­cell,­γδ­T­cell­and­stem­and­progenitor­cell.­Assignment­of­samples­into­lineages­is­in­Supplementary Table 2.­One-way­analysis­of­variance­was­done­for­each­of­the­6,997­genes­with­an­expression­value­above­log2(120)­in­at­least­one­lineage,­followed­by­post-hoc­analysis­(functions­anova1­and­multcompare­in­MATLAB­software).­For­each­of­the­11­lineages,­a­gene­was­considered­induced­if­it­had­significantly­higher­expression­in­that­lineage­than­in­all­other­lineages.­A­gene­was­considered­repressed­if­it­had­significantly­lower­expres-sion­in­that­lineage­than­in­all­other­lineages.­A­false-discovery­rate­(FDR)­of­10%­was­applied­to­the­analysis­of­variance­P­values­of­all­genes.

Definition of modules.­ Modules­ were­ defined­ by­ clustering.­ For­ coarse-grained­ modules,­ clustering­ was­ done­ by­ superparamagnetic­ clustering­(SPC)27,­a­principled­approach­for­choosing­stable­clusters­from­a­hierarchi-cal­setting.­SPC­was­used­because­it­does­not­require­a­predefined­number­­of­ clusters­ but­ instead­ identifies­ the­ number­ inherently­ supported­ by­ the­data.­The­clusters­defined­by­SPC­are­stable­across­a­range­of­parameters,­although­ they­ can­ have­ variable­ degrees­ of­ compactness.­ SPC­ was­ run­­with­default­parameters,­which­resulted­in­80­stable­clusters­(coarse-grained­modules­ C1–C80);­ the­ remaining­ unclustered­ genes­ were­ grouped­ into­ a­­separate­cluster­(C81).

Each­ coarse-grained­ module­ was­ further­ partitioned­ into­ fine-grained­modules­by­affinity­propagation­clustering28,­with­correlation­as­the­affinity­measure.­The­‘self-responsibility’­parameter­(which­indicates­the­propensity­of­the­algorithm­to­form­a­new­cluster)­was­set­at­0.01.­Affinity­propagation­was­used­because­SPC­and­hierarchical­clustering­did­not­further­break­the­coarse­modules.­Affinity­propagation­could­not­be­used­for­clustering­of­all­genes,­because­it­must­work­with­a­‘sparsified’­affinity­matrix.

Clustering­resulted­in­334­fine-grained­modules­(F1–F334).­On­average,­3.9­ fine-grained­ modules­ were­ nested­ in­ a­ single­ coarse-grained­ module.­The­minimum­number­of­fine­modules­nested­in­a­coarse-grained­module­was­ 1­ (23­ coarse-grained­ modules)­ and­ the­ maximum­ was­ 11­ (7­ coarse-­grained­modules).

Choice of candidate regulators.­ Candidate­ regulators­ were­ curated­ from­the­following­sources:­mouse­orthologs­of­all­the­genes­encoding­molecules­used­as­candidate­regulators­in­a­published­study­of­human­hematopoiesis3;­genes­annotated­with­the­gene-ontology­term­‘transcription­factor­activity’­in­mouse,­human­or­rat;­genes­for­which­there­is­a­known­DNA-binding­motif­in­TRANSFAC­matrix­database­(version­8.3)29,­the­JASPAR­database­(version­2008)30­and­experimentally­determined­position­weight­matrices­(PWMs)31,32;­and­genes­with­published­data­obtained­by­ChIP­followed­by­deep­sequenc-ing­or­ChIP­followed­by­microarray­(Supplementary Table 11).­Regulators­that­were­not­measured­on­ the­array­or­whose­expression­did­not­ change­sufficiently­(s.d.­<­0.5­across­the­entire­data­set)­to­be­included­in­the­cluster-ing­were­removed,­unless­they­were­highly­correlated­(>­0.85)­with­another­regulator­ that­passed­ the­cutoff.­This­ resulted­ in­578­candidate­ regulators­(Supplementary Table 12).

Hematopoietic tree building.­The­hematopoietic­tree­(Fig. 1)­was­built­by­the­members­of­the­ImmGen­Consortium.­Each­group­created­its­own­sub-lineage­tree,­and­the­sublineage­trees­were­connected­on­the­basis­of­the­best­knowledge­available­at­present,­although­some­edges­are­hypothetical­(dashed­lines,­Fig. 1).­There­are­two­roots­to­the­tree:­long-term­stem­cells­from­adult­bone­marrow,­and­long-term­stem­cells­from­fetal­liver.­Each­population­is­a­node­in­the­tree­(square,­Fig. 1).­Edges­indicate­a­differentiation­step,­an­activation­step,­time­(as­in­the­activated­T­cells)­or­a­general­assumption­of­similarity­in­regulatory­program­(Supplementary Table 13).­Some­intermedi-ate­inferred­nodes­were­added­to­group­cell­populations­that­were­assumed­to­have­a­common­progenitor­or­common­regulatory­program­but­for­which­this­hypothetical­population­was­not­measured­(for­example,­granulocytes­and­macrophages).­For­the­populations­that­connected­to­more­than­one­parent­population,­one­of­the­edges­was­manually­pruned,­either­the­less­likely­one­or­arbitrarily­(Supplementary Table 13).

Module regulatory program.­Ontogenet­takes­the­following­as­input:­gene-expression­profiles­across­many­different­cell­types;­a­partitioning­of­the­genes­into­modules­(the­coarse-grained­and­fine-grained­clusters­described­above);­a­predefined­set­of­candidate­regulators;­and­an­ontogeny­tree­relating­the­cell­types.­It­then­constructs­a­regulatory­program­for­each­module­consisting­of­a­linear­combination­of­regulators­with­possibly­distinct­activity­weights­for­each­regulator­in­each­cell­type.­A­module’s­regulatory­program­is­the­linear­sum­of­the­regulators’­expression­multiplied­by­each­regulator’s­activity­weight,­which­approximates­the­expression­pattern­of­the­module.­Each­regulatory­program­aims­to­explain­as­much­of­the­gene-expression­variance­in­the­module­as­possible­while­remaining­as­simple­as­possible­and­being­consistent­across­related­cell­types­in­the­ontogeny.­In­a­regular­linear­model,­the­activity­weights­are­constant­across­all­conditions.­Here,­we­allow­a­change­of­activity­weights­between­cell­types­(Fig. 3).

Notably,­all­regulators­are­considered­as­potential­regulators­for­each­mod-ule.­That­includes­regulators­that­are­members­of­the­module.­Thus,­a­module­can­be­assigned­regulators­that­are­its­members­and­regulators­that­are­not­its­members,­but­regulators­that­are­members­of­the­module­will­not­necessarily­be­assigned­to­it.

More­formally,­we­model­the­expression­of­a­gene­in­a­module­as­a­(noisy)­linear­combination­of­the­expression­of­the­regulators.­We­denote­the­activity­of­a­regulator­r­in­a­cell­type­t­as­ar,t.­We­model­the­expression­of­a­gene­i,­a­member­of­module­m,­in­cell­type­t­as

X w ai t m r t r t m tr, , , , ,= +∑ e

where­each­εm,t­is­a­Gaussian­random­variable­with­0­mean­and­variance­sm t,2 ­

specific­to­a­combination­of­a­module­m­and­a­cell­type­t.­Hence­the­regula-tory­program­learned­by­Ontogenet­is­represented­in­terms­of­wm,r,t­activity­weights­specific­to­a­combination­of­module,­regulator­and­cell­type.­Because­of­parameter­tying­enforced­by­the­model,­the­effective­number­of­parameters­is­much­smaller­than­the­nominal­size­of­the­regulatory­program­representation­(modules)­×­(regulators)­×­(cell­types).

Module cell-type specific variance estimation.­ The­ module­ variance­in­a­given­cell­ type­sm t,

2 ­ is­ estimated­ from­ the­expression­of­ the­module’s­­

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunologydoi:10.1038/ni.2587

member­genes­across­all­replicates­of­the­cell­type.­Although­we­use­an­unbi-ased­estimator,­we­make­special­considerations­for­the­modules­with­less­than­10­members.­For­these­modules­the­variance­estimate­sm t,

2 ­is­computed­by­a­pooled­variance­estimator­across­modules­with­more­than­ten­members­but­still­specific­to­the­cell­type.­The­estimated­variances­in­a­fine-grained­module­are­typically­smaller­than­the­variances­in­its­parent­coarse-grained­module.

Regulatory program fitting as a penalized regression problem.­Estimation­of­the­activity­weights­wm,r,t­takes­the­form­of­a­regression­problem,­but­because­of­ ‘over-parameterization’­of­ the­problem,­ it­must­be­ ‘regularized’­with­an­extension­of­ the­ fused­Lasso­ framework33,­which­gives­rise­ to­a­penalized­regression­problem­of­the­form

1 12 2

2

nx w a J w

m m ti t m r t r tri t s ,, , , ,, ( ),−( ) +∑∑

where­J(w)­is­a­chosen­penalty.­In­our­case,­this­penalty­is­composed­of­two­parts,­one­that­promotes­sparsity­and­selection­of­correlated­predictors­and­another­that­promotes­consistency­of­regulatory­programs­between­related­cell­types.

We­assume­that­only­a­small­number­of­regulators­are­actively­regulating­any­one­module.­A­standard­approach­to­promoting­such­sparsity­in­regres-sion­problems­is­to­introduce­an­L1­penalty,­the­sum­of­absolute­values­Σm­Σr­Σt­|­wm,r,t­|.­However,­this­penalty­tends­to­be­overly­aggressive­in­inducing­sparsity­and­thus­prunes­many­highly­correlated­predictors­and­selects­only­a­single­representative.­Such­aggressive­pruning­may­be­inappropriate,­as­the­correlated­regulators­may­all­be­biologically­relevant­because­of­‘redundancy’­in­densely­interconnected­regulatory­circuits.­That­can­be­counteracted­by­the­addition­of­squared­terms­12

2Σ Σ Σm r t m r tw( ), , ,­which­yields­a­composite­penalty­known­as­‘Elastic­Net’11,­as­proposed­before6,

l k| | ( ), , , ,w wm r tt m r ttrr∑ ∑∑∑ +2

2

which­we­write­compactly­as

l k|| || || ||w wm m1 2 22+

An­important­input­to­our­regulatory­program­fitting­procedure­is­the­ontog-eny­(differentiation)­tree­(Supplementary Table 13).­This­tree­is­encoded­as­an­edge­list­(f ),­and­with­(t1,­t2)∈f­we­denote­that­cell­type­t1­is­a­parent­of­cell­type­t2.­The­similarity­of­the­regulatory­programs­for­a­particular­module­in­two­related­cell­types­(t1,­t2)∈f­can­be­assessed­as­a­sum­of­the­absolute­value­of­the­difference­of­activity­weights­in­the­two­programs,­Σr m r t m r tw w| |, , , ,1 2− .­The­key­observation­is­that­| |, , , ,w wm r t m r t1 2− ­=­0­if­the­regulatory­relationship­between­regulator­r­and­module­m­is­the­same­in­cell­type­t2­and­its­parent­type­t1.­More­generally,­the­total­difference­of­the­regulatory­programs­can­be­written­as­Σ Σ( , ) , , , ,| |t t f r m r t m r tw w1 2 1 2∈ − .­We­will­write­this­term­in­a­compact­form­as­||Dwm||,­where­wm­ is­a­vector­of­activity­weights­for­all­regulators­across­all­cell­types­concatenated­together­and­D­is­a­matrix­of­size­(RE)­×­(RT),­where­R­is­the­number­of­regulators,­T­is­the­number­of­cell­types­and­E­is­the­number­of­edges­in­the­tree.­We­note­that­multiplication­by­the­matrix­D­computes­the­differences­between­relevant­entries­of­the­vector­wm.­The­less­the­regulatory­programs­change­throughout­differentiation,­the­smaller­the­term­||Dwm||.­Thus,­with­this­term­as­a­penalty­will­promote­the­preservation­of­a­consistent­regulatory­program­throughout­differentiation.

Combining­all­the­considerations­above,­the­complete­objective­for­fitting­a­regulatory­program­of­a­module­m­is­given­by

1 12 22

21 2

2n

x w a w w Dwm m t

i t m r t r tr m m msl k g

,, , , , || || || || || |−( ) + + +∑ || ., 1i t∑

Optimization­of­this­objective­is­somewhat­complicated­by­the­fact­that­absolute­­value­is­a­non-smooth­function­and­hence­direct­optimization­by­methods­such­as­gradient­descent­is­not­feasible,­as­these­work­only­on­smooth­prob-lems.­Alternative­methods,­such­as­projected­gradients,­can­be­used,­but­their­convergence­is­relatively­slow.­We­therefore­opted­to­use­a­primal­dual­interior­point­method34.­Different­choices­of­the­parameters­λ,­κ­and­δ­yield­different­­

regulatory­ models­ as­ solutions,­ with­ different­ data-fitting­ and­ model-­complexity­properties.­We­scanned­sets­of­parameters­in­the­range­(the­sched-ule­for­each­of­the­parameters­λ,­κ­and­δ­was­geometric,­e−7,­e−6,…,­e3­spanning­values­between­0.001­and­20)­and­chose­the­optimal­set­of­parameters­with­the­Bayesian­information­criteria­(described­below).

To­simplify­ the­discussion­of­ the­optimization,­we­ introduce­ the­sparse­predictor­matrix­A­of­size­(RT)­×­(T),­where­A at r T t r t mt,( ) , /− + =1 s ­and­=­0­­otherwise.­Furthermore,­we­note­that­the­optimal­wm­depends­only­on­the­mean­expression­profile­of­the­module’s­genes­and­we­can­introduce­variable­y

nxt

mti m

mi t= ∈

1 1s

Σ , .­Hence­we­can­rewrite­the­objective­as

12 22

21 2

21|| || || || || || || || .y Aw w w Dwm m m m− + + +l k g

Finally­we­can­absorb­the­term­k2 22|| ||wm ­into­the­first­term­as­follows:

12 0

2

2

1 1y A

Iw w Dwm m m

−

+ +k

l g|| || || || .­­

Regulatory program transfer between coarse-grained and fine-grained modules.­The­fine-grained­modules­are­‘encouraged’­to­have­a­program­similar­to­that­of­the­coarse-grained­module­in­which­they­are­nested.­This­is­accom-plished­by­the­introduction­of­an­additional­penalty­term.­We­will­denote­the­already­learned­regulatory­program­of­a­coarse-grained­module­as­w0­and­the­regulatory­program­of­a­fine-grained­module­that­we­wish­to­learn­as­wm.­The­coarse-to-fine­version­of­our­objective­is­then

12 2 22

21 2

21 0 2

2|| || || || || || || || || ||y Aw w w Dw w wm m m m m− + + + + −l k g t

where­the­last­term­ties­the­programs­of­the­coarse-grained­and­fine-grained­modules.­This­objective­can­be­transformed­into

12

0

0 2

2

1

y

w

A

I

I

w w Dwm m m

t

k

t

l g

−

+ +|| || || ||| .1

­­

Solving the prototypical optimization problem.­We­note­that­regulatory-program-fitting­problems­for­both­coarse-grained­and­fine-grained­module­have­been­expressed­in­the­following­general­form

minimizew

y Xw w Dw12 2

21 1|| || || || || || .− + +l g

We­reformulate­that­optimization­problem­by­adding­variables­that­decouple­the­penalties:

minimize

subject to

w, , || || || ||

, ,

z d r r z d

r y Xw z w d Dw

12 1 1′ + +

= − = =

l g

..

This­reformulation­enables­straightforward­derivation­of­a­primal­dual­interior­point­method34.

Model selection with Bayesian information criterion.­The­formulation­of­our­optimization­problem­above­is­dependent­on­the­set­of­parameters­λ,­κ­and­δ;­we­obtain­a­model­by­solving­the­convex­problem­above­for­a­particular­combination­of­λ,­κ­and­δ.­Different­combinations­of­these­parameters­will­yield­regulatory­programs­of­different­quality.­One­way­to­identify­the­optimal­λ,­κ­and­δ­is­through­the­use­of­held-out­data­or­through­cross-validation.­However,­a­search­for­these­parameters­with­cross-validation­would­be­pro-hibitively­expensive.­As­an­alternative,­we­use­a­model­selection­approach­based­on­the­Bayesian­information­criterion­(BIC)­to­compare­models­result-ing­from­different­choices­of­these­three­parameters­and­select­the­best­one.­

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunology doi:10.1038/ni.2587

This­criterion­compares­models,­here­‘encoded’­by­regulatory­programs,­based­on­their­tradeoff­between­data­log­likelihood­and­degrees­of­freedom.­The­log­likelihood­for­our­model­is

LL const( ) .,

, , , ,,w x w am t

i t m r t r tri tm= − −( ) +∑∑∑ 12 2

2

s

The­computation­of­the­degrees­of­freedom­is­somewhat­technically­involved­but­ intuitively­simple:­an­activity­weight­ that­ remains­ the­same­through­a­particular­connected­portion­of­the­differentiation­tree­is­counted­as­a­sin-gle­degree­of­freedom.­To­make­this­more­formal,­we­consider­matrix­A­and­construct­its­counterpart,­B.­We­use­Ar,t­to­denote­a­column­of­matrix­A.­We­then­construct­a­graph­in­which­nodes­correspond­to­columns­of­matrix­A.­Given­ two­ nodes­ corresponding­ to­ Ar t, 1­ and­Ar t, 2,­ the­ graph­ will­ have­ an­edge­between­these­two­nodes­if­cell­type­t1­is­a­parent­of­cell­type­t2,­and­w wm r t m r t, , , ,1 2= .­The­matrix­B­will­have­columns­that­are­sums­of­columns­corresponding­to­connected­components­in­the­graph.­We­eliminate­all­col-umns­of­B­that­are­zeros­and­the­final­degrees­of­freedom­are­given­by­df(w)­=­Trace(B(B′B­+­κ­diag(c))−1­B′),­where­diag(c)­is­a­diagonal­matrix­with­entries­being­the­number­of­columns­of­A­in­the­connected­component­associated­with­a­column­of­B.

Hence­we­can­compute­the­BIC(w)­as

BIC w LL df

df

( ) ( ) ( )log

( )l,

, , , ,

= − +

= −( ) +∑2

12

2

w w T

x w a wm t

i t m r t r trsoog ., Ti tm∑∑

­­

Post-processing of regulatory programs.­Once­we­obtain­an­optimal­regula-tory­program­in­terms­of­BIC,­we­use­post-processing­to­remove­regulatory­relationships­for­underexpressed­regulators.­We­placed­a­low­expression­cutoff­of­5.5­on­the­log2­scale.­At­this­cutoff,­the­correlation­between­the­predictor­and­the­target­module­may­very­well­be­due­to­noise­and­hence­the­relation-ship­could­be­spurious.

Systematic query for known functions of regulators.­ For­ each­ lineage-­specific­module,­we­automatically­queried­its­regulators­in­PubMed­with­the­name­of­the­lineage(s).­For­each­module­that­was­upregulated­or­downregu-lated­with­differentiation,­we­queried­its­regulators­with­‘hematopoietic­dif-ferentiation’.­All­PubMed­queries­were­done­on­30­July­2012.

Choice of lineage regulators.­For­each­lineage,­we­collected­the­regulators­deemed­active­by­having­a­nonzero­activity­weight­and­significant­expression­in­excess­of­9.0­on­the­log2­scale.­For­a­given­lineage,­we­deemed­a­regulator­a­ lineage­activator­ if­ its­average­activity­weight­across­all­ cell­ types­ in­ the­lineage­and­all­modules­was­positive.­Analogously,­a­regulator­was­deemed­a­lineage­repressor­if­its­average­activity­weight­was­negative.­We­subsequently­ranked­the­regulators­on­the­basis­of­their­average­expression­across­cell­types­in­which­the­regulator­had­a­role.­Hence,­the­regulators­that­were­frequently­active­in­a­lineage­and,­when­active,­had­higher­expression­were­ranked­higher­than­were­regulators­that­were­infrequently­active­or­had­low­expression.­The­regulators­with­the­highest­expression­typically­were­given­the­highest­total­activity­weight­across­lineages.

Notably,­this­procedure,­although­straightforward,­will­not­reflect­all­the­lineage­regulators­identified­by­the­model.­First,­those­lineage­regulators­that­act­only­during­a­limited­window­(for­example,­early­in­differentiation)­would­be­under-represented­by­this­analysis­yet­would­be­captured­in­the­overall­model­in­the­window­in­which­they­act.­Second,­because­of­the­post-processing­­step­(described­above),­regulators­with­high­baseline­expression­can­have­a­constant­activity­weight­even­if­their­expression­is­very­lineage­specific­(for­example,­GATA-3)­and­thus­be­under-represented­in­the­recruitment­analysis­(although­they­too­are­chosen­as­regulators­in­the­model).

Motif scanning.­We­scanned­promoters­of­mouse­genes­for­enriched­motifs.­We­ downloaded­ promoter­ sequences­ for­ mouse­ (mm9)­ from­ the­ genome­browser­website­of­the­University­of­California­Santa­Cruz­(http://hgdownload.cse.ucsc.edu/downloads.html).­For­each­gene,­we­scanned­the­region­starting­

from­position­–1,000­(base­pairs­upstream­of­the­transcription­start­site)­and­ending­at­position­+200­(base-pairs­downstream­of­ the­ transcription­start­site).­We­represented­the­nucleotide­at­position­j­(relative­to­–1,000­bp­from­the­transcription­start­site)­for­gene­i­as­Si,j.­We­represented­each­cis-regulatory­element­by­a­PWM.­We­compiled­a­set­of­1,651­PWMs­from­the­TRANSFAC­matrix­database­(version­8.3)29,­the­JASPAR­database­(version­2008)30­and­experimentally­determined­PWMs31,32.­We­denote­the­PWM­of­the­‘k-th’­motif­by­Pk,­a­matrix­of­size­4­×­Lk,­where­Lk­is­the­length­of­the­motif­and­Pk(i,j)­rep-resents­the­probability­of­encountering­the­nucleotide­j­(that­is,­A,­C,­G­or­T)­­at­the­‘i-th’­position.­For­each­gene­i,­a­position­along­the­promoter­j­and­a­PWM­k,­we­computed­the­local­motif-matching­score­LOD(i,j,k),­defined­as­the­ log­ likelihood­ratio­(LOD­score)­ for­observing­the­sequence­given­the­PWM­versus­a­given­random­genomic­background:

LOD i j k P r S P Sk i j rr

L

b i j rk

( , , ) log ( , ) log ( ), ,= −

+ −=

+ −∑ 2 11

2 1

Genomic­background­was­determined­as­Pb(‘A’)­=­Pb(‘T’)­=­0.3,­Pb(‘C’)­=­Pb(‘G’)­ =­ 0.2,­ which­ represents­ the­ nucleotide­ composition­ of­ the­ mouse­genome.­ We­ then­ found­ the­ best­ motif­ instance­ over­ the­ entire­ promoter­region,­defined­as­MAX-LOD(i,k)­=­maxjLOD(i,j,k).

Motif-scoring threshold.­We­automatically­computed­a­PWM-specific­thresh-old­by­using­the­information­content­of­each­motif.­The­information­content­for­the­‘k-th’­motif­is­defined­as

IC k L P i j P i jk k kji

Lk( ) ( , ) log ( , )= +

==∑∑ 21

4

1

We­defined­the­PWM-specific­threshold­for­the­‘k-th’­motif­as­τk,­the­1­–­2−IC(k)­quantile­of­the­PWM­LOD­distribution­across­all­genes’­promoters.­We­con-sidered­a­‘hit’­for­the­‘k-th’­motif­at­the­‘i-th’­gene­if­the­best­score­(MAX-LOD(i,k))­exceeded­the­threshold­τk.

Motif enrichment in modules.­For­each­module­of­genes­M,­and­each­motif­k,­we­computed­the­P­value­for­enrichment,­pe(M,k)­of­the­motif­in­the­mod-ule­relative­to­that­of­the­entire­set­of­genes­assigned­to­modules­serving­as­background.­An­enrichment­of­a­motif­ in­a­module­results­ in­higher­than­expected­MAX-LOD­scores­for­the­genes­in­this­module;­to­capture­this­effect,­we­computed­the­P­value­by­comparing­the­scores­MAX-LOD(i,k)­for­all­genes­i­in­the­module­M­and­the­scores­for­the­entire­set­of­genes­assigned­to­mod-ules­by­a­one-sided­rank-sum­test.­We­then­used­an­FDR­of­5%­on­the­entire­matrix­of­P­values­pe(M,k)­and­declared­all­P­values­that­passed­this­procedure­significant­‘hits’.­The­FDR­was­calculated­separately­for­coarse-grained­and­fine-grained­modules.

Binding events enrichment.­ The­ public­ data­ sets­ obtained­ by­ ChIP­­followed­by­deep­sequencing­and­ChIP­followed­by­microarray­(Supplementary Table 11)­ were­ downloaded­ from­ the­ GEO­ (Gene­ Expression­ Omnibus)­­database­ repository,­ supplementary­ material­ and­ designated­ sites­ in­ the­­original­ publications­ (Supplementary Table 11;­ 360­ experiments­ of­ 109­unique­regulators).­The­target­list­defined­in­each­original­publication­was­used­whenever­available.­Otherwise,­genes­that­had­a­binding­event­reported­from­the­position­1,000­base­pairs­upstream­of­the­transcription­start­site­to­the­position­200­base­pairs­downstream­of­the­transcription­start­site­were­listed­as­targets.­In­data­sets­obtained­with­human­samples,­gene­symbols­were­replaced­by­the­mouse­gene­symbol­wherever­a­one-to-one­ortholog­exists­accord-ing­to­the­phylogenetic­resource­EnsemblCompara35.­Only­genes­included­in­the­clustering­were­considered­targets­for­the­purpose­of­the­calculation­­of­enrichment.

The­hypergeometric­P­value­was­calculated­for­the­size­of­intersection­of­each­module­with­each­target­ list.­An­FDR­of­10%­was­used­for­the­entire­table­of­P­values­of­all­modules­and­all­targets­lists.­The­FDR­was­calculated­separately­for­coarse-grained­and­fine-grained­modules.

©20

13 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature immunologydoi:10.1038/ni.2587

Estimating the significance of regulatory program overlap.­We­report­two­­P­ values­ for­ each­ overlap­ of­ the­ three­ regulation­ models­ (from­ ChIP,­­cis-elements­and­Ontogenet).­First,­we­calculated­the­hypergeometric­test­for­two­or­three­groups­for­which­the­‘universe­sizes’­were­the­number­of­possible­regulatory­interactions­including­the­overlapping­regulators.­For­example,­for­estimation­of­the­significance­of­the­overlap­of­ChIP­and­Ontogenet­regulatory­interactions,­the­‘universe­size’­is­the­number­of­regulators­that­were­candi-dates­for­Ontogenet­and­had­ChIP­information­multiplied­by­the­number­of­modules.­The­ChIP­interactions­are­the­enriched­modules­according­to­the­ChIP­data­set,­and­the­Ontogenet­interactions­are­the­regulators­chosen­for­each­module.­Second,­we­calculated­an­empirical­P­value­from­10,000­permuta-tions­of­the­regulators­in­the­regulatory­interactions,­including­the­overlapping­regulators.­The­last­P­values­were­calculated­to­account­for­the­fact­that­some­modules­have­more­regulators­than­others.­The­hypergeometric­P­values­and­the­empirical­P­values­are­similar­for­the­overlap­of­each­two­methods­but­dif-fer­in­significance­for­the­three-method­overlap,­because­the­hypergeometric­score­for­three­groups­explicitly­takes­into­account­the­overlap­between­each­two­groups,­whereas­the­empirical­P­value­does­not.

Functional enrichment.­Curated­gene­sets­(C2),­motif­gene­sets­(C3)­and­gene­ontology­(GO)­gene­sets­(C5)­from­the­Molecular­Signatures­Database­(version­V.3)­were­downloaded­from­the­Broad­Institute­website­(http://www.broadinstitute.org/gsea)36.­For­each­group,­gene­symbols­were­replaced­by­the­mouse­gene­symbol­wherever­a­one-to-one­ortholog­exists­according­to­EnsemblCompara.­ Only­ genes­ included­ in­ the­ clustering­ were­ considered­functional­group­members­for­the­purpose­of­the­calculation­of­enrichment.

A­hypergeometric­P­value­was­calculated­for­the­size­of­intersection­of­each­module­with­each­functional­group.­An­FDR­of­10%­was­used­for­the­entire­table­of­P­values­of­all­modules­and­all­functional­groups.­The­FDR­was­cal-culated­separately­for­coarse-grained­and­fine-grained­modules,­and­for­the­different­classes­of­functional­annotation­(C1,­C2,­C3­and­C5).

Identification of differentiation steps with a change in activity weight of regulators.­For­each­module­and­each­edge­(differentiation­step)­of­the­hemato-poietic­tree,­the­activity­weight­of­the­‘parent’­was­compared­with­the­activity­weight­of­the­‘child’,­which­resulted­in­one­of­the­following­classifications:­no­change­(activity­weights­are­the­same);­activator­recruitment­(parent­activity­weight­is­0;­child­activity­weight­is­positive);­activator­strengthening­(parent­activity­weight­is­positive­and­is­smaller­than­that­of­the­child);­activator­dis-posal­(parent­activity­weight­is­positive­and­child­activity­weight­is­0);­repres-sor­recruitment­(parent­activity­weight­is­0;­child­activity­weight­is­negative);­repressor­strengthening­(parent­activity­weight­is­negative­and­is­larger­than­that­of­the­child);­repressor­disposal­(­parent­activity­weight­is­negative­and­child­activity­weight­is­0).­For­simplicity,­we­omitted­the­‘regulator­weakening’­option.­Those­lineage-specific­regulators­that­are­assigned­constant­activity­

weight­across­all­cell­ types­ (such­as­GATA-3)­will­not­be­captured­by­ this­analysis­but­are­part­of­the­model.

Mice.­ Mice­ with­ loxP-flanked­ Etv5­ alleles­ (Etv5fl/fl)37­ were­ crossed­ with­C57BL/6­mice­with­a­transgene­encoding­Cre­recominase­driven­by­the­pro-moter­of­the­gene­encoding­CD2­to­generate­mice­with­T­cell–specific­ETV5­deficiency­(CD2p-CreTg+Etv5fl/fl,­backcrossed­three­times­to­the­C57BL/6­strain).­The­loxP-flanked­Etv5­locus­is­specifically­deleted­from­the­genome­starting­ in­ CD25+CD44−CD3−CD4−CD8−­ thymic­ precursors­ (DN3)­ with­~80%­deletion­efficiency­in­γδ­thymocyte­subsets,­as­inferred­from­the­analy-sis­of­Cre-activity-reporter­mice­(CD2p-CreTg+Rosa-STOPfl/fl-EYFP).

Flow cytometry.­Intracellular­staining­(Cytofix/Cytoperm­Kit;­BD­Biosciences)­and­ intranuclear­ staining­ (FoxP3­ Staining­ Kit;­ eBioscience)­ were­ done­ as­described24.­ The­ following­ antibodies­ were­ used:­ anti-TCRδ­ (GL3),­ anti-CD24­(HSA,­M1/69),­anti-CD44­(IM7),­anti-CD62l­(MEL-15),­anti-IL-17A­(ebio17B7)­and­anti-RORγt­(AFKJS-9;­all­ from­eBioscience);­and­anti-Vγ2­(UC3-10A6),­ anti-Vδ6.3­ (8F4H7B7),­ anti-CCR6­ (140706)­ and­ anti-CD27­(LG.3A10;­ all­ from­ BD­ Biosciences).­ Anti-Vγ1.1­ (2.11)­ was­ purified­ from­culture­supernatant­and­was­biotinylated­with­the­FluoReporter­Mini-Biotin-XX­Labeling­Kit­(Invitrogen).­Data­were­acquired­on­an­LSRII­(BD)­and­were­analyzed­with­FlowJo­software­(Treestar).

27. Blatt, M., Wiseman, S. & Domany, E. Superparamagnetic clustering of data. Phys. Rev. Lett. 76, 3251–3254 (1996).

28. Frey, B.J. & Dueck, D. Clustering by passing messages between data points. Science 315, 972–976 (2007).

29. Matys, V. et al. TRANSFAC and its module TRANSCompel: transcriptional gene regulation in eukaryotes. Nucleic Acids Res. 34, D108–D110 (2006).

30. Sandelin, A., Alkema, W., Engström, P., Wasserman, W.W. & Lenhard, B. JASPAR: an open-access database for eukaryotic transcription factor binding profiles. Nucleic Acids Res. 32 (suppl. 1), D91–D94 (2004).

31. Badis, G. et al. Diversity and complexity in DNA recognition by transcription factors. Science 324, 1720–1723 (2009).

32. Berger, M.F. et al. Variation in homeodomain DNA binding revealed by high-resolution analysis of sequence preferences. Cell 133, 1266–1276 (2008).

33. Tibshirani, R., Saunders, M., Rosset, S., Zhu, J. & Knight, K. Sparsity and smoothness via the fused lasso. J. R. Stat. Soc. Series B Stat. Methodol. 67, 91–108 (2005).

34. Boyd, S.P. & Vandenberghe, L. in Convex Optimization, Ch 11.7 (Cambridge University, Cambridge, UK, 2004).

35. Vilella, A.J. et al. EnsemblCompara GeneTrees: complete, duplication-aware phylogenetic trees in vertebrates. Genome Res. 19, 327–335 (2009).

36. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA 102, 15545–15550 (2005).

37. Zhang, Z., Verheyden, J.M., Hassell, J.A. & Sun, X. FGF-regulated Etv genes are essential for repressing Shh expression in mouse limb buds. Dev. Cell 16, 607–613 (2009).