1.Inferring Regulatory Pathways:Bayesian NetworksModel Based Approach Step 1: define class of potential models Step 2: reconstruct a specific model Step 3: visualization & testable hypotheses Emphasis on the choice of model and how to use it The data manipulation steps are derived from the modelKin ZTF XGene ATF Y TF XGene AUSV1Hap4Msn2 Hap4Atp16ReconstructionInterpretationAdditionalBiologicalKnowledgeDataKin ZTF XGene ATF YTF XGene AKin ZTF XGene ATF YTF XGene AKin ZTF XGene ATF YTF XGene AWhat is a model?“A description of a process that could havegenerated the observed data”stochastic2From Data to Signaling PathwaysRafErkp38PKAPKCJnkPIP2PIP3PlcγAkt...Cell xCell1Cell2Cell3Cell4Automated de-novo reconstructionfrom high throughput data.FlowcytometryComputationalanalysis Datasets of cells• condition ‘a’• condition ‘b’•condition…‘n’RafMek1/2Erkp38PKAPKCJnkPIP2PIP3PlcγAkt12 Color Flow Cytometryperturbation aperturbation nperturbation bConditions (96 well format)Primary Human T-Lymphocyte DataSimultaneous information measured individualprimary human cells.Thousands of data points.3Interventions on a known map…MEK3/6MAPKKKPLCγErk1/2Mek1/2RafPKCp38AktMAPKKKMEK4/7JNKLATLckVAVSLP-76RASPKA1 2 3CD28CD3PI3KLFA-1CytohesinZap70PIP3PIP2JAB-1Activators 1. α-CD3 2. α-CD28 3. ICAM-2 4. PMA 5. β2cAMPInhibitors 6. G06976 7. AKT inh 8. Psitect 9. U0126 10. LY29400210 546798Using CorrelationsPKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3 Phospho-Proteins Phospho-Lipids4 Node: Measured level/activity of protein Edge: Influence between proteinsProtein AProtein BProtein CProtein DProtein ECorrelations between cellular componentsrepresent interactions and influencesBayesian Network ApproachStatistical DependenciesBut, how can statistical dependencies lead to causality?ABC DE5No ManipulationsA inhibitedB inhibitedPhospho APhospho BB ACausality: Who Influences Whom?B A?Praf PmekNo interventionsMek InhibitedSome Real Data6 What would happen if B was not measured?ACPhospho APhospho CIndirect InfluencesBContext SpecificityPhospho BPhospho D E is high B and D seem unrelated Relationship is revealed byconsidering simultaneousmeasurement of EE DB7 Dataset:Per Condition:600 cellsTotal:5400 cellsRafMek1/2Erkp38PKAPKCJnkPIP2PIP3PlcγAktMulti-Color Flow Cytometryperturbation aperturbation nperturbation bConditions (96 well format)BayesianNetworkAnalysisInfluencenetwork ofmeasuredvariablesHuman CD4+ T-Lymphocyte DataPKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3 Phospho-Proteins Phospho-Lipids Perturbed in dataInferred Network8PKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3 Phospho-Proteins Phospho-Lipids Perturbed in dataKnown Direct InteractionDirect InteractionsPhosphorylationHydrolysisAdvantagesIn VivoCaptures variety of mechanismsDirectionality of influencePKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3 Phospho-Proteins Phospho-Lipids Perturbed in dataKnown Indirect SignalingIndirect Signaling9Indirect Signaling Indirect signaling Dismissing edgesRaf Mek ErkPKC JnkPKCMapkkkJnkNot measuredMek4/7Indirect connections can be found even when theintermediate molecule(s) are not measured Phospho-Proteins Phospho-Lipids Perturbed in dataPKCRafErkMekPlcγPKAAktJnkP38PIP2PIP31Reversed3Missed17/17Reported15/17T CellsInferred T cell signaling map[Sachs et al, Science 2005]10Prediction &ProofErk1/2 unperturbed• Erk influence on Akt previously reported in coloncancer cell linesPredictions:• Erk1/2 influences Akt• While correlated, Erk1/2 does not influence PKACD4EGFP49siRNA transfected101102103101102103ErkPKAAktPKCRafMeksiRNA101102103UnstimulatedUnstimulatedCD3/CD28stimulatedCD3/CD28stimulatedp=0.000094 p=0.28101102103Phospho-AKTRelative # of cellsPhospho-PKA C.ControlErk1 siRNAPower of InterventionsPKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3N/AReversed10Missed1Unexplained1/10Reported8/10ExpectedA.LackingIntervention data1CompleteDataset302/1715/17Dataset: 1200 samples:• 2 conditions• no interventions11Power of Large Dataset2N/AReversed 1010Missed61Unexplained 1/151/10Reported 8/158/10Expected B.TruncateddataA.LackingInterventiondata1CompleteDataset302/1715/17PKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3Dataset: 420 samples:• 14 conditions• 30 random cells each Power of Single Cell32N/AReversed12 1010Missed861Unexplained1/16 1/141/10Reported6/16 8/158/10ExpectedC.“Westernblot” B.TruncateddataA.LackingInterventiondata1CompleteDataset302/1715/17PKCRafErkMekPlcγPKAAktJnkP38PIP2PIP3Simulated western blot: 420 samples:• 14 conditions• Each point average of 20 random cells12 Proof of principle: Automatedreconstruction of signaling pathway inhuman cells. Advantages: In-vivo (context specific signaling) Directed edges (causality) Detects direct and indirect influencesConnecting the dots…Correlative data from single cells can definesignaling mapsBayesian Networks by ExampleExample: Pedigree A node representsan individual’sgenotypeJoint distributionHomerBartMargeLisa Maggie ,...),,,,(MargeHomerMaggieLisaBartGGGGGPL,...),,|(,...),,,|(MargeHomerMaggieLisaMargeHomerMaggieLisaBartGGGGPGGGGGP=13Bayesian Networks by ExampleModeling assumption: Ancestors can effectdescendants' genotypeonly by passing geneticmaterials throughintermediategenerations,...),,,,(MargeHomerMaggieLisaBartGGGGGPHomerBartMargeLisa Maggie L,...),,|(,...),,,|(MargeHomerMaggieLisaMargeHomerMaggieLisaBa r tGGGGPGGGGGP=Bayesian Networks by ExampleModeling assumption: Ancestors can effectdescendants' genotypeonly by passinggenetic materialsthrough intermediategenerations,...),,,,(MargeHomerMaggieLisaBartGGGGGPHomerBartMargeLisa Maggie L,...),,|(),|(MargeHomerMaggieLisaMargeHomerBartGGGGPGGGP=14Bayesian Networks by ExampleModeling assumption: Ancestors can effectdescendants' genotypeonly by passing geneticmaterials throughintermediategenerations,...),,,,(MargeHomerMaggieLisaBartGGGGGPHomerBartMargeLisa Maggie L),|(),|(MargeHomerLisaMargeHomerBartGGGPGGGP=Markov Dependencies: Graph structure encodesthe Markov Assumption: A variable is independentof its non-descendantsgiven its parentsOften a natural assumptionfor causal processes if we believe that wecapture the relevant stateof each intermediate stageGrandpaHomerBart LisaMarge15Bayesian NetworksP(A,B,C,D,E) = P(A) P(B | A )P(C | A, B)P(D | A, B, C)P(E | A, B, C, D )P(A,B,C,D,E) = P(A) P(B | A )P(C | A, B)P(D | A, B, C)P(E | A, B, C, D )Bayesian NetworksABCDEXInd( C ; B | A)Ind( D ; A, B | C)Ind( E ; B,