Prediction Of Interactions Between HIV-1 And Human .

Prediction of Interactionsbetween HIV-1 and Human Proteinsby Information IntegrationOznur Tastan1, Yanjun Qi1,Ŧ, Jaime G. Carbonell1 andJudith Klein-Seetharaman1,2,*1 LanguageTechnologies Institute,School of Computer Science, Carnegie Mellon University2 Departmentof Structural Biology,School of Medicine, University of PittsburghŦ NECLaboratories America, Inc.*University of London, Royal HollowayThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Human Immunodeficiency Virus-1 (HIV-1)q Causative agent of AIDS Destructs the immune system Leads to opportunistic infections and malignanciesq Current antiviral therapy prolonged the patients’ survival rates Not accessible to everyone Cannot eradicate HIV from the body Drug resistance problemsq No vaccineGlobal Summary of AIDS epidemic, December 2007Number of people livingwith HIV in 2007TotalChildren under 15 years33 million2 millionAIDS related deathsin 2007TotalChildren under 15 years2.0 million270 000The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

HIV-1 Genome and Life CyclePeterlin and Trono Nature Rev. Immu.(2003) 3: 97-107The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

AimPredict novel direct physical interactionsbetween HIV-1 and human proteinsThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Prediction of Host Pathogen Interactionsq Dyer et al. Bioinformatics (2007) 23(13): i159-66 Human Plasmodium falciparum Co-occurrence of domain sequence signaturesq Davis et al., Protein Sci (2007) 16(12): 2585-96 Inter-PPI of human with 10 pathogens (does not include HIV) Comparative modelingqKonig et al. Cell (2008) 135(1): 49-60 Functional siRNA knockout screen filtered by multiple evidencesNo work to date to predict global interactome of directphysical interactions between HIV-1 and human proteinsThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Our ApproachThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Supervised LearningqHIV-1 human protein pair is described with a feature vectorand a class label :!( xi , y )y Î {'Interact','Not Interact'}Each feature summarizes a biological informationqGiven data learn a function that would map feature spaceinto one of the two classes:f : X YThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Random Forest ClassifierTraining DataN examplesM featuresQi et al. Proteins. (2006) 63: 490-500The Pittsburgh Center for HIV Protein InteractionsBreiman Machine Learning (2001) 5-32www.hivppi.pitt.edu

Random Forest ClassifierCreate bootstrap samplesfrom the training data. N examplesM featuresThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Random Forest ClassifierConstruct a decision treeUse Gini Gain for splitting the nodes. N examplesM featuresThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Random Forest ClassifierAt each node in choosing the split featurechoose only among m M features. N examplesM featuresThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Random Forest ClassifierCreate decision treefrom each bootstrap sampleThe Pittsburgh Center for HIV Protein Interactions. . N examplesM featureswww.hivppi.pitt.edu

Random Forest ClassifierN examplesM featuresThe Pittsburgh Center for HIV Protein Interactions. . Majorityvotewww.hivppi.pitt.edu

Interaction DataThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

HIV-1 Human Protein InteractionsqNIAID database of human HIV-1 protein interactionscurated from teractionsSanders-Beer et al. NAR (2008) doi: 10.1093/nar/gkn708The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

HIV-1 Human Protein InteractionsKeywords: “Nef binds hemopoietic cell kinase isoform p61HCK”qGroup 1: more likely direct interactionsacetylated by, acetylates, binds, cleaved by, cleaves,degraded by, dephosphorylates, interacts with,methylated by, myristoylated by, phosphorylated by,phosphorylates, ubiquitinated by1063 interactions, 721 human proteins, 17 HIV-1 proteinsqGroup 2: could be indirect interactionsactivated by, activates, antagonized by, antagonizes,associates with, causes accumulation of, co-localizeswith, competes with, cooperates with .etc1454 interactions, 914 human proteins, 16 HIV-1 proteinsHIV-1 proteinHuman proteinThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.eduwww.hivppi.pitt.edu

Training and Testing DataThe ‘interaction’ class:Group 1, the more likely direct interactions1063 interactions, 721 human proteins, 17 HIV-1 proteinsThe ‘non-interaction’ class:Select randomly from the pairs that are not reported in NIAIDdatabaseThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

FeaturesThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

35 Featuresq Differential gene expressionin HIV infected vs uninfectedcells (4)q HIV-1 protein type (17)q ELM-ligand feature (1)q Human protein expressionin HIV-1 susceptible tissues(1)q Human PPI interactomefeatures (8)q Similarity of the two proteins interms of (4) Cellular locationMolecular processMolecular functionSequenceThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

ELM-Ligand Featureq Functional interaction motifs obtained Eukaryotic Linear Motif (ELM)database[RKY]XXPXXPmotif involved in protein-protein interactionmediated by SH3 domains* http://elm.eu.orgThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Motif-Ligand Featureq Functional interaction motifs obtained Eukaryotic Linear Motifdatabase[RKY]XXPXXPmotif involved in protein-protein interactionmediated by SH3 domainsIs the motif conservedin HIV-1 sequences?Does the human protein containthe ligand domain or belongs to the ligandprotein class?human proteinSH3 domainf motif q, where 0 q 1* http://elm.eu.orgThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Making Use of the Human PPI Interactome:Mimicry of Human Protein Interaction Partners’NAP-22/CAP-23The N-terminals resembleand are both myristoylatedCalmodulinThe Pittsburgh Center for HIV Protein InteractionsNefwww.hivppi.pitt.edu

Making Use of the Human PPI Interactome:Mimicry of Human Protein’s Interaction Partnersf neigh (i, j ) maxkÎS j {k1 ,k2 }f pairwise (i, k )q Similarity of HIV-1 protein to human protein’s interaction partner SequencePost translational modificationCellular locationMolecular processMolecular functionThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Making Use of the Human PPI Interactome:Human Protein’s Topological Properties the Human PPI networkDegreeNumber of neighborsClustering coefficientThe extent the neighbors areconnected with each otherBetweenness CentralityThe fraction of shortest pathspass through the nodeThe Pittsburgh Center for HIV Protein Interactionskv2nvkv (kv - 1)s uw (v)å su , wÎVuwu , w¹ vwww.hivppi.pitt.edu

EvaluationThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Performance MeasuresqPrecision Recall Curve Precision Recall (Sensitivity): TP/(TP FP): TP/(TP FN)The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Performance MeasuresqPrecision Recall Curve Precision Recall (Sensitivity): TP/(TP FP): TP/(TP FN)qThe Mean Average Precision (MAP): Mean of the average precisions where each average precision iscalculated when recall increases.The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Performance MeasuresqPrecision Recall Curve Precision Recall (Sensitivity): TP/(TP FP): TP/(TP FN)qThe Mean Average Precision (MAP): Mean of the average precisions where each average precision iscalculated when recall increases.qArea Under the Receiver Operating Curve (AUC):ROC curveTP rateAUC Partial AUC scores :Area under the curveuntil reaching N false positivesFP rateThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Performance Evaluationq10 repeated 3-fold cross validationAvgStdMAP0.230.02The Pittsburgh Center for HIV Protein 0.170.02R3000.220.02www.hivppi.pitt.edu

Feature ImportanceGini importance: Normalized sum of improvement in the"Gini gain" due a given feature in the forestThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Feature ImportanceMajority of the human interactome featuresare highly informativeThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Feature ImportanceThe network topology features are highly rankedThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Network FeaturesqEpstein–Barr virus targets high degree human proteinsCalderwood et al., PNAS (2007) 104: 7606-11qPathogens tend to interact with host proteins with highdegrees and betweenness centralityNumber of human partnersDyer et. al. PLoS Pathog (2008) 4, e32The Pittsburgh Center for HIV Protein InteractionsHIV-1 human interactionsRandomly paired interactionsDegree,dwww.hivppi.pitt.edu

Feature ImportanceTop 6 featuresHow can we perform using only the top 6 features?The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

When only the Top Ranked Features UsedTop 6 Gini Features:1. Degree2. Betweenness centrality3. Neighbor process similarity4. Clustering coefficient5. Neighbor function similarity6. Neighbor location similarityThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

When only the Top Ranked Features UsedTop 6 Gini Features:1. Degree2. Betweenness centrality3. Neighbor process similarity4. Clustering coefficient5. Neighbor function similarity6. Neighbor location similarityPTF: Protein type featuresThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

When only the Top Ranked Features UsedTop 6 Gini Features:1. Degree2. Betweenness centrality3. Neighbor process similarity4. Clustering coefficient5. Neighbor function similarity6. Neighbor location similarityPTF: Protein type featuresThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

Predicted InteractionsThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

PredictionsqApply the model to all possible HIV-1, human protein pairsIncreasing NoveltyHigh RecallScore CutoffTotal PairsGroup 1Group 2NovelRecallPrecision 0.003372104023221000.510.20 0.50194210341417670.370.29 1.0014401023683490.260.36 1.501085894341570.180.41 2.0062253815690.130.47 2.502792438280.090.47Used in modelconstructionThe Pittsburgh Center for HIV Protein InteractionsPredictionsHigh Precisionwww.hivppi.pitt.edu

Functionally Interesting Interactionsq 304 cellular proteins detected in virion Ott Rev Med Bio (2008 17: 159-75)q 273 genes that had an effect in the Brass siRNA screen Brass et al, Science (2008) 319: 921-6q 295 genes that had an effect in the Konig siRNA screen Konig et al. Cell (2008) 1: 49-60q The interactors of the siRNA genesRecallPrecisionin 0.47246101481784Brass et al. siRNA screenGenesInteractors461064134415212299149025The Pittsburgh Center for HIV Protein InteractionsKonig et al. ivppi.pitt.edu

Tat interacts with Pin1www.cs.cmu.edu/ HIV/hivPPI.htmldetected in viriontatPin1Ubqln4detected in siRNAscreenPin1 interacts with and reducesexpression of APOBEC3G.Watashi JV (2008) 82: 9928-36The Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

ConclusionsqCollected data from multiple biological information sourcesand encoded as featuresqDeveloped a model to predict HIV-1,human proteininteraction networkqFeatures containing human proteome knowledge is highlyinformativeqSpecific protein interactions are being testedqPredictions availablewww.cs.cmu.edu/ HIV/hivPPI.htmlwww.hivppi.pitt.eduThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu

AcknowledgementsThanks to :-PSB organizersYanjun QiNEC Laboratories America, Inc -National Institutes of Health-Pittsburgh Center for HIV ProteinInteractionsJaime G. CarbonellCarnegie Mellon UniversityJudith Klein-SeetharamanCarnegie Mellon UniversityUniversity of PittsburghThe Pittsburgh Center for HIV Protein Interactionswww.hivppi.pitt.edu