Illuminating the Druggable Genome with Knowledge Engineering and Machine Learning
1 like229 views
This document discusses the Illuminating the Druggable Genome initiative, highlighting the need for research on understudied proteins ('tdark') and the application of machine learning to discover possible drug targets. It presents the metapath-ML approach, which utilizes diverse data sources to prioritize protein-disease associations through machine learning models. The work aims to bridge knowledge gaps and improve therapeutic development amid challenges posed by limited data on rare diseases.
Illuminating the Druggable Genome with Knowledge Engineering and Machine Learning
- 1. Giovanni Bocci, Cristian Bologa, Daniel Byrd, Jayme Holmes, Stephen Mathias, Oleg Ursu, Anna Waller, Jeremy Yang & Tudor Oprea 03/15/2019 INBRE-NMBIST Symposium Santa Fe, NM Funding: NIH U24 CA224370 & NIH U24 TR002278 ILLUMINATING THE DRUGGABLE GENOME WITH KNOWLEDGE ENGINEERING AND MACHINE LEARNING datascience.unm.edupharos.nih.gov/idg/ druggablegenome.net
- 2. 75% of protein research still focused on 10% genes known before human genome was mapped AM Edwards et al, Nature, 2011 This prompted NIH to start the Illuminating the Druggable Genome Initiative (U54, Common Fund) HGP
- 3. 3 "If I have seen further it is by standing on the shoulders of Giants." - Isaac Newton, ~1675 Organization of this talk: 1. Shoulders development (knowledge engineering). 2. Seeing further efforts (machine learning).
- 4. 4 IDG: AN INTERNATIONAL CONSORTIUM
- 7. ML READY: PHAROS 11/13/18 revisionhttps://pharos.nih.gov/idg/targets/GRIN2A 7
- 8. ML READY: DRUGCENTRAL 10/16/18 revision http://drugcentral.org/drugcard/1679 8
- 10. COMPONENTS OF IDG https://druggablegenome.net/ DRGC RDOC IT KMC RFA-RM-16-026 (DRGC) GPCR U24 DK116195: Bryan Roth, M.D., Ph.D. (UNC) Brian Shoichet, Ph.D. (UCSF) Ion Channel U24 DK116214: Lily Jan, Ph.D. (UCSF) Michael T. McManus, Ph.D. (UCSF) Kinase U24 DK116204: Gary L. Johnson, Ph.D. (UNC) RFA-RM-16-025 (RDOC) U24 TR002278: Stephan C. Schürer, Ph.D. (UMiami) Dusica Vidovic, Ph.D. (UMiami) Tudor Oprea, M.D., Ph.D. (UNM) Larry A. Sklar, Ph.D. (UNM) RFA-RM-16-024 (KMC) U24 CA224260: Avi Ma’ayan, Ph.D. (ISMMS) U24 CA224370: Tudor Oprea, M.D., Ph.D. (UNM) RFA-RM-18-011 (CEIT) Awards starting date March 2019 Further information Email: idg.rdoc@gmail.com Follow: @DruggableGenome URLs: https://druggablegenome.net / https://commonfund.nih.gov/i dg/ IDG Knowledge User-Interface Email: pharos@mail.nih.gov Follow: @IDG_Pharos URL: https://pharos.nih.gov/ 10
- 11. TARGET DEVELOPMENT LEVEL (TDL) ▪ Most protein classification schemes are based on structural and functional criteria. ▪ For therapeutic development, it is useful to understand how much and what types of data are available for a given protein, thereby highlighting well-studied and understudied targets. ▪ Tclin: Proteins annotated as drug targets ▪ Tchem: Proteins for which potent small molecules are known ▪ Tbio: Proteins for which biology is better understood ▪ Tdark: These proteins lack antibodies, publications or Gene RIFs 3/23/18 revision T. Oprea et al., Nature Rev. Drug Discov. 2018, https://www.nature.com/articles/nrd.2018.14 11
- 12. TDL LEVELS: Tclin and Tchem ▪ Tclin proteins are associated with drug Mechanism of Action (MoA) – NRDD 2017 ▪ Tchem proteins have bioactivitis in ChEMBL and DrugCentral, + human curation for some targets ▪ Kinases: <= 30nM ▪ GPCRs: <= 100nM ▪ Nuclear Receptors: <= 100nM ▪ Ion Channels: <= 10μM ▪ Non-IDG Family Targets: <= 1μM 10/19/16 revision Bioactivities of approved drugs (by Target class) ChEMBL: database of bioactive chemicals https://www.ebi.ac.uk/chembl/ DrugCentral: online drug compendium http://drugcentral.org/ R. Santos et al., Nature Rev. Drug Discov. 2017, https://www.nature.com/articles/nrd.2016.230 12
- 13. TDL LEVELS Tbio and Tdark ▪ Tbio proteins lack small molecule annotation cf. Tchem criteria, and satisfy one of these criteria: ▪ protein is above the cutoff criteria for Tdark ▪ protein is annotated with a GO Molecular Function or Biological Process leaf term(s) with an Experimental Evidence code ▪ protein has confirmed OMIM phenotype(s) ▪ Tdark (“ignorome”) have little information available, and satisfy these criteria: ▪ PubMed text-mining score from Jensen Lab < 5 ▪ <= 3 Gene RIFs ▪ <= 50 Antibodies available according to antibodypedia.com 13
- 14. TDL: EXTERNAL VALIDATION Tdark parameters differ from the other TDLs across the 4 external metrics cf. Kruskal-Wallis post-hoc pairwise Dunn tests 2/23/18 revision T. Oprea et al., Nature Rev. Drug Discov. 2018, https://www.nature.com/articles/nrd.2018.14 14
- 15. WHY FUND TDARK RESEARCH? 2/23/18 revision T. Oprea et al., Nature Rev. Drug Discov. 2018, https://www.nature.com/articles/nrd.2018.14 Typically, it takes 15-20 years for a Tdark protein to become druggable 15
- 16. IMPC BOLDLY GOES WHERE NO ONE HAS GONE BEFORE 95% of eligible IDG genes (339/356) have plans, attempts, or models 384 genes were prioritized by IDG KMC (2014-2016) 17 28 17 1 63 24 50 79 168306 Tbio genes 90 Tdark genes 42 Tchem genes 11/29/17 revision Slide from Steve Murray, Jackson Lab 16
- 17. TAKE HOME MESSAGE: THERE IS A KNOWLEDGE DEFICIT 3/12/18 revision ~35% of the proteins remain poorly described (Tdark) ~11% of the Proteome (Tclin & Tchem) are currently targeted by small molecule probes Choosing to work on dark genes is a high-risk endeavor (Funders are less likely to award grants for Tdark)
- 18. CHALLENGE: RANKING & SCORING PROTEIN-DISEASE ASSOCIATIONS https://pharos-beta.ncats.io/targets/GRIN2A The IDG KMC tracks more ~10 information channels for protein-disease associations, accessible via the Pharos portal. Our challenge is to harmonize disease concepts, and to enable computational use: e.g., GRIN2A with GRIN1 form the Glutamate NMDA receptor, MoA drug target for memantine (Alzheimer’s). The challenge for ML & AI: How to prioritize targets? i.e., which protein-disease associations are clinically actionable? 10/07/18 revision 18
- 19. WHAT DO WE KNOW ABOUT DISEASES? ▪ There are between 9,000 and 25,000 disease concepts ▪ Pharos/TCRD tracks ~11,000 disease via Disease Ontology, and ~10500 rare disease via eRAM, OrphaNet and the Monarch Initiative MONDO system 19
- 20. PROTEIN KNOWLEDGE GRAPHS ▪ IDG KMC2 seeks knowledge gaps across the five branches of the “knowledge tree”: ▪ Genotype; Phenotype; Interactions & Pathways; Structure & Function; and Expression, respectively. ▪ We can use biological systems network modeling to infer novel relationships based on available evidence, and infer new “function” and “role in disease” data based on other layers of evidence ▪ Primary focus on Tdark & Tbio O. Ursu, T Oprea et al., IDG2 KMC 2/01/18 revision 20
- 21. THE METAPATH-ML APPROACH▪ A metapath is a sequence of relations defined between different object types. ▪ Our metapaths encode type-specific network topology between the source node (Protein) and the destination node (Disease/Phenotype). ▪ This approach enables the transformation of assertions/evidence chains of heterogeneous biological data types into a ML ready format. SOME REFS: G. Fu et al., BMC Bioinformatics 2016. D Himmelstein & S Baranzini, PLOS Comp Bio, 2015. Similar assertions or evidence form metapaths (white). Instances of metapath (paths) are used to determine the strength of the evidence linking a gene to disease/phenotype/function. 21
- 22. 22 SOME EARLY ACKNOWLEDGMENTS ... Abstract: We hear a lot about machine learning and its role in health care, but these methods require large amounts of training data. Using these and other related method to study rare diseases poses substantial challenges: how can we get tens of thousands of training examples when there are tens or hundreds of people with a disease? (Abstract from this conference) Scarce training data our problem too. But with genes instead of people. Our MetapathML method similar to Himmelstein-Baranzini. (Daniel is now post-doc in Greene Lab.)
- 23. METAPATH-ML DATA SOURCES O. Ursu et al., manuscript in preparation Data source Data type Data points CCLE Gene expression 19,006,134 GTEx Gene expression 2,612,227 Protein Atlas Gene & Protein expression 949,199 Reactome Biological pathways 303,681 KEGG Biological pathways 27,683 StringDB Protein-Protein interactions 5,080,023 Gene ontology Biological pathways & Gene function 434,317 InterPro Protein structure and function 467,163 ClinVar Human Gene - Disease/Phenotype associations 881,357 GWAS Gene - Disease/Phenotype associations 54,360 OMIM Human Gene - Disease/Phenotype associations 25,557 UniProt Disease Human Gene - Disease/Phenotype associations 5,365 JensenLab DISEASE Gene - Disease associations from text mining 44,829 NCBI Homology Homology mapping of human/mouse/rat genes 70,922 IMPC Mouse Gene - Phenotype associations 2,153,999 RGD Rat Gene - Phenotype associations 117,606 LINCS Drug induced gene signatures 230,111,315 We developed automated methods for data collection (TCRD), visualization (Pharos) and data aggregation. These aggregated datasets were used to build machine learning models for 20+ disease and 73 mouse phenotype. Each knowledge graph contains ~22,000 metapaths and 284 million path instances. 10/07/18 revision 23
- 24. METAPATH-ML WORKFLOW ▪ A meta-path encodes type-specific network topology between the source node (e.g., Protein target) and the destination node (e.g., Disease or Function) ▪ Target –– (member of) → PPI Network ← (member of) –– Protein –– (associated with) → Disease ▪ Target –– (expressed in) → Tissue ← (localized in) –– Disease O. Ursu, T Oprea et al., IDG2 KMC 2/01/18 revision 24
- 25. METAPATH-ML @ UNM one protein-disease association at the time O. Ursu, T Oprea et al., IDG2 KMC 2/01/18 revision Genes associated with a disease/phenotype are positive examples, whereas genes lacking the same association are negative examples. The Metapath approach transforms assertions/evidence chains into classification problems that can be solved using suitably designed machine learning algorithms. 25
- 26. Use of XGBoost (XGBoost = eXtreme Gradient Boosting) ● https://xgboost.ai/ ● GitHub ● Documentation ● R package ● Exceptional interpretability MetapathML employs XGBoost via the R package API. The inputs to XGBoost are datasets specific to each disease or phenotype. For each disease/phenotype some known associated genes correspond with the positive Y labels in the dataset. XGBoost parameters are optimized via grid search, i.e. iterative testing over discrete parameter value combinations.
- 27. 27
- 29. ALZHEIMER’S DISEASE (AD) METAPATH ML MODEL ▪Build data matrix from “Alzheimer’s disease” in TCRD subset ▪ protein knowledge graph along metapaths: ▪ Protein – Protein Interactions ▪ Pathways ▪ GO terms ▪ Gene expression ▪ Etc. ▪ Training set: 53 genes associated with Alzheimer’s disease (positives); 3,952 genes associated with other pathologies from OMIM were assumed to be negative ▪ Test set: 23 genes associated with Alzheimer's (positives) and 200 genes not associated with Alzheimer's (negatives) ← from Text Mining ▪ “Complete forest” binary classifier using XGBoost & 5-fold cross-validation. 2/14/18 revisionML work by Oleg Ursu Predicted Actual Pos Neg Pos 20 3 Neg 41 159 29
- 30. AD XGBOOST CLASSIFIER: VARIABLE IMPORTANCE PLOT ▪ The top most important features are interactions with proteins mediating inflammatory processes (JAK2/Tclin, IL10 & IL2 / Tchem), response to oxidative stress (GSTP1/Tchem), nervous system development (BDNF/Tbio) and glycolysis (GAPDH/Tchem). ▪ LINCS drug-induced gene expression perturbations are the largest category of features for these predictions. ▪ Brain cortex expression is a necessary requirement. ▪ One Reactome pathway (AU-rich mRNA elements binding proteins) is also important. ▪ Weighted approached showed better performance in the test set for Alzheimer's Disease, Schizophrenia, and Dilated Cardiomyopathy. 4/23/18 revisionML work by Oleg Ursu 30
- 31. EXPERIMENTAL VALIDATION: AD ▪ SHSY5Ys pTau siRNA test ▪ Measured pTau levels after knock-down of gene expression • Human iPSNs qPCR ▪ Measuring endogenous gene expression levels, AD vs Ctrl ▪ Western blot or ICC to characterize AD phenotype versus control • Human Tissue qPCR ▪ Measuring endogenous gene expression levels, AD vs Ctrl ▪ Western blot or ICC to characterize AD phenotype versus control 11/14/18 revision AD validation work by Jessica Binder & Kiran Bhaskar (UNM), funded by U24CA224370-S2 supplement 31
- 32. 2/14/19 revisionAD validation work by Jessica Binder & Kiran Bhaskar (UNM), funded by U24CA224370-S2 supplement ▪Validation on the 20 predicted genes: AKNA, BC02, CCNY, CRTAM, FAM92B, FOXP4, FRRS1, GRIN2C, 1L17REL, LILRA3, LM04, NDRG2, PIBF1, RAB40A, SCGB3A1, SLC44A2, SPOP, STARD3, TMEFF2, TXNDC12 ▪The most obvious effects based on the combined Cellomics & qPCR of iPSNs & autopsy brains suggests that AKNA, LILRA3, NDRG2 and TXNDC12 significantly increased pTau (as tracked by two different antibodies for T180 , S202 and S205 ) ▪For now, it appears that machine learning models may have identified between 4 and 7 new genes that have previously not been associated with Alzheimer’s Disease 32 EXPERIMENTAL VALIDATION: AD
- 33. 33 EXPERIMENTAL VALIDATION: MORE DISEASES AND COLLABORATORS Disease Experimental Collaboration Prostate cancer Work by Art Cherkasov, Kriti Singh & Mike Hsing (UBC, Vancouver). Of the top 50 ML predicted genes, 19 commonly upregulated in YZ Wang Transdifferentiation PDX model and Beltran dataset 2016. Ovarian cancer Spheroid tumor & patient-derived xenograft (PDX) work by Mara Steinkamp (UNM). From the top ML predicted 63 genes, 12 genes show significant changes in cancer cells. NEXT STEPS: ● In vivo experiments. ● More diseases and phenotypes.
- 34. ML LEARNINGS IN TARGET AND DRUG DISCOVERY 1. Model quality is limited by data quality. Good data → good models. 2. ML can identify hidden patterns in big data. For example, the central node(s) in PPI network(s) that are a playing critical role in disease pathology. 3. Deep learning not so applicable to our task (better for tall datasets, well defined good solutions, less need for interpretability). 4. XGBoost (decision tree algorithm) excels in performance & interpretability. 5. Shows real promise in Target Repurposing. 34
- 35. 35 IN CLOSING... 35 ● IDG platform for knowledge discovery about the "dark genome." ● ML provides new insights by integrating multi-omics knowledge graphs. ● Hard questions should be directed to Tudor Oprea!