Informatics in Computational Medicine

Transcription

1 Informatics in Computational Medicine Inderjit S. Dhillon Director, Center for Big Data Analytics UT Austin Computer Science, Mathematics & ICES ICES Computational Medicine Day May 13, 2014

2 Introduction Computational Medicine Quantitative approach to understanding, detecting and treating diseases Need to understand and control biological processes involved in human diseases via systemic measurements and computational analyses Calzolari D, Bruschi S, Coquin L, Schofield J, Feala JD, et al. (2008) Search Algorithms as a Framework for the Optimization of Drug Combinations. PLoS Comput Biol 4(12): e

3 Medical Genetics Analysis begins at the level of genes study of human genetics & its application to medical care Eventual goals: Gene therapy for cure personalized medicine Winslow, Raimond L., et al. "Computational medicine: translating models to clinical care." Science Translational Medicine (2012) 3

4 Gene Networks Multiple genes collectively influence the likelihood of developing many common and comple diseases. Important to understand how genes interact with each other (coepression gene networks) Xiao, X. et al., Multi-tissue analysis of co-epression networks by Higher-Order generalized singular value decomposition identifies functionally coherent transcriptional modules. PLoS Genetics 10 (1), e

5 Predicting gene-disease links AVPR2? AQP1 Diabetes insipidus? AQP2 AVP MYBL2 Goal: Discover human gene-disease associations Biologists prefer a short list of potentially relevant genes for further studies 5

6 Gene Network AVPR2 APQ6 AQP1 APQ5 AQP2 GK MIP AVP MYBL2 Functional interactions between genes (e.g. HumanNet) 6

7 Gene-Phenotype Networks Abnormal kidney physiology AVPR2 Decreased urine osmolality AQP1 AQP2 MYBL2 Response to salt stress Orthologous phenotypes in other model species can shed light on gene functions and in turn disease-causing genes 7

8 The Prediction Problem?? 8

9 Other data sources: Genomic data Genes ( probes ) Conditions (e.g. diseased vs normal tissue) Abundant epression data available (sources such as BioGPS) Co-epression reveals gene function modules ( Eigengenes ) Langfelder, P., & Horvath, S. (2007). Eigengene networks for studying the relationships between co-epression modules. BMC systems biology, 1(1), 54. 9

10 Other data sources: Tet data Descriptions of diagnosis, clinical features and management in tet articles provide information on diseases Represent diseases by document term frequencies Tet data on diseases (e.g. OMIM web pages) endometrial Disease 1 anti-estrogen Disease 2 heterozygotes 10

11 Other data sources: Co-morbidity Similarity network between diseases Prune Belly Syndrome Neurofibromatosis 0.31 Achondroplasia Similarities can be computed from data available on diseases (such as tet, symptoms, drug responses, etc) 11

12 Problem Formulation Genes Diseases??????? We want to predict missing associations denoted by? 12

13 The Netfli Problem 13

14 Inductive Matri Completion Orthologous phenotypes APQ6 Microarray PCA AQP1 AQP5 Genes Diseases????.? Gene network AQP2 GK MIP? 14

15 Inductive Matri Completion Microarray OMIM pages tf-idf 0.12 NF2 EGBRS Disease similarities ACH Orthologous phenotypes PCA Diseases? Genes???. APQ6 AQP1 AQP5? Gene network AQP2 GK MIP? 15

16 Inductive Matri Completion Microarray OMIM pages tf-idf 0.12 NF2 EGBRS Disease similarities ACH Orthologous phenotypes PCA????. APQ6 AQP1 AQP5? Gene network AQP2 GK MIP Genes? 16

17 Inductive Matri Completion Microarray OMIM pages tf-idf 0.12 NF2 EGBRS Disease similarities ACH Orthologous phenotypes PCA????.? Gene network GK MIP Genes? Natarajan N, Dhillon IS. Inductive Matri Completion for Predicting Gene-Disease Associations. To appear in Bioinformatics,

18 Results on OMIM data P(hidden gene among genes looked at) Inductive Matri Completion CATAPULT Biased Matri Completion ProDiGe [Mordelet et al. 2011] Katz LEML [Yu et al. 2014] Matri Completion (baseline) Number of genes looked at 1. Online Mendelian Inheritance in Man: 2. F. Mordelet and J.-P. Vert. ProDiGe: PRioritization Of Disease Genes with multitask machine learning from positive and unlabeled eamples. BMC Bioinformatics Yu, Hsiang-Fu, Prateek Jain, and Inderjit S. Dhillon: Large-scale Multi-label Learning with Missing Labels. ICML

19 Results on OMIM data Inductive Matri Completion CATAPULT Biased Matri Completion P(hidden gene among genes looked at) 0.1 ProDiGe [Mordelet et al. 2011] Katz LEML [Yu et al. 2014] P(hidden gene among genes looked at) Number of genes looked at Number of genes looked at New genes New diseases 19

20 Positive-Unlabeled Learning The prediction problem gives rise to a novel machine learning problem called PU learning - learning in the absence of negative eamples Methods such as biased SVM and biased matri completion shown to perform well empirically Can analyze theoretically using random noise models For biased matri completion, we can show Theorem: Let be the noise rate. With probability at least 1 - δ, solution X to the biased matri completion is close to the true (n n) matri Y: 1 X n 2 (X ij Y ij ) 2 1 = O (1 )n ij 20

21 Conclusions Informatics for Computational Medicine is crucial Need novel statistical techniques Analysis of etremely high-dimensional data from high-throughput studies Learning from diverse information sources Interpretable computational models Can obtain better gene-disease associations than state-of-the-art 21