Data analytics approaches to enable EWAS

Transcription

1 Data analytics approaches to enable EWAS Chirag J Patel and Nam Pho Emory Exposome Workshop 06/16/16

2 Extensible & open-source analytics software library (XWAS R package) Freely available exposome data for your research (NHANES: 40,000 individuals and 1,000 variables) Computer environment to conduct EWASs (Docker container in RStudio) Materials for teaching and demonstration XWAS + + = goodness!

3 Please let us know if you are using the resources (or provide feedback)!

4 Real quick: What is the exposome? What is the phenome? exposome phenome internal external function diseases lead (serum) geography expression diabetes nutrients (serum) air pollution telomeres cancer infection (urine) income metabolome heart disease metabolome

5 Exposome associated with the phenome? and vice versa? exposome phenome Analytic tools and big data infrastructure required to associate exposome with phenome!

6 We can learn a thing or two from genomics investigation genome phenome e.g., GWAS

7 Big data approaches fueled discovery of genetic variants in disease (example: genome-wide association [GWAS]) 50 Locus established previously Locus identified by current study Locus not confirmed by current study 40 TCF7L2 Unconditional analysis log 10 (P) BCL11A THADA NOTCH2 ADAMTS9 PPAR IRS1 IGF2BP2 WFS1 ZBED3 CDKAL1 JAZF1 HHEX/IDE CDC123/CAMK1D CHCHD9 CDKN2A/2B SLC30A8 TP53INP1 KLF14 KCNQ1 (2 signals*: ) KCNJ11 CENTD2 MTNR1B HMGA2 TSPAN8/LGR5 HNF1A ZFAND6 PRC1 FTO HNF1B DUSP9 0 GWAS in Type 2 Diabetes Voight et al, Nature Genetics 2012 N=8K T2D, 39K Controls A search engine for robust, reproducible genotypephenotype associations

8 There are non-trivial data analytic challenges in searching for exposome-phenome associations! JAMA 2014 Pac Symp Biocomp 2015

9 Dense correlational web! what causes what? confounding bias? JAMA 2014 Pac Symp Biocomp 2015

10 Carotenoids Vitamins Minerals A B C Phytoestrogens D Cotinine E Hydrocarbons Volatile compounds Virus Bacteria Phenols Phthalates Polyfluorochemicals Heavy metals PCBs Dioxins Diakyl Furans Pesticides Chloroacetanilide Organochlorine Organophosphate Phenol Pyrethyroid Carotenoids Minerals Phytoestrogens Cotinine Hydrocarbons Vitamins A B C D E Volatile compounds Virus Bacteria Phenols Phthalates Polyfluorochemicals Heavy metals Colour key and histogram PCBs Count Value Dioxins Diakyl Furans Pesticides Chloroacetanilide Organochlorine Organophosphate Phenol Pyrethyroid IJE 2013

11 Interdependencies of the exposome: Correlation globes paint a complex view of exposure Pac Symp Biocomput JECH. 2015

12 Multiplicity: how to determine signal from noise? type 1 error (spurious findings) JAMA 2014

13 Suppose you are testing 1000 exposures in case-control study (disease vs. healthy) and there were no difference between the cases and controls how many findings would be significant at a p-value threshold of 0.05 (due to chance)?

14 Regime of multiple tests and signal to noise : Histogram of p-values in 2 scenarios: no difference and 5% different Frequency A Frequency B pvalues pvalues.ewas No difference (no true associations) 5% exposures different (5% true associations)

15 Estimating the deviation from null: QQplot: -log10(pvalues) in the null and EWAS distributions B 10% different log10(pvalue) expected log10(pvalue) A Bonferroni False discovery rate

16 The tension between type 1 and type 2 errors: Power and replication for robust associations! Discovery N=1K Replication N=1K Discovery sample sizes must be large to overcome multiple testing and mitigate winner s curse Replication sample size must be large to detect association

17 What will the exposome data structure look like?: a high-dimensioned 3D matrix of (1) exposure measurements on (2) individuals as a function of (3) time individual i individuals genome (static) mixtures of exposures nutrient value for individual i exposome factors pollutants diet metabolites xenobiotics drugs gut flora... in review

18 What will the exposome data structure look like?: a high-dimensioned 3D matrix of (1) exposure measurements on (2) individuals as a function of (3) time (A) exposome individuals genome (static) mixtures of exposures system genome pollutants diet metabolites xenobiotics drugs gut flora... time longitudinal in review

19 A schematic of a data-driven search for exposome-phenome associations (A) exposome (C) phenome (B) EWAS, PheWAS..etc. individuals genome (static) mixtures of exposures integrative EWAS GWAS, RVAS, pathway analysis..etc. mixtures of phenotypes pollutants diet metabolites xenobiotics drugs gut flora... time CVD BP cancer weight height T2D... time in review

20 Time for you to give it a try!

21 Extensible & open-source analytics software library (XWAS R Package) Freely available exposome data for your research (NHANES: 40,000 individuals and 1,000 variables) Computer environment to conduct EWASs (Docker container in RStudio) Materials for teaching and demonstration XWAS + + = goodness!

22 Fully merged dataset: National Health and Nutrition Examination Survey since the 1960s now biannual: 1999 onwards 10,000 participants per survey >250 exposures (serum + urine) >85 quantitative clinical traits (e.g., serum glucose, lipids, body mass index) Death index linkage (cause of death) Ready to analyze! N=41K with >1000 variables (let us know; we can give you a DOI) in review

23 13 EWAS-related manuscripts preterm birth type 2 diabetes type 2 diabetes genetics lipids blood pressure income mortality telomere length methodology (5)

24 Associations in Telomere Length: Can you identify the associations in this graph? 4?? 3 log10(pvalue) 2 FDR<5% shorter telomeres effect size longer telomeres median N=3000; N range: IJE, 2016

25 Associations in Telomere Length: Can you identify the associations in this graph? Nam will show you how!

26 Please let us know if you are using the resources (or provide feedback)!

27 Acknowledgements RagGroup Nam Pho Chirag Lakhani Adam Brown Danielle Rasooly Arjun Manrai Grace Mahoney Matthew Roy Harvard DBMI Isaac Kohane Susanne Churchill Stan Shaw Jenn Grandfield Michal Preminger Stanford John PA Ioannidis Gary Miller Kristine Dennis NIH Common Fund Big Data to Knowledge Chirag J Patel