How to build a genetic dose-response portal in 7 years

[I am an employee of Celgene. All opinions expressed here are my own.]

A meeting was recently convened to discuss a roadmap for understanding the genetics of common diseases (search Twitter for #cdcoxf18). I presented my vision of a genetics dose-response portal (slides here; link to related 2018 ASHG talk here). The organizers (@RachelGLiao, @markmccarthyoxf, @ceclindgren, Rory Collins [Oxford], Judy Cho [New York], @NancyGenetics, @dalygene, @eric_lander) asked participants to share their vision. I thought I would blog about my mine.

You’ll notice my vision is ambitious. Nonetheless, I believe these objectives are feasible to accomplish within a 3-year (Phase 1) and 7-year (Phase 2) time frame. Phase 1 would start immediately and would guide projects for Phase 2. In reality, many aspects of Phase 1 are already underway today (e.g., GWAS catalogue at EBI; Global Alliance for Genomics and Health [GA4GH] data sharing methods). Phase 2 consists of two parts: federation of global biobanks and experimental validation of variants, genes and pathways. Some components of Phase 2 could start today (e.g., exome sequencing in >100,000 cases selected from existing case-control cohorts and biobanks; human knockout project). As with Phase 1, many components of Phase 2 are already underway (federation of existing biobanks [e.g., UK Biobank with FinnGen], technologies for high-throughput CRISPR mutagenesis and single-cell eQTL analysis).

You’ll also notice my vision is biased towards target-driven drug discovery. However, the resources generated would enable much more than just a dose-response portal for drug discovery (e.g., genetically-validated cellular readouts for phenotypic screens; catalogue of nodes not targeted by existing therapies; hypothesis-generation for novel therapeutic modalities). Perhaps more importantly, the resources would enable much, much more than just drug discovery.

• Phase 1 – All x All Association study (AAAS)

1. organize summary statistics for all existing GWAS data (including disease traits and quantitative traits such as mRNA expression, protein levels, peripheral blood counts)
2. fine-map signals of association for each trait
3. nominate “causal” variant(s) and gene(s) for each trait (using epigenetic data, human cell atlas, coding variants, etc.)
4. co-localize all signals of association across all traits (i.e., all x all association study)
5. for disease traits co-localized with quantitative traits (e.g., eQTLs, pQTLs), estimate magnitude of clinical effect size relative to quantitative trait effect size
6. for all disease traits, nominate pathogenic cell types using epigenetic, gene expression data, etc.
7. for all traits, generate polygenic scores (PS) and perform all x all PS association study
8. integrate all confirmed rare mutations associated with Mendelian phenotypes
9. for up to 1000 “solved genes” with more than one independent associated allele (common or rare, coding or non-coding), estimate genotype-phenotype dose-response curves
10. perform Mendelian randomization for disease-associated variants from 1000 “solved genes”
11. create searchable database, including visualization tools, for all results, using platforms such as the Open Targets genetics portal
12. implement Global Alliance for Genomics and Health (GA4GH) data sharing methods and create data sharing infrastructure for all future GWAS studies
13. create modular, open-source, automated pipeline for all components above to incentive future investigators to deposit data
14. nominate studies for Phase 2 (e.g., single-cell eQTL studies in pathogenic cell types in >500 individuals, exome sequencing in >100,000 cases from up to 10 selected diseases and/or quantitative traits, high-throughput CRISPR perturbations in relevant cellular systems, human knock-out project)

Rationale for Phase 1: summary statistics are only available for ~30% of GWAS data; poor incentive structure for investigators to make summary statistics and individual level genotype data available; no standardize data sharing format; most large-scale omics data resides in silos; as a consequence, current approach for data integration, interpretation and visualization is bespoke.

Phase 2a – Biobank federation

1. federate global biobank data on up to 50 million individuals from diverse ancestries
2. apply PRS from Phase 1 across biobanks to probe phenotype definitions and trans-ethnic heritability
3. perform Phenome-wide Association Study (PheWAS) for all trait-associated alleles from Phase 1
4. for associated alleles within “solved genes”, perform PheWAS to quantitatively model pleiotropic consequences for dose-response curves
5. create searchable database, including visualization tools for dose-response portal, for all results
6. implement GA4GH data sharing methods and create data sharing infrastructure for all future biobank studies
7. create modular, open-source, automated pipeline for all components above to incentive future biobanks to federate data

Rationale: As with GWAS data, current biobanks are silo’d, which limits ability to conduct cross-biobank analyses; most clinical traits are not represented in existing cohort-based genetic studies; no standardize data sharing format.

Phase 2b – Variant to function mapping

1. generate single-cell eQTL data in pathogenic cell types in >500 individuals; compare magnitude of eQTLs across cell types; incorporate data into dose-response portal
2. generate exome sequencing and perform rare-variant association tests (RVAS) in >100,000 cases (selected from existing case-control cohorts and federated biobanks) in up to 10 selected diseases and/or quantitative traits; define genetic architecture using common variant association studies (CVAS) and RVAS; expand list of “solved genes” based on RVAS; perform PheWAS for disease-associated rare variants.
3. perform high-throughput CRISPR perturbations in relevant cellular systems to experimentally validate causal variants from “solved genes”; reverse-engineer network wiring and define critical nodes within regulatory pathways; estimate magnitude of biological effect of causal variants; validate cellular readouts for future phenotypic screen
4. conduct a “human knock-out project” via whole genome sequencing in >20,000 individuals from consanguineous populations and annotation all putative loss-of-function mutations
5. utilize all data above to refine dose-response portal
6. nominate genetically-validated cellular systems and perform phenotypic screen with all approved drugs; map all “solved genes” to regulatory nodes; estimate fraction of solved genes that are perturbed by approved drugs; catalogue new therapeutic modalities that would be required to target novel nodes.

Rationale: experimental validation is required to confirm pathogenic cell types, and to estimate quantitatively magnitude of biological effect of causal variants; need to understand complete genetic architecture (common to rare variants) of a selected set of diseases; need to understand phenotypic consequences of human knockouts; there is no map of genetic nodes and approved therapies.

So what are we waiting for? Let’s get started!