PhenomeXcan
Summary¶
PhenomeXcan is a large, pre-computed gene-to-phenome association resource built by running S-PrediXcan and S-MultiXcan on 4,091 publicly available GWAS traits against 49 GTEx v8 tissue expression models, then refining the resulting 8.87-million-variant, 22,515-gene association matrix with a newly developed Bayesian colocalization method, fastENLOC (fast enrichment estimation aided colocalization analysis) [1]. Unlike variant-level PheWAS catalogs that assign genes to loci by genomic proximity, PhenomeXcan organizes associations by gene, aggregated across tissues, and queryable at phenomexcan.org.
Method¶
- Association: S-PrediXcan/S-MultiXcan applied across 49 GTEx v8 tissue expression models to 4,091 traits, testing 8.87 million variants' predicted mediating effect through 22,515 genes.
- Fine-mapping: Bayesian fine-mapping (DAP-1/fgwas via TORUS) of both eQTL and GWAS signals to identify candidate causal signal clusters.
- fastENLOC: a novel Bayesian hierarchical colocalization method (implemented in C++) that precomputes signal clusters from the fine-mapping step, making genome/phenome-scale colocalization computationally tractable — output is a locus regional colocalization probability (locus RCP) per gene-trait pair, with putative causal contributors defined as locus RCP > 0.1 (chosen conservatively to limit false negatives, given colocalization methods' general tendency toward false negatives).
Across the full genome/phenome matrix, 72,994 gene-trait associations were Bonferroni-significant (P < 5.49×10⁻¹⁰), of which 22,219 (30.5%) also had locus RCP > 0.1.
Validation¶
Three independent validations, each using PrediXcan significance and fastENLOC colocalization as joint predictors:
- PheWAS catalog (3,144 variants, 1,358 traits, gene assignment by proximity): of 2,202 comparable gene-trait pairs, 1,005 PhenomeXcan associations replicated at P < 0.01 (AUC = 0.62), with significant replicability over random classifiers (empirical P < 0.01).
- OMIM (107 traits mapped via Human Phenotype Ontology, 7,809 curated gene-trait pairs): AUC = 0.64; combining PrediXcan and fastENLOC improved precision over either alone, though absolute precision was limited by OMIM's bias toward rare, large-effect variants that transcriptome-mediated methods are not designed to capture.
- AMP T2D Knowledge Portal (causal/strong/moderate curated type 2 diabetes genes vs. seven UK Biobank T2D-related traits): AUC = 0.67.
Illustrative Findings¶
- For chronotype ("morning/evening person"), PhenomeXcan's top 15 genes recovered several established circadian genes (PER3, FBXL3, RASL10B) and one gene (VIP) independently supported by a 23andMe GWAS, alongside two previously unreported candidates (RP11-220I1.5, RASL10B).
- For TPO (thyroid peroxidase), the strongest associations recapitulated known thyroid phenotypes (hypothyroidism, levothyroxine treatment; locus RCP = 0.99 for both) and surfaced weaker, less-established associations with respiratory medication use, cognitive-task completion time, depression, and an antihypertensive — flagged as hypothesis-generating rather than established.
- The resource was further cross-referenced against ClinVar to identify shared candidate genes linking common complex traits to rare Mendelian diseases (e.g., red blood cell indices and Parkinson's disease).
Availability¶
Web application: phenomexcan.org (query by gene or trait; also includes a PhenomeXcan × ClinVar cross-reference table). Code: github.com/hakyimlab/phenomexcan. Full results archived on Zenodo.
See Also¶
- MetaXcan, sPrediXcan, sMultiXcan — the underlying association methods PhenomeXcan is built on.
- Statistical Colocalization — the broader methodology fastENLOC belongs to.
- RatXcan — uses PhenomeXcan's human gene-trait associations to test cross-species conservation of trait genetics in rats.