Which TWAS Catalog Do You Recommend for Target Discovery?

General
1

I’m looking for community recommendations on TWAS resources for target discovery and prioritization.

There are now quite a few TWAS catalogs and atlases available (e.g., GTEx/PrediXcan, TWAS Hub, Open Targets Genetics, eQTL Catalogue, GTEx v8/v9-based resources, and various disease-specific TWAS databases), each differing in tissues, molecular models, sample size, ancestry, and statistical framework.

For those actively using TWAS in pharmaceutical target discovery:

  • Which TWAS catalog or atlas do you rely on most?

  • Which resource has the best coverage, quality, and ease of integration into target prioritization pipelines?

  • Are there any community “gold standards” emerging for large-scale analyses?

  • If you had to recommend a single resource (or a combination of resources) for systematic target discovery, what would it be, and why?

I’m particularly interested in resources that are well-maintained, regularly updated, and suitable for genome-wide, production-scale analyses.

Thanks in advance for sharing your experience and recommendations!

2

Not a daily TWAS user, so I’ll leave the “which one is best” to people who run them, but one thing from building target-prioritisation pipelines on top of these catalogs that I’d put ahead of the choice of resource:

Track what each catalog didn’t test, per gene, per tissue. A gene missing from a TWAS Hub or PhenomeXcan result set is almost always a gene with no expression model that cleared the prediction-r² threshold in that tissue — not a gene that was tested and came up null. If the pipeline treats both as “no association”, it silently penalises every gene with low cis-heritability, which is a lot of good targets. Same with ancestry: most GTEx-trained models are European-ancestry; a null in a non-European GWAS is often a transferability gap, not biology.

Practically, for each catalog I’d record three columns next to every gene–trait cell: model available (y/n, r²), tissues tested (count of N), and training ancestry. Then absences become “not in universe” versus “measured null”, and you can combine catalogs — e.g. GTEx v8 MASHR models via S-PrediXcan for breadth, eQTL Catalogue for harmonised per-study effects and colocalisation, Open Targets L2G as the gene-prioritisation layer rather than as a TWAS source — without the union quietly inheriting each one’s blind spots.

Beyond that: whatever you pick, pair it with colocalisation (coloc/SuSiE) before calling a gene, since TWAS hits share a locus with the causal gene more often than they are it; and check the resource’s last update date, because several of the precomputed atlases froze around 2020–2022.

Happy to share the per-gene coverage schema if useful.