GEMS framework links non-coding GWAS variants to specific cell types using sequence-to-function modelling

Researchers have posted a preprint to bioRxiv describing GEMS (GWAS Effect Mechanisms via Sequence-to-Function), a computational framework aimed at one of the central outstanding problems in human genetics: determining through which cell types and cellular states the thousands of non-coding variants identified by genome-wide association studies (GWAS) exert their effects.

The majority of GWAS signals fall in non-coding regions of the genome, where the relevant regulatory context — including which cell types have open chromatin at the locus in question — is critical to interpreting biological meaning. GEMS addresses this by training supervised classifiers on multimodal variant effect predictions derived from the sequence-to-function model Borzoi, which integrates ATAC-seq (chromatin accessibility) and RNA-seq (gene expression) predictions. SHAP-based attribution — a method from machine learning used to assign feature importance — is then applied to link individual variants to their most relevant cell types.

The approach offers a more interpretable alternative to existing methods, which the authors argue often operate in an opaque manner that limits biological inference. GEMS is positioned as a tool for researchers seeking to generate mechanistic hypotheses about GWAS loci, particularly for complex common diseases.

This is a preprint and has not undergone peer review. Results and claims should be regarded as preliminary pending independent evaluation.