Year: 2026
Type: preprint
Abstract: Sequence-to-function (S2F) models predict molecular phenotypes from DNA sequence and are increasingly applied to variant effect prediction (VEP), but their accuracy varies across variants, genes, and tissues. Current practice declares predictions as trustworthy by thresholding on predicted effect magnitude, a crude heuristic that discards low-effect variants where most GWAS signal is expected to act. We present gRely, a meta-model that estimates a score that an individual VEP is correct, from 1,161 features derived from the target variant, gene, tissue, and model. Trained on GTEx fine-mapped eQTLs and Borzoi, gRely reaches 97% sign concordance in its top 20% of scores on held-out variants, genes, and tissues. Critically, it remains discriminative among low-magnitude variants, which effect-size filtering discards; SHAP attribution shows that reliability in the low-effect regime is driven by chromatin context and model cross-replicate agreement rather than effect size. Furthermore, the Borzoi trained meta-model generalizes across other architectures, such as AlphaGenome and Enformer. gRely is the first framework to provide per-prediction confidence scores for S2F models and marks a step toward trustworthy deployment of S2F models in genomic research and clinical applications.
Institutions: University of British Columbia, Genentech
Language: English
Subfield: Genetics
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7162571144