gRely: Genentech team adds reliability scores to genome sequence-to-expression variant predictions
anshulkundaje · x · 2026-09-09
A Genentech computational biology team has released gRely, a bioRxiv preprint introducing a meta-modeling framework for the reliability of sequence-to-function (S2F) models in variant effect prediction (VEP).
- Problem: S2F models predict molecular phenotypes from DNA sequence, but their accuracy varies widely across variants, genes, and tissues. Current practice relies on crude magnitude thresholding, discarding most variants where models could still provide signal.
- Method: gRely estimates the probability that a given Borzoi VEP prediction correctly recovers eQTL direction, producing per-prediction confidence scores.
- Authors include Gokcen Eraslan from Genentech's AI Biology group; commenters note such meta-modeling tools can attach confidence scores to individual VEP calls.
More from Research
- A Curated Paper Collection on Harness Engineering from YC Paper Club — dair_ai · 2026-09-09
- Split-LLM Training Privacy Failure: Zero-Gradient Patterns Leak Real Training Rows — Setloop · 2026-09-09
- SlateDB WAL protocol gets an open-source TLA+ spec — blaizedsouza · 2026-09-09
- Harvey post-trains RLM agents for M&A diligence, lifting benchmark pass rate from 23% to 62% — lateinteraction · 2026-09-09
- Every number corrected for LLM judge bias—even the effect size — IanArawjo · 2026-09-09
- Terence Tao posts insightful thread reacting to AI's Navier-Stokes claim — fchollet · 2026-09-09