Preprint: AlphaGenome pervasively underestimates causal variant effects, undermining fine-mapping
anshulkundaje · x · 2026-10-30
A bioRxiv preprint by Shiron Drusinsky and Katherine Pollard (Gladstone Institutes, UCSF) identifies an overlooked flaw in deep learning sequence-to-function (S2F) models for causal variant fine-mapping.
Key findings:
- S2F models classify putatively causal eQTL SNVs reasonably well, but dramatically underperform linear baselines at ranking individuals' gene expression from whole genome sequence.
- Systematic analysis with state-of-the-art AlphaGenome shows it pervasively underestimates the effects of most causal variants, letting nearby non-causal SNVs randomly drive predictions.
- As a result, AlphaGenome fails to fine-map eQTLs in most loci.
- Since sequence models aren't confounded by linkage disequilibrium, the authors attribute the underperformance to causal effect underestimation rather than LD.
More from Research
- Alice & Bob says RSA-2048 falls in 3.14 days with 23,000 cat qubits — jedisct1 · 2026-10-30
- FlexGlue's 0.6 px average keypoint shift matches SIFT's subpixel refinement threshold — ducha_aiki · 2026-10-30
- Thread Author Proposes New Apprenticeship: Judge First, Then Compare With AI — krishnan · 2026-10-30
- NBER RCT: AI Drafting Assistant Boosts Patent Lawyers' Output, Juniors Gain Most — krishnan · 2026-10-30
- Backprop-free pretraining hits 4,000x speedup in 10 days, now 1.6 seconds per run — SonglinYang4 · 2026-10-30
- TriFHE progress stalls as existing FHE schemes already use dense ternary secrets — fractalmoto · 2026-10-30