AlphaGenome2 ablations: early encoder conditioning nearly doubles eQTL recall
anshulkundaje · x · 2026-11-12
adamgayoso shares AlphaGenome2 architecture ablations: conditioning only at the decoder or output heads barely beats an unconditioned multitask model like AG1 — sequence representations must be modulated early, in the encoder and transformer tower. This yields much better cross-tissue and cell-type specificity, carrying over to variant effect prediction: on fine-mapped single-cell eQTLs (TenK10K), AG2 nearly doubles recall over fine-tuned AG1 and Scooby at 90% precision (30% vs. 16%) and at 90% sign accuracy (43% vs. 23%).
More from Research
- MIT's Tegmark coins 'formal-washing': Signal protocol proof is far from complete — tegmark · 2026-11-12
- Signal formal verification hits 80%, Tegmark says Wired overstated completion — tegmark · 2026-11-12
- Category-theoretic framework asks: when does a physical system actually compute? — neurovium · 2026-11-12
- Radical Numerics brings genome language model Omnii-1 to the edge for biodefense — exnx · 2026-11-12
- Against 'verifiable vs. fuzzy': the real axis is how expensive tasks are to grade — dfrsrchtwts · 2026-11-12
- Toby Ord flags 'all-things-considered' AGI forecast table is actually conditional on no pacing — tobyordoxford · 2026-11-12