Enformer, Borzoi, NTv3, Cerberus & AlphaGenome 2 benchmarked: distal enhancer effects remain unsolved
anshulkundaje · x · 2026-11-13
CSHL researchers Alan Murphy and Peter Koo updated their living, versioned Genomics×AI benchmark post, testing Enformer, Borzoi, NTv3, Cerberus and the new AlphaGenome 2 on K562 CRISPRi enhancer data, extending Karollus 2023.
Key points:
- Newer seq2func models keep improving: AlphaGenome 2 (Google DeepMind, Nov 2026) conditions predictions on a chosen cell type/state and is served via API; Cerberus (Calico, Borzoi's successor) replaces attention + U-Net with bidirectional Mamba-2 blocks and reads 786 kb of sequence; NTv3 competes by post-training on the same functional genomic tracks.
- Head-to-head: Cerberus ties Borzoi overall but is worse on proximal and possibly better on distal enhancers.
- Shared weakness: all models still systematically underpredict distal CRE effect magnitudes — distal perturbations remain an open problem.
More from Research
- FLASH-ALOHA open-sources leader-arm-only teleop hardware, CAD, and code — philfung · 2026-11-13
- InferenceBench: agents beat vLLM defaults but lose to simple hyperparameter search — maksym_andr · 2026-11-13
- Inkling joins Eval³ leaderboard that scores models on auditing evals for reward hacks — simonguozirui · 2026-11-13
- Google names 148 PhD Fellows from 38 countries with $13M in research funding — sivareddyg · 2026-11-13
- Incremental research is now riskier than going all in on big ideas — fkasummer · 2026-11-13
- MIT & Meta paper: ReHarness revises agent memory notes with later outcomes, beats baselines — rohanpaul_ai · 2026-11-13