CSHL benchmark: AlphaGenome 2 leads only within ~100kb of TSS on distal enhancer effects
anshulkundaje · x · 2026-11-13
A new benchmark from CSHL researchers Alan Murphy and Peter K. Koo uses CRISPRi screens to evaluate how well sequence-to-function genome models predict distal enhancer effects, covering AlphaGenome, AlphaGenome 2, Borzoi, Cerberus, Enformer and NTv3.
Key findings:
- AlphaGenome 2 (released by Google DeepMind in Nov 2026) now leads, but the gain is incremental: it's better on enhancers within 100kb of the TSS, and level with the original AlphaGenome beyond that.
- AlphaGenome 2 conditions predictions on a chosen cell type/state and is served via API; Calico's Cerberus replaces Borzoi's attention/U-Net with bidirectional Mamba-2 blocks reading 786kb of sequence.
- The benchmark targets known weaknesses of this model class — predicting distal CRE effect magnitudes and personalised genome expression — the questions biologists actually care about.
More from Research
- Spaghetti Splatting real-time viewer renders a stunning 3D Gaussian splat of a bee — jonstephens85 · 2026-11-13
- Erik Hoel: conflating computational functionalism with behaviorism is misinformation — erikphoel · 2026-11-13
- ITIHAS: database shows Indian state GDP first estimates correlate only ~0.3 with final figures — paulnovosad · 2026-11-13
- Apple paper probes the theoretical limits of language model alignment — jiqizhixin · 2026-11-13
- Johns Hopkins paper shows hidden preferences transfer between VLMs via unrelated visual tasks — JohnsHopkins · 2026-11-13
- FLASH-ALOHA open-sources leader-arm-only teleop hardware, CAD, and code — philfung · 2026-11-13