STRAND predicts gene perturbation responses by DNA coordinates, NeurIPS 2026 accepted
量子位 · wechat · 2026-11-22
A Harvard Medical School / Johns Hopkins / Merck team presents STRAND, accepted to NeurIPS 2026 with code released. It fixes a blind spot in single-cell perturbation models that take gene IDs as input: they can't distinguish different regulatory entry points of the same gene or reach noncoding regions (98.5% of the genome).
- Input design: interventions are specified as DNA sequence + coordinate (128bp units, 95% genome coverage), using frozen Flashzoi/Borzoi representations, a mask over the targeted region, and cross-attention with relative distance biases to build the condition.
- Generation: SCimilarity encodes control-cell states, Optimal Transport statistically pairs control and perturbed cell populations, and an I²SB diffusion bridge learns the perturbed response distribution, decoded into multi-gene expression changes.
- Case studies: virtual interventions at FLVCR1's alternative transcription start sites correctly rank the experimentally validated FLVCR1b entry first (0.683 vs 0.564); scanning near e-HDAC6 shows a peak within 400bp of the validated enhancer.
- Results: few-shot (5 perturbed cells per gene) average rank of 1.86 across seven metrics (runner-up 3.14); 32% relative gain in perturbation discrimination on K562; 58.7% direction agreement on held-out cell lines vs 50.0% (STACK) and 48.1% (STATE); ablations show the DNA module lifts Pearson correlation 16.6% and the generative module cuts Energy Distance 38.9%.
- Use case: researchers can scan candidate intervention sites in silico before running CRISPRi experiments.
More from Research
- Researchers find the first genetic code reading CUU/CUA as proline, in four yeast mitochondria — willdepue · 2026-11-22
- Blogger calls for an open-ended AI Cambrian explosion of mathematical diversity — burny_tech · 2026-11-22
- Open-source visloc-rs turns 360° camera street footage into a robot simulator via 3D Gaussian Splatting — rsasaki0109 · 2026-11-22
- Inside a 26B Looped Transformer: One Vector Swap Lifts 4th-Pass Accuracy from 34% to 74% — pandeyparul · 2026-11-22
- Sandberg: weak concept generation means AI proofs will be long 'elementary' paths — anderssandberg · 2026-11-22
- Will 'compressing vast proofs' become an art? Sandberg thinks it might — anderssandberg · 2026-11-22