The abstract is read at the publisher; the summary is JClub's.
GENECraft, a function- and prior-constrained discrete dynamical system, achieves macro-averaged full-joint success rates of 62.70% (promoters), 65.23% (multi-constraint enhancers), and 94.56% across 18 TFs.
- Why it matters: Regulatory DNA design is hard because gLMs capture genomic syntax without ensuring targeted activity, while function predictors lack native-like sequence plausibility.
- What they did: GENECraft performs training-free inference-time optimization by evolving latent nucleotide logits, then scoring strictly decoded DNA with a composite vector field combining AlphaGenome functional utility, Carbon prior guidance, and compositional quality penalties.
- The result: On 2,185 loci it substantially outperforms seven baselines (up to 2.15% on enhancers), and in 336-locus ablations Carbon guidance reduces native-relative dinucleotide divergence by 7.8–11.5% and context/TF mutation sensitivities by 10.5% and 11.0%.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.