Regulatory DNA design stops chasing activity alone: GENECraft folds a sequence prior into the loop, reaching 94.56% joint success on TF sites
Synopsis
Regulatory DNA design has split into two paths: genomic language models produce native-like DNA without guaranteeing target activity, while sequence-to-function predictors supply activity gradients without enforcing sequence plausibility. GENECraft places a frozen Carbon prior and AlphaGenome functional gradients inside one training-free discrete dynamical process, updating continuous nucleotide logits at each step and selecting among all visited states the sequence that first satisfies hard composition constraints. Across 2,185 human regulatory loci, full-joint success reaches 62.70% for promoters, 65.23% for enhancers and 94.56% for 18 transcription factors, with activity checked by held-out Enformer and Borzoi.
Interpretation
Regulatory DNA synthesis can now satisfy functional objectives and sequence-plausibility constraints within a single inference process instead of generating first and filtering afterward. Previously, genomic language models sampled unconditionally and sequence-to-function predictors supplied only activity gradients, each lacking what the other provided. On 512 K562 promoter loci, GENECraft reaches 62.70% full-joint success, above Fast SeqProp at 45.31% and Evo2 at 5.27%.
The method ties updates on continuous logits to exact evaluation of discrete DNA sequences, scoring the same real sequence at every step. A straight-through interface with fixed Gumbel noise decodes forward by argmax and routes gradients backward through softmax, turning discrete search into a differentiable iteration. AlphaGenome base-resolution gradients and Carbon 6-mer gradients pass through a parameter-free relaxation bridge, are RMS-normalized into one update direction, and advance under a linear step-size schedule.
The readout checks constraint satisfaction first, then takes the highest functional utility among feasible states. Discrete trajectories over rugged multi-constraint landscapes are non-monotonic, so the final step is not necessarily the best candidate. The readout pools all states visited across restarts, filters for generation feasibility F=1, ranks by exact AlphaGenome utility, and otherwise keeps the highest-utility sequence with its failure flag.
Perspective
Evaluation covers 2,185 human regulatory loci from the ATGC-Gen benchmark: 512 K562 promoters, 512 melanocyte enhancers and 1,161 GM12878 binding sites across 18 transcription factors; generation and readout use only AlphaGenome and composition constraints, with Enformer and Borzoi reserved for post-generation validation.
All activity conclusions come from cross-oracle prediction rather than wet-lab measurement reported here; enhancer risk thresholds are calibrated from predicted quantiles of 50 reference sequences, and the authors also report a threshold sensitivity analysis.
