GENECraft: A Function- and Prior-Constrained Discrete Dynamical System for Controllable Regulatory DNA Design
In the authors' words
Designing regulatory DNA requires reconciling two complementary modeling paradigms: genomic language models (gLMs) capture evolutionary syntax without guaranteeing targeted regulatory activity, whereas sequence-to-function predictors provide explicit functional supervision without enforcing native-like sequence plausibility. We introduce GENECraft, a training-free Function- and Prior-Constrained Discrete Dynamical System (FP-DDS) that couples a genomic language model (Carbon) and a functional oracle (AlphaGenome) during inference-time optimization to balance predicted regulatory activity and consistency with genomic sequence priors. GENECraft evolves continuous latent nucleotide logits while evaluating strictly decoded discrete DNA sequences under a time-dependent composite vector field that fuses normalized functional utility, bidirectional contextual prior guidance, and compositional quality penalties, followed by feasibility-first trajectory readout. Across 2,185 loci spanning promoters, multi-constraint enhancers, and 18 transcription factors (TFs), GENECraft achieves full-joint success rates of 62.70%, 65.23%, and 94.56% (macro-averaged), substantially outperforming seven optimization, evolutionary, and generative baselines (which reach at most 2.15% on enhancers). In matched 336-locus ablations, in-loop Carbon guidance reduces native-relative dinucleotide divergence by 7.8--11.5%, enhancer context sensitivity by 10.5%, and TF mutation sensitivity by 11.0%, establishing FP-DDS as a principled framework for controllable regulatory DNA design.
Appeared: Thursday, November 19. bioRxiv. Preprint, not yet peer-reviewed.