GrAdaBeam: Combining model gradients with evolutionary search for generalizable nucleic acid design

Abstract

We introduce GrAdaBeam, a hybrid model-based optimization algorithm that combines gradient-derived attention maps with an adaptive beam search to navigate complex nucleic acid fitness landscapes. By unifying the broad exploration of evolutionary methods with the precise guidance of gradient descent, GrAdaBeam overcomes a central limitation of existing approaches: no single optimization strategy performs robustly across the full spectrum of genomic design tasks. We rigorously evaluate GrAdaBeam and seven other design algorithms using NucleoBench, a novel benchmark covering 17 diverse genomic tasks that introduces a paired-start-sequence design for superior statistical comparisons. GrAdaBeam statistically outperforms all other algorithms across over 600,000 experiments, never ranking lower than second across all 17 benchmark tasks, while baseline methods often struggle on large models or long sequences. Critically, GrAdaBeam sequences generalize most reliably to independent predictive models and recover canonical transcription factor binding motifs de novo, providing evidence of biological signal capture beyond the optimization target. GrAdaBeam and NucleoBench are freely available as an open-source package.