BoltzEvo: Evolving All-Atom Protein Binders with Predicted Structural Inheritance

Modern binder-design pipelines generate and evaluate candidates largely independently: even when a structure predictor identifies a promising complex, its predicted geometry and local confidence are typically discarded before the next proposal. We introduce \textbf{BoltzEvo}, a training-free evolutionary algorithm that instead allows promising molecular designs to evolve across generations. Each evaluated binder is treated as an individual whose predicted complex provides both a fitness signal and a heritable structural state. BoltzEvo selects promising parents, preserves confidence-supported regions of their predicted complexes, regenerates the remaining sequence and structure through partial diffusion, and reevaluates the resulting descendants before the next generation. Across ten targets, BoltzEvo shifts the full proposal population toward high-confidence, design-consistent solutions: $66.5%$ of protein and $43.5%$ of nanobody descendants satisfy a joint computational criterion, compared with $4.0%$ and $0.4%$ under a larger independent BoltzGen sampling controls. Matched controls show that inheriting predictor-refined geometry and iteratively updating the parent population are the strongest contributors to this improvement, while confidence-guided inheritance introduces a yield--coverage tradeoff. These results show that structure prediction can serve not only as a terminal filter, but as a source of heredity for evolutionary search, enabling iterative improvement of all-atom molecular designs without retraining either model.