RosettaTrace: Human interpretable protein sequence design using rollout-guided search with Large Language Models

We introduce RosettaTrace, an inference-time optimization approach that uses large language models as a policy within a rollout-based search algorithm, ranking candidate sequences using scalarized rewards computed from structure prediction models such as RosettaFold3. Unlike training-based sequence design methods, RosettaTrace performs optimization entirely at inference time and uses structured textual feedback. Across 500 protein redesign tasks with low designability metrics, RosettaTrace obtains between 17 to 65\% improvement in the metrics. When refining candidates generated by LigandMPNN, RosettaTrace achieves a 2× improvement in design success rates, and matches LigandMPNN when initialized from suboptimal native sequences. Finally, we show that rollout trajectories produced by RosettaTrace can be reused as preference data for continual model adaptation.