AgentFold Uses AI Agents for Autonomous Protein Folding Model Design

While scientific large language model (LLM) agents have shown promise in tasks like literature review and tool use, their ability to autonomously enhance complex scientific machine learning systems, particularly through executable code modifications and computationally intensive validation, has been less clear. This research investigates this capability within the domain of protein folding, a field requiring coordinated architectural changes, multi-objective evaluation, and deep domain interpretation. The proposed solution is AgentFold, a multi-agent framework that conceptualizes protein folding model development as a closed-loop search process over various executable code configurations. Starting from an existing model like ESMFold, AgentFold operates by generating hypotheses, implementing and debugging code-level adjustments, evaluating the performance of these modified model variants, and then analyzing the experimental results. Both successful and unsuccessful interventions are stored in a structured memory for future learning. AgentFold employs an MCTS-style policy to efficiently allocate computational resources across promising search branches. In experiments on an engineering-scale protein-folding codebase, AgentFold explored approximately 80 model variants, utilizing substantial GPU-hours and LLM tokens. It achieved a 7.5% improvement in lDDT over independent Codex proposals and outperformed a random-search control under the same computational budget. The intervention traces also revealed recurring design patterns, such as the effectiveness of early, soft priors and gated refinement for stable gains.