Scientific large language model (LLM) agents have shown promise in literature reasoning, tool use, and experiment planning, yet it remains unclear whether they can autonomously improve large, tightly coupled scientific machine‑learning systems through executable code changes and costly validation.
This work focuses on protein folding, a domain that demands coordinated architectural modifications, multi‑objective evaluation, and domain‑aware interpretation of results.
We introduce AgentFold, a multi‑agent framework that treats folding‑model development as a closed‑loop search over executable code variants. Starting from the open‑source model ESMFold, the system iterates through hypothesis generation, code‑level implementation and debugging, model variant evaluation, outcome analysis, and storage of both successful and failed interventions in a structured memory.
Computational resources are allocated across high‑scoring search branches using an MCTS‑style policy, balancing depth and breadth under a fixed budget.
On an engineering‑scale protein‑folding codebase of over 2,000 lines, AgentFold explored roughly 80 model variants, consuming about 5,000 GPU‑hours and 170 million LLM tokens.
Within the same computational budget, AgentFold improved the best $lDDT$ by 7.5%, surpassing independent Codex proposals and outperforming a random‑search control.
Analysis of intervention traces revealed recurring empirical design patterns: early, soft, learnable priors and gated refinement tend to yield stable gains, whereas direct geometric perturbations and geometry‑conditioned feedback often destabilize training.
The code and experimental resources are publicly available at https://github.com/lmqfly/AgentFold.
Blogger's Review: AgentFold demonstrates the feasibility of LLM agents operating within real scientific software loops, especially by leveraging structured memory and search strategies for efficient model iteration, offering valuable insights for future automated research.