Large language model (LLM) agents have shown promise in driving scientific discovery, but their effectiveness in complex, real-world biological problems remains underexplored. We ask whether a general-purpose LLM agent can drive semi-autonomous development of a model for a genuinely hard biological problem, RNA 3D structure prediction. We designed a development loop where, under a fixed budget and with human supervision, the agent iteratively proposed, implemented, trained, and evaluated model changes. Over 297 iterations, the model evolved from a randomly-initialised baseline to QuickFold, an 8.9M-parameter folding trunk that matches the strongest open-source baselines (RhoFold+, NuFold) on lDDT and TM-score within noise on a held-out test set at a fraction of their inference cost. We frame this less as a new predictor than as a case study in feedback-driven, agent-led model development.
Majewski, M., Malo, L., Montero-Blay, A., Marengo, M., Gkeka, P., Minoux, H.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 27
- Comments 0