MutexaGPT: LLM agents plus MD simulations hit 40% on enzyme design, 4x the baseline
bravo_abad · x · 2026-09-11
MutexaGPT, built by Shao and coauthors, demonstrates a more useful division of labor for LLMs in science: translating a scientist's intuition into the right calculation rather than predicting the answer.
- How it works: a researcher starts with a qualitative goal (e.g., "make the active site larger"); LLM agents clarify the request, map it onto a physical quantity, choose mutations to test, and configure the workflow. Actual ranking comes from molecular-dynamics simulations, not the language model.
- Results: in one case, the system translated that intuition into cavity-volume calculations for 600 enzyme variants, each sampled with MD, and returned a shortlist of ten candidates.
- Hit rate: four of the ten were experimentally known hits (40%), roughly four times the baseline.
More from Research
- Steerable Visual Representations Presented as ICML Long Oral — y_m_asano · 2026-09-11
- OpenCVL: a satellite-to-photo registration dataset at ECCV 2026 — ducha_aiki · 2026-09-11
- Diverse VPR work submitted to ECCV 2026 — ducha_aiki · 2026-09-11
- EASE: evidence-anchored spatial attention lifts multimodal RLVR by up to 3.1 points, EMNLP 2026 — jiqizhixin · 2026-09-11
- P=NP Explained: Why Class Schedules and Circuit Routing Are the Real Hard Problems — thesaraharminta · 2026-09-11
- Hypothesis: ASI Has a Mathematical Incentive to Preserve Human Diversity — No_Cause_2731 · 2026-09-11