Skip to content
WeChat · Synced (Jiqizhixin)·Indexed by this site · Originally published· AI updates →· Its place in the scan record →SelectedAttention38

MolRSI recursive self-improving molecular discovery workflow, with a wet-lab hit rate of 15.9% in antibacterial screening, higher than the 8.7% of the human expert workflow

AI自己改科研策略,能过湿实验这一关吗?他们把RSI带入分子发现

AI summary

The research team proposed the recursive self-improving agent MolRSI, which in an antibacterial molecule discovery task raised the wet-lab hit rate from 8.7% for the human expert workflow to 15.9%, a relative increase of about 82%. Both workflows used the same training data and a screening library of about 12.1 million compounds, and hits were judged by the same criterion of a minimum inhibitory concentration against methicillin-susceptible Staphylococcus aureus (MSSA) of no higher than 32 μg/mL. The system uses experience-guided Monte Carlo tree search to optimize the workflow; the inner loop improves the current task, and the outer loop carries accumulated experience into subsequent searches. After adding cross-task experience, the average test AUPRC improved on all 6 antibacterial prediction tasks. On the task of de novo designing pMHC-binding proteins, MolRSI obtained a total of 8 binding hits meeting the experimental specificity criteria across the two targets MART-1 and SARS-CoV-1, compared with 0 for the Claude Science control under the same settings and computational budget. The paper was published on arXiv (arXiv:2610.39714).

Why it was selected

This work applies recursive self-improvement to molecular discovery and compares wet-lab hit rates with a human expert workflow, providing an experimentally validated case of AI autonomously improving the scientific research process.

Related papers
Read originalmp.weixin.qq.com