A new framework called DrugMCTS pairs multiple AI agents with retrieval and tree search to hunt for new uses of existing drugs.
Researchers built DrugMCTS around five specialized agents that each retrieve and analyze molecular and protein data, then combined that with retrieval-augmented generation and Monte Carlo Tree Search to drive iterative, structured reasoning. Tested on the DrugBank and KIBA datasets, the system beat both general-purpose large language models and deep learning baselines on recall and robustness. The paper, posted on arXiv, is now in its fourth revision.
Drug repurposing matters because reusing an approved drug for a new condition is far cheaper and faster than developing one from scratch. Plain LLMs tend to stall once a question moves past what they memorized during training, and simple retrieval add-ons do not fully exploit structured scientific data like protein-binding profiles. This result suggests multi-agent collaboration with a built-in search-and-feedback loop may do more for scientific reasoning than just feeding an LLM more documents.
It is a promising benchmark result, not a clinical one. Beating other models on two curated datasets is a long way from a drug actually working in a patient, and a paper that has already been revised four times is still very much a work in progress.