Drug-design agents reuse edit signals more than pocket identity
Original title: What Transfers Across Protein Pockets? Experience Reuse in Molecular Optimization for Agentic Structure-Based Drug Design
On one CrossDocked2020 run, keys that recur for unseen pockets mostly encode edit size, action, and round.
The paper tests whether docking probes from LLM agents optimizing ligands in protein pockets can be stored and reused on pockets the experience base has never seen. Using edited-site contacts, the action, and the round as a lookup key separates improving from worsening edits, but that split is largely explained by molecule-size change, action, and round. Naming the target separates outcomes more sharply, yet specific exact-match descriptions rarely recur and coarser ones separate poorly. Under the representations studied, the base is framed as more useful for prioritizing an agent's own docking proposals than for prescribing what to build.
Unconfirmed: Conclusions apply only to the representations, lookup scheme, and single CrossDocked2020 run log analyzed.
- Guidance from controlled probe edits is reported to beat the model's existing chemical knowledge.
- Probes are usually discarded with the pocket instance that produced them.
- Target identity separates improving and worsening edits several times more sharply than the studied key.
- Signals that transfer describe the edit and resulting molecule, not the target.
- Products
- CrossDocked2020
From the original article
1 min read at arxiv.orgLLM agents optimize a ligand in a protein pocket by editing it step by step and docking each edit. Recent agents first spend extra dockings probing the complex with controlled edits; guidance induced from these probes is reported to help more than the chemical knowledge the model already carries.
By Zaifei Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.