Title: Investigate Prior to Modification: Probing-Driven Molecular Optimization for LLM Agents in Structure-Based Drug Design

Abstract: While LLM agents are increasingly utilized in structure-based drug design to iteratively refine ligands within target pockets, creating a viable candidate requires balancing two frequently contradictory goals: binding affinity and druggability. Single optimization steps rarely succeed in enhancing both metrics simultaneously. To measure this challenge, we present two diagnostic metrics: one tracks the frequency with which a single edit improves both objectives, while the other records instances where progress on one goal results in a decline in the other. When applied to existing LLM-agent workflows, these diagnostics reveal a persistent failure pattern: agents modify molecules without understanding how the pocket-ligand complex reacts to local changes, making joint improvements uncommon. Drawing inspiration from medicinal chemists, who test controlled analog edits on the pocket-ligand complex before determining an optimization strategy, we introduce PROBE, an optimization framework centered on edit-response probing. PROBE begins by breaking down the ligand into editable sites and generating a pocket-specific site map that identifies areas where joint gains are possible, where the two objectives conflict, and where liability substructures require alteration. It then executes controlled probe edits, distilling the responses into an EditManual. Leveraging the site map and EditManual, PROBE employs an iterative multi-agent loop involving an affinity agent, a druggability agent, and a co-optimization agent to collaboratively generate edits. Evaluated on the CrossDocked2020 benchmark, PROBE delivers state-of-the-art results and significantly reduces the failure modes identified by our diagnostic metrics.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC