What Transfers Across Protein Pockets? Experience Reuse in Molecular Optimization for Agentic Structure-Based Drug Design

TL;DR

This paper investigates whether experience from prior protein-ligand optimization runs can transfer to new pockets via a memory base using edit history and context as keys. It finds that while specific pocket identities improve separation of good vs. bad edits, coarse descriptions of edits and molecular changes transfer better than target-specific features, suggesting memory is more useful for prioritizing agent-generated proposals than prescribing exact molecules.

Why It Matters

The findings guide the design of agentic drug discovery systems by showing that experience reuse is most effective when focusing on generalizable edit patterns rather than target-specific knowledge, enabling more efficient exploration in unseen protein pockets.

Key Findings

• Using edit context (contacts, action, round) as a key enables effective separation of improving vs. worsening edits in unseen pockets.
• The transferability of experience is primarily driven by molecular changes and edit characteristics, not target-specific features.
• An experience base is more valuable for prioritizing agent-generated candidates than for directly recommending specific ligands.