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

What it claims

The paper tests whether docking measurements collected on one protein pocket stay useful when reused on a pocket never seen before. It builds an experience base from a single full run of the PROBE agent over the CrossDocked2020 test-set pockets, yielding 14,341 records (5,398 from the probe phase, 8,943 from design-make-test-analyze edits). The finding: the keys that still separate good edits from bad ones on a held-out pocket mostly describe the edit and the molecule it produced, not the target. No evaluated key satisfies all three requirements at once: use pocket information, recur on an unseen pocket, and still separate outcomes.

How it works

Each record stores the pocket, ligand, round, editing action, four fragment constraints, pocket-ligand contact features, and the outcome (ΔA affinity change, ΔN heavy-atom change). Affinity is A(m) = −Vina(m). PROBE optimizes 10 MolCRAFT-generated ligands per pocket through a probe phase (round 0) and five DMTA rounds (1–5). A "key" is a tuple of record properties; the base pools records matching a candidate edit and predicts improve when improving records outnumber worsening ones. Every held-out test splits at the pocket, so a record is always labeled by a base that never saw its pocket. Separation is measured two ways: a Cochran-Mantel-Haenszel stratified risk difference, and a bias-corrected ω² on outcome residuals after fitting out ΔN and round.

Evidence

The records' built-in key κ0 = (contact features, action, round) separates edits by +53.3pp. But holding the action, ΔN, and round equal for the two Replace actions shrinks the gap to +0.3pp [−5.1, +6.3], so the contact features (the only pocket field) add nothing detectable. In the cross-pocket table, ligand-change keys win: Δrings beats a ΔN baseline on 91% of pockets (+8.1pp lift), Δrotatable bonds on 85% (+8.0pp), Δaromatic atoms on 70% (+6.8pp), fragment constraints on 63% (+5.1pp). Pocket-description keys either fail to recur (all six descriptors: 17% recurrence) or fail to separate (burial, hydrophobic fraction: 100% recurrence, below baseline).

Limits

The authors recommend using the base to rank an agent's own proposals for docking, not to prescribe what to build; ranking gains and docking savings are not measured. Evidence is one pipeline (PROBE), one initial-ligand generator (MolCRAFT), one scorer (Vina, not experimental binding), one pocket-descriptor set, and one key family: categorical exact-match lookups. Similarity-based pooling and combining pocket descriptors with ligand-change keys are untested. Intervals are unadjusted for multiple comparisons.