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

1 Introduction

• PROBE spends extra dockings measuring controlled edit responses, yet its resulting guidance remains with the instance that produced it.The paper proposes accumulating those measurements so a new optimization can use evidence from earlier runs.Source: §1 · pp. 1–2
• A useful retrieval key should combine pocket and ligand information, recur on an unseen pocket, and still separate edit outcomes.The analysis asks whether a full PROBE run log can supply such a key.Source: §1 · p. 2
• The contact, action, and round key separates improving from worsening edits, but evaluated transferable keys do not jointly provide recurrence and strong separation.Keys that transfer most consistently describe the edit or its product more than the pocket.Source: §1 · p. 2

2 Related work

• CIDD, LIDDIA [1], and PROBE [17] use different LLM roles or workflows, but retain learned pocket information within the current optimization instance.CIDD proposes edits, LIDDIA plans while an algorithm executes, and PROBE measures responses to controlled edits.Source: §2 · p. 2
• Agents outside chemistry reuse run knowledge as verbal accounts, skill libraries, or workflows distilled from trajectories [13] [15] [16].Those systems also face transfer questions.Source: §2 · p. 2
• The paper studies whether edit outcomes measured on one protein pocket remain informative on another.This is its specific form of experience transfer in structure-based drug design.Source: §2 · p. 2

3 Setup

• Each optimization step edits a ligand in a pocket, with affinity scored as negative Vina and outcomes recorded as affinity and heavy-atom changes.The setup defines ∆A as the affinity difference after an edit and ∆N as the change in heavy-atom count.Source: §3 · p. 2
• A single full PROBE run on CrossDocked2020 test pockets produced 14,341 successfully built-and-docked edit records.The run optimized 10 MolCRAFT-generated initial ligands per pocket through round 0 probing and five design-make-test-analyze rounds.Source: §3 · p. 2
• For each held-out pocket, the base is built from other pockets’ records and used to label that pocket’s edits.A key selects records agreeing on specified properties; the evaluation splits by pocket so the base has not seen the evaluated pocket.Source: §3 · p. 3

4 Transferable keys primarily describe the edit

• The evaluated keys test recurrence and separation together; the initial key recurs and separates, but its separation is largely associated with edit properties.Pocket descriptions show a recurrence–separation trade-off.Source: §4 · p. 3

4.1 The initial separation largely reflects edit properties

• κ0 separates improving from worsening edits, but much of its gap is explained by heavy-atom change, action, and round rather than pocket contacts.For the two Replace actions, the remaining contact-feature separation has an interval covering zero after matching those properties.Source: §4.1 · p. 3

4.2 The evaluated pocket keys trade recurrence for separation

• To assess pocket information beyond edit properties, the analysis fits heavy-atom change and round out of the outcome and scores keys on the residual.The fit is ordinary least squares and is refitted in each held-out fold.Source: §4.2 · p. 4
• Pocket identity separates outcomes several times more sharply than κ0, but identity-based partitions are in-sample references, not estimates for unseen pockets.Refining identity partitions by action or starting ligand separates outcomes further.Source: §4.2 · p. 5
• The six pocket descriptors approach identity’s in-sample separation but form nearly one group per pocket; coarser descriptors recur broadly but explain little variation.The evaluated descriptor keys do not outperform the baseline in this evaluation.Source: §4.2 · p. 5
• Ligand and edit keys recur almost everywhere and beat the ΔN baseline on most pockets; four separate outcomes more sharply than κ0.These keys describe the edit or the ligand it produced, not the pocket.Source: §4.2 · p. 5

5 Discussion and limitations

• Exact-match pocket descriptions trade recurrence against separation: fine partitions approach pocket identity’s in-sample separation but recur on few held-out records, while coarser descriptions recur broadly but explain little outcome variation.The conclusions apply to the representations and lookup family evaluated, not to cross-pocket experience reuse in general.Source: §5 · p. 5
• The results favor ranking an agent’s own proposals for docking, but no candidate-ranking, closed-loop optimization gains, or docking savings were evaluated.The authors did not identify a key among the evaluated representations that effectively conditions direct edit prescription on a previously unseen pocket.Source: §5 · p. 5
• Evidence is limited to one PROBE run setup, one initial-ligand generator, one scoring function, CrossDocked2020 test pockets, six pocket descriptors, and categorical exact-match keys.Similarity-based lookup and combining pocket descriptors with the strongest ligand-change keys remain untested; the endpoint is Vina score rather than experimental binding.Source: §5 · p. 5

A The PROBE workflow

• PROBE scores each molecule for affinity from its Vina score and druggability from a score built from QED and synthetic accessibility.The workflow begins with a generative-model ligand docked into the pocket and measures controlled edits before choosing an optimization direction.Source: A The PROBE workflow · p. 8
• A diagnosis LLM builds a fixed site map from protein contacts, geometric fit, and ligand-property signals, labeling editable fragments SYNERGY, TENSION, or LIABILITY.The labels characterize expected trade-offs between the two objectives or identify a liability to repair.Source: A The PROBE workflow · p. 8
• The probe stage submits twelve controlled edits per input ligand—three strategies at high, medium, and low magnitude, each with a counterfactual—and retains successfully built-and-docked edits.These twelve edits constitute round 0.Source: A The PROBE workflow · p. 8
• An analyzer assesses probe response patterns and counterfactuals; a manual-construction LLM records each site’s pharmacophore role, allowed edits, fragment constraints, blacklist, and site dependencies.A strategy is credible only when forward probes improve and its counterfactual does not.Source: A The PROBE workflow · p. 8
• Across five DMTA rounds, affinity and druggability agents draft localized edits, cross-review them against the EditManual, and a co-optimization agent reconciles the drafts.Edits are built, docked, and labeled against their intent; outcomes can tighten the EditManual.Source: A The PROBE workflow · p. 8
• At round 0 and after each DMTA round, PROBE promotes one candidate; candidates failing to improve both affinity and druggability receive a joint-gain score of zero.After the last round, it returns the highest-scoring molecule from the input, evaluated probes, and optimization products.Source: A The PROBE workflow · p. 8
• One assembly engine builds edits in both probing and optimization, drawing needed fragments from 8,505 BRICS fragments and filtering them by shape category.It ranks the matching subset by 0.5 QED + 0.5 SA − 0.15 Nchiral, then attaches and sanitizes the top-ranked fragment.Source: A The PROBE workflow · pp. 8–9

B Corpus record schema

• A record is a structured edit built by the fragment-assembly engine and scored by Vina; it stores the parent molecule, product, pre-product edit fields, and outcomes.Only the affinity change among the listed outcomes requires docking; ligand counts can be read from the assembled product.Source: B Corpus record schema · p. 9
• Each record identifies a pocket, its starting ligand, and the round; the pocket–ligand pair defines an instance, and pocket descriptors join by pocket.The study uses 93 pockets and ten starting ligands per pocket, for 930 instances.Source: B Corpus record schema · p. 9
• The action is one of Delete_Group, Grow, Replace_Terminal_Group, or Replace_Sidechain_or_Ring; the edit atom index is recorded but not used as a key.An atom index does not carry meaning across different molecules.Source: B Corpus record schema · p. 9
• Fragment-adding actions carry categorical size, polarity, flexibility, and shape constraints; Delete_Group carries none, and the engine’s fixed fragment is determined by constraints and edit atom.The engine has a formal-charge field, but the agent does not set it and the corpus does not record it.Source: B Corpus record schema · p. 9
• The contact key joins PLIP contact types at the edited atom, while ligand outcome keys include changes in heavy atoms, rotatable bonds, aromatic atoms, rings, and hydrogen-bond donors and acceptors.Contact labels omit residue identity, contact counts, distances, and angles; NONE occurs in 61% of records.Source: B Corpus record schema · p. 9

• Six pocket descriptors use the receptor structure and reference ligand, rather than run-generated information, so a key can be read on an unseen pocket.Source: C Pocket descriptors · p. 9
• The residue set includes residues with a heavy atom within 10 Å of a reference-ligand heavy atom, summarized by residue count and four overlapping class fractions.The classes omit GLY, so their fractions do not sum to one.Source: C Pocket descriptors · p. 10
• Burial counts receptor heavy atoms within 8 Å of each reference-ligand heavy atom, and all six descriptors are cut into tertiles over 93 pockets.Residue count ranges from 24 to 106 and burial from 3.5 to 25.8; every class fraction varies at least twofold.Source: C Pocket descriptors · p. 10
• Table 3 compares pocket-only and pocket-plus-edit keys with edit keys and pocket identity, reporting each key's group count and scorable held-out pockets.Source: Table 3 · p. 10
• The (k −1)s2 correction in ω2 addresses apparent informativeness from keys forming many groups; permuted-label checks leave ω2 at 0.1% or below.This check supports the group-count correction but does not establish held-out predictive value.Source: C Pocket descriptors · p. 10