SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

Summary

The paper introduces specificity optimization (SpecOpt), a molecular design task defined as making constrained structural edits to an existing drug so it preferentially binds an intended target over known off-targets. The authors build a ChEMBL-derived benchmark of 1,848 compounds (median 3 off-targets each; 915 pass docking filters and complete evaluation) and define the objective as a Vina docking-score gap between target and tightest off-target. Their agentic loop docks the compound against each protein, classifies per-atom contacts within 7 Å (including shared-contact-different-residues, the key class), detects growth opportunities, aggregates over BRICS fragments, and lets GPT-5.4 propose single-fragment edits; candidates are accepted only if they keep Tanimoto above 0.4, stay within ADMET tolerances, and improve the gap. The mean gap shifts from -0.72 to +0.47 kcal/mol, 776/915 compounds (84.8%) improve at mean Tanimoto 0.72, and 190 of 575 negative-baseline compounds reach positive gaps. An ablation on 29 compounds shows replacing residue-aware contacts with a naive three-way classification eliminates all improvement. Key limitations: all conclusions rest on Vina scores whose agreement with measured affinity the authors themselves report as weak (median Spearman rho = -0.075 over 19 proteins) and which barely separate targets from off-targets; candidate acceptance uses the same docking signal as evaluation; there is no experimental or free-energy validation; ablation samples are small (n≈29) with several p-values above 0.05.