Diversity-Seeking Docking-Guided Molecular Design Discovers Broader Predicted Interaction Profiles across Protein Families

Structure-based discovery pipelines commonly prioritize a scalar docking score, concentrating candidates into narrow chemotypes and contact patterns. We formulate docking-guided de novo molecular design as a diversity-seeking exploration problem and compare goal-directed, quality-diversity, Pareto-optimization, affinity-focused, novelty-based, and random search policies. Molecules are generated by fragment-based CReM edits and scored by rigid-receptor GNINA/Vina docking. We introduce a residue-level multi-interaction fingerprint encoding hydrophobic contacts and hydrogen-bond donor and acceptor interactions, together with the reachability ratio, an empirical diagnostic of goal-retrieval discriminability. Across seven targets from five protein families, with 500 docking evaluations per run, Curiosity-IMGEP discovers 15.0% more interaction profiles than random search (d = 1.01, pBonf < 10–6) and produces +12% more Bemis–Murcko scaffolds and +21% more binary residue-contact patterns. A newly added budget-matched NSGA-II baseline achieves the strongest predicted binding affinity in the benchmark (−11.03 vs Genetic algorithm’s −10.98 kcal/mol; d = −1.02 vs Random, pBonf < 10–8), tied within noise with the Genetic algorithm, and leads on Bemis–Murcko scaffold count (23 ± 12), but underperforms QD/goal-directed methods on interaction-profile counts (259 ± 46, −12% vs Random). Representation ablations compare continuous and binary fingerprints, charged-interaction channels, distance cutoffs, and inclusion of docking quality in the search vector. Target-specific ChEMBL neighborhood and pose-robustness analyses provide orthogonal computational triage. The results establish that diversity-aware search broadens the set of docking-predicted interaction hypotheses generated under a fixed budget; they do not experimentally establish distinct binding modes or biochemical activity. Code, configurations, and raw results are publicly available.