From Model Scores to Discovery Decisions: Decision-Contract Evaluation of Tools for Molecular Generation and Docking

Modern computational drug-discovery pipelines chain multiple stages, with competing tools at each. Yet tools are typically benchmarked in isolation using stage-specific metrics—such as molecular validity, uniqueness, property scores, or pose RMSD—that may not predict which tool best supports a downstream discovery decision. We introduce DecisionBench-DD v1.0, a decision-contract evaluation methodology that assesses tools within a frozen, auditable pipeline rather than a ready-made multi-target benchmark suite. We demonstrate it on two connected stages—molecular generation and docking—in a single ambroxol-conditioned glucocerebrosidase (GCase/GBA1) case study comparing GenMol with MolMIM and GNINA with ligand-centered 15 Å-cropped DiffDock-L. Although MolMIM led on uniqueness, diversity, QED, and synthetic accessibility, only 3/300 primary outputs passed the frozen chemical-qualification gates, versus 66/300 for GenMol. Post-hoc MolMIM reruns at min_similarity 0.5 and 0.7 yielded 2/300 and 10/300 passers without reversing this ordering; neither rerun reached its requested threshold under independent ECFP4 recomputation. In redocking, the primary Top-1 endpoint did not reliably separate the dockers (3/14 versus 2/14), but GNINA recovered a near-native pose within five predictions for 8/14 versus 2/14, achieved lower best-of-five RMSD in all 14, and produced geometry-valid poses (70/70 versus 45/70). After downstream GNINA screening, 15/300 GenMol and 1/300 primary MolMIM outputs passed the structural proxy—a separation driven primarily by chemical qualification rather than demonstrated differential structural retention—while alternative rerankers changed redocking and screen counts in directions that cannot be signed without candidate-level ground truth. Tool rankings can therefore change when component metrics are replaced by task-specific constraints, frozen gates, and auditable advancement decisions.