SWAN-MPO: A Safety-Aware, Affinity-Normalized Multi-Parameter Optimization Framework for Natural-Product Scaffold Triage in Oncology

<title>Abstract</title> <p> Multi-parameter optimization (MPO) balances heterogeneous drug-discovery objectives, but early scaffold triage can remain dominated by affinity when binding, absorption/distribution/metabolism/excretion (ADME), safety, and scaffold-specific liabilities are not integrated on a common scale — especially for natural-product scaffolds in oncology, where strong predicted binding often coexists with structurally-linked safety liabilities. SWAN-MPO transforms docking, safety, ADME, and liability inputs into bounded 0.01–1.00 desirability values and integrates them using an equal-weight geometric mean, limiting strong domain performance from compensating severe weakness elsewhere. The scoring implementation is hash-locked with one-command recomputation from packaged canonical predictor inputs. Binding was summarized using median energies across 20 docking grid centers, target-specific reference calibration, and BestNode panel aggregation, with multiple robustness and sensitivity analyses. In pilot colorectal, prostate, and renal cell carcinoma panels and an expanded application to 59 <italic>Annona muricata</italic> natural products across nine targets (10,620 docking jobs), SWAN-MPO produced rankings distinct from binding-only and generic property-based approaches. In a broad ChEMBL benchmark comparing 1,097 natural-product-like actives with 3,291 synthetic descriptor profiles, binding-plus-ADME integrated scoring improved global discrimination relative to binding alone (ROC-AUC 0.810 vs. 0.612) and performed comparably to an ADME-only proxy (0.811); safety and liability were held constant across classes and therefore provided no discriminatory information in this benchmark. A stricter target-matched benchmark using real active and weak/inactive records across seven targets showed greater global discrimination for an RDKit-derived liability proxy than for the ADME proxy (ROC-AUC 0.583 vs. 0.464). SWAN-MPO achieved the highest global ROC-AUC among adapted descriptor-compatible implementations of CNS-MPO, LE–QED, QED, and QEPPI, with mixed early-enrichment performance. A retrospective stress test showed automated predictors missed several documented hazards while literature-curated annotations encoded them, treated as a sensitivity test because the same hazards informed both cohort and annotations. Among 738,801 COCONUT compounds scored on physicochemical properties alone, the representative scaffold ranked at the 45.24th percentile, not exceptional by drug-likeness alone. SWAN-MPO does not establish biological efficacy or therapeutic readiness; it provides a reproducible framework for prioritizing natural-product scaffolds for experimental evaluation. <bold>Scientific Contribution.</bold> Existing MPO frameworks such as QED, CNS-MPO, and QEPPI emphasize general drug-likeness or property-window optimization with uncalibrated binding. SWAN-MPO instead integrates target-specific calibrated multi-grid docking, general safety, ADME, and panel-context-aware liability within a common bounded desirability framework. The scoring implementation and target-calibration resource are hash-verifiable, with one-command recomputation of frozen manuscript scoring outputs from packaged canonical predictor inputs. </p>