Turning P-glycoprotein substrates into predicted non-substrates: SF-GA guides fragment edits by synergy scores and leads all tested methods on retrosynthetic accessibility

Molecular editing can now be steered toward evading P-glycoprotein recognition rather than only scoring existing molecules for a property. P-gp is a transporter that actively effluxes structurally diverse small molecules and, at the blood-brain barrier, limits drug entry into the brain, so reducing substrate liability is a prerequisite for CNS drug design. Machine-learning models already distinguished P-gp substrates from non-substrates with good accuracy, but those predictions were rarely translated into actionable molecular modifications, and the rational conversion of substrates into non-substrates was largely unexplored. On the PMO benchmark, starting from a pool of 100 randomly selected P-gp substrates with a budget of 5,000 function evaluations, SF-GA reached an AUC top-50 of 0.948±0.004, against 0.965±0.002 for GB-GA, 0.964±0.001 for REINVENT and 0.955±0.004 for FragGT, while SF-GA's mean RAscore of 0.932±0.037 exceeded REINVENT's 0.900±0.029 and FragGT's 0.888±0.107.

Every accepted fragment replacement is recorded together with the discarded fragment, the inserted fragment and their synergy scores, so each structural change carries a fragment-level rationale. Earlier generative methods returned molecules largely as black boxes, and attribution maps or feature-importance scores did not translate readily into concrete, synthesizable transformations that could be tested. The synergy score is defined from co-occurrence between MACCS fingerprint bits and bounded between -1 and 1; against F-GA, which picks fragments at random, and Single-bit-GA, which ranks fragments only by SHAP importance, SF-GA's AUC was significantly higher (p<0.001 for both), and the advantage persisted as the fragment count rose to seven.

The demonstrated scope is computational: comparison with GB-GA, REINVENT, FragGT and SMILES-GA on the PMO benchmark under a fixed evaluation budget, scaffold-preserving and peripheral editing under a Bemis-Murcko scaffold constraint with Tanimoto similarity to the starting molecule of at least 0.5, and docking of 1,570 generated molecules passing NIH, Brenk and PAINS filters against two human P-gp structures, 6QEX and 7O9W. The authors state that the docking results provide structural characterization of the generated molecules but do not directly establish non-substrate behavior, and they motivate future experimental evaluation of P-gp transport.

The generated molecules are hypothetical, the non-substrate probability is estimated by an ensemble model, and the observed trends should be interpreted as model-based predictions rather than experimentally validated results. Synergy groups are derived from SHAP values of substructural features, so the method assumes the target property can be represented by explicit substructural contributions and may be less suitable for properties dominated by whole-molecule physicochemical descriptors; the guidance also depends on the reliability of those SHAP values, particularly when features are highly correlated. Dataset size and class imbalance should also be considered when applying SF-GA to new molecular properties.