The study asks when an explicit but fallible scientific rule should be trusted, corrected, or treated cautiously under molecular chemical-space shift.
AGRR combines graph and descriptor information with an explicit rule expert. The final prediction uses molecule-specific rule reliance rather than enforcing the rule globally. The scientific-rule pathway, applicability signal, learned trust/reliance, learned molecular expert, ensemble uncertainty, and independent rule-failure score are retained as separately auditable quantities.
- ESOL: regression rule-guided task.
- BBBP: classification task using a Clark-derived passive-diffusion prior rather than treating the Clark relationship as a ground-truth binary mechanism.
- Lipophilicity and FreeSolv: sensitivity/no-rule tasks.
- B3DB: external BBB transfer corpus after exact standardized-SMILES de-overlap against BBBP.
Primary internal evaluation uses:
- random splits;
- Bemis-Murcko scaffold splits;
- Butina similarity-cluster splits;
- five frozen seeds: 20260711, 20260719, 20260727, 20260804, 20260812.
The primary ESOL/BBBP benchmark uses all five seeds and all three split families. The FreeSolv/Lipophilicity no-rule sensitivity analysis uses the first three frozen seeds and random/scaffold splits only. B3DB is evaluated as a separate external transfer experiment after exact standardized-SMILES de-overlap against BBBP.
Molecular auditing and duplicate/conflict handling occur before splitting. Training-only transformations and fitting are isolated from test labels. Validation data are used for model selection/calibration where applicable. Test/external labels are not used to tune the reported predictors.
The study includes fixed scientific rules/priors, descriptor/fingerprint classical models, graph/descriptor learned models, unrestricted or ungated correction variants, AGRR, and rule/gate control variants where supported by the frozen package.
A central interpretation constraint is that the primary AGRR predictor is a three-member ensemble while multiple comparator/control models are single networks. Therefore, differences between those rows cannot be attributed uniquely to the applicability-gating mechanism.
The frozen evaluation includes:
- predictive metrics under multiple split families;
- calibration and ensemble uncertainty;
- selective prediction/risk-coverage analyses;
- rule reliance/applicability diagnostics;
- a separately trained rule-failure score;
- ablation/control analysis;
- external B3DB transfer;
- explanation-repeatability sensitivity analysis.
Mixed findings are retained. The study does not claim universal model superiority, does not attribute all gains to gating, does not present the ESOL equation as an independent external prior, and does not make a deployment-safety guarantee.