Method Profile
Compare molecular-structure classifiers and class-balancing strategies against a labeled BBB reference dataset.
Explore early drug-discovery classification errors and descriptor regions needing additional validation.
MoleculeNet BBBP dataset with 1,955 compounds: 1,492 BBB-positive and 463 BBB-negative; RDKit descriptors and fingerprints, with 743 retained features.
Stratified tenfold baseline cross-validation; separate single-run resampling comparisons for Logistic Regression and Random Forest; Random Forest feature analysis.
Structure-based inputs
SMILES-derived molecular fingerprints and physicochemical descriptors; paper describes logBB = −1 as the class threshold.
Class-balancing comparisons
Unbalanced training, SMOTE, Borderline-SMOTE and the approach labeled undersampling; training-only resampling according to the Methods.
Keep positive-class recall distinct from minority-class recognition. Compare tenfold baseline averages separately from the 587-compound single-run tables; interpret feature associations within the training chemical space.
Internal model/resampling comparison and exploratory feature-importance analysis; no external scaffold or prospective compound validation.
1,955 reference compounds; single-run comparison 587 test compounds comprising 448 positive and 139 negative labels.
Classifiers, resampling strategies, majority-label baseline and descriptor-defined subgroups.
Random Forest cross-validation accuracy/F1 = 0.877/0.924. In the single-run table, no resampling versus SMOTE gave accuracy 0.903/0.906 and recall 0.989/0.987. Undersampling reduced false positives 52→42 but increased false negatives 5→20. NH/OH-count-three subgroup accuracy was 67.9%.
An analysis of 1,955 reference compounds compared molecular-descriptor classifiers and class-balancing methods. Random Forest performance changed modestly with resampling, while errors clustered in some descriptor groups; external chemotype validation and active transport were outside the study.
