Fingerprint–Descriptor Fusion Improves Scaffold-Stratified External Performance in Blood–Brain Barrier Permeability Prediction

Analysis code, partition assignments, held-out predicted probabilities and derivedresults for a factorial study of molecular representation and learning algorithm inblood–brain barrier permeability prediction. Three representations (2,048-bit Morgan fingerprints, ten physicochemical descriptors,and their 2,058-dimensional concatenation) are crossed with two tree ensembles under asingle protocol with separate training, validation and test partitions, and validatedexternally on 2,201 B3DB compounds sharing no Bemis–Murcko scaffold with the trainingdata. Robustness checks added at peer review cover a feature-subsampling sweep, aTanimoto-kernel support vector machine, a directed message-passing network, structurestandardisation, target-population recalibration, block-level SHAP attribution, andreproduction in a second software environment. Includes a pinned requirements file. Scripts 00–10 reproduce the primary analysis;11–22 reproduce the analyses added at revision.