Fingerprint–Descriptor Fusion Improves Scaffold-Stratified External Performance in Blood–Brain Barrier Permeability Prediction
Year: 2026
Type: article
Abstract: Background: The blood–brain barrier restricts entry of circulating compounds into the brain, so computational permeability triage is valuable in early drug discovery. Whether fingerprints and computed physicochemical properties carry complementary information, and whether any advantage survives a change in algorithm, is not settled by the single-algorithm comparisons common in this literature. Methods: Using 2039 curated molecules, we compared three representations (2048-bit circular fingerprints, ten physicochemical descriptors, and their concatenation) across two tree ensembles and a Tanimoto-kernel support vector machine under one protocol, then validated on 2201 external compounds sharing no skeleton with training data. Robustness was assessed by sweeping feature subsampling, by standardising structures, and by repeating the comparison in a second software environment. Results: Fusion improved external ranking over either single representation in every learner and subsampling setting, all intervals excluding zero. The best configuration reached an external area under the receiver operating characteristic curve of 0.782 (95% CI 0.763–0.801), falling to 0.740 on compounds whose scaffolds were absent from training. The 248-molecule internal test set could not separate most models. Conclusions: Across the learners tested, representation contributed more than algorithm choice, and small test sets cannot support model-selection claims. Reliable triage supports Sustainable Development Goals 3 and 9.
Source: BioMedInformatics
Institution: Amrita Vishwa Vidyapeetham
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $1,207
OpenAlex ID: W7236628669
