Controlled evaluation of descriptor-graph fusion for blood-brain barrier permeability prediction: code, splits, predictions and figures
Year: 2026
Type: software
Abstract: Code, exact scaffold fold assignments, decontaminated external evaluation sets, per-molecule out-of-fold predictions and the manuscript figures for a controlled comparison of nine architectures for blood-brain barrier permeability prediction under a single frozen nested cross-validation protocol. The statistical analyses and figures of the associated manuscript can be recomputed from the deposited predictions and result files without retraining: the six TOST equivalence tests, the max-statistic family-wise correction over all 36 pairwise contrasts, the repeated-measures ANOVA and Tukey HSD of the method-comparison protocol recommended by Ash et al. (2025), the minimum detectable effect on each decontaminated external set, and all fifteen figures. Two inputs are not part of the deposit: the prediction files of the original submission, used only for the re-scoring reported in Section 4.2 of the manuscript and available from the author on request, and the within-B3DB label-disagreement measurement that sets the upper end of the 0.017 to 0.032 threshold envelope; the four cross-database measurements are deposited. Source datasets (B3DB; MoleculeNet BBBP) are not redistributed; the deposit contains the identifiers, labels and input checksums needed to reproduce the decontamination.
Author: Kayaoglu, Mazhar
Institution: Bingöl University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7214218802