Method Profile
Standardize and deduplicate experimental chemical records, combine seven classifiers, and retain disagreement-based inconclusive calls alongside a separate LogBB regression.
Prioritize neurological assay hits for brain-exposure testing while reporting prediction coverage and uncertainty.
B3DB-derived classification set of 3,968 compounds and numerical LogBB subset of 917; seven-model consensus and graph-neural-network regression.
80:20 scaffold split and fivefold cross-validation; class-balanced training; majority voting; CONS-STD below 0.46 defines accepted consensus calls.
Data and representations
Published B3DB experimental labels; standardized two-dimensional chemical structures, two structural-fingerprint families and molecular-property descriptors.
Algorithms
Random forest and gradient boosting across three descriptor sets, plus a Chemprop atom/bond graph model; predictions combined by voting.
Keep scaffold-based testing separate from ordinary cross-validation, and report the accuracy–coverage trade-off when low-agreement predictions are withheld. Classification and continuous-ratio regression have distinct performance.
Evaluate chemically separated test scaffolds, then apply the classifier to existing dopamine D2 receptor screening hits.
3,968 classification compounds; 917 with LogBB; 6,031 curated receptor-screen hits, including 52 already present in the modeling dataset.
Unbalanced versus balanced training; all test compounds versus accepted consensus calls; individual versus consensus regression.
Scaffold-test balanced accuracy 0.81/ROC area 0.89. Excluding 97 inconclusive test compounds raised sensitivity/specificity to 0.89/0.79. Best LogBB regression R² 0.48; consensus regression did not improve it. Application yielded 3,307 permeable, 1,644 nonpermeable and 1,080 inconclusive calls.
A curated chemical dataset supported a consensus blood–brain barrier classifier with balanced accuracy of 0.81 on a scaffold-based test. Predicting continuous brain-to-blood concentration ratios was harder, and a dopamine-receptor screening application left 1,080 of 6,031 compounds inconclusive.
