DeepBBB: A Data-Composition-Aware Graph Screening Workflow for BBB-Focused CNS Library Construction and Prospective...

Background/Objectives: Blood-brain barrier (BBB) permeability is a major practical obstacle in central nervous system (CNS) drug discovery, because only a small fraction of drug-like molecules achieve sufficient brain exposure. Methods: We present DeepBBB, a graph-based, data-composition-aware screening workflow for predicting BBB permeability and for constructing BBB-focused screening libraries from commercial chemical space. Rather than introducing a new graph-learning architecture, the workflow combines standard graph convolutional and graph-transformer models with deliberate control of training-set composition, commercial-library filtering, chemical-space profiling, and prospective experimental evaluation. Three model variants were trained on the Blood-Brain Barrier Database (B3DB): a baseline classifier/regressor pair (DeepBBB_V1_BC/RG), a variant trained with a more strongly negative-enriched configuration (DeepBBB_V2_BC), and a graph-transformer counterpart (DeepBBB_trans_BC/RG). Because the sample-level split assignments and per-compound predictions from the original runs were not recoverable, the archived summary metrics are reported descriptively in the main text and are not used to support calibration, scaffold-level validity, or generalization. Applying the workflow to the ChemDiv collection (~1.5 million compounds) and the Enamine REAL lead-like space (~1.7 billion compounds) produced three progressively more stringently filtered BBB-focused libraries (21,991; 4,808,885; and 151,790 compounds). Results: Analysis of available processed data indicated that the predicted BBB-permeable set occupies a compact, BBB-compatible property region. Physicochemical, fragment, and scaffold summaries were interpreted descriptively at the constructed-library level. In a first prospective campaign, one of 12 tested candidate compounds was PAMPA-BBB-positive (all-tested molecular-level positive fraction 8.3%; exact 95% CI 0.2-38.5%). In a second campaign, five of 35 tested candidate compounds were PAMPA-BBB-positive (14.3%; exact 95% CI 4.8-30.3%); 13 compounds were not quantifiable and were not treated as ordinary CNS-negative measurements, and the two campaigns differed in compound source, selection strategy, and assay setting, so the numerical difference is reported descriptively rather than causally. Conclusions: Together, these results support the feasibility of BBB-focused computational filtering and a PAMPA-BBB evaluation workflow for CNS-oriented discovery.