Graph Neural Networks Versus Descriptor-Based Machine Learning for Blood–Brain Barrier Permeability Prediction: A Matched-Split Evaluation

Background/Objectives: Predicting blood–brain barrier (BBB) permeability from structure remains challenging. We evaluated whether graph neural networks (GNNs) improve experimental logBB prediction over descriptor-based machine learning and whether GNN attention agrees with earlier descriptor-based findings. Methods: Graph Convolutional Network (GCN) and Graph Attention Network (GAT) models were trained on 1058 B3DB compounds and compared with LightGBM using 209 RDKit 2D descriptors under identical random and Murcko scaffold splits. Analyses comprised a primary split, 10 repeated random and scaffold splits with nested hyperparameter optimization and paired tests, descriptor-augmented GNNs, classification, external evaluation on the Martins benchmark, and GAT attention analysis. Results and Discussion: In the primary split, LightGBM achieved the highest R2 (random 0.673; scaffold 0.604) versus GCN (0.514; 0.133) and GAT (0.430; 0.310). Across repeated splits with nested hyperparameter optimization, LightGBM outperformed both GNNs under both protocols (paired t-test p < 0.001). GCN and GAT did not differ significantly under scaffold splitting in either the original repeated-split evaluation (p = 0.49) or the optimized comparison (p = 0.184). Descriptor augmentation improved both GNNs (p < 0.001); under scaffold splitting, augmented GCN (R2 = 0.463 ± 0.084) was not significantly different from LightGBM (0.494 ± 0.093; p = 0.22). Using thresholded regression predictions, LightGBM had the highest mean AUC in every condition (≈0.87 versus 0.77–0.84). On 1627 compounds from the independently curated Martins benchmark, mean AUC was 0.84 (LightGBM), 0.83 (GCN) and 0.81 (GAT). GAT attention was higher on nitrogen/oxygen than carbon in 82 of 98 eligible molecules, but halogen, sulfur, and phosphorus atoms received even more attention. Conclusions: Under the evaluated conditions, 2D descriptors outperformed basic GCN and GAT models across splits; descriptor augmentation substantially narrowed the gap. GAT attention provides hypothesis-generating, not causal, insight.