Augmented Chemical Language Meets Descriptor Space: A Hybrid Deep-learning Pipeline for Predicting Blood-brain Barrier Penetration ofDrug-like Molecules

INTRODUCTION/OBJECTIVE: The brain has a semipermeable interface, the blood-brain barrier (BBB), which protects the brain but greatly limits the absorption of most small-molecule drugs, with only ~ 24% able to cross. METHODS: This study involved 5,412 drug-like molecules from three public BBB datasets and calculated 200 physicochemical descriptors with RDKit and Mordred. With ChemBERTa-2, SMILES strings were tokenized using the ChemBERTa-2 tokenizer and embedded into a 768-dimensional chemical language space. A label-aware SMILES mask-and-replace augmentation expanded the corpus fivefold while retaining chemical validity. Using Bemis-Murcko scaffolds and 5-fold cross-validation on a 4-layer MLP, descriptor and embedding vectors were combined into a 968-dimensional hybrid representation. Statistical reliability was assessed using 2,000-sample bootstrap confidence intervals. RESULTS: The hybrid model (accuracy, 95%) and an AUROC of 0.96 showed an improvement of 7-11% above the descriptors-only (accuracy 0.88; AUROC 0.90) or embedding-only (accuracy 0.91; AUROC 0.93) baselines. Statistically significant (p < 0.001) gains were supported by saliency analyses of descriptors and embeddings, which made unique and orthogonal contributions. DISCUSSION: The findings indicate that integrating traditional descriptors and chemical-language embeddings yields two distinct types of information, enabling strong, understandable predictions. Augmentation strategies increased performance more quickly without causing syntactic invalidity. CONCLUSION: A hybrid AI pipeline shows much better BBB penetration forecasting than single- baseline unimodal baselines. This work facilitates reproducibility by releasing cleaned data, trained weights, and open-source code, and by providing a viable resource for ongoing CNS drug discovery.