Blood–Brain Barrier Permeability Prediction by Integrating Deep Learning Representations with Chemical Compound Knowledge

Abstract

Predicting blood–brain barrier (BBB) permeability is essential for central nervous system drug discovery, but traditional descriptor-based methods of predicting BBB permeability lack structural context, and deep learning-based methods struggle with overfitting. We introduce a multimodel framework that integrates both transformer-derived molecular embeddings with physicochemical descriptors using a late-stage ensemble fusion strategy. Trained on both the B3DB and NeuTox data sets and evaluated on the benchmark BBBP, our model showed strong predictive performance, with data set-dependent performance. Feature fusion improved performance relative to embedding-only models in several settings, whereas descriptor-based models achieved comparable or higher performance in some settings. These results suggest that combining learned structural embeddings with chemical descriptors provides complementary predictive information, while the benefit of their integration may depend on the data distribution or classifier used.