Augmented Chemical Language Meets Descriptor Space: A Hybrid Deep-learning Pipeline for Predicting Blood-brain Barrier Penetration ofDrug-like Molecules
Randomized trial shows improved predictions of blood-brain barrier penetration in drug-like molecules, indicating potential for better CNS drug discovery.
Key Points
This research aims to enhance the prediction of blood-brain barrier penetration for drug-like molecules.
Analyzed 5,412 drug-like molecules from BBB datasets using 200 physicochemical descriptors.
Employed ChemBERTa-2 for tokenization and embedding into a chemical language space, with augmentation to expand the dataset.
Combined descriptors and embeddings in a hybrid model using a 4-layer MLP with 5-fold cross-validation.
The hybrid model achieved an accuracy of 95% and an AUROC of 0.96, outperforming unimodal baselines (accuracy 0.88 and 0.91).
Statistically significant improvements (p < 0.001) were noted, with unique contributions from descriptors and embeddings.