Leveraging quantum chemical properties in transfer learning for predicting blood-brain barrier permeability of drugs

The blood-brain barrier (BBB), crucial for central nervous system (CNS) homeostasis, poses challenges for drug delivery in CNS diseases due to selective permeability. Because of this difficulty, there are limited treatments developed for CNS diseases. As a solution, computational models can be implemented in treatment development to enable rapid screening of drug permeability, saving time and resources. This study explores machine learning, deep learning, and transfer learning models to predict the BBB permeability of drug molecules, validated through an in vitro assay known as Parallel Artificial Membrane Permeability Assay-BBB (PAMPA-BBB). Using the Blood-Brain Barrier Database (B3DB) of ∼ 8,000 compounds of known BBB permeability, classification models including support vector machines (SVMs), deep neural networks (DNNs), direct message passing neural networks (D-MPNNs), and transfer learning with quantum chemical properties were developed. Experimental validation with 18 compounds from the Emory Enriched Bioactive Library (EEBL), a library containing 1,018 FDA-approved pharmacologically active compounds of known activity, highlighted PAMPA-BBB as a robust validation method. The SVM model with combined 2D RDKit and Morgan fingerprint molecular representation achieved high performance (accuracy: 89.08%) on the B3DB test set. The best-performing models for the 18 EEBL compounds were transfer learning models. In particular, the model trained on the QM9-extended polarizability property correctly classified 17 out of 18 compounds, while the model trained on the QM9-extended dipole moment property achieved correct classification across all 18 experimental compounds. Additional analyses demonstrated that QC-based transfer learning provides complementary predictive value beyond traditional molecular descriptors such as LogP and molecular weight. QC-pretrained models achieved higher accuracy and ROC-AUC on both the B3DB and external PAMPA test sets, with performance remaining robust even after descriptor ablation. Moreover, QC-pretrained models outperformed the baseline of P-glycoprotein (P-gp) inhibition, underscoring the unique contribution of quantum-derived representations to BBB permeability prediction. Therefore, this study motivates the synergy of computational and experimental methods in enabling faster, more cost-effective, and targeted identification of CNS-active or CNS-sparing drugs.

Deep learning; Machine learning; PAMPA; Quantum chemical property; Transfer learning.