Manifold Embedding of Quantum Information as Molecule Representation to Predict Blood-Brain Barrier Permeability by Deep Learning.

Abstract

Neurological disorders continue to be a leading global health challenge, with the blood-brain barrier (BBB) presenting considerable obstacles to effective drug delivery for central nervous system (CNS) therapies. Accurately predicting BBB permeability is essential for the early stages of CNS drug design. This study utilizes Manifold Embedding of Molecular Surface (MEMS) as a quantum-informed molecule representation to improve log BB prediction using deep learning models. Employing the B3DB data set, our approach achieved competitive performance, with an average RMSE of 0.49 ± 0.06, MAE of 0.38 ± 0.05, and R² of 0.55. The ability of MEMS to authentically encode molecular interactions facilitates a more direct modeling of log BB compared to traditional descriptors. Still, as expected, model performance is influenced by the size and quality of the data, exhibiting notable variability across different B3DB groups and imbalances in the distribution of the log BB values. Additionally, although chirality significantly influences BBB permeability, the limited stereochemical data in the data set constrain its impact. Future efforts should focus on curating high-quality, stereochemically rich measurements and addressing data imbalances to train predictive models.