Background The blood brain barrier (BBB) is a selective, semi-permeable boundary in the central nervous system (CNS) that plays a critical role in maintaining the homeostasis of the brain microenvironment. While clinical experiments can accurately determine which compounds effectively cross the BBB, these are time-consuming and labor-intensive. Machine Learning (ML) and computational techniques can alleviate this burden by rapidly screening a large set of potential drug candidates for those that should be prioritized for experimentation validation. In this study, we use ML to predict BBB permeability and then use Parallel Artificial Membrane Permeability Assays (PAMPA) as the in vitro method to assess the accuracy and reliability of these predictions.

Methods The dataset used is B3DB, one of the largest public BBB datasets consisting of seven thousand compounds compiled from 50 published resources. We trained and validated models on this dataset with methods including support vector machines (SVMs), deep neural networks (DNNs), and graph convolutional neural networks (GCNNs). The transfer learning model consisted of an initial DNN trained to the task of quantum chemical property before retraining to the task of BBB permeability. After using these models to screen the Emory Enriched Bioactive Library, a test dataset of two thousand compounds, the top 50 compounds will be prioritized for in vitro experimental confirmation with PAMPA permeability assays. These assays consist of an artificial lipid membrane that simulates BBB properties. Samples will be collected from both sides at specific intervals and analyzed spectrophotometrically post-incubation.

Results For fair comparison, the same 75/25 train/test split of the B3DB dataset was used consistently across the various models. The DNN had a similar performance to the SVM with an accuracy of 83.09% and 82.33% respectively. For transfer learning, the most predictive model was of the DNN with target domain of polarizability of 76.89%. The results indicate that the GCNN was the best performing model of 87.14%.

Conclusion Currently, GCCN’s demonstrate the best predictive capability, highlighting their ability to learn relevant molecular features to the predictive task in the message passing phase. After then comparing these predictions with the in vitro experiments, the next step is to conduct in vivo validation using small and large animal models. The intent is for this hybrid computational and experimental approach to accelerate the understanding of drug therapeutic effects on neuro-oncology and provide insights into key chemical properties that will guide the design of future compounds.