Predicting Blood-Brain Barrier Permeability Through Multi-view Graph Neural Network with Global-Attention and Pre-trained Transformer

Assessing the blood-brain barrier (BBB) permeability of drug candidates is crucial in early-stage drug discovery and development, particularly for optimizing CNS drug efficacy. Since traditional experimental approaches are expensive and time-consuming, some computational methods based on Graph Neural Networks (GNNs) have emerged for predicting BBB permeability of small molecule compounds. While these methods have demonstrated considerable success in predicting BBB permeability, the predictive performance still needs improvement. Furthermore, they may only focus on either local details or global overall information, without considering both together. To further improve the prediction performance of BBB permeability, we propose multi-view GNN with global-attention. Firstly, a small molecule compound is represented as a graph structure by treating atoms as nodes, and bonds as edges. Then, the chemical properties of atoms and bonds initialize the features of nodes and edges, respectively. Additionally, the atomic features extracted from SMILES by pre-trained Transformer and the Positional Encoding (PE) of nodes served as additional node features. Next, a multi-view GNN including atom-central and bond-central views is employed to learn molecule representations, which considers both atoms and bonds equally important. Furthermore, a global attention module is introduced to enable iterative local and global interactions. After that, two graph embeddings are produced by a shared self-attention readout. Finally, two graph embeddings are concatenated with molecular descriptors, and fed into two Multi-Layer Perceptions (MLPs) to make predictions, respectively. The final prediction is obtained by averaging these two individual predictions. To assess the effectiveness of our method, hold-out validation and 5-fold cross-validation (5-CV) experiments are conducted on BBBP dataset, respectively. Additionally, we compare it with other methods. The experiment results show that our method is superior to other compared methods and can effectively predict the BBB permeability of small molecule compounds.