Natural-Language Processing (NLP) based feature extraction technique in Deep-Learning model to predict the Blood-Brain-Barrier permeability of molecules

Blood-Brain-Barrier (BBB) permeability is one of the critical factors in the success and failure of CNS drug development. The most accurate method of measuring BBB permeability involves clinical experiments, which are labour-intensive and time-consuming. Thus, numerous efforts were made to use artificial intelligence (AI) to predict molecules′ BBB permeability. Most of the previous models are based on calculated descriptors and molecular fingerprints. In the present work, we have developed an NLP-based feature extraction technique in Deep-Learning models to predict BBB permeability. We have used the B3DB database and generated SELFIES to extract features from the molecules. We have employed word level and N-gram tokenization to represent words into numeric vectors. The extracted features were fed into several Artificial Neural Network (ANN) and Bi-directional Long Short-Term Memory (LSTM) models. The model, ANN-10 built using ANN and 6-gram tokenization, performed best on the independent test set. The accuracy, precision, recall, F1, specificity and AUC of ROC scores were found to be 0.89, 0.91, 0.91, 0.91, 0.85 and 0.90. Thus, the developed model can be used for the early screening of CNS drugs.