Mycobacterium tuberculosis cell wall has complexity and unusual organization. These conditions make the nutrients and antibiotics difficult to penetrate this wall which affects the low activity of several antimycobacterial drugs in mycobacteria cells. Based on this information, the cell wall permeability prediction in some compounds becomes important and would help develop novel antitubercular drugs. Recently, there have been many predictions helped by computational technology using the Simplified Molecular Input Line Entry System (SMILES) input drug compounds. In this study, we applied computational technology to predict the permeability of cell walls to some compounds or drugs. We evaluated several common machine learning models for their ability to predict cell wall permeability. However, none of these models achieved satisfactory performance. We investigated a Graph with Attention Neural Network (GATNN) model to address this challenge. In the case of permeability detection, to the best of our knowledge, the GATNN model is considered a new approach to improve the prediction performance of the penetration ability of some compounds to the cell wall of the mycobacterial. Additionally, we optimized the accuracy value to get the best hyperparameter and the best model by Optuna. After getting the optimal model, by using the benchmark dataset, this model has slightly increased the performance over the previous model in accuracy and specificity to 78.9% and 81.5%. As a complementary, we also provided an ensemble model and generated the interpretability of the model. The code and materials of all experiments in this paper can be accessed freely at this link: https://github.com/asw1982/MTbPrediction.
Read full abstract