Interpretable machine learning and graph attention network based model for predicting PAMPA permeability

Abstract

Parallel artificial membrane permeability assay (PAMPA) is widely used in the early phases of drug discovery as it is quite robust and offers high throughput. It serves as a platform for assessing the permeability and absorption of pharmaceutical compounds across lipid membranes. This study uses machine learning (Random forest or RF, Explainable boosting machine or EBM and Adaboost) and deep learning (Graph attention network or GAT) to build models to predict PAMPA permeability. A curated dataset of 5447 compounds with PAMPA permeability scores (in a scale 10⁻⁶ cm/s) was used to train and validate these models. During validation it was observed that, RF and EBM models could predict with an accuracy of 81 % and 80 % respectively, whereas with Adaboost and GAT, the accuracies were limited 76 % and 74 % respectively. Further, an external dataset was used to screen the predictive capability of these models and results showed that RF, EBM and Adaboost had quite similar accuracies with 91 %, 90 % and 89 % respectively. Interestingly, with this external dataset, the GAT-based model also reached a significant accuracy of 86 %. The overall results show that all the models in this study could well predict PAMPA permeability over the benchmark and covering diverse chemical space. All the datasets and codes for developing these models have been deposited on the GitHub platform (https://github.com/hridoy69/pampa_premeability).