EGFRAP: a predictive machine learning model for assessing small molecule activity against the epidermal growth factor receptor

Abstract

Epidermal growth factor receptor (EGFR) is a membrane-bound protein that interacts with epidermal growth factor, triggering receptor dimerization and tyrosine autophosphorylation, subsequently promoting cell proliferation. EGFR-associated pathways regulate cell housekeeping functions like growth, division, and apoptosis. However, the mutations/overexpression of EGFR cause unrestrained cell differentiation, leading to tumorigenesis. This study proposes a machine-learning-based tool, EGFR^AP, to compute novel molecules' biological activities (pIC₅₀) against EGFR. The tool is based on a robust quantitative structure–activity relationship (QSAR) model, trained on a large dataset of existing EGFR inhibitors using multiple machine learning algorithms. The extra trees regressor (ET) model showed promising results for the training dataset with an R² value of 0.99, an RMSE value of 0.07 and an MAE of 0.009. The Pearson correlation between the observed and predicted pIC₅₀ values of the training set inhibitors was also very substantial, i.e. 0.99. The model was then validated using a test dataset, and the findings were satisfactory. An R² value of 0.67, an RMSE of 0.89 and an MAE of 0.61 were detected for the test dataset, and the Pearson correlation coefficient of observed/predicted pIC₅₀ values was 0.82. The model was probed for overfitting using 10-fold cross-validation, and a series of structure-based drug design experiments were performed to validate the tool's predictions. The findings backed up the model's performance. This tool will be of significant importance to medicinal chemists in identifying promising EGFR inhibitors.