Development of QSAR Models and Web Applications for Predicting hDHFR Inhibitor Bioactivity Using Machine Learning

Human dihydrofolate reductase (hDHFR) is a crucial cellular enzyme in folate metabolic pathway, where it catalyzes the reduction of dihydrofolate into tetrahydrofolate (THF) and an important cofactor involved in DNA, RNA, protein biosynthesis and cell proliferation. Due to its importance, hDHFR has become a promising target for therapeutic development, particularly in treating cancer, bacterial infections, and autoimmune diseases. Its inhibition has found clinical value in antitumor, antimicrobial and antiprotozoal treatment; however, the emergence of resistance to existing hDHFR inhibitors necessitates the development of new and more potent compounds. In the current study, we propose a cheminformatics-based approach using machine learning to develop predictive models of hDHFR bioactivity. We used three types of molecular descriptors in the form of fingerprints, i.e., PubChem, Substructure, and MACCS, to capture structural properties associated with hDHFR inhibition. Predictive models were built using a random forest algorithm optimized through hyperparameter tuning. Feature selection was performed using Recursive Feature Elimination (RFE), and dataset dimensionality was reduced by removing outliers through Principal Component Analysis (PCA) to optimize model performance and reducing overfitting and weak predictivity. The resulting models are validated through external test sets, domain applicability analysis, and interpretation of influential molecular features via random forest feature importance selection plots and correlation matrix analysis. All three models exhibited strong predictive capabilities, with R-squared (R²) values ranging from 0.9849 to 0.9934 for the training set and 0.9381 to 0.9591 for the test set. These final predictive models were further incorporated into an accessible web application, enabling users to estimate the bioactivity of new compounds targeting hDHFR.