AI-Driven Virtual Screening: Machine Learning-Based Prediction of Molecular Activity and Binding Affinity for Drug Discovery

Accurately predicting molecular bioactivity and binding affinity is a cornerstone of modern drug discovery, where early-stage virtual screening significantly reduces time and cost. This study evaluates the performance of multiple machine learning classifiers—Logistic Regression, Support Vector Machine (SVM), Decision Tree, K-Nearest Neighbors (KNN), and Random Forest—in predicting compound activity based on physicochemical and interaction-based features. A comprehensive feature engineering pipeline was applied, including scaling, imputation, and mutual information analysis to identify highly predictive variables such as binding_affinity, logp_pi_interaction, and logp. Among the models, Random Forest emerged as the most effective, achieving a 99.89% accuracy, 100% precision, and 99.82% F1-score, outperforming all baseline classifiers while maintaining generalization. The confusion matrix revealed perfect classification with zero false positives and false negatives, highlighting the model's robustness. Feature importance analysis further confirmed that compound binding strength is the dominant driver of activity classification. While simpler models suffered from overfitting or underfitting, the Random Forest model effectively captured non-linear feature dependencies, making it a reliable tool for virtual screening. Future work will focus on improving interpretability, validating across external datasets, and exploring advanced neural architectures and graph-based models to scale predictive capacity in real-world drug discovery applications.