VIRTUAL SCREENING OF CANDIDATE DRUG COMPOUNDS USING MOLECULAR DESCRIPTORS, PROTEIN PHYSICOCHEMICAL PROPERTIES, BINDING AFFINITY, AND MACHINE LEARNING-BASED ACTIVITY PREDICTION

Abstract

Virtual screening of candidate drug compounds is an important computational strategy for accelerating early-stage drug discovery by identifying compounds with predicted biological activity before experimental validation. This study evaluated candidate drug compounds using molecular descriptors, protein physicochemical properties, binding affinity, and machine learning-based activity prediction. The dataset contained 2,000 compound–protein interaction records and 17 variables, including molecular weight, LogP, hydrogen bond donors and acceptors, rotatable bonds, polar surface
area, protein length, protein isoelectric point, hydrophobicity, binding site size, engineered interaction features, binding affinity, and binary activity status. Data preprocessing involved missing value treatment, duplicate assessment, feature standardization, and stratified data splitting. Exploratory analysis showed that active and inactive compounds differed mainly in binding affinity, LogP, protein pI, and LogP–pI interaction. Multiple supervised learning models were developed, including Logistic Regression, Random Forest, Support Vector Machine, Gradient Boosting, and k-Nearest Neighbors. Random Forest and Gradient Boosting produced the strongest classification performance, while feature
importance analysis identified binding affinity as the dominant predictor, followed by LogP–pI interaction and LogP. The findings indicate that integrated molecular, protein, and interaction-based descriptors can support accurate activity prediction and candidate prioritization, although external validation remains necessary before biological interpretation.