Research on Drug Prediction and Optimization Model Based on Machine Learning

This paper explores the challenge of drug prediction and optimization using data-driven machine learning approaches. We developed a quantitative model for predicting biological activity through the random forest algorithm, along with a binary classification model for ADMET properties using ensemble learning techniques. These models are designed to enhance the optimization of anticancer drug candidates. First, a molecular descriptor importance ranking model was constructed using the random forest algorithm, and the top 20 descriptors with the greatest influence on biological activity were selected. Next, a quantitative model for predicting biological activity was built based on a machine learning approach. Variables showing strong correlations, as determined by the Pearson correlation coefficient, were filtered to ensure representativeness and independence. Following this, biological activity prediction models were trained using BP neural networks with a genetic algorithm, support vector machines, and random forests, focusing on the relationship between 12 molecular descriptors of 1,579 compounds and their PIC50 values. Finally, by integrating five machine learning methods— C4.5 decision tree, K-nearest neighbor, naive Bayes, random forest, and support vector machine—ensemble learning models were developed to predict ADMET properties.