Optimizing Hepatitis C Virus Inhibitor Identification with LightGBM and Tree-structured Parzen Estimator Sampling

Abstract

Identifying potent inhibitors against the Hepatitis C Virus (HCV) is crucial due to the continuous emergence of drug-resistant strains. Traditional drug discovery methods, including high-throughput screening, are often resource-intensive and time-consuming. Machine Learning (ML) approaches, particularly Quantitative Structure-Activity Relationship modeling, have been increasingly adopted to address this. This study utilized LightGBM, an efficient gradient-boosting framework, to predict the activity of potential HCV inhibitors. Additionally, the Tree-structured Parzen Estimator (TPE) was employed for hyperparameter optimization to enhance model performance. The optimized LightGBM-TPE model outperformed other ML models, including standard LightGBM, XGBoost, Random Forest, K-Nearest Neighbors, and Support Vector Machines, achieving an accuracy of 86.27%, a precision of 85.47%, a recall of 87.50%, a specificity of 85.03%, and an F1-score of 86.47%. Feature importance analysis identified critical molecular descriptors contributing to the model's predictive power. The results underscore the potential of advanced ML techniques and robust optimization methods to accelerate drug discovery, particularly for challenging targets such as HCV.