Abstract

This study investigates the application of machine learning models to classify biological interactions and macromolecular structural features, leveraging two distinct datasets: the BioGRID Interaction Dataset and the RCSB PDB Macromolecular Structure Dataset. Four machine learning models, Random Forest, XGBoost, Support Vector Machines (SVM), and Deep Learning, are evaluated for their ability to predict high-throughput interactions and solvent content in macromolecular structures. Key findings show that Random Forest and XGBoost outperform SVM and Deep Learning in both accuracy and interpretability. Specifically, XGBoost was optimized to prioritize recall, achieving a recall rate of 99.98% for highthroughput interaction detection. Random Forest demonstrated high precision, making it ideal for scenarios requiring accurate identification of positive cases. Both models achieved high F1 scores of 96%, indicating a well-balanced performance between precision and recall. Through hyperparameter tuning and threshold adjustment, we were able to enhance XGBoost’s sensitivity to positive cases, highlighting the importance of optimizing models for specific application needs in bioinformatics. The study also identifies critical features, such as Percent Solvent Content and Matthews Coefficient, as key determinants for classification. This research fills a gap in the use of machine learning for bioinformatics by providing a detailed comparison of widely-used models, identifying key factors influencing classification tasks, and demonstrating how model adjustments can improve predictive accuracy. The findings contribute to more effective data-driven approaches in understanding biological interactions and macromolecular structure analysis, with potential applications in drug discovery, molecular biology, and structural bioinformatics.

Keywords:Machine learning; Bioinformatics; High-throughput interactions; Macromolecular structures; Model optimization