A Machine Learning Framework for Predicting Protein-Protein Interactions from Sequence-Derived Physicochemical Features

Abstract

The prediction of protein-protein interactions (PPIs) from primary sequence data remains a fundamental challenge in systems biology. While experimental methods are resource-intensive, computational approaches offer a scalable alternative to map the complex human interactome. This study develops a machine learning framework to not only predict PPIs with high accuracy but also to uncover the underlying biochemical principles that govern these associations. We constructed a balanced dataset of approximately 1.8 million human protein pairs derived from the BioGRID database. For each protein, we engineered a feature set based on its Amino Acid Composition (AAC) and Dipeptide Composition (DPC) to represent its global and local physicochemical properties. An Extreme Gradient Boosting (XGBoost) classifier was trained on these features to distinguish between interacting and non-interacting pairs. The final model demonstrated strong predictive performance on a large, held-out test set, achieving an accuracy of 78.1% and an Area Under the Curve (AUC) of 0.865. To interpret the model’s logic, we employed SHAP (SHapley Additive exPlanations). The interpretability analysis revealed that the model’s predictions were overwhelmingly driven by AAC features. Specifically, the model learned that a high abundance of hydrophobic residues (e.g., Phenylalanine, Isoleucine) increased interaction likelihood, while a high abundance of polar residues (e.g., Serine) decreased it. Our work successfully validates a highly accurate and, critically, interpretable model for PPI prediction. By demonstrating that a machine learning model can independently learn fundamental principles of biophysics, such as the hydrophobic effect, from sequence data alone, we highlight the power of interpretable AI to generate new biological insights from large-scale genomic data.