Heterogeneous fusion of statistical sequence features and ESM2 representations improves protein-protein interaction prediction.

Abstract

Protein-protein interactions (PPIs) are central to biological regulation, and accurate computational PPI prediction is important for systems biology and drug discovery. Although recent deep learning approaches have improved prediction performance, many require complex architectures and high computational cost. This study proposes HeteroFuse-PPI, a feature-driven framework that integrates handcrafted statistical sequence descriptors with contextual embeddings generated by the pretrained ESM2 protein language model. The statistical descriptors capture amino acid composition, physicochemical properties, sequence-order information, residue distribution patterns, and higher-order dependencies, while ESM2 embeddings encode deep contextual and evolutionary information. Each feature block is independently normalized and optimized using dimensionality reduction techniques, including PCA, TruncatedSVD, UMAP, and Kernel PCA (KPCA). KPCA provides the most consistent improvement and is used for the final hierarchical fusion strategy, in which optimized statistical and contextual representations of both proteins are combined and classified using LightGBM. Experiments on Saccharomyces cerevisiae, human, and Helicobacter pylori benchmark datasets show strong predictive performance, with AUC values of 99.34%, 99.54%, and 95.17%, respectively. The results demonstrate that combining biologically informative handcrafted descriptors with pretrained protein language model representations can improve PPI prediction while reducing dependence on computationally expensive deep architectures.