Abstract
MOTIVATION: Proteins carry out most biological processes via interactions with other proteins, known as protein-protein interactions (PPIs). Accurately predicting PPIs is crucial for understanding protein function, yet existing methods often fall short in capturing their complex and hierarchical nature. RESULTS: We propose PF-AGCN, an adaptive graph convolutional network that leverages two distinct graph structures: a function graph representing hierarchical Gene Ontology term relationships and a protein graph modeling direct interactions between proteins. Unlike traditional graph attention networks, PF-AGCN preserves the original biological structures while dynamically learning new relationships, ensuring the retention of essential biological information. Additionally, our framework integrates a protein language model with stacked dilated causal convolutional neural networks, enabling the synergistic fusion of global sequence semantics and local structural patterns. Extensive experiments on a comprehensive protein dataset across three evaluation facets demonstrate PF-AGCN's superior prediction accuracy. AVAILABILITY AND IMPLEMENTATION: The source code is publicly available at https://github.com/smyang107/PFAGCN.