A deep understanding of Protein-protein interactions (PPIs) can provide comprehensive insights into many biological functions, thereby facilitating drug target identification and novel therapeutic design. Recent developments in artificial intelligence (AI)-driven computational methods have enabled the discovery of previously uncharacterized PPIs from large-scale interactome datasets. Almost all existing machine learning methods rely on Subcellular Localization (SL) to construct balanced datasets based on positive interactions to achieve predictions. Despite high fitting accuracy, the generalization ability of these models is questionable. To solve this problem, we analyzed existing methods and found that the high false positives in these methods are due to the bias in data distribution caused by SL. Therefore, we proposed a new strategy for negative instance sampling in PPI prediction and developed a Hybrid Graph Neural Network framework for Protein-protein Interaction Prediction (HGNNPIP). The experimental results showed that HGNNPIP works well on six benchmark datasets. Comparison analysis demonstrated that our model outperformed the other four existing methods. We also used HGNNPIP to explore the molecular contacts involved in the rice-pathogen interaction system. In vivo experiments confirmed multiple regulations related to disease resistance in rice. In summary, this study provides new insights into establishing a computational framework for PPI prediction with high reliability.

A deep understanding of Protein-protein interactions (PPIs) can provide comprehensive insights into many biological functions, thereby facilitating drug target identification and novel therapeutic design. Recent developments in artificial intelligence (AI)-driven computational methods have enabled the discovery of previously uncharacterized PPIs from large-scale interactome datasets. Almost all existing machine learning methods rely on Subcellular Localization (SL) to construct balanced datasets based on positive interactions to achieve predictions. Despite high fitting accuracy, the generalization ability of these models is questionable. To solve this problem, we analyzed existing methods and found that the high false positives in these methods are due to the bias in data distribution caused by SL. Therefore, we proposed a new strategy for negative instance sampling in PPI prediction and developed a Hybrid Graph Neural Network framework for Protein-protein Interaction Prediction (HGNNPIP). The experimental results showed that HGNNPIP works well on six benchmark datasets. Comparison analysis demonstrated that our model outperformed the other four existing methods. We also used HGNNPIP to explore the molecular contacts involved in the rice-pathogen interaction system. In vivo experiments confirmed multiple regulations related to disease resistance in rice. In summary, this study provides new insights into establishing a computational framework for PPI prediction with high reliability.