A Local–Global Graph KAN for Multi-Class Prediction of PPI

Abstract

Traditional experimental methods for identifying protein–protein interactions (PPI) are expensive and time-consuming. Therefore, using machine learning to treat multiple PPI predictions as binary classifications has become an alternative, but there is a problem of data imbalance. The proposed GLGKAN-PPI method integrates features from both global graphs and local subgraphs to capture the complex structural information of PPI networks comprehensively. Specifically, the method utilizes the pre-trained model MASSA to extract multimodal features of proteins. The global graph features are extracted using the GKAN (Graph Kolmogorov-Arnold Network) algorithm. Meanwhile, the local subgraph features are extracted using the MOE-GKAN (Mixture of Experts-Graph Kolmogorov-Arnold Network) algorithm. To mitigate data imbalance, an asymmetric loss function is utilized to better handle minority classes and improve overall prediction accuracy. Experimental results demonstrate that GLGKAN-PPI outperforms a range of existing intelligent approaches across multiple datasets and partitioning strategies.