Gated Graph Attention Networks with Multichannel Fusion for Disease Comorbidity Prediction

Xihan Qin, Li Liao

Keywords: Multichannel fusion, Gating mechanisms, Graph attention networks, Graph neural networks, Disease comorbidity, Human interactome, Protein–protein interaction networks

TL;DR: We propose Gated Graph Attention Networks with multichannel fusion to integrate connectivity and disease association information in protein interaction networks, achieving state-of-the-art performance in disease comorbidity prediction.

Abstract: The co-occurrence of multiple diseases, or comorbidity, significantly complicates clinical management and worsens patient outcomes. Comorbidity is believed to arise from genetic mutations functionally connected through protein-protein interactions (PPIs) within the human interactome. Unraveling these intricate PPI networks is essential for understanding disease progression and addressing the challenges posed by comorbid conditions. In this study, we propose a novel Gated Graph Attention Network (GGAT) framework tailored for disease comorbidity prediction by addressing issues that hinder the existing methods via three key aspects: (1) applying attention over local neighbors rather than global pairwise attention among all protein nodes, enabling more biologically meaningful aggregation; (2) incorporating a gating mechanism to adaptively regulate information flow and enhance representation learning for comorbidity prediction; and (3) introducing a multichannel fusion strategy that integrates connectivity based and disease association based embeddings, both of which have been shown to be important for disease comorbidity prediction. Experimental results on the benchmark dataset demonstrate that GGAT significantly outperforms the Transformer baselines across all metrics (AUROC, AUPRC, accuracy, and MCC), with the multichannel gated fusion variant achieving the best overall performance. These findings highlight the importance of integrating complementary biological features through graph structure and indicate that the proposed GGAT provides a generalizable graph learning framework applicable beyond disease comorbidity prediction.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 12597