FDS-CAP: Modeling Fragmented Disease Subgraphs with Component-Level Attention for Comorbidity Prediction

Abstract

Understanding comorbidity between human diseases is essential for uncovering shared pathophysiological mechanisms and improving diagnostic and therapeutic strategies. Although prior studies have investigated genetic and network-based disease associations, they often overlook the fragmented nature of disease modules within the human interactome. To address this limitation, we introduce Fragmented Disease Subgraphs with Component-Level Attention for Comorbidity Prediction (FDS-CAP), a novel graph-based deep learning framework. FDS-CAP first embeds fragmented disease subgraphs using Subgraph Neural Networks (SUBGNN) with component-level attention, then applies a variational comorbidity predictor built upon a Variational Graph Auto-Encoder that is used to predict comorbid disease associations within the Human Disease Network. SUBGNN encodes disease subgraphs by propagating information at the connected component level across three property-aware channels—capturing positional, neighborhood, and structural roles—and integrates a component-level attention mechanism that weighs each connected component based on its significance to the overall subgraph representation. A core contribution of our method is the attention-based aggregation of connected component embeddings, enabling more accurate and expressive disease representations that reflect the biological complexity in fragmented disease subgraphs for improved comorbidity prediction. FDS-CAP achieves state-of-the-art performance for comorbidity prediction on a benchmark dataset, with an AUROC of 0.966. We further illustrate its biological interpretability through a single representative case study on glioma, showing that attention-weighted subgraph components capture meaningful patterns associated with disease mechanisms.