bioRxiv2025
MotivationAccurate prediction of biomedical relationships, such as chemical-gene interactions, is fundamental to understanding disease mechanisms and advancing drug discovery. With the rapid growth of heterogeneous biological data, modeling large-scale, multi-entity networks has become increasingly challenging. Traditional approaches, including homogeneous GNNs (e.g., GCN, GAT) and meta-path-based random walks, struggle to efficiently capture high-order, diverse neighborhood information in complex biomedical graphs. To address these limitations, we propose a novel multi-view heterogeneous graph attention network (GAT)-based architecture that effectively aggregates rich, heterogeneous interactions across multiple biomedical entity types. The proposed encoder captures comprehensive structural and semantic information while remaining computationally efficient. Through optimized aggregation strategies and multi-processing, the model generates high-quality node embeddings with significantly reduced training time. For relation prediction, multiple decoder architectures were evaluated, with a multilayer perceptron (MLP) identified as the most effective for accurate multi-type relation classification. The resulting network comprises 124,604 unique nodes and 48,482,286 interactions. ResultsExperimental results show that the proposed model consistently outperforms state-of-the-art methods, including CGINet, Node2Vec, and the GCN-based BioNet, achieving an AUROC of 0.91 for chemical-gene interaction prediction. The model further explores its ability to identify top-ranking chemical-gene interactions in cancer and to predict gene-phytochemical relationships. Overall, this work introduces a scalable and powerful framework for biomedical relation prediction, with strong potential applications in drug screening and disease mechanism discovery. Key PointsO_LIWe constructed a large-scale heterogeneous biological interaction network by integrating curated datasets across multiple entity types, including chemicals, genes, pathways, and diseases. C_LIO_LIWe propose a novel graph neural network framework, Optimized Multi-View Network Integration (OMNI), based on an encoder-decoder architecture, which employs a multi-view heterogeneous Graph Attention Network (GAT) to learn entity embeddings from subgraphs and a multilayer perceptron (MLP) decoder to predict chemical-gene interactions (CGIs). C_LIO_LIWe integrated a PyTorch Lightning based parallel training strategy to scale up the learning process, significantly enhancing the models ability to efficiently handle large-scale heterogeneous data. C_LIO_LIWe demonstrated the applicability of OMI by evaluating cancer-related chemical-gene interactions and vitamin D receptor (VDR)-phytochemical interactions, including the prediction of interaction types. C_LI