Multilabel prediction of virus target proteins via multimodal graph representation learning

Identification of virus target proteins (VTPs) is crucial for understanding viral pathogenesis. Existing computational studies have addressed this issue by predicting host-virus protein interactions, typically framed as a single-label problem. However, targets can be identified using only intrinsic information of host proteins. Moreover, a host protein may participate in the infection processes of multiple viruses, a scenario that can be treated as a multilabel prediction problem. Herein, we present MultiVTP, a multilabel framework for VTP prediction that employs graph learning with multimodal information. This algorithm samples subgraphs centered on query proteins to capture topological properties, while multimodal features are extracted to represent proteins from complementary perspectives. A graph transformer integrates and upgrades these attributes, followed by a progressive layered extraction module that captures both shared and virus-specific binding patterns to predict VTPs. Ablation experiments reveal that graph-based attributes and modules are the key contributors to performance, with additional components leading to further improvements in accuracy. Comprehensive evaluations demonstrate that MultiVTP not only surpasses various baseline models but also remains robust under limited training data. Applying our approach to the human proteome enables the systematic identification of novel VTPs for both individual and multiple viruses.