Multilabel prediction of virus target proteins via multimodal graph representation learning

Area of Science:

• Computational biology
• Virology
• Bioinformatics

Background:

• Identifying virus target proteins (VTPs) is essential for understanding viral pathogenesis.
• Current computational methods often treat VTP prediction as a single-label problem, overlooking proteins involved in multiple infections or relying solely on intrinsic host protein data.

Purpose of the Study:

• To develop a multilabel framework, MultiVTP, for predicting VTPs.
• To leverage graph learning and multimodal information for enhanced VTP prediction accuracy.

Main Methods:

• MultiVTP employs graph learning by sampling subgraphs around query proteins to capture topological features.
• Multimodal features are extracted to represent proteins from diverse perspectives.
• A graph transformer integrates these features, followed by a progressive layered extraction module for predicting VTPs.

Main Results:

• Ablation experiments confirmed that graph-based attributes and modules are critical for MultiVTP's performance.
• The framework demonstrated superior performance compared to baseline models.
• MultiVTP showed robustness even with limited training data.

Conclusions:

• MultiVTP offers a powerful multilabel approach for VTP prediction, outperforming existing methods.
• The framework can systematically identify novel VTPs for individual and multiple viruses by analyzing the human proteome.