HV-MultiScaleGNN: A Multi-Scale Graph Neural Network Framework for Host-Virus Protein-Protein Interactions Prediction

HV-MultiScaleGNN predicts host–virus protein-protein interactions from heterogeneous network topology, with the Local view achieving AUROC = 0.9444 and AUPRC = 0.9422 on large-scale HV-PPIs datasets.

• Why it matters: Experimental HV-PPIs discovery is costly and slow, and many computational methods require incomplete sequence, structure, or interaction features that are not always available.
• What they did: The method treats HV-PPIs as heterogeneous link prediction by integrating the human interactome with experimentally validated HV-PPIs, testing Local, Global, and Local+Global graph representations using discrimination, calibration, and ranking metrics.
• The result: Local+Global yields consistently strong performance (AUROC = 0.9378, AUPRC = 0.9356), while Global view has the lowest ECE (0.0379), enabling topology-only relation-specific HV-PPIs prediction.

AI summary of the abstract · may contain errors · not medical advice · read the paper.