ViralMultiNet: A structure-aware multimodal framework for viral protein function prediction in wastewater surveillance

Accurate functional annotation of viral proteins is essential for genomic surveillance, yet rapid viral evolution causes "functional drift" that challenges conventional sequence-only models. These models often lack interpretability and struggle with fragmented sequences from complex environmental samples such as wastewater. We developed ViralMultiNet, a structure-aware multimodal framework that integrates multi-scale k-mer encodings (4-7-mers) with functional semantic embeddings derived from UniProt annotations. Using a curated Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) dataset of 66,011 samples from wastewater metagenomics (NCBI SRA: SRX28474964), we implemented gated multimodal fusion and triple knowledge distillation to transfer structural insights from a teacher to a student model. Model performance was evaluated via 5-fold cross-validation and external validation on emerging variants. Training efficiency was optimized using Low-Rank Adaptation and Flash Attention. ViralMultiNet achieved robust classification performance with a macro F1 score of 0.921 ± 0.004, accuracy of 0.928 ± 0.003, and AUC of 0.983 in cross-validation. The distilled student model matched teacher performance within a negligible margin (<0.003 F1 difference) while reducing training time by 40.4% (from 94.3 to 56.2 minutes per epoch). Interpretability analysis revealed that model attention peaks consistently aligned with experimentally validated functional domains of the SARS-CoV-2 Spike protein, including the receptor-binding domain (residues 319-541), S1/S2 cleavage site (681-685), and fusion peptide (816-835). ViralMultiNet offers a scalable, interpretable solution for viral protein function prediction. Its ability to generalize across variants and map attention to critical biological regions supports deployment in wastewater-based early warning systems, enhancing global pandemic preparedness.