DynML-Net: a porcine enteric virus identification network based on protein language models and a dynamic heterogeneous multi-branch architecture

Porcine Enteric Viruses (PEVs) pose a serious threat to the global pig industry, leading to substantial economic losses. However, traditional clinical diagnostic methods are limited by high costs, long turnaround times, and difficulties in rapidly addressing challenges such as co-infections with multiple viruses and the emergence of different viral species. To overcome the limitations of existing approaches in recognizing viruses under complex clinical conditions, poor adaptability to variant strains, and the insufficient extraction of key features, this study proposes a deep learning network integrating the ESM-2 protein language model and a multi-branch heterogeneous architecture, named DynML-Net. The model adaptively integrates decision outputs representing local architectural patterns, global dependencies, and long-sequence evolutionary features through a dynamic gating mechanism, and enhances classification performance via a deep mutual learning (DML) strategy. Experimental results demonstrate that DynML-Net achieves accuracies of 94.52% and 89.16% for binary and multiclass classification tasks, respectively, outperforming current mainstream models, while exhibiting robust performance under conditions of sample imbalance, data reduction, and heterologous viral datasets scenarios. This study demonstrated the efficacy of DynML-Net in PEVs identification and classification tasks, and based on DynML-Net, a Streamlit-based locally deployable system was developed to facilitate a closed-loop workflow from sequence upload to automated identification, providing an accurate and efficient computational solution for the rapid detection of PEVs.