<p>With the rapid growth of biomedical research literature, automatic knowledge extraction and mining have become increasingly important. In the field of biomedical text mining, Named Entity Recognition (NER) and Relationship Extraction (RE) are essential tasks. However, challenges such as the lack of viral protein text recognition, limited corpora, and poor annotation quality remain significant obstacles. To address these challenges, this study aims to establish a named entity recognition corpus specifically for viral protein entity recognition. We propose a BERT-SPAN model, enhanced with adaptive modularization, to identify SARS-CoV-2 protein entities. The BERT model processes input text, generating high-dimensional feature representations for each token. To further improve entity recognition, we introduce an AM-FFN (adaptive modular feedforward neural network) with an attention mechanism. Each module can independently optimize the network structure and adaptively tune parameters based on task requirements. The SPAN model is then employed to predict the starting and ending positions of entities within the text. Experimental results on a viral protein entity dataset demonstrate that the BER-AM-FFN-SPAN model achieves an F1 score of 93.66%, surpassing the BERT-SPAN model (F1 score of 91.85%) and the BERT-BiLSTM-CRF model (F1 score of 80.81%). These results not only provide a solid foundation for selecting optimal virus protein NER models but also offer valuable insights into the design of NER experimental schemes and processes.</p>