AVP-GPT2: Prompt-conditioned fine-tuning of a GPT-2 protein language model for antiviral peptide identification

Antiviral peptides (AVPs) are promising therapeutic candidates as it can inhibit viral replication, interfere with host-virus interactions, and provide sequence-specific antiviral activity. However, experimental screening of large peptide libraries is time-consuming and costly, creating a need for computational tools that can prioritize promising candidates before laboratory validation. Here, we present AVP-GPT2, a practical protein language model workflow for antiviral peptide identification. In this study, peptide sequences were embedded within structured task-specific prompts and optimized using a causal language modeling objective to distinguish antiviral from non-antiviral peptides. AVP-GPT2 was evaluated using four prompt formats to determine how input design affects prediction performance. The best-performing configuration achieved an accuracy of 0.9421, sensitivity of 0.9625 , specificity of 0.9187 and MCC of 0.8839, outperforming several existing AVP prediction methods. To support biological interpretation, we analyzed sequence regions emphasized by the model and found enrichment of residues and motifs associated with antiviral activity. In experimentally validated SARS-CoV-2 peptide case studies, AVP-GPT2 identified biologically relevant glycine-centered signatures in Nsp9-binding peptides and hydrophobic/aromatic residue patterns in Nsp16-binding peptides, in agreement with reported binding mechanisms. Structural evaluation of representative NOTCH4- and LMTK3-derived peptides further showed spatial agreement between model-prioritized residues and protein-peptide interface contacts. Together, these results support the utility of AVP-GPT2 as an interpretable computational screening workflow for antiviral peptide prioritization and hypothesis generation prior to experimental validation.