AVP-Pro: adaptive multi-representation fusion and contrastive learning for antiviral peptide identification and functional subtype prediction

Antiviral peptides (AVPs) are bioactive short peptides with antiviral activity. Their biological properties are closely related to sequence features such as amino acid composition, net positive charge, hydrophobicity, and conserved sequence motifs. Because different AVPs may show distinct virus-targeting specificities, simple binary discrimination between AVPs and non-AVPs is insufficient for fine-grained functional prediction. However, existing computational methods still face challenges in modeling complex sequence dependencies, integrating multi-source features, and distinguishing highly similar and easily confused samples. To address these challenges, we propose AVP-Pro, a two-stage prediction framework based on adaptive feature fusion and contrastive learning. AVP-Pro integrates ESM-2-derived deep sequence representations with 10 conventional physicochemical descriptors. CNN, BiLSTM, self-attention, and adaptive gating modules are used to capture local fragment-level features and global contextual dependencies in peptide sequences. To reduce the effect of ambiguous decision boundaries caused by the high similarity between positive and negative samples, we introduce BLOSUM62-guided data augmentation and OHEM-based contrastive learning, which help improve the model’s ability to distinguish difficult samples. In the first-stage general AVP identification task, AVP-Pro showed competitive predictive performance on the independent test set. In the second-stage functional subtype prediction task, the framework incorporated transfer learning to predict AVP subtypes associated with six virus families and eight specific viruses, showing stable performance across multiple evaluation metrics. AVP-Pro showed stable performance in both general AVP identification and functional subtype prediction across the evaluated benchmark datasets. These results suggest that AVP-Pro can serve as a computational framework for sequence-level functional annotation and prioritization of candidate antiviral peptides.