PeptideNet: An Integrative Deep Learning Framework for Predicting Diverse Bioactive Peptides Using Protein Language Model Embeddings

Bioactive peptides are multifunctional biomolecules composed of short amino acid sequences that exhibit diverse biological activities, including antioxidative, antihemolytic, anticell-penetrating, antiviral, and antimicrobial effects. Accurate computational prediction of peptide bioactivity is essential for accelerating the discovery and design of peptide-based therapeutics. In this study, we investigated five categories of bioactive peptides using four distinct feature representations, including large protein language model embeddings (ESM1, ESM2, and ProtBert) and physicochemical descriptors. A total of 20 hybrid deep learning models integrating Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (BiGRUs) were developed to capture both local sequence motifs and long-range dependencies. The proposed PeptideNet model achieved robust predictive performance, with accuracies of 0.83, 0.87, 0.89, 0.92, and 0.94 for antioxidative, antihemolytic, anticell-penetrating, antiviral, and antimicrobial peptides, respectively, on the independent data set. Among the evaluated feature sets, ESM-2 embeddings consistently outperformed others across all peptide types, providing rich contextual and evolutionary information. Furthermore, t-SNE visualization of learned representations demonstrated effective generalization across peptide classes, while positional sequence logo analysis revealed conserved residue patterns contributing to peptide bioactivity. The integration of large protein language model embeddings with the PeptideNet architecture enables the model to capture both global contextual information and residue-level features, establishing a generalized and interpretable framework for multipeptide bioactivity prediction.