DeepNPA: A protBert-based framework for residue-level annotation of neuropeptide precursors

Neuropeptides are key signaling molecules that regulate the nervous and endocrine systems, and their maturation depends on precise proteolytic cleavage of precursor proteins. However, residue-level annotation remains essential yet experimentally labor intensive, with limited dedicated computational tools. Here, we present DeepNPA, a residue-level annotation framework that integrates ProtBert, a pretrained protein language model, with focal loss for imbalance-aware optimization. DeepNPA was trained and evaluated on a non-redundant Swiss-Prot-derived dataset annotated using the BIO scheme. On a held-out test dataset, DeepNPA achieved strong performance, with an accuracy of 0.953, an F1-score of 0.927, and an AUPRC of 0.965. On an independent benchmark dataset, DeepNPA outperformed NeuroCS and NeuroPred, achieving an ACC of 0.969 and Matthews correlation coefficient (MCC) of 0.666. External validation further demonstrated its ability to identify neuropeptide regions and their associated cleavage sites. In addition, we provided an online prediction platform to facilitate neuropeptide precursor annotation and cleavage site analysis.