NeuroCL: A deep learning approach for identifying neuropeptides based on contrastive learning

Abstract

Neuropeptides (NPs), a unique class of neuronal signaling molecules, involved in neurotransmission, endocrine regulation, immune response, mood, and appetite control. The identification of neuropeptides provides critical scientific insights for early diagnosis, targeted therapy, and personalized medicine of related diseases. Previous models struggle to capture complex relationships among features and inter-sample connections. In this study, we introduce NeuroCL, a deep learning model harnessing contrastive learning and a cross-attention mechanism to efficiently identify NPs through multifaceted attribute representation. Experimental outcomes demonstrate that NeuroCL effectively captures data nuances, achieving an impressive accuracy of 93.8 % and a Matthews correlation coefficient (MCC) of 87.8 % on an independent test set. Contrastive learning enhances class distinction and coherence, while cross-attention mechanisms integrate pre-trained large models with manually encoded features, synergistically boosting their capabilities and strengthening feature connections. Our model surpasses current state-of-the-art predictors in NPs identification. Visualization via uniform manifold approximation and projection (UMAP) reveals that NeuroCL distinctly segregates positive NPs from negative ones. To facilitate the accessibility and application of our model, we have established a web-based platform available at http://www.bioai-lab.com/NeuroCL.