NeuroPpred-PSCG: A Multimodal Framework Using ProtT5 and Structural Features for Neuropeptide Prediction Based on...

As fundamental signaling molecules, neuropeptides are involved in diverse physiological regulatory activities. Abnormal expression patterns have been closely linked to neurological and metabolic disorders, which underscores the urgent need for precise and rapid identification methods. Here, we develop NeuroPpred-PSCG, a novel multimodal deep learning framework. At its core, the framework employs both a bidirectional cross-attention mechanism and a gated fusion module to integrate global semantic insights from the protein language model (ProtT5) with local conformational data derived from secondary structure. Furthermore, it utilizes a multiscale gated convolutional network to enhance hierarchical contextual information. On the independent test set, NeuroPpred-PSCG surpasses current leading methods on most metrics, achieving ACC of 94.8%, SN of 95.1%, F1-score of 94.8%, and MCC of 0.896, reflecting robust overall discriminative capability. Compared to the previous best-performing model, NeuroPpred-MSN, our method shows improvements of 1.2% and 0.024 in ACC and MCC, respectively, and most notably, a significant 2.8% gain in SN. Ablation studies, robustness analysis, and interpretability analysis further substantiate the effectiveness of each component, as well as the overall stability and interpretability of NeuroPpred-PSCG. The source code and data sets have been made publicly accessible at https://github.com/xinyizhang186/NeuroPpred-PSCG.