ProtT5-MSCRNet: a multi-scale convolutional and channel-recalibrated deep learning framework for anticancer peptide prediction

Introduction Cancer is a leading cause of mortality worldwide. Anticancer peptides (ACPs) are promising therapeutic candidates due to their low toxicity, favorable biocompatibility, and selective anticancer activity; however, experimental ACP identification and screening remain labor-intensive, time-consuming, and costly. Methods We developed ProtT5-MSCRNet, an end-to-end deep learning framework for ACP prediction that integrates ProtT5-based evolutionary representations, multi-scale convolutional feature extraction, channel-wise attention recalibration, and robust optimization strategies. Results Experiments on two independent benchmark datasets showed that ProtT5-MSCRNet outperformed state-of-the-art ACP prediction methods. The model achieved ACC/SN/SP/MCC values of 0.954/0.874/0.983/0.881 on Test Set 1 and 0.984/0.980/0.987/0.967 on Test Set 2. Ablation studies and visualization analyses further supported the effectiveness and interpretability of the proposed model. Discussion ProtT5-MSCRNet provides a robust, accurate, and biologically meaningful computational tool for facilitating ACP identification and accelerating anticancer peptide discovery.