iACP-DPNet: a dual-pooling causal dilated convolutional network for interpretable anticancer peptide identification

Anticancer peptides (ACPs) are acknowledged for their potential in cancer therapy, attributed to their safety, low side effects, and high target specificity. However, the discovery of ACPs is slowed by the high cost and labor-intensive nature of experimental validation, resulting in a limited number of confirmed ACPs. Although various computational methods have been proposed, existing models commonly suffer from three critical limitations: reliance on small-scale datasets, lack of interpretable feature learning mechanisms, and insufficient generalization capability. To address these challenges, this study constructs a larger and more diverse dataset by consolidating data from existing literature and databases, and proposes a novel deep learning predictive model named iACP-DPNet. The model utilizes the protein language model ProtBert with positional encoding to convert protein sequences into feature vectors, then applies a two-step feature selection process via LightGBM and MIC. The selected features undergo processing by a causal dilated convolution network. A dual-pooling mechanism is designed to enhance the model's ability to synergistically model local critical residues and global sequence contexts, integrating parallel GlobalAveragePooling and attention pooling layers. Compared to traditional single-pooling models (e.g., ACP‑MHCNN), this architecture significantly improves feature extraction capability. To understand the model's decision-making process, we employ t-SNE for visualizing key steps, ISM for interpreting sequence regions, and SHAP analysis for evaluating feature importance. These approaches significantly improve the model's interpretability. The model exhibits outstanding performances on the novel dataset, as evidenced by rigorous tenfold cross-validation. Achieving remarkable metrics-including Sp of 96.1%, Sn of 92.91%, Acc of 94.5% and MCC of 89.05%, it significantly outperforms all existing state-of-the-art methods in comparative analyses. Furthermore, to assess its generalizability, we evaluated iACP-DPNet on an additional dataset, where it outperformed other current models. In conclusion, the iACP-DPNet exhibits exceptional performance and generalizability, showcasing its advanced design and effectiveness in ACPs prediction. This research provides a robust and interpretable framework for advancing research in anticancer peptide discovery. HIGHLIGHTS: • We have established a larger and more diverse dataset for ACPs prediction, addressing the limitations of existing datasets and providing a robust foundation for model training and evaluation. • The implementation of a dual-pooling layer mechanism (GlobalAveragePooling and attention pooling) bolsters the model's capacity to learn diverse features, ultimately enhancing its prediction efficiency. • We employed t-SNE visualization and ISM-based interpretability analysis to provide insights into the model's decision-making process, highlighting key regions and amino acids critical for ACPs functionality. • The iACP-DPNet model demonstrates strong generalizability across diverse datasets, making it a reliable tool for ACPs prediction and potentially other peptide-related tasks.