ACP-EPC: an interpretable deep learning framework for anticancer peptide prediction utilizing pre-trained protein language model and multi-view feature extracting strategy.

Abstract

Cancer remains a major global health challenge, as conventional chemotherapy often causes extensive damage to healthy cells and leads to severe side effects. Anticancer peptides (ACPs) have emerged as a promising therapeutic alternative, capable of selectively targeting and eliminating cancer cells while improving patient quality of life and treatment outcomes. Nevertheless, identifying ACPs through traditional biological experiments is both labor-intensive and time-consuming. To address this limitation, we developed ACP-EPC, a deep learning framework which predicts ACPs directly from protein sequences. ACP-EPC integrates contextual representations from Evolutionary Scale Modeling 2 (ESM-2) with handcrafted physicochemical descriptors and employs a Cross-Attention mechanism for multimodal feature fusion. The model was rigorously evaluated using tenfold cross-validation and two test sets, ACP135 and ACP99, achieving accuracy of 0.935 and 0.984, respectively. These results substantially outperform existing models, underscoring the advantages of combining diverse feature representations. To promote accessibility, we have also deployed ACP-EPC as a publicly available web server at http://www.bioai-lab.com/ACP-EPC .