Accurate prediction of anticancer peptides using a stacking ensemble of convolutional and transformer models with...

We present a high-performance predictive framework for anticancer peptide (ACP) identification, based on a stacking ensemble learning approach that synergistically combines convolutional neural networks and transformer models using a random forest as a meta-classifier. This architecture is driven by conjoint sequence representations that integrate both one-hot encoding and pre-trained evolutionary scale modeling embeddings, enabling the extraction of complementary local and global features from peptide sequences. Our proposed model achieved a robust accuracy of 88.9% on the primary ACP data set, while maintaining competitive or superior performance across multiple external benchmark data sets, with accuracies ranging from 83.2% to 95.2%, highlighting its strong generalization capability on par with the state-of-the-art models. To demonstrate translational relevance, the model was applied to a curated set of clinically approved and candidate ACPs, producing probabilistic scores to support experimental prioritization. To further enhance model interpretability, SHapley Additive exPlanations analysis was employed, revealing lysine as a consistently influential residue, alongside other positively charged and hydrophobic amino acids. These findings not only corroborate known mechanistic insights into ACP-membrane interactions but also highlight the utility of model-derived feature importance in guiding peptide design. Taken together, this work introduces a robust, interpretable, and generalizable approach for computational ACP prediction, offering valuable implications for peptide-based anticancer drug discovery. To enhance the accessibility and translational potential of our model, we developed an interactive web-based prediction tool, named ACPredictor, for the identification of ACPs. This platform is freely available at https://acpredictor.streamlit.app/.