Abstract
MOTIVATION: Peptide toxicity is a critical concern in the development of peptide-based therapeutics, as toxic peptides can lead to severe side effects, including organ damage, immune reactions, and cytotoxicity. Predicting peptide toxicity accurately is essential to ensure the safety and efficacy of these drugs. RESULTS: In this study, we propose a novel model, ToxMSRC, to predict peptide toxicity using a combination of the continuous bag of words (CBOW) method from word2vec, synthetic minority over-sampling technique (SMOTE), multi-scale convolutional neural networks (CNN), and bidirectional long short-term memory (BiLSTM). This approach addresses the challenge of data imbalance by augmenting positive samples and improves feature extraction through multi-scale convolution. Furthermore, the model incorporates a residual connection that helps prevent overfitting and enhances generalization ability, improving classification performance. The model is evaluated on benchmark and independent test sets, achieving BACC scores of 92.17% on independent test1 and 86.89% on independent test2, outperforming existing state-of-the-art models. Additionally, ToxMSRC provides valuable insights into the relationship between peptide toxicity and amino acid sequences, demonstrating its potential and practical value in peptide-based drug development. AVAILABILITY AND IMPLEMENTATION: The complete datasets, source code, and pre-trained models are made available at https://github.com/Renjingyi123/ToxMSRC and https://doi.org/10.5281/zenodo.15668530.