ToxMamba: a novel method for toxic peptide prediction based on the fusion of pretrained embeddings and multi-scale state space models.

Abstract

Peptide therapeutics offer great promise in drug discovery, yet the accurate identification of peptide toxicity remains essential for both safety assessment and biodefense. We developed ToxMamba, a toxic peptide prediction framework that integrates Implicit Structure Model (ISM)-derived pretrained embeddings, state space models, multi-scale feature learning, and ensemble inference. Built upon the Mamba architecture, ToxMamba enables efficient sequence modeling and captures contextual patterns across multiple receptive fields from parallel branches. On the independent test set, ToxMamba consistently surpasses existing methods, improving the F1 score and Matthews Correlation Coefficient by 3.9-18.3 and 7.5-46.9 percentage points, respectively. Ablation analyses further showed that Mamba-based models generally achieved stronger overall performance than Transformer-, Convolutional Neural Network-Long Short-Term Memory-Attention-, and Kolmogorov-Arnold Network-based variants. Feature visualization showed that ISM embeddings yielded clearer separation between toxic and non-toxic peptides in the latent space. Intermediate-layer feature analysis highlighted residue-level regions associated with toxicity-related sequence motifs, supporting the biological interpretability of ToxMamba. Overall, this study provides an efficient and interpretable computational framework for high-throughput toxic peptide screening.