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

ToxMamba achieves up to 18.3% higher F1 scores and 46.9% better Matthews Correlation Coefficient than existing methods in toxic peptide prediction.

• Why it matters: Accurate toxicity prediction of peptides is critical for safe drug development and biodefense, yet current methods lack precision and interpretability, creating a significant gap in the field.
• What they did: The authors developed ToxMamba, a framework combining pretrained embeddings, multi-scale state space models, and ensemble inference, tested on independent datasets with ablation analyses comparing various architectures.
• The result: ToxMamba outperforms alternative models, offers clearer biological interpretability through feature visualization, and facilitates high-throughput screening of toxic peptides with improved accuracy and insight.