Mamba-ACP: a Hybrid State-Space and Transformer Framework for Interpretable Anticancer Peptide Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Anticancer peptides (ACPs) show therapeutic promise but face clinical translation challenges.
• Existing ACP prediction models require improvement in accuracy and generalizability.

Purpose of the Study:

• To develop a robust hybrid deep learning framework, Mamba-ACP, for accurate and scalable prediction of anticancer peptides.
• To integrate diverse peptide features for enhanced prediction performance.

Main Methods:

• Developed Mamba-ACP, a hybrid deep learning model combining Evolutionary Scale Modeling (ESM-2) embeddings and handcrafted features (AAindex, BLOSUM62).
• Utilized a Mamba-based architecture for sequence modeling, processing fused token-level representations.
• Trained and validated the model on two benchmark datasets (Set 1 and Set 2).

Main Results:

• Mamba-ACP achieved 87.59% accuracy and 0.9356 AUC on Set 1, and 96.69% accuracy and 0.9922 AUC on Set 2.
• Outperformed state-of-the-art ACP predictors, demonstrating superior accuracy and reduced false positives.
• Provided model-level explanations using saliency maps and feature importance analysis.

Conclusions:

• The hybrid approach effectively combines evolutionary and physicochemical peptide properties for improved ACP classification.
• Mamba-ACP establishes a new benchmark for computational peptide discovery with strong generalizability and efficiency.
• The model offers insights into ACP mechanisms through explainability features, aiding future therapeutic development.