Aegis: a transformer-based deep learning framework for the accurate identification of anticancer peptides

Area of Science:

• Biochemistry
• Computational Biology
• Pharmacology

Background:

• Anticancer peptides (ACPs) show selective cancer cell toxicity but are difficult to identify experimentally.
• Computational methods offer a faster, more efficient alternative for ACP prediction.

Purpose of the Study:

• To introduce Aegis, a novel transformer-based deep learning framework for accurate anticancer peptide identification.
• To evaluate and compare various machine learning and deep learning models for ACP prediction.

Main Methods:

• Systematic evaluation of models using feature extraction methods: CKSAAP, CTDC, CTDT, CTDD, and PAAC.
• Feature importance analysis using ANOVA, ReliefF, and SHAP, followed by incremental feature selection (IFS).
• Development of Aegis, a transformer-based deep learning framework utilizing 103 optimal features identified via SHAP.

Main Results:

• Aegis achieved state-of-the-art performance on an independent dataset, outperforming existing ACP prediction models.
• Compositional analysis indicated ACP sequences are rich in positively charged and hydrophobic residues.
• The study identified an optimal subset of 103 discriminative features for ACP prediction.

Conclusions:

• Transformer-based deep learning shows significant potential for anticancer peptide identification.
• Aegis provides a robust computational tool for screening and developing novel ACPs.
• This work lays the groundwork for future computational drug discovery and clinical development of ACPs.