Enhancing anticancer peptide discovery: A fusion-centric framework with conditional diffusion for prediction and...

Area of Science:

• Computational biology
• Peptide science
• Bioinformatics

Background:

• Anticancer peptides (ACPs) show promise for cancer therapy due to targeted tumor cell selectivity and low toxicity.
• Existing computational models for ACPs struggle with sequence representation and imbalanced datasets.
• Development of robust computational tools is crucial for efficient ACP discovery.

Purpose of the Study:

• To develop a unified fusion-driven framework (UACD-ACPs) for enhanced anticancer peptide prediction and generation.
• To address limitations in sequence representation and class imbalance in existing ACP computational models.
• To facilitate targeted downstream screening of novel ACP candidates.

Main Methods:

• Integrated ProtBERT embeddings and physicochemical descriptors using the Multiscale Embedding Compression Strategy (MECS).
• Employed a diffusion-inspired noise-conditioned classifier for robust ACP prediction.
• Utilized a denoising diffusion-based generative framework with Bitemporal Fusion Module (BFM) and Temporal Feature Attention Module (TFAM) for peptide generation.
• Incorporated cancer-type-aware organization in the generative module.

Main Results:

• UACD-ACPs significantly outperformed state-of-the-art methods in accuracy, F1-score, and AUC-ROC.
• Generated peptides exhibited favorable physicochemical properties, diverse secondary structures, and high structural stability.
• Validated peptide stability and membrane-binding properties using molecular dynamics simulations.

Conclusions:

• Fusion-driven diffusion-based frameworks can effectively mitigate class imbalance and data heterogeneity in ACP modeling.
• UACD-ACPs offers a scalable and biologically grounded approach for discovering novel anticancer peptides.
• This framework holds potential for advancing next-generation cancer therapies.