CCMPIP: Cross-attention and capsule network-based multi-feature fusion for proinflammatory peptide prediction

Area of Science:

• Biochemistry
• Computational Biology
• Immunology

Background:

• Proinflammatory peptides (PIPs) are crucial mediators of immune responses and inflammatory diseases.
• Traditional wet-lab identification of PIPs is resource-intensive due to sequence complexity.
• Efficient computational methods are needed for accurate PIP identification.

Purpose of the Study:

• To develop a computational framework, CCMPIP, for accurate identification of proinflammatory peptides.
• To integrate semantic and physicochemical features for enhanced peptide representation.
• To provide interpretable insights into the biological activity of PIPs.

Main Methods:

• CCMPIP utilizes ProtT5 embeddings and AAindex physicochemical descriptors.
• A cross-attention mechanism fuses dual feature matrices derived from peptide sequences.
• Cascading convolutional neural network (CNN) and capsule network layers refine representations before MLP classification.
• Performance was evaluated using 5-fold cross-validation against existing predictors.

Main Results:

• CCMPIP demonstrated superior predictive performance compared to state-of-the-art ensemble predictors.
• Interpretability analyses, including attention heatmaps and STREME motif enrichment, identified biologically relevant residues.
• The framework provides transparent insights into the mechanisms underlying proinflammatory activity.

Conclusions:

• CCMPIP offers a powerful and efficient computational solution for identifying proinflammatory peptides.
• The integration of diverse features and advanced deep learning architectures enhances predictive accuracy.
• The model's interpretability facilitates a deeper understanding of PIP function in inflammatory processes.