Graph-Contrastive Convolutional Neural Network for Extracting and Classifying Peptide-Based Periodontal Immunomodulatory and Anti-Inflammatory Signatures.

INTRODUCTION

Periodontitis is a chronic inflammatory disease that occurs when the body's immune system fails to respond properly to microbial biofilms. Peptides exhibiting anti-inflammatory and immunomodulatory properties are promising as biomarkers and therapeutic agents; however, identifying functional peptide signatures remains a significant challenge due to sequence variability and limited data availability. The aim of this study was to develop a graph-contrastive convolutional neural network (GCCNN) framework for accurately categorizing periodontitis-related peptides into immunomodulatory and anti-inflammatory classes, to enhance model interpretability, and facilitate customized peptide-based therapies.

METHODS

In this research, we introduce a Graph-Contrastive Convolutional Neural Network (GCCNN) framework for classifying peptide sequences relevant to periodontal health into anti-inflammatory peptides (AIPs) and immunomodulatory peptides (IMPs). Our model combines a 1D CNN encoder with SimCLR-style contrastive learning to obtain motif-level representations from peptide sequences that remain unchanged. To fix a big class imbalance, pre-trained representations are fine-tuned using supervised learning with class-weighted binary cross-entropy.

RESULTS

On a curated peptide dataset, GCCNN performed well with AIPs (F1 = 0.94, ROC-AUC = 0.98), but struggled to generalize IMP predictions due to insufficient data (F1 = 0.04, ROC-AUC = 0.54). Using t-SNE to visualize the latent space revealed that AIP clusters were closely grouped, whereas IMP embeddings were more widely dispersed. SHAP-based interpretation revealed conserved sequence motifs that contribute to AIP predictions, highlighting the model's transparency.

CONCLUSION

This study demonstrates the effectiveness of contrastive learning for peptide signature extraction and provides a foundation for improving minority-class peptide classification in biomedical sequence tasks.