Graph Convolutional Neural Networks for the Classification of Oral and Periodontal Pathogens Based on Protein Sequence Analysis

A BSTRACT Introduction: Oral pathogens, including bacteria, viruses, fungi, and protozoa, are crucial in dental and systemic diseases. Traditional microbiological techniques are time-consuming and limited, presenting an opportunity for advanced computational techniques such as machine learning and deep learning. Graph convolutional networks (GCNs) offer a powerful tool for understanding protein sequences and their classification based on structural and functional properties. Early detection and accurate classification of oral and periodontal pathogens can help develop targeted therapies and prevention strategies, addressing the challenge of efficiently classifying and interpreting large datasets. The study uses a graph convolutional neural network (GCN) to classify oral pathogens based on their protein sequence characteristics, a groundbreaking application of deep learning in oral microbiology. Methods: Pathogen–Host Interactions Database is a comprehensive resource for cataloging the interactions between pathogens and host organisms. Researchers use a graph convolutional network (GCN) architecture to classify protein sequences from pathogens such as Candida albicans , Porphyromonas gingivalis , and Treponema denticola , accessing data on host–bacterial interactions. GCN architecture uses 410 encoded sequences and cosine similarity to construct an adjacency matrix with 2000 encoded protein sequences, 128 neurons, 0.6 dropout rate, 64 neurons, and an output layer for pathogen species predictions. Results: The GCN performed well for Candida albicans but showed poor performance for Porphyromonas gingivalis and failed to identify Treponema denticola, indicating class-dependent limitations. Conclusion: The study uses graph-based neural networks to classify oral and periodontal pathogens based on protein sequence data. While it achieves high accuracy, it faces challenges in handling imbalanced datasets. Improvements in model architecture show promise, but limitations in class imbalance and feature representation need further research. The methodological framework can be a foundation for future pathogen classification and protein sequence analysis studies.