Modeling Host-Microbe Interactions in Periodontal Disease: A GCNN Vs GAT Approach

Abstract

ABSTRACT

Introduction: Periodontitis is an inflammatory condition that affects the tooth supporting structures and causes significant tissue destruction. It is not caused by merely poor oral hygiene but by a complex interplay between microbial agents and the host's immune response, which triggers an inflammatory response. Periodontitis severity and progression vary among individuals, with genetic predisposition, systemic health issues, and environmental factors playing crucial roles. People with compromised immune systems and chronic inflammation are at higher risk. Understanding the relationship between host immune responses and microbial factors is vital for developing effective prevention and treatment strategies. This gap presents an opportunity for future research, potentially leading to advancements in understanding host-bacterial interactions and health management implications.

Methods: The PHI-base Pathogen-Host Interactions Dataset provides detailed information on pathogen-host interactions, including protein and gene data, enhancing our understanding of plant diseases. The initial phase involves data collection from a periodontal pathogen virulence database organized in tab-delimited text files. This data includes gene information, pathogen species, phenotypes, and functional annotations, providing insights into periodontal disease roles. It is then subjected to graph convoluted neural networks for analysis.

Results: Graph Convolutional Neural Network (GCNN) and Graph Attention Network (GAT) models demonstrated high precision metrics and confidence distributions in predicting reduced virulence. They achieved a rate of 84.62% accuracy, with GCNN showing a higher prediction confidence at 86.19% compared to GAT's 84.82%. Both models performed well in predicting the majority class, characterized by reduced virulence, yielding a precision of 0.846, a recall of 1.0, and an F1-score of 0.917. However, they faced challenges in minority classes, particularly those indicating increased virulence and unaffected states. The GAT model reached a final loss of 0.5213, suggesting better performance. Both models achieved an accuracy of 0.8462, indicating they can effectively capture relevant patterns within the trained data.

Conclusion: The study demonstrates the effectiveness of machine learning in predicting host virulence interactions with periodontal inflammation, highlighting the need for future research for improved clinical outcomes.