Comparison of light gradient boosting and logistic regression for interactomic hub genes in Porphyromonas gingivalis and Fusobacterium nucleatum-induced periodontitis with Alzheimer's disease

Abstract

Introduction: Porphyromonas gingivalis and Treponema species have been found to invade the central nervous system through virulence factors, causing inflammation and influencing the host immune response. P. gingivalis interacts with astrocytes, microglia, and neurons, leading to neuroinflammation. Aggregatibacter actinomycetemcomitans and Fusobacterium nucleatum may also play a role in the development of Alzheimer's disease. Interactomic hub genes, central to protein-protein interaction networks, are vulnerable to perturbations, leading to diseases such as cancer, neurodegenerative disorders, and cardiovascular diseases. Machine learning can identify differentially expressed hub genes in specific conditions or diseases, providing insights into disease mechanisms and developing new therapeutic approaches. This study compares the performance of light gradient boosting and logistic regression in identifying interactomic hub genes in P. gingivalis and F. nucleatum-induced periodontitis with those in Alzheimer's disease.

Methods: Using the GSE222136 dataset, we analyzed differential gene expression in periodontitis and Alzheimer's disease. The GEO2R tool was used to identify differentially expressed genes under different conditions, providing insights into molecular mechanisms. Bioinformatics tools such as Cytoscape and CytoHubba were employed to create gene expression networks to identify hub genes. Logistic regression and light gradient boosting were used to predict interactomic hub genes, with outliers removed and machine learning algorithms applied.

Results: The data were cross-validated and divided into training and testing segments. The top hub genes identified were TNFRSF9, LZIC, TNFRSF8, SLC45A1, GPR157, and SLC25A33, which are induced by P. gingivalis and F. nucleatum and are responsible for endothelial dysfunction in brain cells. The accuracy of logistic regression and light gradient boosting was 67% and 60%, respectively.

Discussion: The logistic regression model demonstrated superior accuracy and balance compared to the light gradient boosting model, indicating its potential for future improvements in predicting hub genes in periodontal and Alzheimer's diseases.

Keywords: Aggregatibacter actinomycetemcomitans; Alzheimer's disease; Fusobacterium nucleatum; Porphyromonas gingivalis; hub genes; interactome; light gradient boosting; periodontal disease.

Figure 1

Network analysis of genes associated with endothelial dysfunction. Figure shows interactome visualization employs color-coded nodes to depict entities or data points, edges to illustrate relationships, and a circular layout to organize nodes in a circular arrangement, likely utilized to examine genes involved in endothelial dysfunction.

Figure 2

Roc curve of algorithms. Figure shows the ROC curves of two models, logistic regression and light gradient boosting, are compared. The logistic regression ROC curve, plotted with false positive rate (FPR) and true positive rate (TPR), shows poor performance with an AUC of approximately 0.5.The gradient-boosting ROC curve, which indicates better performance than logistic regression, is smoother and steeper in the right panel (light gradient boosting).

Figure 3

Confusion matrix comparison of logistic regression and light gradient boosting models. Figure shows the comparison of the confusion matrices of two models: logistic regression and light gradient boosting. Left panel (logistic regression): the matrix displays true labels (0, 1) and predicted labels (0, 1), including true negative (TN), false positive (FP), false negative (FN), and true positive (TP) values. Right panel (light gradient boosting): the confusion matrix illustrates true labels (Y-axis) and predicted labels (X-axis), offering insights into the classification performance of two models, with values ranging from 0 to 134.

Figure 4

Lift chart for light gradient boosting model. Figure illustrates the bins. Based on predicted values, average target values, actual values, and the model's performance, the predicted line should closely align with the actual line, indicating an accurate prediction of the target variable.