Exploiting machine learning models to identify novel Alzheimer's disease biomarkers and potential targets

Abstract

We still do not have an effective treatment for Alzheimer's disease (AD) despite it being the most common cause of dementia and impaired cognitive function. Thus, research endeavors are directed toward identifying AD biomarkers and targets. In this regard, we designed a computational method that exploits multiple hub gene ranking methods and feature selection methods with machine learning and deep learning to identify biomarkers and targets. First, we used three AD gene expression datasets to identify 1/ hub genes based on six ranking algorithms (Degree, Maximum Neighborhood Component (MNC), Maximal Clique Centrality (MCC), Betweenness Centrality (BC), Closeness Centrality, and Stress Centrality), 2/ gene subsets based on two feature selection methods (LASSO and Ridge). Then, we developed machine learning and deep learning models to determine the gene subset that best distinguishes AD samples from the healthy controls. This work shows that feature selection methods achieve better prediction performances than the hub gene sets. Beyond this, the five genes identified by both feature selection methods (LASSO and Ridge algorithms) achieved an AUC = 0.979. We further show that 70% of the upregulated hub genes (among the 28 overlapping hub genes) are AD targets based on a literature review and six miRNA (hsa-mir-16-5p, hsa-mir-34a-5p, hsa-mir-1-3p, hsa-mir-26a-5p, hsa-mir-93-5p, hsa-mir-155-5p) and one transcription factor, JUN, are associated with the upregulated hub genes. Furthermore, since 2020, four of the six microRNA were also shown to be potential AD targets. To our knowledge, this is the first work showing that such a small number of genes can distinguish AD samples from healthy controls with high accuracy and that overlapping upregulated hub genes can narrow the search space for potential novel targets.

Figure 1

The study workflow consists of two key paths, via ranking algorithms and feature selection methods.

Figure 2

An ImaGEO generated heatmap of the top-100 DEGs (Red represents the relative upregulated gene expression; green represents the relative downregulated gene expression; black represents no significant change in gene expression).

Figure 3

The prediction performance of the list of DEGs selected by (a) LASSO (L) and (b) Ridge (R) regression algorithms at multiple thresholds for the importance scores.

Figure 4

Comparing the AUC results of different genes lists. The blue bar is for the whole DEGs list (924 genes), the red bar is the overlapping list of LASSO and Ridge (5 genes). The yellow bar is the overlapping list of the six ranking algorithms (28 genes), and the green is the overlapping list of the ranking algorithms excluding MCC (53 genes).

Figure 5

The top-10 enriched diseases retrieved from DisGeNET for upregulated genes in the different gene sets, including the DEGs, LASSO, Ridge, and the overlapped/common hub and feature selection genes.

Figure 6

Network generated by miRNet. It shows six important miRNAs, represented by blue squares, and highlights JUN as the critical transcription factor.