Cross-disease modeling of peripheral blood identifies biomarkers of type 2 diabetes predictive of Alzheimer's disease

Abstract

Type 2 diabetes (T2D) is a significant risk factor for Alzheimer's disease (AD). Despite multiple studies reporting this connection, the mechanism by which T2D exacerbates AD is poorly understood. It is challenging to design studies that address co-occurring and comorbid diseases, limiting the number of existing evidence bases. To address this challenge, we expanded the applications of a computational framework called Translatable Components Regression (TransComp-R), initially designed for cross-species translation modeling, to perform cross-disease modeling to identify biological programs of T2D that may exacerbate AD pathology. Using TransComp-R, we combined peripheral blood-derived T2D and AD human transcriptomic data to identify T2D principal components predictive of AD status. Our model revealed genes enriched for biological pathways associated with inflammation, metabolism, and signaling pathways from T2D principal components predictive of AD. The same T2D PC predictive of AD outcomes unveiled sex-based differences across the AD datasets. We performed a gene expression correlational analysis to identify therapeutic hypotheses tailored to the T2D-AD axis. We identified six T2D and two dementia medications that induced gene expression profiles associated with a non-T2D or non-AD state. Finally, we assessed our blood-based T2DxAD biomarker signature in post-mortem human AD and control brain gene expression data from the hippocampus, entorhinal cortex, superior frontal gyrus, and postcentral gyrus. Using partial least squares discriminant analysis, we identified a subset of genes from our cross-disease blood-based biomarker panel that significantly separated AD and control brain samples. Our methodological advance in cross-disease modeling identified biological programs in T2D that may predict the future onset of AD in this population. This, paired with our therapeutic gene expression correlational analysis, also revealed alogliptin, a T2D medication that may help prevent the onset of AD in T2D patients.

Keywords: Alzheimer’s disease; Type 2 diabetes; computational gene correlation analysis; cross-disease modeling; partial least squares discriminant analysis; transcriptomics.

Figure 1.. Workflow of TransComp-R.

(a) Genes across T2D and AD are selected for analysis. Each AD cohort is individually projected into the T2D PCA space to combine the two diseases. (b) PC translatability from T2D to AD is determined by running a GLM regression against AD outcomes using PCs consistently selected across each AD cohort. (c) Pathway enrichment analysis is performed on the loadings of significant PCs to identify enriched biological pathways. Potential therapeutic candidates are then identified using a correlation analysis framework.

Figure 2.. TransComp-R identifies T2D PCs predictive of AD outcomes.

(a) AD PCs were separated by cohort, with variance explained in AD (b) Selection of PCs using a LASSO model incorporating sex and age demographics from the AD datasets. The model was run across twenty random rounds of ten-fold cross-validation. PCs consistently determined significant across both AD cohorts from the GLM regression were further analyzed. (c) Principal component plots of AD scores on selected T2D PCs separating AD and control outcomes in AD cohort 1 and (d) AD cohort 2. Each T2D PC is represented by the percent variance explained in AD.

Figure 3.. Pathway Enrichment Analysis.

The transcriptomic variance separating AD and control subjects on T2D PC2 was interpreted with GSEA using the (a) KEGG and (b) Hallmark databases. Significantly enriched pathways were determined with a Benjamini-Hochberg adjusted p value less than 0.01. (c) Shared leading edge genes between biological pathways in the KEGG and (d) Hallmark pathways. The node size represents the number of genes contributing to the pathway from GSEA, whereas the edge size is the number of shared genes between each biological pathway. Missing pathways signified that there were no shared genes with other pathways.

Figure 4.. Comparison of global gene expression and AD-predictive T2D PCs

(a) AD and T2D log₂ fold change plot of all shared 11,455 genes (b) AD and T2D log₂ fold change plot filtered by gene expressions with the top 50 and bottom 50 loadings of T2D PC2. (c) Scores of T2D PC2 separated by sex and disease condition. A Mann-Whitney test adjusted by Benjamini-Hochberg was used to determine statistical significance. The distribution of the data is annotated by the mean and interquartile ranges.

Figure 5.. Computational gene expression correlational analysis.

(a) All significant drugs identified from the LINCS database. Drugs filtered by (b) FDA approval status and (c) over-the-counter drugs. (d) FDA-approved T2D drugs (alogliptin and glipizide) associated with control group signatures. (e) FDA-approved T2D drug (orlistat) associated with genes upregulated in AD. (f) FDA-approved medications for cognitive-enhancement (galantamine and donepezil). (g) FDA-approved drug (brexpiprazole) with signatures correlated to genes elevated in AD.

Figure 6.. Translating blood-predictable signatures to the brain.

(a) Method of testing blood-derived data predictability in the brain. (b) Z-score of significant AD-associated genes identified in the human hippocampal dataset (Mann-Whitney adjusted by Benjamini-Hochberg, p adjusted < 0.20). (c) PLS-DA model using significant genes to predict AD status. AD groups are labeled by APOE genotype, Braak stage, and MMSE. (d) Loading variables LV1 and LV2 for the model are presented. A VIP>1 is annotated with a star, and the color of the loading bar represents the highest contribution to the specific class by the respective gene.

Figure 7.. PLS-DA models using blood biomarkers to predict AD status in other brain regions

(a) Z-score of significant genes identified in the human EC dataset. (b) PLS-DA using the significant genes on the EC data with loadings on LV1 and LV2. (c) Z-score of significant genes identified in the human SFG dataset. (d) PLS-DA using the significant genes on the SFG data with loadings on LV1 and LV2. (e) Z-score of significant genes identified in the human PoCG dataset. (f) PLS-DA using the significant genes on the PoCG data with loadings on LV1 and LV2. For all brain regions, the significance of the genes was determined by a Mann-Whitney adjusted by Benjamini-Hochberg (p adjusted < 0.20) across AD and control groups.