Cross-species modeling identifies gene signatures in type 2 diabetes mouse models predictive of inflammatory and estrogen signaling pathways associated with Alzheimer's disease outcomes in humans

Introduction: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by memory loss, confusion, and behavioral changes. The onset for AD is not attributed to one cause but stems from a multitude of risk factors such as age, sex, genetic predisposition, and pre-existing health conditions. Among these precursors, evidence also suggests that individuals with type 2 diabetes (T2D), a chronic metabolic condition that disrupts the body’s ability to regulate blood sugar levels, face an increased susceptibility to develop AD. Some such studies have reported the link to chronic inflammation and dysfunctions to glucose metabolism as potentially shared risk factors. Despite multiple studies supporting a link between T2D and AD risk, the biological mechanisms by which this occurs are not well-understood. A key challenge in understanding the mechanistic links between these conditions is that evidence must be synthesized and translated between experiments in mouse models and human-based clinical studies. Translating information from pre-clinical models to human clinical contexts is often difficult due to discrepancies in interspecies physiology, timeline of disease development, and heterogeneity of the human population. To overcome this challenge, we developed a computational framework termed translatable components regression (TransComp-R) to identify transcriptomic signatures in mouse models predictive of AD condition in human (Figure 1).

Materials and

Methods: Publicly available mouse (GSE152539) and human (GSE48350) microarray hippocampal data were accessed from Gene Expression Omnibus. The mouse data includes male mice of solely AD, only T2D, and simultaneous occurrence of both diseases (ADxT2D), whereas the human data contains sex, age, and transcriptomic data of AD and control subjects.

Mouse and human transcriptomic data were processed by selecting homologous genes and normalizing by z-score. We applied TransComp-R by conducting principal component analysis (PCA) on the mouse data separated into AD, T2D, and ADxT2D groups with controls, such that three groups of PCs encoded transcriptomic variation between healthy controls and AD, T2D, or ADxT2D mice. Human AD and control subjects were projected into the mouse PCA space. Mouse PCs associated with AD outcomes in human were selected by performing LASSO across 100 rounds of 5-fold cross-validation regressing the human positions in mouse PC space against human disease status. Four sets of LASSO models were trained, including main effects for just mouse PCs, PCs and human sex, PCs and human age, and PCs and human age and sex. PCs with a coefficient frequency greater than 50 rounds were selected for generalized linear models with individual PCs and human clinical covariates regressed against human AD outcomes. The PC loadings selected from LASSO were analyzed through Gene Set Enrichment Analysis (GSEA) using the KEGG and Hallmark databases. Pathways were determined significant if the Benjamini-Hochberg adjusted p-value was less than 0.25. Additionally, sex-based differences were compared with a Mann-Whitney test (Benjamini-Hochberg corrected p-value < 0.05).

Results, Conclusions, and Discussions: The TransComp-R model identified two T2D mouse PCs (PC2, 35.6%; PC3, 19.3% variance explained in human by mouse) and one ADxT2D mouse PC (PC3, 17.0%) predictive of human AD disease outcome (Figure 2-3). We performed GSEA on the selected T2D and ADxT2D PCs and identified enriched pathways for inflammatory and estrogen signaling pathways (Figure 4). From the KEGG database, we identified ‘complement and coagulation cascades’ and ‘cytokine-cytokine receptor interaction’ on T2D PC2. There were no significant KEGG pathways on T2D PC3. On the Hallmark database, we identified ‘interferon gamma response’, ‘interferon alpha response,’ ‘IL6 JAK STAT3 signaling’, and ‘inflammatory response’ to be enriched in AD conditions by T2D PC2. Interestingly in the Hallmark database, we identified ‘estrogen response late’ to be enriched for AD in T2D PC3 and ADxT2D PC3, while ‘estrogen response early’ was enriched for the control group in T2D PC2. Expanding upon the potential of sex-based predictability, we separated the scores of each PC by sex and disease status and found that mouse T2D PC2, T2D PC2, and ADxT2D PC3 significantly stratified human female AD and control groups, but not male AD and control groups, despite the mouse data originating from all male mice (Figure 5).

Our work suggests that T2D may affect AD through inflammatory and estrogen pathways. Through literature, we find evidence that chronic inflammation may lead to insulin resistance and cognitive deficits. Furthermore, studies report that estrogen serves a protective role in T2D and AD. With genes defining late response to estrogen enriched in AD, and early response to estrogen enriched in control subjects, hormonal alterations of estrogen may be a potential mechanism to further investigate.

Here, we used TransComp-R to identify potential biological pathways in which T2D may exacerbate AD development. We show that PCs from T2D and ADxT2D were most predictive of AD outcomes in human. Though AD alone was not as predictive, this may indicate that AD is influenced by multiple factors. Indeed, these results encourage future applications of TransComp-R to identify affected biological pathways in AD or different diseases, which may further enhance clinical studies or possible routes of therapeutics.

Acknowledgements (Optional): This work is supported by an award from the Good Ventures Foundation and Open Philanthropy (DKB and BKB). BKB is supported by the National Science Foundation through the Graduate Research Fellowship program under grant number DGE-1842166. BKB acknowledges the support of the NIH T32 predoctoral fellowship T32DK101001 from the National Institute of Diabetes and Digestive and Kidney Diseases.