Large-scale gene expression changes in APP/PSEN1 and GFAP mutation models exhibit high congruence with Alzheimer's disease

Abstract

Alzheimer's disease (AD) is a complex neurodegenerative disorder with both genetic and non-genetic causes. Animal research models are available for a multitude of diseases and conditions affecting the central nervous system (CNS), and large-scale CNS gene expression data exist for many of these. Although there are several models specifically for AD, each recapitulates different aspects of the human disease. In this study we evaluate over 500 animal models to identify those with CNS gene expression patterns matching human AD datasets. Approaches included a hypergeometric based scoring system that rewards congruent gene expression patterns but penalizes discordant gene expression patterns. The top two models identified were APP/PS1 transgenic mice expressing mutant APP and PSEN1, and mice carrying a GFAP mutation that is causative of Alexander disease, a primary disorder of astrocytes in the CNS. The APP/PS1 and GFAP models both matched over 500 genes moving in the same direction as in human AD, and both had elevated GFAP expression and were highly congruent with one another. Also scoring highly were the 5XFAD model (with five mutations in APP and PSEN1) and mice carrying CK-p25, APP, and MAPT mutations. Animals with the APOE3 and 4 mutations combined with traumatic brain injury ranked highly. Bulbectomized rats scored high, suggesting anosmia could be causative of AD-like gene expression. Other matching models included the SOD1G93A strain and knockouts for SNORD116 (Prader-Willi mutation), GRID2, INSM1, XBP1, and CSTB. Many top models demonstrated increased expression of GFAP, and results were similar across multiple human AD datasets. Heatmap and Uniform Manifold Approximation Plot results were consistent with hypergeometric ranking. Finally, some gene manipulation models, including for TYROBP and ATG7, were identified with reversed AD patterns, suggesting possible neuroprotective effects. This study provides insight for the pathobiology of AD and the potential utility of available animal models.

Fig 1. Congruent gene expression in APP/PS1 mice and human AD.

A: An RRHO heatmap [26] of the AD portrait (X axis) with the APP/PS1 mice (top ranked model) (Y axis) indicates a high matching of up-up and down-down genes (arrows). Color is–log transformed hypergeometric p-value showing the strength of the overlap as positive or negative enrichment. In X and Y axes the profiles of upregulated genes are shown in red and downregulated genes in blue (some clipping of highly significant downregulated genes occurred in the model). B: Enrichment analysis of common up and down regulated genes is shown using ShinyGO [28]. C: Congruent genes with the highest levels of protein-protein interaction (determined via STRING [29]) are plotted in Cytoscape [30]. Interactions are highlighted by lines. AD and model upregulated genes are shown in red and AD and model downregulated genes are shown in blue. Increased size of font for gene symbol reflects higher number of connections between genes. Common genes of interest include: GFAP, BDNF, GRIA2, GRIA1, GABRA1, GABRG2, SNAP25, and PTPRC.

Fig 2. Congruent gene expression in GFAP mice and human AD.

A: An RRHO heatmap [26] of the AD portrait (X axis) with the GFAP mutation mice (hippocampus; second ranked model) (Y axis) indicates a high matching of up-up and down-down genes (arrows). B: Enrichment analysis of common up and down regulated genes is shown using ShinyGO [28]. C: Congruent genes with the highest levels of protein-protein interaction (determined via STRING [29]) are plotted in Cytoscape [30]. Interactions are highlighted by lines. AD and model upregulated genes are shown in red and AD and model downregulated genes are shown in blue. Increased size of font for gene symbol reflects higher number of connections between genes. Common genes of interest include: GFAP, EGFR, TLR4, SNAP25, PTPRC, and SNCA.

Fig 3. Congruent gene expression in APP/PS1 and GFAP mice.

A: An RRHO heatmap [26] of the APP/PS1 (X axis) with the GFAP mutation mice (Y axis) indicates a high matching of up-up and down-down genes (arrows). B: Enrichment analysis of common up and down regulated genes is shown using ShinyGO [28]. C: Congruent genes with the highest levels of protein-protein interaction (determined via STRING [29]) are plotted in Cytoscape [30]. Interactions are highlighted by lines. Common upregulated genes are shown in red and AD and model downregulated genes are show in blue. Increased size of font for gene symbol reflects higher number of connections between genes. Common genes of interest include: GFAP, TLR4, PTPRC, ITGAM, and TYROBP. Note that the majority of highly connected genes between the two models are upregulated (in red). D: A Venn diagram [48] highlights overlapping genes in the AD portrait and the top two models.

Fig 4. Heat maps of congruent models with AD and common genes among models.

RRHO heatmaps [26] (human AD is X-axis) of additional congruent models with AD include: 5XFAD, APOE3 mutation with TBI; SOD1G93A strain, CK-p25 model, TDP-43 antisense treatment, MAPT mutation, infection with sporadic Creutzfeld-Jakob tissue, bulbectomized, PWScrm+/-, GNAS conditional KO, GRID2 KO, INSM1 KO, XBP1 KO, and CSTB KO (A). For the models shown and APP/PS1 and GFAP mutations, common genes with AD from at least 9 of the 16 models were identified (see S2 File) and ShinyGO [28] enrichment analysis was performed (B).

Fig 5. UMAP plotting of models with human AD.

UMAP [27] was used to plot (A) all models and (B) just the top 50 models with the AD portrait using 5862 differentially expressed genes. A shorter distance (spatial proximity) of a model with human AD reflects a better match. The X- and Y-axes reflect arbitrary embedding dimensions for A and B. In (A), the top 50 models closely align with the AD portrait. AD portrait circle size was increased 10% and replotted in top surface to emphasize location. In (B), APP/PS1 (orange) and GFAP models (yellow-orange) were congruent with (close spatially to) the human AD (black) as were the 5XFAD (light green), SOD1G93A strain (blue), and the CK-p25 (blue/green) model.