Multiscale network modeling of oligodendrocytes reveals molecular components of myelin dysregulation in Alzheimer's disease

Abstract

Background: Oligodendrocytes (OLs) and myelin are critical for normal brain function and have been implicated in neurodegeneration. Several lines of evidence including neuroimaging and neuropathological data suggest that Alzheimer's disease (AD) may be associated with dysmyelination and a breakdown of OL-axon communication.

Methods: In order to understand this phenomenon on a molecular level, we systematically interrogated OL-enriched gene networks constructed from large-scale genomic, transcriptomic and proteomic data obtained from human AD postmortem brain samples. We then validated these networks using gene expression datasets generated from mice with ablation of major gene expression nodes identified in our AD-dysregulated networks.

Results: The robust OL gene coexpression networks that we identified were highly enriched for genes associated with AD risk variants, such as BIN1 and demonstrated strong dysregulation in AD. We further corroborated the structure of the corresponding gene causal networks using datasets generated from the brain of mice with ablation of key network drivers, such as UGT8, CNP and PLP1, which were identified from human AD brain data. Further, we found that mice with genetic ablations of Cnp mimicked aspects of myelin and mitochondrial gene expression dysregulation seen in brain samples from patients with AD, including decreased protein expression of BIN1 and GOT2.

Conclusions: This study provides a molecular blueprint of the dysregulation of gene expression networks of OL in AD and identifies key OL- and myelination-related genes and networks that are highly associated with AD.

Keywords: Alzheimer’s disease; BIN1; CNP; Causal network; Differential expression; Myelin; Oligodendrocyte; Proteomics; RNA sequencing; co-expression network.

Fig. 1

Workflow of the analyses performed in this study. Human postmortem AD brain tissue samples from multiple brain regions were used to construct coexpression networks (a) using Weighted Gene Coexpression Network Analysis (WGCNA). The oligodendrocyte/myelination enriched coexpressed gene modules from WGCNA were annotated by a variety of external data sets including DNA, RNA, proteomic, cell type, and proteome compartment data (b). Next, Bayesian gene regulatory networks were constructed based on the DNA and RNA postmortem human AD data (c). The Bayesian networks were used to identify key driver genes and several key drivers were perturbed in mouse models to identify their downstream targets (d). The gene signatures in response to the perturbations of the key driver genes were used to validate the network structure (e) and to compare with the differential expression patterns in human AD postmortem brains (f)

Fig. 2

A myelination/oligodendrocyte enriched module in the multi-tissue AD coexpression network is enriched for AD GWAS genes. The left panel shows a heatmap of the enrichments (BH-adjusted -log₁₀ p-values) of multiscale modules with at least 50 members in marker genes from each of the five major brain cell types, i.e. astrocytes, endothelial cells, microglia, neurons, and myelinating oligodendrocytes, derived from a previous study in mice [32]. The bottom panel shows the enrichment (BH-adjusted -log₁₀ p-values) of each corresponding module in the 543 AD risk genes derived from the IGAP AD GWAS study

Fig. 3

Key drivers and AD GWAS risk genes in the AD oligodendrocyte regulatory network. Visualization of the core oligodendrocyte Bayesian regulatory network (COLBN), where arrows refer to the predicted direction of interaction in the AD sample-derived network. Nodes corresponding to genes that are called as one of the top 40 key drivers in the network are larger sized, while nodes corresponding to genes that are one of the AD GWAS risk factors are colored pink

Fig. 4

An oligodendrocyte-enriched protein coexpression module is dysregulated in AD. a Heatmap of transformed correlations in non-AD samples (Braak <= 2; lower left) and AD samples (Braak > = 5; upper right) in the OL-enriched module consisting of 150 proteins. The transformation consists of taking the absolute value of Pearson correlation coefficients raised to the power of β, i.e., the soft-thresholding power coefficient of 3 used in coexpression network construction. b, c Expression levels for MBP (b) and BIN1 (c) in samples classified as non-AD (Braak <= 2), mild AD (Braak 3–4), and AD samples (Braak > = 5). Significance assessed using Student’s t-tests (* = p-value <0.05)

Fig. 5

In vitro and in vivo perturbations of key driver genes in the AD myelin/oligodendrocyte networks validated a number of predicted downstream targets. In these network plots, arrows refer to the predicted direction of interaction in the AD sample-derived core oligodendrocyte Bayesian regulatory network (COLBN). The presence of multiple arrows between two genes is because COLBN was constructed by a union of directed links of three networks from three brain regions. In each plot, the perturbed (i.e., knocked-down or knocked-out) gene is colored yellow, the genes significantly down-regulated in the samples with the driver perturbed are colored green, the genes up-regulated are colored red, and the genes with inconsistent expression changes (i.e., multiple probes corresponding to the same gene show opposite directions of changes in expression) are colored blue. The size of the node is proportional to the number of downstream nodes in the subnetwork. a Validation of the two-layer subnetwork regulated by MYRF using the differentially expressed genes (FDR < 0.3, p < 0.05) derived from a Myrf knockout experiment in cultured mouse oligodendrocytes. b, c, d Validation of the two-layer subnetworks regulated by PLP1, CNP, and UGT8 using the differentially expressed gene signatures (FDR < 0.3, p < 0.05) from the RNAseq data derived from our knockout experiments in mice

Fig. 6

Gene signatures from in vitro and in vivo perturbations of key driver genes in the AD myelin/oligodendrocyte network are significantly enriched in the predicted subnetworks regulated by the driver genes. a, b Fold-enrichment (a) and BH-adjusted -log₁₀ enrichment p-values (b) for the overlap between the Myrf in vitro perturbation signature and each n-layer network neighborhood regulated by MYRF in the core oligodendrocyte Bayesian regulatory network (COLBN). c, d Fold-enrichment (c) and -log₁₀ enrichment p-values (d) for the overlap between each in vivo perturbation signature and each n-layer network neighborhood regulated by the corresponding driver gene. The result was based on the in vivo knockout differentially expressed gene signatures experiments for Cnp, Plp1, and Ugt8 from the cerebellum (CBM) and/or the frontal cortex (FC)

Fig. 7

Perturbation of oligodendrocyte network key drivers in mice recapitulates key gene expression changes in human AD brain samples. a A heatmap of the -log₁₀ gene ontology (GO) enrichment p-values of core oligodendrocyte Bayesian regulatory network key driver knockout differentially expressed gene (DEG) signatures (left panel) and AD DEG signatures from the hippocampus and prefrontal cortex (right panel). The top 3 GO terms with between 100 and 800 gene symbols most enriched in each of the DEG signatures are shown. The p-values for each tested signature were adjusted using the Benjamini-Hochberg method. b, c, d Venn diagrams showing the intersections of genes encoding proteins associated with the GO terms “mitochondrial protein complex” (b), “ribosome” (c), and genes in the myelin proteome (d) with genes downregulated in various DEG signatures are shown

Fig. 8

Downregulation of AD-associated proteins in mice lacking the oligodendrocyte network key driver Cnp. a Representative western blots of BIN1, GOT2, CNP and alpha-TUBULIN stainings on Cnp-WT and Cnp-KO corpus callosum (CC) tissues (n = 3). b Quantification of BIN1 and GOT2 protein expression in Cnp-KO corpus callosum compared to Cnp-WT corpus callosum (protein expression normalized to alpha-TUBULIN expression for each sample, n = 3 per condition; p = 0.0041, p = 0.0007, p = 0.032). c Representative confocal image of post-natal day 60 sections of Cnp-WT and Cnp-KO corpus callosum, stained for OLIG2 (red), BIN1 (green) and DAPI (blue), scale bar = 300 μm. The bottom panel shows higher magnification of the double stainings (white arrowheads indicate double positive for BIN1 and OLIG2), scale bar = 50 μm. d Quantification of the proportion of each cell types expressing BIN1 in Cnp-WT and Cnp-KO corpus callosum (oligodendrocytes – OLIG2+ in gray, neurons – NeuN+ in white, astrocytes – GFAP+ in black; n = 3–4). e Quantification of the percentage of OLIG2+ cells expressing BIN1 in Cnp-WT and Cnp-KO corpus callosum (n = 3–4). f Representative confocal image of post-natal day 60 sections of Cnp-WT and Cnp-KO cortex, stained for OLIG2 or NeuN (red), GOT2 (green) and DAPI (blue), scale bar = 300 μm. The bottom panel shows magnification of the double stainings (white arrowheads indicate double positive for GOT2 and OLIG2 or NeuN, respectively), scale bar = 50 μm. g Quantification of the proportion of each cell types expressing GOT2 in Cnp-WT and Cnp-KO cortex (oligodendrocytes – OLIG2+ in gray, neurons – NeuN+ in white, astrocytes – GFAP+ in black; n = 3–4). h Quantification of the percentage of OLIG2+ cells and of NeuN+ cells expressing GOT2 in Cnp-WT and Cnp-KO corpus callosum (n = 3–4). * p < 0.05, ** p < 0.01, *** p < 0.001