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

We re-analyzed data from the Harvard Brain Tissue Resource Center (HBTRC) consisting of 376 late-onset Alzheimer’s disease brain samples as well as 173 non-demented brain samples, harvested from the cerebellum (CBM), dorsolateral prefrontal cortex (Brodmann area 9; henceforth, PFC), and visual cortex (Brodmann area 7; VC), which has been previously described [26] (GEO: GSE44772). Based upon the HBTRC AD data, we previously built up a multi-tissue weighted gene co-expression network using WGCNA [26, 31]. In order to find the AD OL-specific modules, we tested the enrichment of each module for the gene signatures specifically expressed in OLs. To do this, we employed datasets on cell enrichment and identified the genes most highly expressed in myelinating OLs compared to the non-OL cell types [32]. Three modules, one with probes primarily derived from each of the three investigated brain regions were found to be the most enriched for the OL gene signature. We therefore refer to these as the “region-specific oligodendrocyte co-expression network” modules. We further inferred an interaction network for each OL-enriched and region-specific gene module by integrating gene expression and DNA genotypic data, as previously described [26]. Specifically, relationships between genes in each co-expression module were inferred based on conditional independence tests of gene expression in a Bayesian framework, and the presence of more expression single nucleotide polymorphisms (eSNPs) associated with a gene were used as priors to break Markov equivalence [33]. We refer to each of these networks as the “region-specific OL-enriched Bayesian interaction networks”. For robustness, we combined the genes in the three region-specific co-expression networks into a core OL gene set (COLGS), merged the three brain region-specific Bayesian interaction networks by a set union of directed links into a core OL-enriched Bayesian interaction network (COLBN). The gene symbols from this previous study were updated to current Hugo Gene Nomenclature Committee (HGNC) standards using the R package HGCNhelper (v. 0.3.1), and the two gene symbols in the OL-enriched networks that were still mapped to multiple symbols were mapped to the more common symbol (MARCH1 and LPAR1) for legibility. In order to identify key regulatory genes in this interaction network, we then performed key driver analysis on COLBN [34, 35], which uses network connectivity to infer regulatory importance scores for genes.

In order to find the enrichment of COLGS and other co-expression modules in Alzheimer’s risk factor genes, we converted the International Genomics of Alzheimer’s Project (IGAP) GWAS SNP-level data set [28] to gene-level p-value calls using VEGAS2 [36], and used the genes with significant association at a nominal p < 0.05 for further analysis. We then measured the enrichment of this 543 AD GWAS risk factor gene set in the 62 overall qualifying multiscale AD modules with at least 50 genes using Fisher’s Exact Test (FET).

Grey matter brain samples in 50 mg aliquots were harvested from the prefrontal cortex (PFC; Brodmann Area 10) from the autopsied brains of persons with a wide range of cognitive status at the time of death, ranging from no cognitive impairment to dementia, as well as a wide range of Braak scores, from 0 to 6. Liquid chromatography-tandem mass spectrometry (LC-MS/MS) was used to measure the abundance of peptides in each brain sample, from which a protein-level quantitation was estimated using MaxQuant (v1.5.3.30). We used WGCNA to define modules of proteins in the proteomics data, with a soft-thresholding power coefficient of 3. We next annotated these modules based on their relative FET enrichment in the human homologues of genes specifically expressed in each of the five major mouse brain cell types (see “Estimating brain cell type enrichment”).

Within this OL-enriched module, we performed modular differential connectivity analysis in order to assess the overall difference in between samples classified as non-AD (Braak 0–2) and AD (Braak 5–6), calculated using the mean difference in z-scores option in DGCA [37] (v. 1.0.1), with 10,000 permutation samples to assess significance. We used a Student’s t-test to measure differences in average expression between conditions, using the q value R package [38] to estimate false discovery rates in the context of the multiple hypothesis tests.

We found the enrichment of both mouse key driver knockout DEG signatures, aggregated across significant DEGs (p < 0.05, FDR < 0.3) from both the frontal cortex and cerebellum, as well as two human AD gene expression signatures in the gene ontology pathways. The prefrontal cortex (PFC) human AD DEG signature was estimated from the HBTRC cohort. We used the list of genes identified as previously identified as having significantly different RNA levels in brain samples from persons with high levels of Alzheimer’s neuropathology (Braak = 5–6) compared brain samples from persons with low levels of Alzheimer’s neuropathology (Braak = 0–2) via a Student’s t-test [26]. The second human AD DEG signature is from a previous study that identified genes with a significant trend in RNA expression changes across the severity spectrum of AD (control, incipient, moderate, and severe) in the hippocampus (HIPP) [40]. In this data set, we selected the genes as having an increasing or decreasing trend in expression across AD severity stages and an ANOVA p-value <0.05 as the HIPP AD DEG signature. We used the moduleGO function in DGCA [37] (version 1.0.1) to perform gene ontology (GO) enrichment analysis on these DEG sets, which leverages the GOstats (version 2.34) [41] and org.​Hs.​eg.db GO annotation (version 3.1.2) R packages. We filtered for those GO terms with less than 800 and greater than 100 gene symbols. We adjusted the enrichment p-values for all GO terms in each DEG set using the Benjamani-Hochberg method. In order to determine whether there was a similar pattern of dysregulation in the mouse key driver knockout models as in human AD, we next found the degree to which the compartment overlaps intersected, using the R package SuperExactTest (version 0.99.2) [42].

The primary goal of this study was to interrogate OL-enriched multiscale gene networks constructed from human AD postmortem brain samples (Fig. 1). We first identified co-expressed human gene modules, that were built from data obtained from three regions [e.g. prefrontal cortex (PFC), visual cortex (VC), and cerebellum (CBM)] from postmortem late-onset AD human brains samples from a large cohort of patients [26]. We further annotated these gene modules with additional gene sets to investigate their associations with OLs and AD. We then constructed Bayesian gene regulatory networks for the OL-enriched co-expression modules, identified key regulatory genes in the regulatory networks, and systematically validated the topological structures of the regulatory networks based on a series of gene perturbation experiments in vitro and in vivo. Specifically, in order to identify OL co-expression networks in AD, we tested each of the 62 co-expression modules with at least 50 genes for the enrichment of genes expressed specifically in each of five major brain cell types, i.e., astrocytes, endothelial cells, microglia, neurons, and myelinating OLs [32] (Fig. 2). We identified three co-expression modules with the strongest enrichment of OL genes (Fig. 2), which we subsequently combined, leading to a set of 1631 unique gene symbols, which we henceforth refer to as the core oligodendrocyte gene set (COLGS). Notably, because the three OL-enriched co-expression modules were each primarily identified in one of the three brain regions from the multi-tissue experiment, by combining them we sought to obtain a more sensitive measure of genes whose expression is associated with OLs in the context of AD across brain regions. COLGS is highly enriched for genes encoding proteins identified in the myelin proteome [43] (Fold Enrichment (FE) = 1.92, Fisher’s Exact Test (FET) p = 2.4e-15), and is also enriched for genes specifically expressed in each of oligodendrocyte precursor cells (FE = 2.8, p = 4.2e-17), newly formed oligodendrocytes (FE = 6.2, p = 1.0e-84), and myelinating oligodendrocytes [32] (FE = 7.4, p = 2.5e-87), indicating that genes in COLGS capture a wide spectrum of OL functions.

To systematically evaluate the AD genetics of COLGS, we identified 543 genes with nominally significant (p < 0.05) gene-level associations with AD based on a meta-analysis of AD genome-wide association study (GWAS) data by the International Genomics of Alzheimer’s Project (IGAP) [28]. We identified a significant enrichment of the AD risk genes in COLGS (FE = 1.71, p-value = 0.0004), due to the presence of BIN1, PICALM, NME8, SNX1, and other genes. Of all the individual co-expression modules, one OL-enriched module had the second strongest enrichment for AD GWAS risk genes, second only to an immune- and microglia-enriched module (Fig. 2, Table 1). We further highlighted the specific genes in COLGS that were also identified as AD risk genes (Fig. 3, Table 2). These results suggest that the COLGS contains a relatively large proportion of AD risk genes.

Next, we sought to examine the robustness of COLGS to a variety of sources of variance. We first used data from an independent study that also identified coexpression modules in human AD postmortem brain RNA expression samples [44]. Despite originating from a different brain region (the hippocampus), we found that 83% of the genes in the module with the strongest OL-enrichment in this data set overlapped with the members of COLGS (FE = 18.3, p = 9.1e-79). In order to measure the robustness of the co-expression network with respect to an alternative gene expression modality, we utilized a large AD proteomic data set from the prefrontal cortex (PFC; Brodmann area 10), which we corrected for batch, age of death, and sex. We used WGCNA [31] to identify gene modules in this proteomic data set, and found that the module with the strongest OL-enrichment (FE = 1.7, p = 0.0053), has 50% overlap with COLGS (FE = 10.7, p = 6.6e-58), including the AD risk factor BIN1 (Additional file 1: Figure S1, Additional file 2). These results suggest a robust co-regulation of the COLGS genes at both the transcript and protein levels in AD.

Our previous analysis showed that the three OL-enriched modules were each among the modules with the strongest loss of connectivity in AD samples on the RNA expression level [26]. We sought to extend these results by interrogating gene expression dysregulation in the OL-enriched proteomics module, which includes 150 proteins and is the most OL-enriched module. We found that this module has a significant decrease in correlation in AD samples (mean difference in z-transformed correlation = −0.466, empirical p-value = 0.0489; Fig. 4a), thus validating the loss of coordination among OL network genes in AD at the protein level. To further explore gene expression changes in this module in AD, we measured differences in protein expression of the module members, identifying 17 proteins up-regulated in AD and 7 proteins down-regulated in AD at FDR < 0.3 and p-value <0.05 (Additional file 3). Notably, we found that a BIN1 protein isoform was down-regulated in AD (t = −2.4, p = 0.019, FDR = 0.19), as well as an MBP protein isoform (t = −2.3, p = 0.023, FDR = 0.19). Therefore, we found that there are both variable changes in expression levels for individual proteins as well as a loss of overall coordination of OL network protein expression in AD.

We next sought to predict the gene-gene regulatory relationships among the COLGS genes in the RNA expression data, which profiles a wider set of genes than the proteomics data. Specifically, we constructed a Bayesian gene regulatory network for each of the three OL-enriched co-expression modules we identified by integrating RNA expression and genotype data from autopsied brain samples of persons with AD. As detailed in our previous studies [33, 34], Bayesian networks are a type of probabilistic causal network, providing a natural framework for integrating highly dissimilar genetic and gene expression data to predict regulatory relationships. We then combined the three OL Bayesian networks by a set union of directed links, leading to a more robust core oligodendrocyte-enriched Bayesian network (COLBN), and performed key driver analysis on COLBN to identify master regulatory OL genes in the context of AD (Fig. 1b; Table 3).

To interrogate the topology of the COLBN as well as how the dysregulation of COLBN key drivers could relate to AD, we identified an in vitro experiment that perturbed a key driver gene in the COLBN, MYRF, and examined how the predicted network structures correspond to its identified experimental targets. Specifically, we used data from a previous study that performed transcriptional profiling of cultured mouse OLs with a deletion of the myelination transcription factor Myrf (also known as C11orf9) [45]. The set of genes differentially expressed in the cells with a Myrf deletion compared with the control was significantly enriched in the 5-layer downstream neighborhood of MYRF in the COLBN (FE = 2.5, p = 7.5e-33; Fig. 5a, Fig. 6a-b), thus validating the topology of the subnetwork regulated by MYRF. The validated downstream targets of MYRF include PLD1, which encodes a protein that regulates the shuttling of APP and is associated with APP in AD brains [46, 47], as well as APLP1, which encodes a protein that has been shown to accumulate within neuritic plaques [48].

Next, we validated the network structure of COLBN in vivo using knockout mouse models of three of the top 40 key drivers in COLBN, Ugt8, Cnp, and Plp1, each of which have been found to cause axon pathology without substantial myelin structural alterations [21, 23, 49]. We performed RNAseq (GEO: GSE80437) on tissue samples of postnatal day 20 mice from the frontal cortex (FC) and cerebellum (CBM) in these three mouse models to determine their KO DEG gene signatures at FDR < 0.3 (Additional files 4 and 5, Additional file 1: Figure S2). We profiled RNA from the FC and CBM because these two regions were also profiled in the human AD postmortem brain tissue study, from which the COLBN was generated. We selected this time point because we wanted to focus on gene changes that occurred prior to the development of axonal pathology and therefore more directly would allow us to determine gene changes characteristic of the prodromic stage of the disease.We identified no DEGs for ΔCnp in the CBM and ΔPlp1 in the FC, while we found that all of the other DEG signatures were significantly enriched in the downstream neighborhoods of their corresponding driver genes (Fig. 6c-d). UGT8 encodes an enzyme that transfers galactose to ceramide to generate galactosylceramide, which makes up approximately one-fourth of myelin lipid dry mass [50]. UGT8’s 5-layer neighborhood is most enriched for the Ugt8 KO signature in the CBM (FE = 2.1, p = 2.4e-9). One validated downstream target of UGT8 is LIPA, which encodes a lysosomal cholesterol-metabolizing enzyme associated with genetic polymorphisms that affect plasma 24S–hydroxycholesterol/cholesterol levels in AD patients [51] (Fig. 5d). The key driver CNP encodes a protein that plays a role in OL process outgrowth and promoting axon survival [23, 52]. CNP’s 5-layer neighborhood is significantly enriched for the Cnp KO signature in the FC (FE = 1.7, p = 6.4e-6). One of the validated downstream targets of CNP is SEC14L5, which has been previously found to be down-regulated in CA1 in AD [53] (Fig. 5c). Finally, the key driver PLP1 encodes a protein that is the most abundant protein in CNS myelin sheaths [54], regulates OPC process outgrowth [55], and promotes axonal integrity [56]. PLP1’s 5-layer neighborhood is significantly enriched for the Plp1 KO signature in the CBM (FE = 2.0, p = 1.2e-4). A validated downstream target of Plp1 is Fgf1 (Fig. 5b), which encodes a protein that has been found to stimulate neuronal growth and promote remyelination [57] and has been found to have increased levels in the CSF in AD [58].

Overall, as with the in vitro perturbation data, the in vivo perturbation experiments strongly support the network structure of COLBN and pinpoint specific downstream targets to explain how dysregulation of the key drivers may help mediate AD pathology.

Since the key drivers of COLBN are predicted to orchestrate the expression of a network with a strong genetic association with AD, we hypothesized that the KO signatures of COLBN key drivers in vivo would mimic their gene expression dysregulation in human AD brains. In order to test this hypothesis, we performed gene ontology analysis on DEG signatures from postmortem brain tissue from both key driver KO mice and patients with AD. The DEG signatures in AD were derived based on samples from the PFC in the Harvard Brain Tissue Resource Center cohort [26] and samples from the hippocampus (HIPP) in the University of Kentucky Brain Bank cohort [40], because these regions are among those with the strongest AD pathology. Overall, the enrichment of DEGs from the human AD cases in these gene compartments show several similar dysregulation patterns as that seen in the key driver knockouts (Fig. 7a). For example, we identified a strong enrichment for genes in the gene ontology (GO) category “mitochondrial protein complex” in the down-regulated ΔCnp signature (FE = 10.6, p = 8.2e-22) and the down-regulated AD HIPP signature (FE = 6.7, p = 7.5e-11). We next found that the down-regulated ΔCnp and AD HIPP signatures intersected along with the mitochondrial gene set substantially more than expected due to chance (FE = 34, p = 2.1e-9; Fig. 7b), suggesting that the actual genes dysregulated in the mitochondria are similar between Cnp-KO mice and in the hippocampus of AD patients. For example, COX6A1, a mitochondrial-associated gene in which mutations are causative of peripheral neuropathy [59], was found to be down-regulated in both Cnp-KO mice and in the hippocampus of AD patients. We also identified a strong enrichment for genes in the GO category “ribosome” in the down-regulated ΔCnp signature (FE = 4.8, p = 3e-11), the down-regulated ΔPlp1 signature (FE = 8.1, p = 4e-18), and the down-regulated AD PFC signature (FE = 3.0, p = 0.002). We identified a significant overlap of all three of these down-regulated signatures and the ribosomal gene set (FE = 76, p = 0.013; Fig. 7c), which has limited power due to the small size of the overall ribosomal gene set. Since the Cnp and Plp1 knockout RNAseq profiling occurred prior to the typical age of onset of axon degeneration in these mice, the mitochondrial and ribosomal gene set dysregulation seen is suggestive of “presymptomatic” pathways that precede subsequent axonal and neuronal degeneration.

Finally, we also detected that there was a significant enrichment for genes in myelin sheath GO category in the up-regulated ΔPlp1 signature (FE = 4.3, p = 3e-8), the down-regulated ΔCnp signature (FE = 2.5, p = 0.002), and the down-regulated AD HIPP signature (FE = 5.2, p = 2e-9). However, the myelin sheath GO signature only contains 161 genes, which makes overlap analysis difficult. In order to use a more well-powered gene set, we utilized a larger set of 1778 genes that have been previously reported to be present in the myelin proteome [43]. In this gene set, we found a strong enrichment for the intersection of genes in the myelin proteome with down-regulated genes in Cnp-KO mice and human AD hippocampus (FE = 6.0, p = 2.1e-8; Fig. 7d). This shared set of down-regulated myelin genes includes CDK5, which encodes an enzyme whose activity is associated with tau pathology [60] and is essential for myelination [61]. Both Cnp-KO mice and patients with AD undergo axon damage in the absence dramatic ultrastructural changes in myelin structure [62], and our data shows that both conditions share a down-regulation of myelin-associated genes that may be associated with dysmyelination [22] and subsequent axon damage.

In order to validate the decreased expression of key myelin and mitochondrial genes in Cnp-KO mice, we measured the protein expression of BIN1 and GOT2 in Cnp-KO mice compared to WT mice. BIN1 is primarily expressed in mature OLs and white matter in AD patients [63]. Interestingly, the longer, central nervous system (CNS)-specific isoform of Bin1 has decreased expression levels in AD, although overall Bin1 transcript levels show increased expression [64]. This may reflect a proliferation of monocyte lineage cells in AD and a concomitant decrease in expression of Bin1 in other cell types, including OLs. BIN1 protein expression has been found to decrease in the myelin proteome of Cnp-KO mice compared to WT mice, but not in Plp-KO, Mag-KO, or Sept8-KO mice [65]. Consistent with this previous data, we found a significant down-regulation of BIN1 expression in the corpus callosum in Cnp-KO mice compared to WT mice (p = 0.0041 and p = 0.0007, Fig. 8a-b). We selected to profile the corpus callosum because it is an OL- and myelin-rich tissue, and therefore we reasoned that protein expression in this region more accurately would reflect the effect of Cnp-KO on OLs protein expression. No changes in the subcellular distribution of BIN1 were detected in Cnp-KO mice, which was primarily expressed in the cytoplasm of cells that were also stained by antibodies specific for the nuclear pan-OL lineage marker OLIG2 (Fig. 8c-e). Next, we measured the protein expression of GOT2, which is a mitochondrial protein that has been shown to be down-regulated in many AD gene expression studies [66], and has down-regulated transcript levels in Cnp-KO mice in our RNAseq expression profiles. Consistent with this data, we found a down-regulation of GOT2 protein levels in Cnp-KO mice (p = 0.0332; Fig. 8a-b). Although GOT2 has been identified in the myelin proteome, we found that it was primarily co-expressed with neurons (Fig. 8f–h), suggesting that the loss of Cnp expression may also affect mitochondrial gene expression within neurons. Taken together, these results validate that Cnp-KO mice have gene dysregulation in key myelin and mitochondrial proteins that is similar to that seen in AD brain samples.