Background
Alzheimer’s disease (AD) causes a progressive dementia that affects approximately 1/6th of people in the US age 75 and above [
1]. Although the number one risk factor for AD is advanced age, the reason for this remains unknown [
2,
3]. A wealth of evidence has emerged over the past decade to support the role of non-neuronal cells, especially astrocytes and microglia, in amyloid beta (Aβ) processing and AD pathogenesis [
4‐
6]. While less well studied, several lines of evidence have suggested that dysregulation of oligodendrocytes (OLs) and associated dysmyelination might be important in AD pathology. For example, human neuroimaging studies have shown that white matter changes occur early in AD and are predictive of disease status [
7‐
9]. In particular, MRI studies have detected white matter volume atrophy in multiple brain regions prior to changes in gray matter in AD progression [
10‐
12]. Further, post-mortem human pathological studies have demonstrated that the pattern of neurofibrillary tangle deposition in AD is highly correlated with the developmental pattern of myelination, with late-myelinated axonal tracts substantially more vulnerable to degeneration in AD [
13,
14]. Recent reports have highlighted the importance of OLs and myelin metabolic function for axonal health and transport capacity [
15‐
20], thereby suggesting the possibility that OL-driven axonal damage could precede a secondary demyelination in the pathogenesis of AD. In particular, there are several mouse models of ablation of the OL-associated myelin genes
Ugt8 [
21],
Cnp [
22,
23], and
Plp1 [
24,
25], in which axonal degeneration occurs in the presence of minimal ultrastructural myelin alterations and therefore are well suited to study altered OL gene expression, presumably leading to myeling dysfunction preceding the onset of neurodegeneration. To investigate the hypothesis that OL dysregulation in AD may be part of the underlying mechanism leading to neurodegeneration, we sought to employ a detailed molecular and systems-level analysis to provide a molecular substrate for the potential role of OLs in mediating the initial axonal damage.
In this study, we systematically examined and validated OL-enriched gene networks to uncover key genes and molecular signaling circuits of OLs in AD. We built upon AD-associated and OL-enriched networks constructed in a previous study of genetic, gene expression, and pathophysiologic data in late-onset AD [
26]. We constructed a union of the three OL-enriched modules from a multi-tissue AD co-expression network and found that it was strongly enriched for AD risk factor genes. Our OL-enriched consensus module includes genes encoding proteins associated with Aβ-production
PSEN1 and
BACE1, as well as the AD risk factor genes
BIN1,
PSEN1,
PICALM, and
UNC5C [
27‐
30]. We next built co-expression networks from a large-scale proteomics data set, identifying a strong loss of coordination among proteins in the most OL-enriched network, an interaction of this dysregulation and dementia status, as well as a down-regulation of key OL network genes, including
BIN1. We then used the OL modules to construct regulatory networks and found that the topological structures of our OL-enriched networks were validated through in vitro and in vivo perturbations of the predicted key regulatory genes in the networks. Further, transcriptomic analysis of brain tissue isolated from mice with a genetic ablation of three top key driver genes (
Ugt8,
Cnp, and
Plp1) recapitulated key aspects of the dysregulation in gene pathways related to myelination that are seen in human AD brains. We chose to profile the mice with these genetic ablations at an early stage of development (postnatal day 20), in order to detect alterations in gene pathways occurring during the process of myelination and prior to the onset of widespread axon degeneration in
Cnp-KO [
22] or
Plp1-KO [
24] mice. This approach allowed us to define alterations in OL gene expression occurring during the “pre-symptomatic” or prodromal phase of the disease, as opposed to reactive changes that might be consequent to axonal degeneration. We found that differentially expressed gene (DEG) signature in the
Cnp knockout (KO) mouse mimicked gene expression changes detected in AD brains at the early stages of the pathology, both at the gene pathway and individual protein level, thereby suggesting that dysregulation of OLs in general and
CNP in particular may play a key role in driving AD-associated gene expression changes.
Discussion
In this study, we employed an unbiased systems biology approach to characterize OL-enriched and AD-associated co-expression and regulatory molecular networks. We derived a core OL-enriched gene set (COLGS) that was highly enriched in AD GWAS genes. We found that this set of genes strongly overlapped with the corresponding genes from a protein co-expression module, which we found to be highly dysregulated in AD. We next constructed a core OL Bayesian regulatory network (COLBN) to dissect the causal relationships among the genes in COLGS. By employing a series of in vitro and in vivo perturbations of key driver genes in COLBN, we validated the predicted network structures in COLBN. We further showed that the knockouts of key drivers of COLBN in mice mimic the dysregulation of OL-associated compartments of genes in human postmortem AD brains. In particular, our data revealed a surprisingly strong, convergent gene expression effect of the knockouts
Cnp and
Plp1 on organelle-associated gene expression pathways, specifically in genes annotated for mitochondrial and ribosome functions. These organelles have also been reported to be dysregulated in axons in AD brains [
17,
67‐
69], suggesting that altered OL-axon communication may lead to dysregulated expression of ribosomal and mitochondrial genes and possibly play a role in contributing to AD axonal pathology. It is difficult to study prodromal changes in late-onset AD brains, because we are not able to determine a priori the individuals who would progress to AD. However, in mice with well-defined ablation of myelin genes also detected as key drivers of human gene network in AD, it is possible to define alterations that occur in OL and that precede frank neurodegeneration. For this reason the detection of similar gene changes in brain tissue samples of
Cnp-KO and
Plp1-KO mice prior to the development of any axonal pathology [
22,
24], allowed us to infer that the dysregulated expression of similar genes in murine samples and in human AD brains is suggestive of similar events occurring during the early part of the pathological cascade. Overall, this study improves our understanding of the molecular underpinnings of myelination and OLs in AD by identifying biologically relevant pathways, dissecting the causal relationships among the OL- and myelin-related genes, and implicating key driver genes in AD pathogenesis.
At the individual gene level, many of the genes in COLGS have been described as genetic risk factors associated with late-onset AD, including
BIN1 [
28],
PICALM [
28],
NME8 [
28],
UNC5C [
30], and
PSEN [
27]
. Notably,
BIN1 is the nearest protein-coding gene to the SNP with the second-strongest GWAS signal for AD, following
APOE [
28]. Histologically, BIN1 is primarily found in the brain at the nodes of Ranvier, consistent with its high RNA expression in OLs and its presence in the myelin proteome [
43,
70,
71]. In COLBN,
BIN1 is downstream of
ABCA2, a cholesterol transporter that has been associated with the risk of AD in many study populations [
28,
72‐
75]. Much of the literature about the role of
BIN1 in AD has focused on its roles in neurons and microglia [
76], but our data, in addition to another recent study [
63], suggest that its role in OLs should be explored further. In addition to genetic risk factors, our AD OL network also contains many genes encoding proteins that have been associated with AD pathophysiology (e.g.
, via Aβ production) including
PSEN1 [
77],
BACE1 [
77],
PLD1 [
46,
78], and
APLP1 [
79]. Consistent with the important role of
BACE1 in OLs suggested by our network,
BACE1 has been shown to play a key role in myelination [
80,
81]. The role of
BACE1 in OLs is of high relevance to AD, as mutations in the
BACE1-cleaving region of APP have been associated with a decreased risk of AD [
82], and β-secretase inhibitors intended to treat AD may have side-effects of myelin defects [
83]. A focus on the interaction targets of both
PSEN1 and
BACE1 within OLs using regulatory networks may be a fruitful avenue to identify treatment modalities that decrease deleterious Aβ production without causing off-target effects.
At the gene set level, our enrichment analysis of the AD co-expression modules shows that the three modules significantly enriched for AD risk genes are associated with three different cell types, i.e., microglia, OLs, and neurons (Table
1). Note that gene modules were identified based on the correlation between gene expression profiles in postmortem human AD brains, and they don’t necessarily correspond to any particular known cell types or biological processes due to interactions among cell types and biological processes. Therefore, we denote the co-expression modules by randomly selected color names in addition to their most enriched gene ontology term, to emphasize their multifaceted and highly context-dependent functions. The top ranked module was enriched for immune (microglia/macrophage) genes, consistent with recent reports that immune cells and in particular innate immunity plays a critical role in promoting AD [
76,
84‐
86]. However, it is imprudent to focus on a single cell type and ignore the interactions with other cells. For example,
TREM2, an established AD risk factor that is primarily expressed in immune cells [
87,
88], is also the causative gene of Nasu-Hakola disease, an early-onset subcortical dementia that presents with white matter demyelination [
89]. Mice lacking
Trem2 have been shown to have delayed myelin debris clearance, which may lead to increased microglia activation and thus demyelination and neuronal death [
90]. The dysregulation of myelin proteome genes that we observed in AD may contribute to pathologic inflammation, by increasing the available lipid pool for scavenging by microglia, which can activate microglia into a pro-inflammatory state [
91,
92]. Further investigation of cell type interactions in AD via network biology is a promising approach in addressing the underlying causes of AD.
Existing mouse models of AD tend to focus on Aβ and/or neuronal deficits in AD. For example, several mouse models of AD express genes with familial AD-causative mutations under the
Thy1 promoter [
93‐
95], which is a neuronal marker and will serve to restrict the pathologic changes to neurons. However, the data presented in this study and others suggest that the dysregulation of other cell types, including OLs, may play a role in AD. This opens up a need for mouse models of AD that can recapitulate the OL- and myelin-associated dysregulation in AD. In this study, we found that a mouse knockout model of one of the key drivers in the OL network,
Cnp, demonstrates a strikingly similar myelin and mitochondrial dysregulation pattern as is seen in brain samples of patients with AD. Notably, we did expression profiling prior to the onset of axon degeneration in
Cnp-KO mice [
22], to minimize the possibility that the myelin gene expression changes observed would be reactive to as opposed to premonitory for axon damage. Taken together, our data suggest that the
Cnp-KO mice may be a good model of the OL network gene dysregulation and dysmyelination that occurs in the brains of patients with AD. Therefore, therapeutic agents that are able to mitigate and/or prevent the dysmyelination and axon degeneration seen in
Cnp-KO mice are worthwhile of investigation as potential therapeutic agents for ameliorating cognitive deficits in patients with AD.
In this study, we focused on the molecular networks in AD in a brain cell type, OLs, which have not been widely studied in AD. Our network modeling approach uncovered a network of OL-associated genes that is enriched for AD GWAS genes, pathways through which dysregulation of this OL network may promote AD pathology, and key driver genes that orchestrate these pathways. Further work on our model for the role of OLs in AD may help to address why aging is the major risk factor for AD, since myelin maintenance and plasticity are known to become progressively less robust in normal aging [
62,
96]. In particular, preventing or reversing the dysregulation of key OL driver genes such as
CNP and downstream targets such as
BIN1 deserve further research as interventions to help to alleviate the progression of cognitive deficits in AD.
Acknowledgments
We would like to thank members of the Zhang and Casaccia labs for many fruitful discussions.