Alzheimer's Integrative Biology

Published on January 2017 | Categories: Documents | Downloads: 46 | Comments: 0 | Views: 279
of 14
Download PDF   Embed   Report

Comments

Content


Resource
Integrated Systems Approach Identifies
Genetic Nodes and Networks in
Late-Onset Alzheimer’s Disease
Bin Zhang,
1,2,3,4,14,
* Chris Gaiteri,
4,14
Liviu-Gabriel Bodea,
5,14
Zhi Wang,
4
Joshua McElwee,
6
Alexei A. Podtelezhnikov,
7
Chunsheng Zhang,
6
Tao Xie,
6
Linh Tran,
4
Radu Dobrin,
6
Eugene Fluder,
6
Bruce Clurman,
8
Stacey Melquist,
6
Manikandan Narayanan,
6
Christine Suver,
4
Hardik Shah,
1,2
Milind Mahajan,
1,2,3
Tammy Gillis,
9
Jayalakshmi Mysore,
9
Marcy E. MacDonald,
9
John R. Lamb,
10
David A. Bennett,
11
Cliona Molony,
6
David J. Stone,
7
Vilmundur Gudnason,
12
Amanda J. Myers,
13
Eric E. Schadt,
1,2,3
Harald Neumann,
5
Jun Zhu,
1,2,3
and Valur Emilsson
12,
*
1
Department of Genetics and Genomic Sciences
2
Icahn Institute of Genomics and Multi-scale Biology
Icahn School of Medicine at Mount Sinai, 1425 Madison Avenue, New York, NY 10029, USA
3
Graduate School of Biomedical Sciences, Icahn School of Medicine at Mount Sinai, One Gustave L. Levy Place, New York, NY 10029, USA
4
Sage Bionetworks, 1100 Fairview Avenue North, Seattle, WA 98109, USA
5
Neural Regeneration Group, Institute of Reconstructive Neurobiology, University of Bonn, Sigmund-Freud-Str. 25, 53127 Bonn, Germany
6
Merck Research Laboratories, Merck & Co. Inc., 33 Avenue Louis Pasteur, Boston, MA 02115, USA
7
Merck Research Laboratories, Merck & Co. Inc., 770 Sumneytown Pike, WP53B-120 West Point, PA 19486, USA
8
Fred Hutch Cancer Research Center, 1100 Fairview Avenue North, Seattle, WA 98109, USA
9
Center for Human Genetic Research, Massachusetts General Hospital, 185 Cambridge Street, Boston, MA 02114, USA
10
GNF Novartis, 10675 John Jay Hopkins Drive, San Diego, CA 92121, USA
11
Rush Alzheimer Disease Center, Rush University Medical Center, Chicago, IL 60612, USA
12
Icelandic Heart Association and University of Iceland, Holtasmari 1, IS-201 Kopavogur, Iceland
13
Department of Psychiatry and Behavioral Sciences, Division of Neuroscience, Miller School of Medicine, University of Miami, Miami,
FL 33136, USA
14
These authors contributed equally to this work
*Correspondence: [email protected] (B.Z.), [email protected] (V.E.)
http://dx.doi.org/10.1016/j.cell.2013.03.030
SUMMARY
The genetics of complex disease produce alterations
in the molecular interactions of cellular pathways
whose collective effect may become clear through
the organized structure of molecular networks.
To characterize molecular systems associated with
late-onset Alzheimer’s disease (LOAD), we con-
structed gene-regulatory networks in 1,647 post-
mortem brain tissues from LOAD patients and
nondemented subjects, and we demonstrate that
LOAD reconfigures specific portions of the molecu-
lar interaction structure. Through an integrative
network-based approach, we rank-ordered these
network structures for relevance to LOAD pathology,
highlighting an immune- and microglia-specific
module that is dominated by genes involved in
pathogen phagocytosis, contains TYROBP as a key
regulator, and is upregulated in LOAD. Mouse micro-
glia cells overexpressing intact or truncatedTYROBP
revealed expression changes that significantly over-
lapped the human brain TYROBP network. Thus the
causal network structure is a useful predictor of
response to gene perturbations and presents a
framework to test models of disease mechanisms
underlying LOAD.
INTRODUCTION
Complex diseases such as late-onset Alzheimer’s disease
(LOAD) arise from the downstream interplay of DNA-sequence
variants and nongenetic factors that act through molecular
networks to confer disease risk (Schadt, 2009). Despite
decades of intensive research, the causal chain of mechanisms
behind LOAD remains elusive. In fact, there are no effective
disease-modifying or preventive therapies, and the only
available treatment remains symptomatic; meanwhile, the inci-
dence of LOAD is expected to double by 2050 (Brookmeyer
et al., 2007). Progress in LOAD research is fundamentally
limited by our reliance on mouse models of severe familial/
early-onset Alzheimer’s disease; therefore, our primary knowl-
edge of LOAD is in actuality based on the downstream effects
of three rare mutations in APP, PSEN1, and PSEN2 (Bertram
et al., 2010). Although such mouse models are necessary and
helpful, the cognitive deficits in these transgenic mice are
less severe than those in humans, and they do not exhibit
equivalent neurodegeneration, which is the most accurate
clinical marker of cognitive disease progression in humans.
Correspondingly, attrition rates from early discovery to late
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 707
drug development have been very high (Scha¨ fer and Kolkhof,
2008).
In contrast to the plethora of potential disease mechanisms
detected in humans with LOAD, the search for LOAD-modifying
interventions has focused primarily on compounds targeting
the amyloid-b pathway. Both biological risk factors, often related
to vascular health and psychosocial factors (Cechetto et al.,
2008; Qiu et al., 2010), as well as genetic susceptibility play a
critical role in the underlying pathophysiology of LOAD
(Bertramet al., 2010). APOE is still the best validated susceptibil-
ity gene accounting for at least 30% of the genetic variance in
LOAD (Corder et al., 1993). Genome-wide association studies
(GWAS) have identified several additional genetic risk loci for
LOAD that seem to cluster in patterns that suggest immunity
(CLU, CR1, CD33, EPHA1, MS4A4A/MS4A6A), lipid processing
(APOE, ABCA7), and endocytosis (PICALM, BIN1, CD2AP) as
important causal biological processes (Bettens et al., 2013).
More recently, low-frequency missense variants in APP and
TREM2 were found to confer strong protection or elevated risk
of LOAD (Guerreiro et al., 2013; Jonsson et al., 2012, 2013).
However, the overall contribution of these new common and
low-frequency variants to the heritability of LOAD is very small,
suggesting that a large fraction of the genetic variance beyond
the APOE risk still remains hidden. Can we clarify the pathology
of LOAD by zooming out to the pathway level to search for emer-
gent risk of many genomic contributions? If so, howcan we iden-
tify the key causal genes in these pathways?
In light of the complexity and elusiveness of LOAD pathogen-
esis, new approaches are needed to boost the probability of
identifying causal genes and pathways. Recently, we have
leveraged the molecular network structure that is reflected in
genotypic and gene-expression data to uncover biologically
meaningful gene modules involved in the development of com-
plex disease (Chen et al., 2008; Emilsson et al., 2008). Targeting
such causal networks in ways that restore them to a normal
state has been proposed as a path to treat disease (Schadt
et al., 2009), but this potential has never been realized for
LOAD. However, the complexity of these networks makes it
difficult to distinguish the causal from correlated disease effects
or how the causal regulators propagate their effects. To better
address this, we constructed molecular networks based on
whole-genome gene-expression profiling and genotyping data
in 1,647 autopsied brain tissues from hundreds of LOAD
patients and nondemented subjects. We identified numerous
modules of distinct functional categories and cellular specificity,
many showing a massive remodeling effect in the LOAD brain.
Next, we applied an integrative network-based approach to
rank-order these modules for relevance to LOAD pathology
and used a Bayesian inference to identify the key causal regula-
tors of these remodeled networks. For instance, we identified
eight causal regulators of the top-ranked immune/microglia
module, including TYROBP (a.k.a. DAP12) as the highest
ranking in terms of regulatory strength and differential expres-
sion in LOAD brains. We demonstrate through mouse microglia
cells overexpressing intact or truncated dominant-negative
TYROBP that downstream expression changes significantly
overlapped the human TYROBP brain network. This study pre-
sents many of the network advantages useful in identifying
and prioritizing pathways and gene targets involved in the
pathophysiology of LOAD.
RESULTS
Leveraging a Systems Approach to LOAD
We developed and applied an integrative network-based
approach to identify modules of genes associated with neurode-
generative disease (Figures 1A–1C). We processed 1,647 autop-
sied tissues fromdorsolateral prefrontal cortex (PFC), visual cor-
tex (VC), and cerebellum(CB) in 549 brains of 376 LOADpatients
and 173 nondemented healthy controls (Figure 1A). All subjects
were diagnosed at intake, and each brain underwent extensive
LOAD-related pathology examination. We note that the known
APOE genotype exposure was confirmed in the Harvard Brain
Tissue Resource Center (HBTRC) sample, showing an odds ratio
of 3.74 per copy ε4 allele (p = 4.1 3 10
À13
). Each tissue sample
was profiled for 39,579 transcripts representing 25,242 known
and 14,337 predicted gene-expression traits, and each subject
genotyped for 838,958 unique SNPs (Figure 1A). Unless other-
wise noted, gene-expression analyses were adjusted for age and
sex, postmorteminterval (PMI) in hours, and sample pHand RNA
integrity number (RIN). In the overall cohort of LOAD and nonde-
mented brains, the mean ± standard deviation (SD) for sample
PMI, pH, and RIN were 17.8 ± 8.3, 6.4 ± 0.3, and 6.8 ± 0.8,
respectively. Extensive analysis of the effect of covariates on
gene-expression variation in LOAD and nondemented brains
was carried out, as shown in Figure S1 (available online) and
described in the Extended Experimental Procedures. Here, we
used a robust linear regression model for covariate corrections
in all our gene-expression analyses (Experimental Procedures).
Results of traditional differential expression analysis demon-
strate that subsets of genes were up- or downregulated in
LOAD (Figure 2A). Consistent with the known progression and
regional severity of LOAD pathology (Braak and Braak, 1991),
we observed that the PFC region contained the greatest number
of differentially expressed genes (Figure 2B). Figure 2C summa-
rizes the clustering or colinearity of the various LOAD pathology
traits and age within the HBTRC cohort, resulting in distinct
groups of clinical pathology and age as separate clusters. For
instance, the number of significant correlations of expression
traits to neuropathology like Braak stage within the LOADpatient
group was highest in the PFC region (Figure 2D). Given these
observations and the fact that PFC is more commonly affected
in LOAD than CB and VC (Braak and Braak, 1991), a particular
attention was paid to this region in our strategy to rank-order
modules for relevance to LOAD. These massive data sets were
the basis of further method development with the aim to identify
and rank-order network modules and gene targets associated
with LOAD pathology (Figures 1A–1C). Results of these various
analysis steps are discussed in the sections that follow, and a
more detailed description of methods and statistical procedures
is found in the Extended Experimental Procedures.
Remodeling of the Molecular Interaction Structure in
LOAD Brains
For simultaneously capturing the intra- and interregional gene-
gene interactions in the LOAD or nondemented state, we
708 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
constructed multitissue coexpression networks consisting of the
top one-third (n = 13,193) of the most variable gene-expression
traits per brain region in individuals donating tissues from all
three regions (Extended Experimental Procedures). The multitis-
sue coexpression network in LOAD brains indicated strong
structurally segregated regions of the human brain molecular
interactome (Figure 3A), consisting of 111 modules and each
containing between 30 and 1,446 gene members (Figure 3A),
whereas the network generated from nondemented samples
has 89 modules ranging in size from30 to 2,278 genes. Figure 3B
highlights a direct comparison of the two topological overlap
matrices corresponding to the LOAD or nondemented associ-
ated network for a subset of 16 modules, demonstrating that
LOAD reconfigures specific portions of the molecular interaction
structure. To analytically detect and quantify this network reor-
ganization across the demented and nondemented states, we
developed a metric that we refer to as modular differential con-
nectivity (MDC) (Extended Experimental Procedures). MDC is
Figure 1. Sample Processing and the Integrative Network-Based Approach
(A) Five hundred and forty-nine brains were collected through the Harvard Brain Tissue Resource Center (HBTRC) from376 LOADpatients and 173 nondemented
subjects, and tissues extracted from three brain regions, the commonly affected PFC in LOAD and the less affected VC and CB (1). Each brain went through
extensive neuropathology examination, and all tissues were profiled for 39,579 transcripts, and every subject genotyped for 838,958 SNPs (2). These data sets
were the basis of the method development in the present study (3).
(B) From the microarray RNA expression data, we identified gene-expression traits showing individual variability in gene-expression traits as per brain region (1).
Next we computed the coregulation (connectivity) strength between genes, defined the appropriate connectivity threshold (2), and performed hierarchical
clustering analysis to construct the undirected coexpression network (3). Finally, we used brain eSNPs (Q) as causal anchors in the construction of directed
Bayesian networks to infer a causal relationship between nodes in the network (4). A variant of the underlying causality-scoring process here can be applied to
relationships among thousands of nodes to infer genome-scale networks.
(C) Comparison of LOAD and nondemented networks was performed to explore any effect on the molecular interaction structure associated with the disease.
Differentially connected modules in LOAD were investigated for their functional organization (1), module relevance to clinical outcome, as well as the enrichment
of brain eSNPs (2). Modules were rank-ordered (this figure does not showthe true rank-order) for their strength of the functional enrichment, module correlation to
neuropathology, and eSNP enrichment (3).
See also Figure S1 and Table S1.
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 709
the ratio of the average connectivity for any pair of module-
sharing genes in LOAD compared to that of the same genes in
the nondemented state and is a continuous measure ranging
from 0 to infinity. This module-centric measure of differential
connectivity between the two states is therefore fundamentally
different from the gene-centric analysis of previous studies that
applied hard cutoffs (Mani et al., 2008). Given the nature of
the coexpression network analysis, MDC > 1 indicates gain of
connectivity (GOC) or enhanced coregulation between genes,
whereas MDC<1 indicates loss of connectivity (LOC) or reduced
coregulation between genes. In extreme cases where MDC >>1,
e.g., the glutathione transferase (GST) module (Figure 3B), or
MDC << 1, e.g., the nerve myelination module (Figure 3B), the
corresponding genes do not form a coherent cluster in the
nondemented state or LOAD, respectively. Thus, new modules
are created in LOAD, whereas in other cases, a portion of the
network is completely disrupted. The statistical significance of
the MDC metrics was computed through the false discovery
rate (FDR) procedure as described in the Extended Experimental
Procedures. Based on 10% FDR, 54% of all modules showed
GOC, whereas 4.5% of modules exhibited LOC. The structures
of the remaining 41.5% of the modules were found to be
conserved across the LOAD and nondemented states by this
MDC measure. We note a negligible overlap of only 6%between
signatures of differential connectivity and standard differential
gene expression in LOAD brains, implying that the observed
disruption in coregulation of genes reflects a previously un-
tapped marker of LOAD neuropathology.
Functional Organization of the Network and Its
Relevance to LOAD Pathology
As observed in previous network-based studies (Chen et al.,
2008; Emilsson et al., 2008; Zhang and Horvath, 2005), we
find that brain gene expression is organized into modules of
distinct functional categories (Figure 3C). Overrepresentation
of canonical pathways and biological processes in modules
was measured through Fisher’s exact test (FET) and corrected
for number of modules and functional categories tested. Fig-
ure 3C highlights significant overrepresentation of functional
categories in modules showing GOC, LOC, or conserved con-
nectivity and containing at least 50 genes. The multifactorial
basis of LOAD neuropathology involves biological processes
active in both the central nervous system (CNS) and the meta-
bolic and vascular peripheral system that have often progressed
silently for many years (Huang and Mucke, 2012; Murray et al.,
2011). In fact, we find that multiple functional categories,
including the immune response, unfolded protein, vascular sys-
tem, extracellular matrix, neurogenesis (brain development),
glucose homeostasis, synaptic transmission, and olfactory sen-
sory perception categories in the GOC modules, are highly
enriched in the LOAD-associated modules (Figure 3C), whereas
the LOC modules are enriched for genes involved in nerve mye-
lination, cell cycle, g-aminobutyricacid (GABA) metabolism, and
neurotrophin signaling (Figure 3C). Many of these functional
categories have previously been implicated in LOAD and/or
CNS-related function (Ansari and Scheff, 2010; Cechetto et al.,
2008; Dodel et al., 2003; Luchsinger, 2008; Morawski et al.,
Figure 2. Differential Gene Expression in
LOAD Brains and Expression Correlation to
Braak Stage
(A) The heatplot shows the genes (n = 6457),
absolute mean-log ratio > 1.5 for each profile,
which most significantly differentiate disease
status in PFC. The legend to the right shows the
arrangement of samples with blue points denoting
LOAD (A), and red points denoting nondemented
state (N).
(B) The number of differentially expressed genes
in LOAD compared with controls per brain
region using Bonferroni adjusted p < 0.05 by
correcting for the number of probes tested (p %
2.46 3 10
À7
).
(C) Clustering analysis where the rows and
columns represent age, and 25 LOAD pathology
traits are arranged in a symmetric fashion and
sorted by the hierarchical clustering tree of the
correlation matrix. The color intensity signifies the
correlation strength between two traits (red posi-
tive and green negative). AT, atrophy; WMAT,
white matter atrophy; EL, enlargement.
(D) Number of genes showing significant expres-
sion correlation to Braak stage as measured per
brain region using Bonferroni adjusted p < 0.05
by correcting for the number of probes tested
(p %2.46 3 10
À7
).
See also Table S1.
710 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
2012; Schiffman et al., 2002), again reinforcing the complex
multifactorial basis of the underlying pathophysiology. The func-
tional categories enriched in the conserved modules included
‘‘muscle contraction’’ (actin-related system), coated vesicle,
cadherin, and zinc ion metabolism (Figure 3C).
CNS cell-type-specific gene signatures, from the Allen Brain
Atlas (http://www.brain-map.org/), were enriched in distinct
network modules as previously observed (Oldham et al.,
2008): neurons in the synaptic transmission modules (11-fold,
p = 3.7 3 10
À24
), astrocytes in the GABA biosynthesis module
(22 fold, p = 1.5 310
À15
), oligodendrocytes in the nerve myelina-
tion module (30 fold, p = 2.5 3 10
À30
), choroid plexus cell types
in the extracellular matrix module (35 fold, p = 3.9 3 10
À15
), and
microglia signatures responding to amyloid-b treatment (Walker
et al., 2006) in the immune module (10-fold, p = 4.5 310
À20
) (Fig-
ure 3C). In contrast to the GOC and LOC modules, conserved
modules were not enriched for any CNS-specific cell types (Fig-
ure 3C). Pathways enriched in the brain modules and not previ-
ously implicated in LOAD may therefore represent novel disease
mechanisms including, for instance, the glucuronosyl trans-
ferase activity and the dynein complex (Figure 3C). Moreover,
the comprehensive representation of gene-gene interactions
in the LOAD-associated networks can uncover novel gene
members in pathways already implicated in LOAD, thus allowing
us to work out a known pathologic mechanism in more detail
than ever before. In summary, the immune module shows the
statistically most significant functional enrichment of all modules
(Figure 3C) and as such may have a more comprehensive repre-
sentation of its respective pathway members.
Table S1 contains extensive information regarding the func-
tional enrichment and gene membership of modules containing
at least 50 unique gene symbols. We highlight some specific
findings of interest from Figure 3C: (1) The enrichment of path-
ways related to olfactory sensory perception in a LOAD-associ-
ated module is of interest given that the processing of olfactory
function is affected in subjects who are genetically at risk of
developing LOAD long before the symptoms of dementia are
manifested (Schiffman et al., 2002). (2) The APOE transcript is
located in the LOC module enriched for astrocyte signatures
and GABA metabolism, consistent with the observation that
astrocytes are the major source of APOE in the CNS (Boyles
et al., 1985). The close connectivity of APOE and GABA meta-
bolism in the brain network may therefore have some bearing
on the observation that GABA interneuron dysfunction is partic-
ularly severe in APOE4 carriers (Li et al., 2009). (3) The previously
identified macrophage-enriched metabolic network (MEMN) in
peripheral tissues and strongly supported as causal for a number
of metabolic and vascular traits related to obesity, diabetes,
and heart disease (Chen et al., 2008; Emilsson et al., 2008) is
remarkably enriched within the brain immune/microglia module
(3.9-fold, p = 2.4 3 10
À46
). This is of interest given the strong
epidemiological evidence for metabolic- and vascular-based
exposure on LOAD (Huang and Mucke, 2012; Murray et al.,
2011). (4) The postsynaptic density proteome in the human
neocortex of 748 proteins overrepresented with risk loci known
to underlie cognitive, affective, and motor phenotypes (Baye´ s
et al., 2011) is significantly enriched in the synaptic transmission
module (3-fold, p = 1.6 310
À32
). It is still unclear how and which
of these different biological processes mentioned above interact
to affect LOAD; however, it is likely that only a few downstream
mechanisms on which many diverse effects converge are caus-
ally related to LOAD (Huang and Mucke, 2012; Murray et al.,
2011). The accumulated data show a strikingly coherent organi-
zation of molecular processes in the LOAD-associated network.
The coexpression network structure, its changes between
nondemented and LOADbrains, and the genetic loci responsible
for the expression covariation behind these networks collectively
reflect molecular processes associated with LOAD. By linking
the network modules to clinical outcome or LOAD neuropa-
thology via a multiple regression analysis (Extended Experi-
mental Procedures), we can infer key molecular processes
associated with LOAD. A covariance matrix of the average
expression correlation (jrj) between 49 modules, comprised of
at least 100 probes, and 25 LOAD-related traits is shown in Fig-
ure 4A. We performed principal component analysis (PCA) to
estimate the module-trait correlation and used the FDR method
to assess the significance (see Extended Experimental Proce-
dures). Of all modules, the immune/microglia showed correlation
to the greatest number of LOAD-related neuropathology traits
(Figure 4B). Expression of the PFC immune/microglia module
correlated to atrophy levels in multiple brain regions, including
frontal cortex (r = 0.27, FDR = 0.018) and parietal (r = 0.20,
FDR = 0.016), temporal (r = 0.19, FDR = 0.022), and neostriatum
regions (r = 0.28, FDR = 3.3 3 10
À9
), as well as ventricular
enlargement (r = 0.17, FDR = 0.031). Several modules, however,
showed correlation to a more restricted type of neuropathology,
including the modules characteristic for the glucuronosyl trans-
ferase correlated to Braak stage (r = 0.18, FDR = 9.8 3 10
À5
),
NAD(P) homeostasis to Braak stage (r = 0.25, FDR = 1.4 3
10
À7
), neurogenesis to ventricular enlargement (r = 0.19, FDR =
5.1 3 10
À5
), and GST to ventricular enlargement (r = 0.22,
FDR = 4 3 10
À6
). The significance of functional enrichment in
modules and the number of neuropathology traits correlated
with modules were considered important criteria in rank-
ordering modules for their potential to affect LOAD.
Bayesian Networks and the Immune Module as an
Effector in LOAD
Causal probabilistic Bayesian networks were constructed and
used as an alternative approach to delineate potential regulatory
mechanisms. In order to establish a causal relationship or
dependency between nodes in the network, we constructed a
directed probabilistic Bayesian network through the application
of brain cis expression (e)SNPs as causal anchors. Because cis-
eSNPs are in linkage disequilibrium(LD) with causal variants that
affect the expression levels of a neighboring gene or they are the
causal variant themselves, they serve as an excellent source of
natural perturbation to infer causal relationships among genes
and between genes and higher-order phenotypes like disease
(Chen et al., 2008; Emilsson et al., 2008). We detected a total
of 11,318 unique cis-eSNP transcripts in the three brain regions,
at FDR of 10% (Figure S2A), which is the largest number of brain
eSNP transcripts detected to date in a single study (Webster
et al., 2009). The methodology to identify cis- and trans-eSNPs
is detailed in Extended Experimental Procedures, whereas Table
S1 lists all cis- and trans-acting eSNPs detected in the present
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 711
Figure 3. Multitissue Gene Coexpression Network in LOAD Brains
(A) The topological overlap matrix (TOM) plot corresponds to the LOAD multitissue coexpression network. The rows and columns represent the same set of the
top one-third (13,193) of the most variably expressed genes in each of the three brain tissues and states, expressed in a symmetric fashion and sorted by the
hierarchical clustering tree of the LOAD network.
(B) Individual TOM covariance matrices of 15 differentially connected and one conserved modules in LOAD (the upper right triangle of each module) versus that
in the nondemented state (the lower left triangle of each module). Differential connectivity (MDC) and FDR estimate is specified in each panel in parenthesis
(MDC, FDR).
(legend continued on next page)
712 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
study at FDR of 10%. There was between 70%and 80%sharing
of cis-eSNP transcripts across different brain regions, and
37% overlapped all brain regions (Figure S2A). Importantly, we
find a variable and often strong enrichment of brain eSNPs in
many of the LOAD-associated modules compared to all probes
on the array, suggesting the possibility that these variants
determine the differential connectivity observed in LOAD.
For instance, in the PFC region (Figure 4C), there were five
modules showing significant enrichment for cis-eSNPs, in-
cluding the unfolded protein (3.8-fold, p = 3.8 3 10
À81
), nerve
myelination (2.5-fold, p = 2.9 3 10
À40
), immune function (2.2-
fold, p = 4.3 3 10
À30
), GABA metabolism (2.7-fold, p = 2.3 3
10
À13
), and extracellular matrix (1.6-fold, p = 2.3 3 10
À7
)
modules (Figure 4C). The enrichment of cis-eSNPs in the differ-
entially connected LOAD modules in the VC and CB regions is
shown in Figure S2B. For the present study, however, a partic-
ular attention was paid to the cis-eSNPs for their applicability
as priors in the construction of Bayesian networks (Extended
Experimental Procedures and schematic Figure S3).
We constructed Bayesian networks for each coexpression
module. Although many of the LOAD-associated network
modules are of potential interest, the reconstruction of the
Bayesian network for the immune/microglia module is high-
lighted given that it has the strongest disease association based
on clinical covariates and network-associated properties: (1) sig-
nificant differential connectivity of the cortex-specific immune
modules in LOAD (MDC between 49% and 100% GOC at
FDR < 0.001); (2) the immune/microglia module showed the
most significant enrichment of functional categories; (3) the
highest degree of gene-expression correlation to several mea-
sures of LOAD neuropathology; (4) the PFC version of the
module was highly enriched for brain eSNPs. To increase the
predictive power of inflammation-related regulatory networks,
we further built up the directed Bayesian network for the inflam-
mation modules derived from the individual brain regions. Fig-
ure 5 highlights the interactions within and between the five
predominant immunologic families in the PFC-based putative
microglia module. To generate this roadmap to the complex
structure of the immune/microglia module, genes that were not
direct members of one of these five core pathways were
assigned to the family with which they have the greatest number
of causal interactions. The immune module was dissected into
five families representing functional immune pathways that
were labeled according to their main function as ‘‘complement,’’
‘‘Fc’’ for Fc-receptors, ‘‘MHC’’ for major histocompatibility com-
plex, ‘‘cytokines’’ for cytokines/chemokines, and ‘‘toll-like’’ for
toll-like receptors (Figure 5).
Highlighting the Microglia Pathway with TYROBP as
Causal Regulator
The Bayesian inference enabled us to compute the causal
regulators of the differential connectivity in individual modules,
defined as the genes controlling many downstream nodes in
the respective network (see Extended Experimental Proce-
dures). The causal regulators of the highest scoring immune/
microglia module were rank-ordered based on the number of
downstream nodes, i.e., the power of regulating other genes,
as well as differential expression in LOAD brains. Here, we
used a combined score as G
j
=
Q
i
g
ji
, where, g
ji
is the dis-
criminant value of a j in the case i and is defined as
ðmax
i
ðr
ji
Þ +1 À r
ji
Þ=
P
j
r
ji
(Duda et al., 2000). In comparison to
the average gene/node in a given network, the causal regulators
are expected to have a stronger effect on the clinical outcome as
they direct the expression of a significant portion of the network
module they reside in. The size of the gene membership for the
different regional-specific immune modules ranges from 386 in
CB to 1,108 in the PFC, with 247 of the genes in the CB detected
in all regions (p = 1 3 10
À19
). The identity of the key causal
regulators is somewhat variable across each brain-regional
version of the microglia module of which CTSC, HCK, TYROBP,
SERPINA1, S100A11, LY86, DOCK2, and FCER1G were com-
mon to all immune modules, regardless of brain region. Through
the combined ranking score based on regulatory strength and
differential expression in PFC of LOAD brains, TYROBP scored
the highest (Figure S4A). Table 1 lists the 20 top-ranking PFC
modules and their respective key causal regulators. Expression
of TYROBP is restricted to cells involved in the innate immunity,
including the microglial cells in the brain (Schleinitz et al., 2009).
Here, TYROBP was significantly upregulated in LOAD brains in
the HBTRC sample (1.18-fold, p = 0.028), and the direction of
this effect was replicated (1.17-fold, p = 5.1 3 10
À5
) in an inde-
pendent multicenter cohorts study (see Extended Experimental
Procedures and Figure S4B). Additionally, we observed a pro-
gression of TYROBP expression changes across mild cognitive
impairment (MCI) in the replication study (Figure S4B). Esti-
mating what constitutes a ‘‘large’’ or ‘‘small’’ change in gene-
expression levels is challenging in microarray analyses. We
note, however, that TYROBPwas the 124
th
most differentially ex-
pressed probe out of 48,803 probes assayed in the replication
study cohort. Moreover, TYROBP was more differentially ex-
pressed in LOAD brains than the classical markers of microglia,
AIF1 and CD68, indicating that there was not a relative downre-
gulation of TYROBP despite elevated microgliosis in LOAD
brains (Perry et al., 2010).
The majority of the common causal regulators were located
either in the ‘‘Fc’’ pathway and associated/clustered genes
(HCK, SERPINA1, S100A11, DOCK2, and FCER1G) or the ‘‘com-
plement’’ pathway (TYROBP) in the immune/microglia network
(Figure 5). Recent reports (we note that our submission predates
these reports) show a striking association of a low-frequency
DNA variant in TREM2 to LOAD (Guerreiro et al., 2013; Jonsson
et al., 2013). More specifically, TREM2 is known to associate and
signal via TYROBP, the key regulator of the immune/microglia
network activated in LOAD. Thus, our data-driven, network-
based approach places both TREM2 and TYROBP in a gene
network that literally unifies them with previous top GWAS risk
(C) Significant (FET p value after correcting for number of modules and functional categories/pathways tested) enrichment of functional categories in conserved
modules (left most panel), LOC modules (center panel), or GOC modules (right most panel). The y axis represents the Àlog(p value) of enrichment, whereas the
x axis denotes the number of genes per module. Each module contains at least 50 unique gene symbols.
See also Table S1.
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 713
Figure 4. Module Relevance to LOAD Pathology and Enrichment of Brain eSNPs
(A) Aheatmap of the correlations (jrj) between 49 module principal components (PCs) and 25 LOAD-related neuropathology traits. These modules contain at least
100 probes. AT, atrophy; WMAT, white matter atrophy; EL, enlargement.
(B) Number of significant module-dependent correlations to LOAD-related neuropathology of all differentially connected modules with at least 100 members and
showing significant correlation to at least a single neuropathology trait (see Extended Experimental Procedures). The total number of traits associated with a
module was used to rank-order modules for relevance to LOAD pathology.
(C) We tested the enrichment of brain eSNPs in the differentially connected modules of the multitissue coexpression network in LOADas per brain region. Here we
present a significant enrichment of brain eSNPs in many of the PFC modules. We used the FET analysis to access the significance of the overlap between each
module and cis-eSNPs, correcting for the number of modules tested.
See also Figure S2 and Table S1.
714 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
loci including MS4A4A, MS4A6A, and CD33 (Figure 5). These
newresults provide exciting convergent evidence for the specific
microglia network that we had directly implicated as activated in
LOAD and reinforce the potential causality of this pathway in
LOAD pathology. In fact, the dissection of the immune/microglia
module into distinct families and key causal regulators points
toward an important function of the microglia pathways involv-
ing genes of the ‘‘complement’’ and/or ‘‘Fc’’ network clusters.
EIF4EBP1
APOBEC3C
GLT25D1
MYO7A
SLC15A3
GRN
PFN1
DOK1
ADORA3
HHEX
ARPC1B
DOCK2 VAV1
CSF1R
SYK
FCER1G
BLNK
SUSD3
LILRA2
TREM2
LYN
TGFBR1
RBM47
HLA−DRB3
HLA−DRB1
HLA−DRB4
CD74
CMTM6
ADAM28
PTGS1
REL
PARVG
FYB
HLA−DPA1
HLA−DOA
HLA−DRA
HLA−DMA
HLA−DMB
HLA−DRB6
AIF1
FPR1
MYO1F
CMTM7
S100A11
PTPN6
APOB200R
RAC2
ALOX5
CSF3R
LAPTM5
SERPINA1
APOC1
IFI30
HCK
FPR2
ARHGAP30
CCRL2
IL17RA
CXCL16
TSPO
WDFY4
IL10RB
CCR1
TNFRSF1B
TNFSF13B
CCR5 MFNG
ABCC4
IL18
IL13RA1
IL16
IL10RA
CSTA MYD88
CLIC1
IRF8
C3AR1
TYROBP
ITGAM
SLC7A7
ALOX5AP
C3
LAT2
C1QB
C1QA
VSIG4
CASP1
CD33
ITGAX
PYCARD
CD4
TLR10
DOCK8
TLR7
RUNX1
CD14
TLR5
TLR2
SLC2A5
TBXAS1
TLR1
PTPRC TLR8
CTSS
CD86
BTK
NCKAP1L
TNFA
CX3CR1
C1QC
TCIRG1
ITGB2
MYO1G
IP8L2
HMOX1
complement
cytokines
MHC
toll-
like
Fc
MS4A4A
higher-opacity nodes are core members of group
square nodes are previously implicated in AD
node size ~ number of links
(inner circle) edge width ~ number of connections
GPX1
MS4A6A
Figure 5. The Bayesian Brain Immune and Microglia Module
A module that correlates with multiple LOAD clinical covariates and is enriched for immune functions and pathways related to microglia activity (PFC module
shown). (Inner networks) The PFC module is enriched in genes that can be classified as members of the complement cascade (‘‘complement’’), toll-like receptor
signaling (‘‘toll-like’’), chemokines/cytokines (‘‘chemokine’’), the major histocompatibility complex (‘‘MHC’’), or Fc-receptor system (‘‘Fc’’). The direction and
strength of interactions between these pathways are collected across all gene-gene causal relationships that span different pathways. The minimum line width
corresponds to a single interaction (MHC to toll-like) and scales linearly to a maximum of 17 interactions (Fc to complement). (Outer networks) Each color-coded
group of genes consists of the core members of the different families and genes that are causally related to a given family. Core family members of each pathway
are shaded darkly, whereas square nodes in any family denote literature-supported nodes (at least two PubMed abstracts implicating the gene or final protein
complex in LOAD or a model of LOAD). Labeled nodes are either highly connected in the original network, literature-implicated LOAD genes, or core members of
one of the five immune families. Node size is proportional to connectivity in the module. See also Figure S5.
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 715
Figure S5 (genes marked in red) highlights many of the key genes
in the pathogen phagocytosis pathways found in the immune/
microglia module. It is notable how comprehensive representa-
tion of specific signal transduction pathways is observed within
the two immune families of this module. The strategic network
position of TYROBP as a causal regulator of many genes mirrors
its bottleneck position in several microglia activation-signaling
cascades. Extrapolating from this data-driven interaction, it is
possible that TYROBP may be associated with neuronal pruning
activity of the complement system that may be reawakened in
LOAD via amyloid-b and tau aggregates (Perry et al., 2010). In
this manner, the network structure can become a data-driven
hypothesis generator for disease-relevant interactions.
Structure of Causal Networks Guides Differential
Expression in a Distance-Dependent Manner
To test our prediction that TYROBP can direct LOAD-associated
gene networks, we contrast both the molecular function and
genome-wide effects of TYROBP with those predicted by the
structure of causal networks inferred from human LOAD brains.
For this, microglia cells derived from mouse embryonic stem
cells were genetically modified by lentiviral vectors to overex-
press either full-length or a truncated version of Tyrobp that
lacks both intracellular immunoreceptor tyrosine-based activa-
tion motif (ITAM) motifs (Extended Experimental Procedures
and Figure S6). To assess the genome-wide gene-expression
changes in response to the perturbation of Tyrobp, we derived
gene-expression data from the RNA sequencing of mouse
microglia cell lines overexpressing (1) vehicle, (2) the full-length
Tyrobp, or (3) dominant-negative truncated Tyrobp. We identi-
fied 2,638 and 3,415 differentially expressed genes for the
overexpression of full-length Tyrobp and truncated Tyrobp,
respectively (Table S1), at FDR < 2.5%. Roughly one-third
(858 to 1,092) of these genes are found in the most variable
gene set in the brain data set used for the network reconstruc-
tion. The PFC variant of the human immune/microglia module
was highly enriched for genes that are differentially expressed
in the full-length or truncated Tyrobp experiments (p < 1 3
10
À15
) (Figure 6A). We projected results of RNA-sequencing
experiments onto a large Bayesian brain network of 8,000
nodes that contains the microglia module as well as many other
modules. In this large network, we could track differential
expression of genes that are predicted to be downstream of
TYROBP at various network path distances (Figure 6B). The
highest predictive power for differential expression is in the
primary neighborhood of the perturbed gene, and this power
decreases for genes that are farther away in the network. The
enrichment for differentially expressed genes in the network
neighborhood of TYROBP and strong negative correlation
between the fraction of confirmed targets and path distance
(r = À0.82, p = 4 3 10
À7
) (Figure 6B) show that our causal
network structure is a significant and useful predictor of
response to gene perturbations, even in a challenging cross-
species setting. Thus, both the structure and direction of links
Table 1. Top 20 Modules in PFC Ranked for Relevance to LOAD Pathology
Module Rank Top Functional and Cellular Category N PFC Genes MDC Highlighted Causal Regulators
a
Yellow 1 immune and microglia 1,102 1.49 TYROBP, DOCK2, FCER1G
Pink 2 glutathione transferase 113 92.67 GSTA4, ABCC2, TIMELESS
Gray 1 3 cell junction 51 0.82
b
ACBD5, LMAN1, MLL3S
Seashell 4 coated vesicle 278 1.29
b
KIFAP3, PCTK2, SNCA
Red 3 5 ribosome 50 24.93 RPS27, RPS18, PCBP2
Green yellow 6 unfolded protein 721 4.50 STIP1, HSPA1A, DOPEY1
Red 7 nerve myelination and oligodendrocytes 987 0.68 ENPP2, PSEN1, GAB2
Gold 2 8 axon growth repulsion 80 3.27 TUBB4, ACTL9, ACTG1
Tan 9 extracellular matrix and choroid plexus cells 700 2.88 SLC22A2, AGTR1, ZIC2
Gold 3 10 dynein complex 67 12.12 TEKT1, FANK1, HYDIN
Light yellow 11 mRNA cleavage 96 6.01 MED6, STATIP1, SFRS3
Brown 2 12 olfactory perception 77 25.51 PPP2R5A, C1ORF143, RNASE11
Dark cyan 13 steroid biosynthesis 110 1.39
b
LAMP2, P2RX7, MID1IP1
Khaki 14 GABA biosynthesis and astrocytes 267 0.29 GJA1, STON2, CST3
Grey 60 15 Ser/Thr kinase receptor 495 4.64 CREBBP, ABCC11, MDGA1
Purple 16 synaptic transmission and neurons 805 1.22 SNAP91, BSN, GLS
Green 4 17 cell cycle 50 0.33 DTL, UBE2C, BUB1
Honey dew 18 muscle contraction 128 1.10
b
RFX4, DGCR6, AQP4
Red 2 19 zinc homeostasis 83 1.17
b
MT1M, MT1JP, MT1P3
Beige 20 glucose homeostasis 95 12.64 AMPD1, EGR2, PDGFB
This table lists the top 20 rank-ordered modules consisting of at least 50 genes fromPFC, or if majority of genes are fromPFCin mixed modules with a
total of 50 genes or more. See also Table S1 and Figure S4.
a
Selected set of maximum three causal regulators per module.
b
MDC FDR > 10% and therefore not considered significant.
716 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
in these causal networks provide significant information on the
effects of complex signal transduction mechanisms.
The inferred network structure has significant predictive
power for nodes that are several links away from TYROBP. We
studied the enrichment of functional categories in the gene
sets responding to the Tyrobp perturbation experiments and
applied Bonferroni-corrected p values for statistical significance
(Extended Experimental Procedures). Approximately 99% of
the differentially expressed genes from the microglia overex-
pressing intact Tyrobp were downregulated compared to the
control vehicle. This set was enriched for genes involved
in RNA metabolism (p = 6.2 3 10
À5
) and cell-cycle mitosis (p =
2.7 3 10
À3
). In the microglia cells overexpressing the domi-
nant-negative truncated Tyrobp, 2,856 upregulated genes
were enriched for the vacuole/autophagy (p = 1.7 3 10
À8
) and
mitochondrion (p = 4.6 3 10
À4
), whereas 559 genes involved in
RUNX3
RPS6KA1
PPP1R18
ITGAM
SAMSN1
SLC1A5
C3
TCIRG1
TYROBP
GPX1
ADAP2 CXCL16 PYCARD
LHFPL2 MAF
KCNE3 TNFRSF1B SFT2D2 RGS1 IL18
HCLS1
NCF2 GAPT
SH2B3 GIMAP2
SLC7A7
RNASE6
ITGB2
APBB1IP C1QC RBM47
BIN2
HLX
CREB3L2
ITGAX ELF4
SPP1 ZFP36L2
LYL1LGALS9C FKBP15 TMEM106A LRRC33 CD4 PLEK
CD84
GAL3ST4
LOXL3 IGSF6 DPYD IL13RA1
ABCC4 CAPG
NPC2
STAT5A
CD37 TGFBR1
CYTL1
square nodes are previously implicated in AD
node size ~ number of links in full network
P
B
O
R
Y
T
m
o
r
f
m
a
e
r
t
s
p
U
P
B
O
R
Y
T
m
o
r
f
m
a
e
r
t
s
n
w
o
D
complement associated genes
cytokine associated genes
MHC associated genes
FC associated genes
toll-like associated genes
IL10RA
framed nodes DE (p<.05) TYROBP truncated
framed nodes DE (p<.05) TYROBP full-length
framed nodes DE (p<.05) both experiments
NCKAP1L
0 5 10 15 20 25
0
0.1
0.2
0.3
0.4
0.5 10
17
15
32
32
37
35
75
79
73
70
74
62
53
36
37
25
10
Distance from TYROBP in global Bayesian network (steps)
F
r
a
c
t
i
o
n

o
f

t
r
u
e

t
a
r
g
e
t
s
Digits
linear fit: R = -.82 p = 4.1e−7
significant at p<.01
65
39
78
50
35
19
4
number of genes at step-distance
A
B
Figure 6. Structure of Causal Networks
Guides Differential Expression in a Dis-
tance-Dependent Manner
(A) Within the microglia module, we showall genes
that receive direct or indirect causal inputs to/from
TYROBP. Genes that were differentially expressed
in either full-length or truncated Tyrobp experi-
ments are circled (p value < 0.05, n = 4/4/4
for control/truncated/full-length RNA-sequenced
samples). Possible reasons for differentially ex-
pressed (DE) predicted upstream genes are
mouse-human network differences, network in-
accuracy, or presence of feedback loops, which
are not represented in a Bayesian framework.
(B) We mapped results of RNA-sequencing ex-
periments onto a large Bayesian network of
8,000 nodes that contains the microglia module
as well as many other modules. In this large
network, we could track differential expression of
genes that are predicted to be downstream of
TYROBP at various network distances (link dis-
tances). There was a strong negative correlation
(r = À0.82, p = 4 310
À7
) between the differentially
expressed genes in the microglia and the path
distance from TYROBP in the brain immune
network.
See also Figure S6 and Table S1.
histone assemply (p = 1.6 3 10
À31
) were
downregulated. Moreover, the Tyrobp-
regulatory effect reflects a degree of
symmetry as 658 genes, related to the
vacuole/autophagy (p = 5 3 10
À3
), were
downregulated by active Tyrobp but
upregulated in cells expressing domi-
nant-negative truncated Tyrobp. These
findings are of interest because they
link the far downstream effects of
TYROPB to known molecular pathology
in LOAD, such as abnormalities in the
cell cycle, mitochondrion, and autophagy
(Coskun et al., 2004; Webber et al.,
2005). The accumulated data suggest
that TYROBP may be a therapeutic target in prevention of
neuronal damage in LOAD.
DISCUSSION
The construction of gene-regulatory networks in a large
sampling of human brain specimens has revealed many facets
of the molecular-interaction structure in LOAD, when compared
to that in nondemented brains. A comprehensive characteriza-
tion of gene-network connectivity and its regulation and
association to disease can provide critical insights into the
underlying mechanisms and identify genes that may serve as
effective targets for therapeutic intervention. For instance, tar-
geting genes that are the most central (highly connected)
may be more effective in disrupting disease-related networks
for the purpose of therapy, but that could be at the cost of
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 717
more adverse effects. In summary, the utility of network-based
approaches to complex disease includes the following: (1)
elucidating the biological function and molecular context of a
particular set of causal genes, (2) establishing a framework to
map interaction between genes and network modules, (3)
providing an objective filter for rank-ordering genes based on
connectivity or other network features, (4) defining dynamic
changes and corresponding causal regulators of the altered
network structure associated with disease condition, (5) identi-
fying modules and pathways causally related to disease, and
(6) revealing tissue-to-tissue interactions that can aid in the
identification of key target tissues for disease (Dobrin et al.,
2009). The present study utilizes many of these network advan-
tages to highlight and prioritize pathways and gene targets
causally related to LOAD.
Our network-based integrative analysis not only highlighted
the immune/microglia module as the molecular system most
strongly associated with the pathophysiology of LOAD but
also identified the key network regulators, including TYROBP.
In a separate in vitro study, we have found that the microglia-
expressed TYROBP is directly involved in amyloid-b turnover
and neuronal damage (our unpublished data). Of interest,
mutations in TYROBP or TREM2 cause Nasu-Hakola disease
(Bianchin et al., 2010), a rare Mendelian disease characterized
by bone reabsorption dysfunction and chronic inflammatory
neurodegeneration, leading to death in the fourth or fifth
decade of life. The exact pathomechanism underlying Nasu-
Hakola disease is still unclear, but it was hypothesized that
failure of proper microglial clearance is causal for the lethal
effect of neurodegeneration. Thus, dysfunctional immune/
microglia pathways might not be unique to LOAD. To test the
generalization of this concept, we explored the connection of
the immune/microglia module to Huntington disease (HD),
another neurodegenerative disease. HD pathology, caused by
expanded alleles of a variable stretch of trinucleotide (CAG)
repeat length in HTT (The Huntington’s Disease Collaborative
Research Group, 1993), features astrogliosis and neurodegen-
eration of the striatum, prefrontal cortex, and hippocampus.
We constructed molecular networks in the PFC from 194
HD patients genotyped for CAG allele size (see Extended
Experimental Procedures) and found that the PFC version of
the immune/microglia module was well conserved between
LOAD and HD in terms of gene annotation (75% overlap,
p value < 1 3 10
À300
). This module, however, did not show
any alteration in connectivity in HD brains compared to the
disease-free controls used in our LOAD study. Moreover,
through a PCA, we did not detect any gene-expression correla-
tion of the HD brain immune/microglia module to expanded
CAG repeat length (r = À0.05, FDR = 56%), a key biomarker
for predicting HD severity (Gusella and MacDonald, 2006).
Thus, based on the comparison to HD, the disease-related
effect of the immune/microglia module appears to be specific
to LOAD (and possibly Nasu-Hakola disease).
Immune activation in LOAD may have multifaceted activity:
long-term use of nonsteroid anti-inflammatory drugs (NSAIDs)
before onset of the disease decreases risk (Etminan et al.,
2003), and microglia effector function via interfering with reac-
tive oxygen production, cytokines, and complement cascade
members has been postulated to damage healthy neurons and
synapses (Cameron and Landreth, 2010). Close association
and positive feedback between amyloid-b and microglia
(Meyer-Luehmann et al., 2008) further cloud the cause and effect
relationships of inflammation to disease progression. Without a
causal framework for these observations, it is difficult to find
optimal molecular targets that direct LOAD inflammation. There-
fore, we integrated clinical factors with whole-genome genotype
and molecular trait data to identify a network module containing
several microglia-signaling cascades functionally related to
the reactive oxygen burst during pathogen phagocytosis. We
highlight the causal regulator TYROBP that exerts control over
multiple genes within this module and pathways involved in
LOAD, thus validating our network structure and its relevance
to LOAD pathology. This approach appears to offer insights for
drug-discovery programs that can affect neurodegenerative dis-
eases, such as LOAD.
EXPERIMENTAL PROCEDURES
Raw gene-expression data together with information related to demo-
graphics, disease state, and technical covariates are available via the GEO
database (GEO accession number GSE44772; GSE44768, GSE44770, and
GSE44771). A brief description of key methods and sample description are
provided below, whereas complete details are found in the Extended Experi-
mental Procedures.
Data Sets and Sample Processing
We compiled six disease- and tissue-specific gene expression data sets
consisting of 1,647 postmortem specimens from three brain regions (PCF
[BA9], VC [BA17], and CB) in LOAD and nondemented subjects recruited
through the HBTRC. Each subject was diagnosed at intake and via exten-
sive neuropathology examination. Tissues were profiled on a custom-made
Agilent 44K array of 39,579 gene-specific DNA probes, and each subject gen-
otyped for 838,958 SNPs.
Molecular Networks and Causal Regulators
We constructed both multitissue and single-tissue coexpression networks
from the top one-third (n = 13,193) of the most variably expressed genes in
each tissue and condition. We computed the MDC in LOAD brains as:
d
U
ðx; yÞ =
P
NÀ1
i =1
P
N
j =i +1
k
x
ij
P
NÀ1
i =1
P
N
j =i +1
k
y
ij
;
where k
ij
is the connectivity between two genes i and j in a given network, and
assessed the statistical significance through the FDRmethod. We constructed
causal probabilistic Bayesian networks from individual coexpression modules
and used brain cis-eSNPs as priors to infer directionality between nodes (see
Figure S3). For this, we identified 11,318 unique cis-eSNPs transcripts at
FDR of 10% (Extended Experimental Procedures), all listed in Table S1. The
Bayesian inference allowed us to compute the causal regulators of the differ-
ential connectivity in individual modules by examining the number of N-hob
downstream nodes.
Mouse Microglia Cultivation, Cell Transduction, and RNA
Sequencing
Genome-wide gene expression of messenger RNA (mRNA) from cultivated
microglia cells overexpressing intact or genetically modified TYROBP was
sequenced with a TruSeq Kit for RNA capture and HiSeq 2000 for the
sequencing. Read mapping was done with the TopHat (Trapnell et al., 2009)
RNA-seq aligner.
718 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.
ACCESSION NUMBERS
Raw gene-expression data together with information related to demo-
graphics, disease state, and technical covariates are available via the GEO
database (GEO accession number GSE44772; see also accession numbers
GSE44768, GSE44770, and GSE44771).
SUPPLEMENTAL INFORMATION
Supplemental Information includes Extended Experimental Procedures, six
figures, and one table and can be found with this article online at http://dx.
doi.org/10.1016/j.cell.2013.03.030.
ACKNOWLEDGMENTS
H.N. and L.-G.B. were supported by the Hertie-Foundation and the Deutsche
Forschungsgemeinschaft (FOR1336, SFB704, KFO177). H.N. is a member of
the DFG-funded excellence cluster ImmunoSensation. We thank Jessica
Schumacher and Rita Hass for technical assistance. We are grateful to the
HBTRCfor the generous gift of human postmortembrain samples. The authors
are also grateful to the participants in the Religious Orders Study and the
Memory and Aging Project. This work is supported by the National Institutes
of Health (R01 AG034504, R01 AG030146, P30 AG10161, R01 AG17917,
R01 AG15819, K08 AG034290, P30 AG10161, R01 AG11101, and
NS032765), the Illinois Department of Public Health, and start-up funding
from the University of Miami, Miller School of Medicine. J.M., A.A.P., C.Z.,
T.X., R.D., E.F., S.M., M.N., C.M., and D.J.S. are employees and shareholders
of Merck & Co., Inc. J.R.L. is an employee and shareholder of Novartis.
Received: April 11, 2012
Revised: October 17, 2012
Accepted: March 22, 2013
Published: April 25, 2013
REFERENCES
Ansari, M.A., and Scheff, S.W. (2010). Oxidative stress in the progression of
Alzheimer disease in the frontal cortex. J. Neuropathol. Exp. Neurol. 69,
155–167.
Baye´ s, A., van de Lagemaat, L.N., Collins, M.O., Croning, M.D., Whittle, I.R.,
Choudhary, J.S., and Grant, S.G. (2011). Characterization of the proteome,
diseases and evolution of the human postsynaptic density. Nat. Neurosci.
14, 19–21.
Bertram, L., Lill, C.M., and Tanzi, R.E. (2010). The genetics of Alzheimer
disease: back to the future. Neuron 68, 270–281.
Bettens, K., Sleegers, K., and Van Broeckhoven, C. (2013). Genetic insights in
Alzheimer’s disease. Lancet Neurol. 12, 92–104.
Bianchin, M.M., Martin, K.C., de Souza, A.C., de Oliveira, M.A., and Rieder,
C.R. (2010). Nasu-Hakola disease and primary microglial dysfunction. Nat.
Rev. Neurol. 6, 193–201.
Boyles, J.K., Pitas, R.E., Wilson, E., Mahley, R.W., and Taylor, J.M. (1985).
Apolipoprotein E associated with astrocytic glia of the central nervous system
and with nonmyelinating glia of the peripheral nervous system. J. Clin. Invest.
76, 1501–1513.
Braak, H., and Braak, E. (1991). Neuropathological stageing of Alzheimer-
related changes. Acta Neuropathol. 82, 239–259.
Brookmeyer, R., Johnson, E., Ziegler-Graham, K., and Arrighi, H.M. (2007).
Forecasting the global burden of Alzheimer’s disease. Alzheimers Dement.
3, 186–191.
Cameron, B., and Landreth, G.E. (2010). Inflammation, microglia, and
Alzheimer’s disease. Neurobiol. Dis. 37, 503–509.
Cechetto, D.F., Hachinski, V., and Whitehead, S.N. (2008). Vascular risk
factors and Alzheimer’s disease. Expert Rev. Neurother. 8, 743–750.
Chen, Y., Zhu, J., Lum, P.Y., Yang, X., Pinto, S., MacNeil, D.J., Zhang, C.,
Lamb, J., Edwards, S., Sieberts, S.K., et al. (2008). Variations in DNA elucidate
molecular networks that cause disease. Nature 452, 429–435.
Corder, E.H., Saunders, A.M., Strittmatter, W.J., Schmechel, D.E., Gaskell,
P.C., Small, G.W., Roses, A.D., Haines, J.L., and Pericak-Vance, M.A.
(1993). Gene dose of apolipoprotein E type 4 allele and the risk of Alzheimer’s
disease in late onset families. Science 261, 921–923.
Coskun, P.E., Beal, M.F., and Wallace, D.C. (2004). Alzheimer’s brains harbor
somatic mtDNA control-region mutations that suppress mitochondrial tran-
scription and replication. Proc. Natl. Acad. Sci. USA 101, 10726–10731.
Dobrin, R., Zhu, J., Molony, C., Argman, C., Parrish, M.L., Carlson, S., Allan,
M.F., Pomp, D., and Schadt, E.E. (2009). Multi-tissue coexpression net-
works reveal unexpected subnetworks associated with disease. Genome
Biol. 10, R55.
Dodel, R.C., Hampel, H., and Du, Y. (2003). Immunotherapy for Alzheimer’s
disease. Lancet Neurol. 2, 215–220.
Duda, R.O., Hart, P.E., and Stork, D.G. (2000). Pattern Classification, Second
Edition (New York: John Wily & Sons, Inc.).
Emilsson, V., Thorleifsson, G., Zhang, B., Leonardson, A.S., Zink, F., Zhu, J.,
Carlson, S., Helgason, A., Walters, G.B., Gunnarsdottir, S., et al. (2008).
Genetics of gene expression and its effect on disease. Nature 452, 423–428.
Etminan, M., Gill, S., and Samii, A. (2003). Effect of non-steroidal anti-
inflammatory drugs on risk of Alzheimer’s disease: systematic review and
meta-analysis of observational studies. BMJ 327, 128.
Guerreiro, R., Wojtas, A., Bras, J., Carrasquillo, M., Rogaeva, E., Majounie, E.,
Cruchaga, C., Sassi, C., Kauwe, J.S., Younkin, S., et al.; Alzheimer Genetic
Analysis Group. (2013). TREM2 variants in Alzheimer’s disease. N. Engl. J.
Med. 368, 117–127.
Gusella, J.F., and MacDonald, M.E. (2006). Huntington’s disease: seeing the
pathogenic process through a genetic lens. Trends Biochem. Sci. 31,
533–540.
Huang, Y., and Mucke, L. (2012). Alzheimer mechanisms and therapeutic
strategies. Cell 148, 1204–1222.
Jonsson, T., Atwal, J.K., Steinberg, S., Snaedal, J., Jonsson, P.V., Bjornsson,
S., Stefansson, H., Sulem, P., Gudbjartsson, D., Maloney, J., et al. (2012). A
mutation in APP protects against Alzheimer’s disease and age-related cogni-
tive decline. Nature 488, 96–99.
Jonsson, T., Stefansson, H., Steinberg, S., Jonsdottir, I., Jonsson, P.V., Snae-
dal, J., Bjornsson, S., Huttenlocher, J., Levey, A.I., Lah, J.J., et al. (2013).
Variant of TREM2 associated with the risk of Alzheimer’s disease. N. Engl. J.
Med. 368, 107–116.
Li, G.F., Bien-Ly, N., Andrews-Zwilling, Y., Xu, Q., Bernardo, A., Ring, K.,
Halabisky, B., Deng, C., Mahley, R.W., and Huang, Y. (2009). GABAergic inter-
neuron dysfunction impairs hippocampal neurogenesis in adult apolipoprotein
E4 knockin mice. Cell Stem Cell 5, 634–645.
Luchsinger, J.A. (2008). Adiposity, hyperinsulinemia, diabetes and Alzheimer’s
disease: an epidemiological perspective. Eur. J. Pharmacol. 585, 119–129.
Mani, K.M., Lefebvre, C., Wang, K., Lim, W.K., Basso, K., Dalla-Favera, R., and
Califano, A. (2008). A systems biology approach to prediction of onco-
genes and molecular perturbation targets in B-cell lymphomas. Mol. Syst.
Biol. 4, 169.
Meyer-Luehmann, M., Spires-Jones, T.L., Prada, C., Garcia-Alloza, M., de
Calignon, A., Rozkalne, A., Koenigsknecht-Talboo, J., Holtzman, D.M.,
Bacskai, B.J., and Hyman, B.T. (2008). Rapid appearance and local toxicity
of amyloid-beta plaques in a mouse model of Alzheimer’s disease. Nature
451, 720–724.
Morawski, M., Bru¨ ckner, G., Ja¨ ger, C., Seeger, G., Matthews, R.T., and
Arendt, T. (2012). Involvement of perineuronal and perisynaptic extracellular
matrix in Alzheimer’s disease neuropathology. Brain Pathol. 22, 547–561.
Murray, I.V., Proza, J.F., Sohrabji, F., and Lawler, J.M. (2011). Vascular and
metabolic dysfunction in Alzheimer’s disease: a review. Exp. Biol. Med.
(Maywood) 236, 772–782.
Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc. 719
Oldham, M.C., Konopka, G., Iwamoto, K., Langfelder, P., Kato, T., Horvath, S.,
and Geschwind, D.H. (2008). Functional organization of the transcriptome in
human brain. Nat. Neurosci. 11, 1271–1282.
Perry, V.H., Nicoll, J.A., and Holmes, C. (2010). Microglia in neurodegenerative
disease. Nat Rev Neurol 6, 193–201.
Qiu, C., Xu, W., and Fratiglioni, L. (2010). Vascular and psychosocial factors
in Alzheimer’s disease: epidemiological evidence toward intervention.
J. Alzheimers Dis. 20, 689–697.
Schadt, E.E. (2009). Molecular networks as sensors and drivers of common
human diseases. Nature 461, 218–223.
Schadt, E.E., Friend, S.H., and Shaywitz, D.A. (2009). A network view of
disease and compound screening. Nat. Rev. Drug Discov. 8, 286–295.
Scha¨ fer, S., and Kolkhof, P. (2008). Failure is an option: learning from unsuc-
cessful proof-of-concept trials. Drug Discov. Today 13, 913–916.
Schiffman, S.S., Graham, B.G., Sattely-Miller, E.A., Zervakis, J., and Welsh-
Bohmer, K. (2002). Taste, smell and neuropsychological performance of indi-
viduals at familial risk for Alzheimer’s disease. Neurobiol. Aging 23, 397–404.
Schleinitz, N., Chiche, L., Guia, S., Bouvier, G., Vernier, J., Morice, A., Hous-
saint, E., Harle´ , J.R., Kaplanski, G., Montero-Julian, F.A., and Ve´ ly, F. (2009).
Pattern of DAP12 expression in leukocytes from both healthy and systemic
lupus erythematosus patients. PLoS ONE 4, e6264.
The Huntington’s Disease Collaborative Research Group. (1993). A novel
gene containing a trinucleotide repeat that is expanded and unstable on
Huntington’s disease chromosomes. Cell 72, 971–983.
Trapnell, C., Pachter, L., and Salzberg, S.L. (2009). TopHat: discovering splice
junctions with RNA-Seq. Bioinformatics 25, 1105–1111.
Walker, D.G., Link, J., Lue, L.F., Dalsing-Hernandez, J.E., and Boyes, B.E.
(2006). Gene expression changes by amyloid beta peptide-stimulated human
postmortem brain microglia identify activation of multiple inflammatory pro-
cesses. J. Leukoc. Biol. 79, 596–610.
Webber, K.M., Raina, A.K., Marlatt, M.W., Zhu, X., Prat, M.I., Morelli, L.,
Casadesus, G., Perry, G., and Smith, M.A. (2005). The cell cycle in Alzheimer
disease: a unique target for neuropharmacology. Mech. Ageing Dev. 126,
1019–1025.
Webster, J.A., Gibbs, J.R., Clarke, J., Ray, M., Zhang, W., Holmans, P.,
Rohrer, K., Zhao, A., Marlowe, L., Kaleem, M., et al.; NACC-Neuropathology
Group. (2009). Genetic control of human brain transcript expression in
Alzheimer disease. Am. J. Hum. Genet. 84, 445–458.
Zhang, B., and Horvath, S. (2005). A general framework for weighted gene
co-expression network analysis. Stat. Appl. Genet. Mol. Biol. 4, e17.
720 Cell 153, 707–720, April 25, 2013 ª2013 Elsevier Inc.

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close