Divergence of human and mouse brain transcriptome
highlights Alzheimer disease pathways
Jeremy A. Millera,b, Steve Horvathc,d, and Daniel H. Geschwindb,c,1
aInterdepartmental Program for Neuroscience, Departments ofcHuman Genetics anddBiostatistics, andbProgram in Neurogenetics and Neurobehavioral
Genetics, Department of Neurology and Semel Institute, University of California, Los Angeles, CA 90095-1769
Edited by Joseph S. Takahashi, Howard Hughes Medical Institute, University of Texas Southwestern Medical Center, Dallas, TX, and approved June 1, 2010
(received for review December 10, 2009)
Because mouse models play a crucial role in biomedical research
related to the human nervous system, understanding the similari-
ties and differences between mouse and human brain is of
fundamental importance. Studies comparing transcription in hu-
To better characterize gene expression differences betweenmouse
gene coexpression network analysis on more than 1,000 micro-
arrays from brain. We find that global network properties of the
brain transcriptome are highly preservedbetweenspecies. Further-
more, all modules of highly coexpressed genes identified in mouse
were identified in human, with those related to conserved cellular
functions showing the strongest between-species preservation.
Modules corresponding to glial and neuronal cells were sufficiently
preserved between mouse and human to permit identification of
cross species cell-class marker genes. We also identify several robust
human-specific modules, including one strongly correlated with
whose hubs are poorly-characterized genes likely involved in Alz-
heimer disease. We present multiple lines of evidence suggesting
links between neurodegenerative disease and glial cell types in hu-
man, including human-specific correlation of presenilin-1 with oligo-
dendrocyte markers, and significant enrichment for known neuro-
degenerative disease genes in microglial modules. Together, this
work identifies convergent and divergent pathways in mouse and
human, and provides a systematic framework that will be useful for
understanding the applicability of mouse models for human brain
neurodegenerative disease|systems biology|evolution|metaanalysis|
in the case of neurological disorders. For example, whereas mouse
knock-in models of Huntington disease (HD) display many of the
same behavioral and pathological features seen in human HD (1),
none of the three highly penetrant, dominantly transmitted causes
of Alzheimer disease (AD), alone, produce AD-like pathology
and psychiatric diseases, such species differences are unlikely to be
molecular basis of phenotypic differences between mouse and hu-
man caused by even a single mutation is likely to require using
a systems biology approach and a genome-wide view.
has a reproducible, higher-order organization, and that knowing
this structure permits significant functional insights (3). So, to sys-
we created and compared separate transcriptional networks from
mouse and human brain tissue. Such a characterization in mouse
brain could orient future brain-related studies, by providing
a priori knowledge regarding the similarity of specific gene ex-
pression patterns between human and mouse. We created net-
lthough certain disease mutations cause comparable pheno-
types in mouse and human, the effects in mouse of any given
works by merging data from many microarray studies, and then
applying weighted gene coexpression network analysis (WGCNA)
(4–7). WGCNA elucidates the higher-order relationships between
of biologically related genes and permitting a robust view of
transcriptome organization (3, 5, 8, 9). Within groups of highly
coexpressed genes (“modules”) that comprise the core functional
units of the transcriptional network, WGCNA also identifies
the most highly connected, or most central genes within each
module, referred to as “hubs.” We find that both gene expression
and connectivity—the summation of coexpression relationships
for each gene with all other genes—tend to be preserved between
species.Furthermore, we find that many modules in human show
preserved expression patterns in mouse. This includes, for ex-
ample, all modules associated with core cellular processes, such
as ribosomal and mitochondrial function, consistent with pre-
vious results (10, 11).
We also find many between-species differences that provide
insight into human disease. First, we identify a human-specific
module that was originally associated with AD progression in an
earlier study of aging and AD (7), as well as another related
human-specific module containing GSK3β and tau, both of
which are implicated in AD and other dementias (12). Next,
we find that significant changes in network position between
the species may reflect a gene’s relationship to human-specific
disease phenotypes. For example, the AD-associated gene pre-
senilin 1 (PSEN1) exhibits poor between-species network pres-
ervation, showing strong transcriptional coexpression with
oligodendrocyte markers in human alone, suggesting that its
role in adult human brain has significantly diverged from its role
in mouse. We also find evidence for clustering of neurodegener-
ative disease related genes within microglial modules, highlighting
the potential role of this glial cell type in human neurodegene-
ration. To the best of our knowledge this is the first metaanalysis
to focus solely on brain-specific data, and can therefore provide
unique insight into similarities and differences in transcriptional
patterns between the human and mouse brain.
Constructing the Mouse and Human Networks. We reasoned that
comparison of coexpression networks between mouse and hu-
man could provide valuable insight into human brain disorders.
We sought to compile inclusive coexpression networks repre-
senting a general survey of brain transcription in both species
(Fig. 1A shows a schematic of network creation, and the SI Text
includes a glossary of network related terms). After careful data
filtering and preprocessing to eliminate outliers (3) (Materials and
Methodsand SI Text),our analysisincluded 1,066 samples from 18
Author contributions: J.A.M., S.H., and D.H.G. designed research; J.A.M. performed re-
search; and J.A.M. and D.H.G. wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
Freely available online through the PNAS open access option.
1To whom correspondence should be addressed. E-mail: email@example.com.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
| July 13, 2010
| vol. 107
| no. 28 www.pnas.org/cgi/doi/10.1073/pnas.0914257107
human and 20 mouse data sets, representing various diseases,
brain regions, study designs, and Affymetrix platforms (Table S1).
For each species, we created a network from these data, first by
calculating weighted Pearson correlation matrices corresponding
to gene expression, then by following the standard procedure of
WGCNA to create the two networks. Briefly, weighted correla-
tion matrices were transformed into matrices of connection
strengths using a power function (5). These connection strengths
were then used to calculate topological overlap (TO), a robust
and biologically meaningful measurement that encapsulates the
similarity of two genes’ coexpression relationships with all other
genes in the network (5, 13). Hierarchical clustering based on TO
was used to group genes with highly similar coexpression rela-
tionships into modules. In all, we found 15 modules in the human
network (Fig. 1B) and nine modules in the mouse network (Fig.
1C), which were used to guide our final module characterizations
(Materials and Methods).
Global Similarities Between Mouse and Human Brain Transcription.
We first compared general network properties to ensure that our
networks were reasonably matched. Both gene expression and
connectivity between networks were significantly preserved be-
tween the species (R = 0.60, P < 10−400for expression; R = 0.27,
P < 10−70for connectivity; Fig. S1). Expression levels were more
preserved than connectivity, consistent with our previous results,
indicating that connectivity is a more sensitive measure of evo-
lutionary divergence than differential expression (4). We also
found greater between-species expression correlations than pre-
vious microarray studies of mouse and human brain (14) as well
as liver, testes, and muscle (15) (R of approximately 0.45 for all
studies). These results suggest that our large data sets allow us to
recognize interspecies transcriptional similarities as well as, or
better than, previous methods. As a further validation of pre-
dicted network interactions on a global level, we compared gene–
gene connectivity based on TO to known protein–protein inter-
actions (PPIs), demonstrating a linear relationship between TO
and PPIs in both human and mouse (Fig S2 and SI Text). Genes
with high TO were much more likely to interact, consistent with
previous results from multiple species (3, 16).
Many Mouse and Human Network Modules Are Highly Similar. To
assess gene coexpression preservation between the species on
a module-by-module basis, we first calculated the module mem-
bership (MM)—a measure of how well each gene correlates with
the first principal component of gene expression within a module,
We then imposed a threshold based on MM values (R > 0.2, P <
10−13) to make final module assignments (Materials and Methods).
Using this method, each module contained an exact number of
assigned genes, and many genes were assigned to multiple mod-
ules, albeit with different strengths. We observed a high degree of
between-species module preservation (Fig. 2 and Materials and
Methods). In fact, all mouse modules showed significant overlap,
in terms of gene members, with at least one human module,
whereas there were multiple human-specific modules (Table 1 and
Fig. 1 B and C). Gene-by-eigengene tables containing MM and
initial module characterizations for all genes in both networks
are available at our website (www.genetics.ucla.edu/labs/horvath/
Modules identified here in mouse and human were both vali-
dated and annotated by comparison with a previous analysis of
thehuman brain transcriptome in cortex (3). This analysis showed
not only that all modules in our human network overlapped sig-
nificantly with those identified previously, but also that these
modules consist of biologically meaningful gene groups (Table 1,
SI Text, and Materials and Methods). Annotations using Gene
Ontology(GO) and Ingenuity pathway analysis (IPA)also showed
high concordance between matched human and mouse modules
(Table S2). Furthermore, modules associated with basic cellular
components showed the most significant between-species preser-
vation, as measured both by the number of overlapping genes and
by a summary measure of module preservation between networks
(Table 1 and SI Text) (18). Finally, markers for at least one neu-
ronal module (M11h) showed higher between-species preserva-
creation. Data sets were collected from GEO (1) and preprocessed separately
(2) creating similarly scaled expression files with the best probe set (P.S.)
chosen for each gene and the outlier samples removed (3) (SI Text). After
calculating Pearsoncorrelation matrices separatelyfor eachdataset(4), these
matrices were combined to form a single weighted correlation (corr.) matrix
for each species (5). Networks were created from the weighted correlation
matrices using WGCNA, by first calculating adjacency matrices (6), then cal-
culating TO (7) and using these values to hierarchically cluster genes into
coexpression modules (8) (Materials and Methods). Final module assignments
were made based on MM (9). (B and C) (Upper) Cluster dendrograms in the
human (B) and mouse (C) metaanalyses group genes into distinct modules
(step 8). The y-axes correspond to distance (1 − TO). (Lower) Dynamic tree
cutting was used to determine modules, generally by dividing the dendro-
gram at significant branch points. Modules with significant overlap were
assigned the same labels.
Creation of mouse and human networks. (A) Flowchart for network
thermore, modules with significant overlap tend also to have similar func-
tional characterizations (also see Table 1 and Table S2). Dots correspond to
modules from the mouse (light) and human (dark) networks. Line widths are
scaled based on the significance of overlap between corresponding modules.
Position of the dots and length of the lines are arbitrary to aid visualization.
There is high module overlap both within and between-species. Fur-
Miller et al.PNAS
| July 13, 2010
| vol. 107
| no. 28
tion than markers for astrocytes or microglia, consistent with cur-
rent knowledge of glial divergence between the species, in that
astroglia in human are more numerous and of higher complexity
than in other mammals (19).
Interspecies Convergence (Marker Genes) and Divergence for Cell
Types. Previous work from our group has found that coexpression
is a viable method for determining cell type markers (3). To es-
tablish interspecies markers in this analysis, in essence providing
additional confirmation of human–mouse preservation at the cel-
on annotation (Fig. 2, Table 1, and Table S2). We then identified
the top ranked genes based on the significance of MM in both
oligodendrocyte (M2h), neuron (M13h), astrocyte (M3h), and
microglia (M10h) were chosen as a starting point for comparison
(Table S3; a more extensive table is presented as TableS4). These
(3, 23) marker genes for each cell type, and validated by show-
ing highly significant overlap (P < 10−10for each cell type). To
provide further validation we also ran the converse assessment,
measuring to what extent our networks correctly identified well-
established marker genes, again finding generally positive results
allow readers to further screen for interspecies marker genes
(Table S4) or identify genes with significant between-species di-
vergence (Table S6).
Differences in Mouse and Human Modules Provide Insight into AD.
There is great precedent for the importance of studying the mo-
lecular evolutionary basis of phenotypic differences between
humans and other mammalian species on the transcriptional level
(24, 25). We reasoned that differences in network organization
human populations, such as AD. With this in mind, we identified
one highly human-specific module, M9h, showing significant
overlap with a module recently identified to correlate highly with
AD progression in another human data set (P < 10−13; compare
Fig. 3A vs. figure 3D in ref. 7). Both of these modules had four
matching hub genes (FBXW12, LOC152719/ZNF721, FLJ12151,
and ZNF160). Data from the Celsius database, a unique resource
encompassing many Affymetrix microarrays, confirms high coex-
pression of probe sets representing these four hubs (Fig. 3B) (26).
Although three of these hubs are of unknown function, ZNF160
is a known transcriptional repressor of TLR4, which contributes
to amyloid peptide–induced microglial toxicity (27), suggesting
a possible known molecular link between M9h and AD.
We performed two confirmatory analyses to ensure that the
correlation of this module with AD progression was not an artifact
of the specific samples or microarray platforms chosen: (i) an
analysis of human aging in the hippocampus, entorhinal cortex,
superior frontal gyrus, and postcentral gyrus using data from
Affymetrix microarrays (28); and (ii) an analysis of human AD in
the temporal cortex using Illumina arrays(29) (SI Text). We found
that, not only were modulescorresponding to M9h present in both
analyses (Fig. S3), but that they also showed significant positive
correlation with both age and AD progression (Fig. 3 C and D).
Further, we found that CXXC1 is a commonhub to both M9h and
on different platforms, and in different labs, these results confirm
the likely role of this module and its reproduced hubs in
human disease. A list of the top genes across M9h and its related
modules is presented in Table S7.
dementia (FTD) in humans, GSK3β and tau (12), were also rela-
tively central genes in another human specific module, M7h (Fig.
(3) based on number of overlapping genes (Table 1), but GSK3B
wasalso found to bea hub gene in both modules(see figureS4Kin
ref.3),further implicatingboththis hub andthis module indisease
processes. Thus, although both M7h and M9h fail to show signifi-
cant functional annotation (Table 1 and Table S2) and contain
many genes whose functions are unknown in the nervous system,
these two human-specific modules provide key targets for fur-
thering our understanding of neurodegenerative dementias.
Next, we assessed how transcriptional differences between
mouse and human brain networks can provide insight into disease
at the level of individual genes. Orthologous genes showing dis-
cordant expression patterns may indicate divergent regulation or
novel functions between species (32), and may be important reg-
ulators of brain function (14). We identified 67 validated, human-
Table 1. Characterization and preservation of mouse and human network modules
Top CTX module
Columns 5–7 represent the module from the CTX network in ref. 3 showing the highest overlap with each module in the human network, along with
associated characterization. Significance of overlap with the corresponding mouse module is presented in column 8. Column 9 measures module preservation
(bolded Z-scores indicate significant preservation). P values are corrected for multiple comparisons.
*The expected number of overlapping genes is presented in parentheses.
†Module characterization was inferred using other means.
‡Overlap with a network other than CTX in ref. 3 (M8m = CN network, M15m = CTX95 network).
| www.pnas.org/cgi/doi/10.1073/pnas.0914257107 Miller et al.
PSEN1—one of three known genes whose mutation causes fa-
milial AD (2)—is centrally positioned in the human, but not the
mouse, oligodendrocyte module (Fig. 4A). To quantify this ob-
servation, we measured the correlation between PSEN1 and my-
elin oligodendrocyte glycoprotein (MOG; the top interspecies
marker for oligodendrocytes and a known myelin sheath surface
(Fig. 4B). PSEN1 and MOG showed consistent positive correla-
tion only in human, providing strong evidence for species-specific
function and regulation of PSEN1. Expression patterns of other
prominent AD-related genes are presented in SI Text.
Systematic Evaluation of Neurodegenerative Disease and AD Genes.
more systematically, we accessed a public curated list (Jackson
Labs) (33) of approximately 5,000 known DGs (more precisely,
whenmutated).Fromthislist, wefoundthe subsetofDGsrelated
to neurodegeneration or dementia (dementia DGs; SI Text). We
then compared the module assignments of these dementia DGs
between species. In human, we found that most modules associ-
ated with cell types or basic cellular processes showed significant
over-representation of dementia DGs (Fig. 4C). While most cor-
responding mouse modules showed similar enrichments, both
genes. We confirmed this result for AD by measuring the overlap
between each module and a published list of AD genes (34). Only
the two human microglial modules showed significant enrichment
for AD genes (M8h, P = 0.002; M10h, P = 0.02), providing evi-
dence of another important species difference in glial cells. This
result is particularly striking given increasing evidence of a causal
role for neuroinflammation in AD pathogenesis, and considering
that the microglia is the resident innate neural immune cell (35).
the progression of AD, and likely other dementias.
Observed Network Differences Are Not Due to Confounding Factors.
Because the initial human samples involved several brain regions
that the human-specific modules were inadvertently related to
these disease samples. To control for this possibility, we con-
structed additional mouse and human networks using only the
subset of “control” (nondisease) microarrays from cortex of both
species, finding similarly preserved modules and the same be-
tween-species differences identified in the larger network analysis
(SI Text). Another potential confounder was agonal state, which
differs between mice and humans. To control for this factor,
we assessed the modules identified here for enrichment in genes
previously associated with agonal state in humans (SI Text) (36).
Agonal state genes were not concentrated in the human-specific
modules; rather, highly preserved modules between mouse and
human (including the mitochondrial modules) showed the most
significant enrichment for agonal state genes, indicating that ag-
onal state is not a significant source of the observed between-
species differences (SI Text).
Implications for AD and Neurodegenerative Disease Research. We
have used a systems biology approach to find a number of
between-species transcriptional differences relevant to neuro-
Text). This includes a human specific module related to AD pro-
gression, which implicates many genes of previously unknown
nervous system function in dementia. These genes include at least
three hub genes with zinc finger motifs (ZNF160, CXXC1, and
LOC152719/ZNF721) that are likely involved in transcriptional
regulation—in some cases with disease-related genes—providing
a new window of investigation into AD pathophysiology.
At the level of genes with known function, we highlight PSEN1
as an example, as (i) mutations in human PSEN1 cause a domi-
nant, highly penetrant form of AD, but only limited pathology is
seen in mouse mutants; and (ii) in the present study, PSEN1
shows high correlation with oligodendrocyte markers only in
human (Fig. 4 A and B). Recent studies suggest myelin dysfunc-
tion contributes to a wide range of psychiatric disorders, such as
schizophrenia and depression, and may be involved in normal
cognitive function, learning, and IQ (reviewed in ref. 37). In AD,
there is growing genetic evidence that myelin integrity may play
an early role, both in humans and in animal models (reviewed in
ref. 38)—evidence that is supported by recent imaging data (39).
In particular, oligodendrocytes located in brain regions most
vulnerable to AD pathology myelinate many axonal segments,
and are thought to be susceptible to AD risk factors, such as head
injury and high cholesterol (38). Dysfunction of these cells could
then lead to a progressive disruption of cell communication fol-
lowed by neurodegeneration in a predictable sequence. The cur-
rent results are consistent with this theory, suggesting that one of
Network depiction of M9h shows that M9h shares four hubs in common
with the red module from ref. 7, which contained genes whose expression
increased with AD progression; and one hub in common with a corre-
sponding module in a second study of AD (29) (Fig. S3A). Dots correspond
to genes and lines to connections, with the top 250 connections in M9h
shown. Larger genes correspond to hubs, which have at least 15 con-
nections. (B) These hubs show extremely high coexpression in both studies
and in the large Celsius database. Error bars represent SD of between-hub
correlation values. (C) The corresponding module from ref. 29 shows signifi-
cantly higher expression in AD than in control (CT). Bars represent mean ME
values over all CT/AD data sets and error bars represent SE. ***P < 10−11.
(D) M9h is also reproduced in a study of aging (28), in which the corre-
sponding module (Fig. S3B) shows positive correlation with age. Points rep-
resent ME expression for individual samples plotted against age. The gray
line indicates the line of best fit of the data. P < 10−5, R = 0.37. (E) Network
depiction of M7h, which is poorly characterized but contains both MAPT
and GSK3B (a prominent tau kinase). Labeling is as in A.
Human-specific modules provide insight into AD mechanisms. (A)
Miller et al.PNAS
| July 13, 2010
| vol. 107
| no. 28
the likely many distinctions between AD in mouse models and
humans is related to evolutionary changes in expression patterns
of PSEN1 in the context of neuron–oligodendrocyte interactions.
Combining WGCNA and Coexpression Metaanalysis in Brain. Several
studies comparing human and mouse transcription have been
published (10, 15, 40–43), coming to different conclusions about
divergence of the transcriptome. One reason for this may be that
changes in gene expression levels are not as sensitive as network
position (connectivity) to evolutionary divergence (3, 4). This
comprehensive network-based metaanalysis thus has a number of
advantages over traditional transcriptional analyses, leading to
more reliable results than in previous studies. First, we limit our
morethanone tissue (43,44); therefore, suchfiltering emphasizes
transcriptional correlations based on brain-specific gene func-
tions. Second, we include data from multiple studies across array
platforms, resulting in more functionally relevant coexpression
relationships (26, 45). Our unbiased preprocessing steps lead to
much higher comparability than previous studies, as measured by
correlation of ranked expression between species (Fig. S1). Third,
we compare data across species. Between-species coexpression
preservation has been shown to prioritize DG selection under
genetic disease loci (40) and to categorize the function of poorly-
characterized genes better than coexpression in a single species
(11). Finally, our data are combined at the level of correlation
matrices (rather than gene expression levels), which minimizes
coexpression relationships (47). Following this approach, we con-
structed networks using WGCNA, a method proven to produce
functionally relevant modules in a wide variety ofsituations (3–8).
Together, these strategies result in highly reproducible networks,
lending credence to the claim that our results are biologically
relevant and may provide important insights into disease.
Limitations and Future Work. Because this study represents a rela-
First, we were able to include only the 4,527 genes in common
between networks, whereas nearly 80% of all gene transcripts are
thought to be expressed in mouse and human brain (20). With
improvements in sequencing technology, future data sets should
allow for a more complete comparative analysis. Second, differ-
ences between human and mouse networks may be a result of
to be genuinely biological, rather than caused by confounders
such as agonal state or unbalanced sample selection. We have
addressed these issues by creating smaller networks using only
control microarrays from cortex, and have also shown that dif-
ferences in agonal state do not account for the between-
species differences (Results and SI Text).
Ourresults suggest that, alongside behavioral andphysiological
profiling, gene expression analysis could be a useful tool for eval-
uating mouse models of human neurodegenerative and neuro-
psychiatric disease. First, we identify several similarities between
the mouse and human transcriptomes, providing useful resources
for the study of mouse model systems. We also find multiple hu-
man-specific modules associated with dementia, including one
tau gene, which is mutated in the related condition, FTD (48).
and extensive between-species comparisons for any gene in which
expression data are available. For example, Creutzfeld–Jakob
disease is caused by mutations in the prion protein (PrP), which,
when inoculated into mice, recapitulate human neurodegenera-
tive phenotypes with more fidelity than single AD mutations,
consistent with the strong interspecies module preservation ob-
served here for PrP (PRNP—the gene for PrP—is in M14 in both
networks). Conversely, several genes involved in autism (49) show
significant between-species differences [i.e., CNTNAP2 is in the
human-specific module M7h (Fig. 3E) and CYFIP1 is a human-
specific hub in the microglia module (Table S6)], whereas others
disorders could be conducted (Tables S4 and S6). Overall, our
results suggest that information is present in transcriptional data
that should be used to aid in the understanding of neurodegen-
erative and neuropsychiatric disorders, and the corresponding
mouse models developed to study these diseases.
Materials and Methods
Data Set Acquisition and Network Formation. Mouse and human microarray
datasetswere downloadedfrom theGeneExpression Omnibus (GEO) (50). As
our goal was to compile an extensive set of comparable data, we collected as
many relevant data sets as we could find, and then subjected these data to
a stringent but unbiased filtering process (SI Text); we included only brain
samples from experiments run on Affymetrix platforms in our analysis, then
removed data sets with disproportionately low average within-species ex-
pression or connectivity correlation. These studies represent various diseases,
brain regions, study designs, Affymetrix platforms, and sample sizes, and
therefore represent a general survey of brain transcription (Table S1). From
these expression data we followed the protocols of WGCNA (3, 5) to create
within-species consensus networks for human and mouse (as described in
Results and SI Text; modified from ref. 47). This left a total of 9,778 genes in
the human analysis and 6,368 genes in the mouse analysis. Fig. 1A summa-
rizes this entire methodology.
Module Formation and Characterization. For the initial module characteriza-
tion, all but the top 5,000 connected genes in the human network (3,000 in
mouse) were excluded to decrease noise, leaving the most informative genes
for network formation (SI Text). Genes were hierarchically clustered and
modules were determined by using a dynamic tree-cutting algorithm (Fig. 1 B
and C) (51). Module identifiers in the mouse network were then changed to
For both networks, MM was calculated, then a threshold (R > 0.2; P < 10−13) was
used to establish final module assignments as described in Results. These
specific values were chosen such that on average each module contains ap-
proximately 5% of the genes present in our networks. This thresholding
hypergeometric distribution. Modules were graphically depicted using the
program VisANT (52) as previously described (4, 7). These network depictions
perties. (A) PSEN1 has high connectivity in M2h (the human oligodendrocyte
module). Labeling is as in Fig 3A. (B) PSEN1 and MOG show strong positive
correlation in all human data sets, but minimal correlation in any mouse data
set. Error bars represent SD of correlation values in each network. (C) Most
modules contain excess dementia-related genes, with microglial modules
showing the greatest between-species differences. The x axis corresponds to
mouse and human module labels (“X” indicates no mouse module). The y axis
corresponds to the percent of observed dementia DGs relative to the number
expected by chance (100%). *P < 0.05; **P < 0.001.
Multiple glial cell types show disease-relevant, human-specific pro-
| www.pnas.org/cgi/doi/10.1073/pnas.0914257107Miller et al.
15 depicted connections. It is important to note that with the module defi-
nitions based on MM, some genes can be members of multiple modules.
Network Comparisons. For all between-species network comparisons, human
orthologs of the mouse gene were used as proxy, and only the 4,527 genes
common to both networks were included. Nearly all of our comparative
analyses involving cross-tabulations were done using a hypergeometric dis-
tribution, whereby we tested if the number of overlapping genes between
one category and a given module was significantly large. For example, this
strategy was used to assess module overlap (Fig. 2 and Table 1), to confirm
interspecies markers for cell type (Table S3), and to compare modules
against a list of known dementia DGs (Fig. 4C and SI Text) (33). In the case of
our comparisons with modules from ref. 3, CTX modules were defined with
a cutoff MM significance of P < 0.05.
ACKNOWLEDGMENTS. We thank Peter Langfelder for providing valuable
Oldham for valuable discussions, and Brent Bill, Jeff Goodenbour, Jennifer
was supported by National Research Service Award F31 AG031649 from the
National Institute on Aging (NIA) (to J.A.M.); National Institutes of Health
(NIH) Grants U19 AI063603-01 and P01 HL028481 (to S.H.); NIH/National
Institute of Mental Health Merit Award R37 MH 60233-09S1 (to D.H.G); NIH/
NIA Award R01 AG26938-05 (to D.H.G); and Consortium for Frontotemporal
Dementia Research Award 108400 (to D.H.G).
1. Lin CH, et al. (2001) Neurological abnormalities in a knock-in mouse model of
Huntington’s disease. Hum Mol Genet 10:137–144.
2. Oddo S, et al. (2003) Triple-transgenic model of Alzheimer’s disease with plaques and
tangles: Intracellular Abeta and synaptic dysfunction. Neuron 39:409–421.
3. Oldham MC, et al. (2008) Functional organization of the transcriptome in human
brain. Nat Neurosci 11:1271–1282.
4. Oldham MC, Horvath S, Geschwind DH (2006) Conservation and evolution of gene
coexpression networks in human and chimpanzee brains. Proc Natl Acad Sci USA 103:
5. Zhang B, Horvath S (2005) A general framework for weighted gene co-expression
network analysis. Stat Appl Genet Mol Biol 4:e17.
6. Horvath S, et al. (2006) Analysis of oncogenic signaling networks in glioblastoma
identifies ASPM as a molecular target. Proc Natl Acad Sci USA 103:17402–17407.
7. Miller JA, Oldham MC, Geschwind DH (2008) A systems level analysis of transcriptional
changes in Alzheimer’s disease and normal aging. J Neurosci 28:1410–1420.
8. Konopka G, et al. (2009) Human-specific transcriptional regulation of CNS development
genes by FOXP2. Nature 462:213–217.
9. Winden KD, et al. (2009) The organization of the transcriptional network in specific
neuronal classes. Mol Syst Biol 5:291.
10. Bergmann S, Ihmels J, Barkai N (2004) Similarities and differences in genome-wide
expression data of six organisms. PLoS Biol 2:E9.
11. Stuart JM, Segal E, Koller D, Kim SK (2003) A gene-coexpression network for global
discovery of conserved genetic modules. Science 302:249–255.
12. Geschwind DH (2003) Tau phosphorylation, tangles, and neurodegeneration: The
chicken or the egg? Neuron 40:457–460.
13. Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabási AL (2002) Hierarchical organization
of modularity in metabolic networks. Science 297:1551–1555.
14. Strand AD, et al. (2007) Conservation of regional gene expression in mouse and
human brain. PLoS Genet 3:e59.
15. Xing Y, Ouyang Z, Kapur K, Scott MP, Wong WH (2007) Assessing the conservation of
mammalian gene expression using high-density exon arrays. Mol Biol Evol 24:1283–
16. Ge H, Liu Z, Church GM, Vidal M (2001) Correlation between transcriptome and
interactome mapping data from Saccharomyces cerevisiae. Nat Genet 29:482–486.
17. Horvath S, Dong J (2008) Geometric interpretation of gene coexpression network
analysis. PLOS Comput Biol 4:e1000117.
18. Langfelder P, Horvath S (2008) WGCNA: An R package for weighted correlation
network analysis. BMC Bioinformatics 9:559.
19. Oberheim NA, Wang X, Goldman S, Nedergaard M (2006) Astrocytic complexity
distinguishes the human brain. Trends Neurosci 29:547–553.
20. Lein ES, et al. (2007) Genome-wide atlas of gene expression in the adult mouse brain.
21. Cahoy JD, et al. (2008) A transcriptome database for astrocytes, neurons, and
oligodendrocytes: A new resource for understanding brain development and
function. J Neurosci 28:264–278.
22. Gan L, et al. (2004) Identification of cathepsin B as a mediator of neuronal death
induced by Abeta-activated microglial cells using a functional genomics approach. J
Biol Chem 279:5565–5572.
23. Albright AV, González-Scarano F (2004) Microarray analysis of activated mixed glial
(microglia) and monocyte-derived macrophage gene expression. J Neuroimmunol
24. Khaitovich P, Enard W, Lachmann M, Pääbo S (2006) Evolution of primate gene
expression. Nat Rev Genet 7:693–702.
25. King MC, Wilson AC (1975) Evolution at two levels in humans and chimpanzees.
26. Day A, Carlson MR, Dong J, O’Connor BD, Nelson SF (2007) Celsius: A community
resource for Affymetrix microarray data. Genome Biol 8:R112.
27. Takahashi K, Sugi Y, Hosono A, Kaminogawa S (2009) Epigenetic regulation of TLR4 gene
expression in intestinal epithelial cells for the maintenance of intestinal homeostasis. J
28. Berchtold NC, et al. (2008) Gene expression changes in the course of normal brain
aging are sexually dimorphic. Proc Natl Acad Sci USA 105:15605–15610.
29. Webster JA, et al.; NACC-Neuropathology Group (2009) Genetic control of human
brain transcript expression in Alzheimer disease. Am J Hum Genet 84:445–458.
30. Keshava Prasad TS, et al. (2009) Human Protein Reference Database—2009 update.
Nucleic Acids Res 37 (database issue):D767–D772.
31. Reid SJ, et al. (2004) TBP, a polyglutamine tract containing protein, accumulates in
Alzheimer’s disease. Brain Res Mol Brain Res 125:120–128.
32. Pao SY, Lin WL, Hwang MJ (2006) In silico identification and comparative analysis of
differentially expressed genes in human and mouse tissues. BMC Genomics 7:86.
33. Bult CJ, Eppig JT, Kadin JA, Richardson JE, Blake JA; Mouse Genome Database Group
(2008) The Mouse Genome Database (MGD): Mouse biology and model systems.
Nucleic Acids Res 36 (database issue):D724–D728.
34. Bertram L, McQueen MB, Mullin K, Blacker D, Tanzi RE (2007) Systematic meta-
analyses of Alzheimer disease genetic association studies: The AlzGene database. Nat
35. Meda L, et al. (1995) Activation of microglial cells by beta-amyloid protein and
interferon-gamma. Nature 374:647–650.
36. Atz M, et al. (2007) Methodological considerations for gene expression profiling of
human brain. J Neurosci Methods 163:295–309.
37. Fields RD (2008) White matter in learning, cognition and psychiatric disorders. Trends
38. Bartzokis G (2004) Age-related myelin breakdown: A developmental model of
cognitive decline and Alzheimer’s disease. Neurobiol Aging 25:5–18.
39. Ringman J, et al. (2007) Diffusion tensor imaging in preclinical and presymptomatic
carriers of familial Alzheimer’s disease mutations. Brain 130:1767–1776.
40. Ala U, et al. (2008) Prediction of human disease genes by human-mouse conserved
coexpression analysis. PLOS Comput Biol 4:e1000043.
41. Chen J, Xu H, Aronow BJ, Jegga AG (2007) Improved human disease candidate gene
prioritization using mouse phenotype. BMC Bioinformatics 8:392.
42. Tsaparas P, Mariño-Ramírez L, Bodenreider O, Koonin EV, Jordan IK (2006) Global
similarity and local divergence in human and mouse gene co-expression networks.
BMC Evol Biol 6:70.
43. Yanai I, Graur D, Ophir R (2004) Incongruent expression profiles between human and
mouse orthologous genes suggest widespread neutral evolution of transcription
control. OMICS 8:15–24.
44. Su AI, et al. (2004) A gene atlas of the mouse and human protein-encoding tran-
scriptomes. Proc Natl Acad Sci USA 101:6062–6067.
45. Lee HK, Hsu AK, Sajdak J, Qin J, Pavlidis P (2004) Coexpression analysis of human
genes across many microarray data sets. Genome Res 14:1085–1094.
46. Berg J, Lässig M (2006) Cross-species analysis of biological networks by Bayesian
alignment. Proc Natl Acad Sci USA 103:10967–10972.
47. Langfelder P, Horvath S (2007) Eigengene networks for studying the relationships
between co-expression modules. BMC Syst Biol 1:54.
48. Lee VM, Goedert M, Trojanowski JQ (2001) Neurodegenerative tauopathies. Annu
Rev Neurosci 24:1121–1159.
49. Bill BR, Geschwind DH (2009) Genetic advances in autism: heterogeneity and con-
vergence on shared pathways. Curr Opin Genet Dev 19:271–278.
50. Edgar R, Domrachev M, Lash AE (2002) Gene Expression Omnibus: NCBI gene ex-
pression and hybridization array data repository. Nucleic Acids Res 30:207–210.
51. Langfelder P, Zhang B, Horvath S (2008) Defining clusters from a hierarchical cluster
tree: The Dynamic Tree Cut package for R. Bioinformatics 24:719–720.
52. Hu Z, Mellor J, Wu J, DeLisi C (2004) VisANT: An online visualization and analysis tool
for biological interaction data. BMC Bioinformatics 5:17.
Miller et al.PNAS
| July 13, 2010
| vol. 107
| no. 28