Genomic responses in mouse models greatly mimic
human inflammatory diseases
Keizo Takaoa,b and Tsuyoshi Miyakawaa,b,c,1
a
Section of Behavior Patterns, Center for Genetic Analysis of Behavior, National Institute for Physiological Sciences, Okazaki, Aichi 444-8585, Japan;
Core Research for Evolutional Science and Technology, Japan Science and Technology Agency, Kawaguchi, Saitama 332-0012, Japan; and cDivision of Systems
Medical Science, Institute for Comprehensive Medical Science, Fujita Health University, Toyoake, Aichi 470-1192, Japan
b
The use of mice as animal models has long been considered essential in modern biomedical research, but the role of mouse
models in research was challenged by a recent report that genomic
responses in mouse models poorly mimic human inflammatory
diseases. Here we reevaluated the same gene expression datasets
used in the previous study by focusing on genes whose expression
levels were significantly changed in both humans and mice.
Contrary to the previous findings, the gene expression levels in
the mouse models showed extraordinarily significant correlations
with those of the human conditions (Spearman’s rank correlation
coefficient: 0.43–0.68; genes changed in the same direction: 77–
93%; P = 6.5 × 10−11 to 1.2 × 10−35). Moreover, meta-analysis of
those datasets revealed a number of pathways/biogroups commonly regulated by multiple conditions in humans and mice. These
findings demonstrate that gene expression patterns in mouse
models closely recapitulate those in human inflammatory conditions and strongly argue for the utility of mice as animal models of
human disorders.
transcriptome analysis
| inflammation | sepsis | burn | trauma
T
he use of mice as animal models of human disorders has long
been considered essential for elucidating the underlying
mechanisms of disease, as well as for translational research from
bench to bedside. This notion was seriously challenged by the recent
report by Seok et al. (1) that genomic responses to different acute
inflammatory stressors are highly similar in humans but very poorly
reproduced in the corresponding mouse models. Seok et al. (1)
investigated gene expression changes in individuals with trauma,
burns, and endotoxemia and compared them with those in mouse
models of these conditions. Surprisingly, in their study there were
few correlations between the gene expression changes in the human
conditions and those in mouse models, whereas the gene expression
patterns were similar between humans with trauma, burns, and
endotoxemia. Based on their findings, Seok et al. (1) concluded that
higher priority should be focused on studying complex human
conditions directly rather than relying on mouse models to study
human inflammatory diseases. The study has drawn much attention
from researchers (2–10) and mass media (11–14) and has been cited
more than 360 times since its publication only 18 months ago. Altmetric analysis of the article indicates that it is among the top 1% of
all publications tracked by the system (as of January 20, 2014). Most
of the reactions were an expression of general concern regarding the
utility of mouse models for biomedical research.
In the present study, we reevaluated the same gene expression
datasets analyzed in Seok et al. (1) using more conventional
statistical methods. There are a few critical differences in the
analysis methods used between the previous study and ours.
First, we focused on genes whose expression levels were significantly changed in both humans and mice. The previous study
analyzed sets of genes that were significantly changed in the
human conditions regardless of the significance of the changes in
mouse models for comparison, which is not a conventional
method of comparing two gene expression datasets. Assuming
that mouse models would mimic only partial aspects of human
disorders, inclusion of genes that showed no significant response
www.pnas.org/cgi/doi/10.1073/pnas.1401965111
to the stimulus would generally decrease the sensitivity to detect
the responses shared by the disorders and their models. For this
reason, we excluded such genes from our analysis. Second, we
compared each of the conditions in a single mouse study independently with the human reference conditions. Mouse studies,
such as GSE7404 and GSE19668, included multiple conditions or
gene sets. For example, GSE19668 contains multiple datasets,
including those for two different mouse strains and multiple timecourse data points after infection. Because humans and mice are
expected to be quite different, the optimal conditions/parameters
that most closely mimic human conditions should be rigorously
searched and considered the best model when trying to establish
any animal model of a human disorder. Therefore, among such
multiple conditions in a mouse study we chose the gene set with the
highest similarity to the human reference condition and used this
set for further analyses. Third, we mainly used Spearman’s rank
correlation coefficient (or Spearman’s ρ), instead of Pearson’s correlation coefficient (R) or correlation coefficient of determination
(R2), because there is no reason to assume linearity and normal
distribution of fold changes or log-twofold changes of gene expression levels. Fourth, we used a bioinformatics tool, NextBio (15),
to conduct more sophisticated nonbiased statistical analyses of the
similarity between gene sets. NextBio employs a normalized ranking
approach, which enables comparability across data from different
studies, platforms, and analysis methods by removing dependence on
absolute values of fold change and minimizing some of the effects of
the normalization methods used and platform effects. We also used
NextBio to conduct meta-analyses of the human and mouse gene sets
to determine the pathways/biogroups that were commonly changed
in the human and mouse gene sets.
Significance
The role of mouse models in biomedical research was recently
challenged by a report that genomic responses in mouse
models poorly mimic human inflammatory diseases. Here we
reevaluated the same gene expression datasets used in the
previous study by focusing on genes whose expression levels
were significantly changed in both humans and mice. Contrary
to the previous findings, the gene expression patterns in the
mouse models showed extraordinarily significant correlations
with those of the human conditions. Moreover, many pathways were commonly regulated by multiple conditions in
humans and mice. These findings demonstrate that gene expression patterns in mouse models closely recapitulate those in
human inflammatory conditions and strongly argue for the
utility of mice as animal models of human disorders.
Author contributions: T.M. designed research; K.T. and T.M. performed research; K.T. and
T.M. analyzed data; and K.T. and T.M. 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.
1
To whom correspondence should be addressed. Email: miyakawa@fujita-hu.ac.jp.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1401965111/-/DCSupplemental.
PNAS Early Edition | 1 of 6
MEDICAL SCIENCES
Edited by Ruslan Medzhitov, Yale University School of Medicine, New Haven, CT, and approved June 11, 2014 (received for review January 31, 2014)
Human
Burn
GSE37069
15
ρ = 0.88
ρ = 0.82
ρ = 0.68
ρ = 0.43
ρ = 0.57
ρ = 0.51
N = 13225
P < 0.0001
N = 9149
P < 0.0001
N = 13223
P < 0.0001
N = 9145
P < 0.0001
N = 13222
P < 0.0001
N = 13222
P < 0.0001
Human
Trauma
0
GSE36809
-15
15
ρ = 0.72
ρ = 0.63
ρ = 0.50
ρ = 0.54
ρ = 0.51
N = 12566
P < 0.0001
N = 13212
P < 0.0001
N = 13210
P < 0.0001
N = 13215
P < 0.0001
N = 13219
P < 0.0001
Human
Sepsis
0
GSE28750
-15
15
ρ = 0.56
ρ = 0.48
ρ = 0.48
ρ = 0.52
N = 13223
P < 0.0001
N = 13216
P < 0.0001
N = 13224
P < 0.0001
N = 13215
P < 0.0001
Mouse
Burn
ρ = 0.47
ρ = 0.54
ρ = 0.57
N = 13224
P < 0.0001
N = 13224
P < 0.0001
N = 13212
P < 0.0001
0
GSE7404
-15
15
Mouse
Trauma
0
GSE7404
-15
15
ρ = 0.50
ρ = 0.51
N = 13224
P < 0.0001
N = 13211
P < 0.0001
Mouse
Sepsis
0
GSE19668
-15
15
Mouse
Infection
0
-15
-15
ρ = 0.57
N = 13210
P < 0.0001
GSE20524
0
15 -15
0
15 -15
0
15 -15
0
15 -15
0
15 -15
0
15
Fig. 1. Correlations of gene changes among human burns, trauma, sepsis,
and the corresponding mouse models. Scatterplots and Spearman’s rank
correlations (ρ) of the fold changes of the genes responsive to both conditions for each pair of interest (P < 0.05; fold change >1.2). Vertical bar and
horizontal bar for each panel represents fold change in right and upper
panels, respectively. Murine models were highly significantly correlated with
human conditions with Spearman’s correlation coefficient (ρ = 0.43–0.68; P <
0.0001 for every comparison between human conditions and mouse models).
The correlations between different mouse models were also significant (ρ =
0.47–0.57; P < 0.0001 for every comparison).
2 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1401965111
Human Burn vs. Mouse Burn
(GSE7404)
50
Overlap P value:
40
3.9 × 10-34
30
20
B
20
15
5
0
0
940 genes 854 genes 267 genes 397 genes
Human↑ Human↓ Human↑ Human↓
Mouse↑ Mouse ↓ Mouse↓ Mouse ↑
Human Burn vs. Mouse Sepsis
(GSE19668)
50
Overlap P value:
40
1.2 × 10-35
30
20
D
20
593 genes 591 genes 482 genes 816 genes
Human↑ Human↓ Human↑ Human↓
Mouse↑ Mouse ↓ Mouse↓ Mouse ↑
Human Burn vs. Mouse Sepsis
(GSE26472)
Overlap P value:
-log (p-value)
60
6.3 × 10-13
10
10
C
Human Burn vs. Mouse Trauma
(GSE7404)
Overlap P value:
-log (p-value)
60
-log (p-value)
gene expression between patients and healthy subjects for each
dataset of human burn (GSE37069), trauma (GSE36809), and
sepsis (GSE28750) conditions and between treated and control
groups for the murine models (GSE7404 for burn and trauma,
GSE19668 for sepsis, and GSE20524 for infection) were obtained
from a bioinformatics platform, NextBio.
Between any two datasets, the agreements of the gene fold
changes (Spearman’s rank correlation between the two datasets
for Fig. 1 and Pearson’s correlation for Fig. 3) as well as the directionality of the changes (P values derived from nonparametric
ranking method by NextBio for Fig. 2 and the percentages of genes
changed in the same direction between the two datasets for Fig. 3)
were compared.
We conducted a Kolmogorov–Smirnov test to check for the
normality of the distribution of gene expression data in human
burn conditions, which were compared with mouse models as
a reference dataset. The assumption of normality was rejected
either for the fold changes or for the log-twofold changes of the
gene expression levels (P < 0.0001). Therefore, we mainly used
nonparametric Spearman’s correlation coefficient (ρ) for the
correlation analyses. The criteria for the selection of the genes of
interest was fold change >1.2 and P < 0.05 within each condition
to be compared. The correlations of the gene changes as assessed
by Spearman’s correlation coefficient indicated that there were
highly significant similarities in gene responses between each of
the human conditions and those of the mouse models (Fig. 1; ρ =
0.43–0.68, P < 0.0001 for every comparison between human
conditions and the corresponding mouse models). There were also
highly significant correlations among different mouse models
A
-log (p-value)
Correlation of Gene Changes to Trauma, Burns, and Endotoxemia
Between Human Subjects and Murine Models. The fold changes of
15
6.5 × 10-11
10
5
10
0
0
771 genes 1266 genes 773 genes 441 genes
Human↑ Human↓ Human↑ Human↓
Mouse↑ Mouse ↓ Mouse↓ Mouse ↑
E
40
-log (p-value)
Results
30
578 genes 221 genes 212 genes 513 genes
Human↑ Human↓ Human↑ Human↓
Mouse↑ Mouse ↓ Mouse↓ Mouse ↑
Human Burn vs. Mouse Infection
(GSE20524)
Overlap P value:
3.4 × 10-35
20
10
0
564 genes 1044 genes 353 genes 237 genes
Human↑ Human↓ Human↑ Human↓
Mouse↑ Mouse ↓ Mouse↓ Mouse↑
Fig. 2. Statistical comparison of the direction of the gene expression
changes between human burns and mouse models. Vertical bar represents
the significance of the overlap between gene sets. Genes whose expression
levels were changed in human burns significantly overlapped with those in
the condition of mouse burn (A, overlap P value = 3.9 × 10−34), mouse
trauma (B, 6.3 × 10−13), mouse sepsis from GSE19668 (C, 1.2 × 10−35), mouse
sepsis from GSE26472 (D, 6.5 × 10−11), and mouse infection (E, 3.4 × 10−35).
Value is expressed as the –log 10 of the P value. Statistical significances
regarding the directionality of the gene expression changes were derived
from the nonparametric ranking method provided by the bioinformatics
platform NextBio.
(Fig. 1; ρ = 0.47–0.57, P < 0.0001 for every comparison between
a pair of mouse models).
Nonparametric ranking analysis by NextBio rejected, with
extraordinarily high confidence, the hypothesis that mouse
models show only coincidental overlap of the directionality of
gene changes with those in human burn conditions [Fig. 2;
overlap P value = 3.9 × 10−34, 6.3 × 10−13, 1.2 × 10−35, 6.5 × 10−11,
and 3.4 × 10−35 for mouse models of burn, trauma, sepsis
(GSE19668), sepsis (GSE26472), and infection, respectively],
demonstrating that the directionality of the changes in mouse
models was highly similar to that in human burn conditions.
Outputs of these analyses using NextBio are available from
the URLs shown in Table S1.
Although the data violate the assumption of a normal distribution, we created a scatterplot for the responses to inflammation
from different etiologies in humans and murine models (Fig. 3)
using Pearson’s correlation coefficient R and percentages of genes
changed in the same direction to compare with figure 4 in Seok
et al. (1). Contrary to the conclusion of Seok et al., there were
significant correlations and similarities of the direction in the gene
expression changes between human burn conditions and mouse
models (Fig. 3; human: fold change >2.0; mouse model: fold
change >1.2; R = 0.26–0.51; P < 0.0001 for all comparisons;
percentage: 48.1–86.2). With more stringent criteria for gene selection (human: fold change >4.0; mouse model: fold change
Takao and Miyakawa
respectively. Significant enrichment of up-regulated genes in five
pathways/biogroups was detected across all seven conditions.
Pathways/biogroups that show enrichment with a P value less than
0.6 in at least one of the conditions are listed in Dataset S1.
Some of the pathways/biogroups with high overlap are shown
in Fig. 4 A–D. The significance of the overlap between each
condition and pathways/biogroups is also shown in the graph.
There was significant overlap between genes annotated in GO
as “innate immune response” and the genes up-regulated in the
mouse models of burn (Fig. 4A, P = 6.3 × 10−24), sepsis (P = 4.6 ×
10−33), and infection (P = 1.3 × 10−16), as well as in human burn
A
Innate immune response
suppressed activated
>1.2), R values and the percentages of genes changed in the same
direction were generally greater [Fig. 3; R = 0.36–0.59; P < 0.0001
for all comparisons except for mouse sepsis (P = 0.0162); percentage: 74.4–93.2]. Detailed information on the analyses is available
in Table S2.
Comparison of the Significantly Regulated Pathways/Biogroups Between
Human Conditions and Mouse Models. Pathways/biogroups commonly
altered across datasets of human diseases and mouse models were
determined through a combination of rank-based enrichment statistics and biomedical ontologies using NextBio. In the human burn
condition, significant enrichment of up-regulated and down-regulated gene expression was identified in 875 and 197 pathways/biogroups, respectively, from a total of 6,176 sets of pathways/
biogroups registered in NextBio [1,454 sets from Gene Ontology
(GO) database, 4722 sets from canonical pathways in Broad
MSigDB]. As a selection criterion, a P value that was corrected by
the Bonferroni correction for multiple comparisons (P < 8.1 × 10−6 =
0.05/6,176) was used for this identification. Among the 875 or 197
pathways/biogroups with significant enrichment of up-regulated or
down-regulated genes in the human burn condition, the pathways/
biogroups that demonstrated significant enrichment in other human and mouse conditions as well were determined with a P value
corrected with the Bonferroni correction (P < 5.7 × 10−5 = 0.05/875
or P < 2.6 × 10−4 = 0.05/197, respectively). Of the 875 pathways/
biogroups with significant enrichment of up-regulated genes in
human burn conditions, 313, 163, 345, and 114 also showed significant enrichment in mouse models of burn, trauma, sepsis, and
infection, respectively. Unexpectedly, only half of the pathways/
biogroups up-regulated in human burn conditions were shared by
the other human disease conditions (trauma 52.1%, sepsis 53.4%).
Significant enrichment of up-regulated genes in six pathways/biogroups was detected across all conditions (three human conditions
and four mouse models). Similarly, among the 195 pathways/biogroups with significant enrichment of down-regulated genes in
human burn conditions, 85, 44, 60, and 130 also showed significant
enrichment in mouse models of burn, trauma, sepsis, and infection,
Takao and Miyakawa
Human Burn
Human Trauma
Human Sepsis
Mouse Burn
Mouse Trauma
Mouse Sepsis
Mouse Infection
-10 0
10 20 30 40 50 60 70 80 90
Significance of overlaps with the GO biogroup (- log)
B Genes involved in Cytokine Signaling in Immune system
suppressed activated
Human Burn
Human Trauma
Human Sepsis
Mouse Burn
Mouse Trauma
Mouse Sepsis
Mouse Infection
-10
0
10
20
30
40
50
60
Significance of overlaps with the Broad MSigDB biogroup (- log)
C
Lymphocyte differentiation
activated
suppressed
Human Burn
Human Trauma
Human Sepsis
Mouse Burn
Mouse Trauma
Mouse Sepsis
Mouse Infection
-10
-5
0
5
10
15
20
25
30
Significance of overlaps with the GO biogroup (- log)
D Genes involved in Translocation of ZAP-70 to Immunological synapse
activated
suppressed
Human Burn
Human Trauma
Human Sepsis
Mouse Burn
Mouse Trauma
Mouse Sepsis
Mouse Infection
-5
0
5
10
15
20
25
Significance of overlaps with the Broad MSigDB biogroup (- log)
Fig. 4. Representative pathways or biogroups shared by the human diseases and mouse models. (A–D) Shown are bar graphs of statistical significance (−log 10 P value) for the representative pathways with significant
regulation in the human and murine models. (A) Genes annotated as “innate immune response (GO)” significantly overlapped with genes up-regulated in the mouse models as well as in human diseases. (B) Significant
overlap was also detected between “genes involved in cytokine signaling in
immune system (canonical pathways, Broad MSigDB)” and genes up-regulated in both the mouse models and human diseases. (C) Genes annotated
“lymphocyte differentiation (GO)” and genes down-regulated in the mouse
models and human diseases significantly overlapped. (D) There was also
significant overlap between “genes involved in translocation of ZAP-70 to
immunological synapse (canonical pathways, Broad MSigDB).” As can be
seen, in the pathways/biogroups shown here, the mouse models are comparable to the human conditions in the significance of enrichment between
each condition and pathways/biogroups. For a complete list of pathways/
biogroups shared by the human diseases and mouse models see Dataset S1.
PNAS Early Edition | 3 of 6
MEDICAL SCIENCES
Fig. 3. Comparison of the genomic response to severe acute inflammation
from different etiologies in human and murine models. The datasets that
were used in Seok et al. (1) and are registered in NextBio were reevaluated
using the criteria of Seok et al. (1) to select the genes of interest. We chose
genes with significant responses in both the human burn dataset and in the
mouse dataset for comparison, whereas Seok et al. chose the 4,918 genes
with significant responses in the human datasets regardless of the significance of the changes in the genes in mice. The datasets used here are listed
in Table S2. Shown are Pearson’s correlations (R; x axis) and directionality (%;
y axis) of the gene response from multiple published datasets in GEO compared with human burns. Note that R2 data taken from Seok et al. were
recalculated to obtain the R values shown here for comparison (blue symbols). Most of the mouse models showed a high directionality score (random
chance is 50%), indicating that their gene expression patterns are similar to
the that in human burns, as long as stringent criteria are applied to the
selection of the genes of interest.
(P = 5.0 × 10−51), trauma (P = 4.1 × 10−90), and sepsis conditions
(P = 6.3 × 10−21). Significant overlap was also detected between
“genes involved in cytokine signaling in immune system (canonical
pathways, Broad MSigDB)” and genes up-regulated in the mouse
models of sepsis (Fig. 4B, P = 4.0 × 10−36), burn (P = 1.6 × 10−11),
and infection (P = 1.3 × 10−8), as well as in human burn (P = 5.5 ×
10−28), trauma (P = 6.5 × 10−52), and sepsis conditions (P = 3.2 ×
10−12). With regard to down-regulated pathways/biogroups, genes
annotated “lymphocyte differentiation (GO)” significantly overlapped with genes down-regulated in the mouse models of burn
(Fig. 4C, P = 2.2 × 10−8), trauma (P = 1.4 × 10−15), sepsis (P = 4.2 ×
10−20), and infection (P = 1.3 × 10−18), as well as in human burn
(P = 1.7 × 10−18), trauma (P = 5.6 × 10−12), and sepsis conditions
(P = 3.3 × 10−30). There was also significant overlap between “genes
involved in Translocation of ZAP-70 to immunological synapse
(canonical pathways, Broad MSigDB)” and genes down-regulated
in all of the human disease conditions and mouse models of these
conditions (Fig. 4D, human burn, P = 1.3 × 10−11; human trauma,
P = 0.0003; human sepsis, 8.7 × 10−25; mouse burn, P = 0.0003;
mouse trauma, P = 5.1 × 10−10; mouse sepsis, P = 1.3 × 10−7; and
mouse infection, P = 1.8 × 10−9).
Discussion
Contrary to the findings reported by Seok et al.(1), the findings
of the present study using the same datasets presented in their
study demonstrated that the expression levels of the genes in
mouse models of human disease conditions were highly significantly correlated with those of the human conditions. The percentage of genes that changed in the same direction between
human diseases and mouse models was greater than chance level
and statistically highly significant (P = 6.5 × 10−11 to 1.2 × 10−35).
Moreover, meta-analysis of those datasets revealed a number of
pathways/biogroups commonly and significantly altered by multiple conditions in both humans and mice. Thus, our findings
were quite different from those reported by Seok et al. (1) using
virtually identical datasets, resulting in almost completely opposite conclusions.
How can such discrepancies be explained? As a research
group that uses mice as a main tool for studying human disorders, we applied commonly used methodologies to the analyses to effectively detect similarities between humans and mice. It
is generally well known and accepted among researchers using
animal models that mouse models mimic only partial aspects of
human disorders. To identify the signals/signs of the similarities
with higher sensitivity, genes that do not show significant responses to the experimental manipulations must be excluded from
the analyses, because those nonresponsive genes would simply
produce noise in the correlation analysis. For example, 13,586 and
3,116 genes are changed (P < 0.05 and fold change >1.2) in human burn conditions and mouse models of infection, respectively,
and 1,992 among them are commonly changed in both humans
and mice. We calculated the correlation coefficients and percentages of genes that changed in the same direction as the 1,992
overlapping genes. However, Seok et al. (1) used 4,981 genes for
the calculation, which included all 3,250 genes that were changed
(false discovery rate <0.001 and fold change >2.0) in human burn
conditions, and the 1,668 genes with significant changes only in
other human conditions, regardless of their responses in mouse
models. With the calculation method used in the previous study,
the biased inclusion of genes that showed a significant response
only in the human conditions and not in mouse models greatly
reduces the value of the overall correlation coefficient/percentage
of the genes that changed in the same direction and misses the
possible responses in the mouse model that recapitulate the human condition. The genes shared by humans and mice (i.e., the
1,992 genes in the example described above) are those that mainly
represent the aspects of human disorders and biological pathways
mimicked by the mouse model. It is not surprising that Seok et al.
(1) found little correlation in each pathway, considering that genes
4 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1401965111
showing no significant response in the mouse models were included in the calculation.
To compare pathways/biogroups that are commonly changed
between human diseases and mouse models, Seok et al. (1) used
Pearson’s correlation coefficient of determination R2, which
should not be used for detecting similarities between different
species. The use of R2 is inadequate in this case for the following
reasons. First, Pearson’s correlation assumes a linear relationship between normally distributed data, and there is no reason
to assume a linear relationship among fold changes or log-twofold changes of gene expression levels. In this case, the gene
expression level of each gene may follow a different function of
the severity of inflammation. Some of the genes may have linearity but others may not, and such differences among genes
could be more apparent between different species. Second, the
normality of the distribution cannot be assumed regarding
fold changes and log-twofold changes of gene expression levels.
Indeed, in the present study, the assumption of a normal
distribution for the datasets was formally rejected by the
Kolmogorov–Smirnov test. Thus, the use of Pearson’s correlation in this analysis is not justified. In addition, R2 provides the
impression that the correlation is small when R is an intermediate value. Supposing that R is 0.14–0.28, which is usually
considered an intermediate or moderate correlation, R2 would
be 0.02–0.08, which seems small to those not familiar with the
use of R2. In the present study, we used Spearman’s rank correlation, a nonparametric statistical method that does not assume linearity or normal distribution of the data.
Another problem with the study by Seok et al. (1) is that they
did not try to find the optimal condition in a mouse study that
has multiple conditions in a dataset and that time-course data
were used in aggregate. In GSE19668, for example, there are
eight conditions that can be compared (2, 4, 6, or 12 h after
infection in C57BL6J or AJ mice). The overlap P values with the
human burn dataset vary from 8.7 × 10−8 (C57BL6J, 2 h after
infection) to 1.2 × 10−35 (C57BL6J, 6 h after infection). In our
analyses, we used the dataset that showed the highest similarity for
further analyses. In general, it is preferable to use conditions that
most highly mimic the human disorders when using animal models. In fact, when researchers develop an animal model of a human
condition, they screen the models to determine which model
most highly mimics the particular aspect of the human condition being studied.
We also used the sophisticated statistical methods provided by
NextBio to find potential similarities between human conditions
and mouse models. Expression profiles vary considerably from
study to study and from species to species. Different studies for
different species can yield different dynamic ranges, distributions
of fold changes, and P values that reflect the conditions used.
To allow for interstudy comparisons, NextBio employs a nonparametric approach, in which ranks are assigned to each gene
signature based on the magnitude of fold change. Ranks are then
further normalized to eliminate any bias owing to varying platform size. This normalized ranking approach enables comparisons across data from different studies, platforms, and analysis
methods by removing the dependence on absolute values of fold
change and minimizing some of the effects of the differences of
normalization methods, platform, and species used. Use of this
method demonstrated highly significant similarities between
human and mouse conditions (P value = 6.5 × 10−11 to 1.2 × 10−35).
Evaluating similarities across microarray studies is an important
topic for contemporary biomedical research. Simple Pearson’s
correlations are widely used, but other sophisticated statistical
measures, including that of NextBio, have been proposed to
quantify the similarity of any two microarray studies (16). It
would be interesting to assess the same datasets with other
advanced analytical measures.
The study by Seok et al. (1) has been interpreted by some
researchers and mass media as robust and persuasive evidence
that there are no correlations or similarities between mouse
models and human conditions, provoking discussions of the
Takao and Miyakawa
Takao and Miyakawa
heterogeneous populations in terms of the site or microbiology
of infection, genetic makeup, age, comorbidity of the patients,
and so forth. Such heterogeneity in human conditions could dilute the positive effect of the treatment, if any, in a subgroup of
patients in a clinical setting. A review of studies using animal
models of TNF neutralization in sepsis suggests that anti-TNF
therapies are most efficacious in models of systemic endotoxemia or Gram-negative infection, ineffective in models of complex polymicrobial infections, and harmful when the challenge
organism is Streptococcus pneumoniae or an opportunistic pathogen (31). This implies that heterogeneity of the efficacy of
a treatment exists even among animal models with relatively
homogeneous populations and that assessment of a treatment
requires rigorous characterization of the heterogeneity of human
patients. Third, the index for the outcome of clinical trials is
focused on patient mortality. If a trial is evaluated based solely
on mortality, it provides no information on the effect on the
clinical or biological events that are not lethal but may be important to patients who survive their sepsis episodes (23). It is
possible that one therapy cannot improve patient mortality but
improves the prognosis of survivors, such as the incidence rate of
amputations (23, 32) and level of organ dysfunction (32, 33).
Taken together, it is not surprising that Seok et al. (1) found
almost no correlation between the genomic responses in human
disease conditions and those of mouse models, because they used
methodologies with limited detection ability, which are rarely
used by researchers who study animal models. With reasonably
sophisticated and commonly used methods, we found highly
significant correlations between human and mouse data. It
should be noted that in our analyses we used the data from the
exact same studies as those used in Seok et al. (1), and still we
found highly significant similarities, demonstrating that their
failure to detect correlations is purely a result of the inappropriately
biased methodologies they used. Efforts to find more optimal age/
time-course/organs/strains/induction protocols that recapitulate human gene expression for each disease condition, might lead to
better mouse models of these disorders than the ones evaluated
here. Given that even these suboptimal models provide highly significant correlations, however, Seok et al. (1) might consider
modifying their conclusions, which have drawn the attention of
the popular media as well as researchers in broad areas of biomedical sciences.
Materials and Methods
Datasets for Human Diseases and Mouse Models. The datasets that we analyzed in the present study were the same as those used in the study by Seok
et al. (1) and are registered in NextBio. In the present study, the following
datasets were used for gene expression pattern analyses: “leukocytes of
patients with severe burns on >20% of total body surface area vs. healthy
controls” from GSE37069 is referred as “Human Burn”; “white blood cells of
severe blunt trauma patients 14 d after injury vs. healthy subjects” from
GSE36809 is referred as “Human Trauma”; “whole blood of sepsis patients
with community-acquired infection vs. healthy subjects” from GSE28750 is
referred as “Human Sepsis”; “WBC from blood at 7 d after burn injury vs.
burn injury sham” from GSE7404 is referred to as “Mouse Burn”; “WBC from
spleen at 3 d after trauma hemorrhage vs. trauma hemorrhage sham” from
GSE7404 is referred to as “Mouse Trauma”; “Blood of C57BL6J mice 4 h after
Staphylococcus aureus infection vs. uninfected” from GSE19668 is referred
to as “Mouse Sepsis”; “Blood from 8-wk-old BALB-c mice 1 d after tail vein
injection - Candida albicans vs. saline control” from GSE20524 is referred
to as “Mouse Infection.” Only genes with a P value of 0.05 or less and an
absolute fold change of 1.2 or greater were considered to be differentially expressed.
Comparison of Gene Response Between Datasets. Fold changes of datasets
used in the present study were calculated by dividing the value of human
diseases/mouse models by that of healthy controls/normal control mice.
Values were converted into the negative reciprocal, or −1/(fold change) if the
fold change was less than 1. In Fig. 1, genes meeting the criteria of P < 0.05
and fold change >1.2 are plotted in the graph. The normality of each
dataset was assessed using the Kolmogorov–Smirnov test, in which none of
the datasets had a normal distribution (P < 0.0001). Because the distribution
PNAS Early Edition | 5 of 6
MEDICAL SCIENCES
validity of animal models as a tool for studying human disease
and for translational research (3, 8). Commentaries in response
to the previous study claim that the implications of the study by
Seok et al. (1) may well go beyond mice and sepsis (10), and that
better definitions of clinical phenotypes, especially in heterogeneous or overlapping conditions such as neurodevelopmental
and psychiatric diseases, as well as more emphasis on rigorously
defining molecular alterations in human patients, is needed (8).
It should be noted that, even for schizophrenia, which could
be considered a uniquely human disorder (17), recent mouse
models have been developed using an approach similar to those
used in the present study that are shown to closely recapitulate
not only the clinical or behavioral phenotypes but also the molecular alterations in transcriptional and protein changes in the
brain (18–21).
The New York Times published an article attempting to
translate the findings in the study by Seok et al. (1) for a general
audience, noting, “But now, researchers report evidence that the
mouse model has been totally misleading for at least three major
killers—sepsis, burns, and trauma. As a result, years and billions
of dollars have been wasted following false leads, they say” (11).
In the present study, however, we provide strong evidence, by
reanalysis of the same datasets that were used by Seok et al. (1),
that mouse models do in fact closely recapitulate certain aspects
of transcriptomic responses in some inflammatory conditions in
humans and are useful for studying human disorders and conducting translational research.
In some commentaries (4, 6–8), the use of only a single strain
(i.e., C57BL/6) has been criticized as a potential caveat of the
study by Seok et al. (1). This may not be the case, at least for
some datasets we analyzed, because the similarity of the data
derived from C57BL/6 with human data are comparable to or
even greater than those from other strains of mice, although our
study does not exclude the possibility that some other mouse
strain mimics the human disorders better. Our results extend
those criticisms about the use of only a single strain and indicate
the importance of searching for optimal conditions in multiple
dimensions, including time course, dose, age, induction protocol,
organ, and so on, by using appropriate and nonbiased datamining methods to establish animal models of human disorders.
Seok et al. (1) pointed out that all of the nearly 150 clinical
trials testing candidate agents intended to block the inflammatory response in critically ill patients have failed (1). Why
have these trials failed when such agents have proven effective in
mouse models? Although it is beyond the scope of the study to
discuss this serious issue in detail, we would like to suggest some
possible reasons. First, it is possible that some of the failed drugs
were targeting genes with distinctly different responses between
humans and mice, although the targeted pathways were indeed
enriched for differentially expressed genes in humans and mice.
Genes that show altered expression with a P value less than 0.05
in at least three of the conditions are listed in Dataset S2. TNFalpha, a major proinflammatory cytokine that activates the
NF-κB pathway, is a therapeutic target for sepsis (22, 23). Candidate agents, such as the TNF antibody, are effective in mouse
models (24) but failed to reduce mortality in clinical trials (25).
In fact, the mRNA of TNF-alpha is significantly increased in the
datasets from mouse models (fold change, 4.8; P = 0.002 in
mouse sepsis and fold change, 1.39, P = 0.0023 in mouse infection) but not in the datasets of human conditions assessed in
the present study (P > 0.05 in all of the four human conditions;
TNF-alpha is not present in Dataset S2, because the table shows
the genes significantly changed in three or more biosets), consistent with previous studies that showed little increase in the
gene (26, 27) or its product (28, 29) in human sepsis. Most experimental therapies for sepsis have focused on attenuating the
proinflammatory response, but the idea that immunosuppression
is a major contributor to the conditions has recently become
more accepted (30). In this sense, the failure of the trials could
be due to an inadequate strategy/target of the therapy. Second,
compared with animal models, human sepsis patients comprise
of the datasets was not normal, Spearman’s rank correlation (ρ) was calculated between each dataset. In nonparametric ranking analysis of the gene
expression signature by NextBio (Fig. 2), genes with a P value of 0.05 or less
and an absolute fold change of 1.2 or greater were used, which is the default criterion used in analyses in NextBio. Pearson correlation (R) between
the human burn condition and other mouse models was calculated using
genes with a P value of 0.05 or less and an absolute fold change greater than
4.0 (Fig. 3, solid red circles) or 2.0 (Fig. 3, open red circles) in the human burn
dataset (as a reference) and genes with a P value of 0.05 or less and an
absolute fold change greater than 1.2 in other conditions of mouse models
(Fig. 3). The percentage of genes changed in the same direction as in the
human burn condition was also calculated using genes with the same criteria. The five datasets of human disease and five datasets of mouse models
that were used in Seok et al. (1) and are registered in NextBio as well are also
plotted in Fig. 3.
vein injection - C. albicans vs. saline control.” Pathways/biogroups from GO
and canonical pathways of Broad MSigDB were included in this analysis.
For identification of the pathways/biogroups showing significant enrichment of the genes changed in Human Burn, a pathway enrichment analysis
was conducted among the 6,176 pathways/biogroups registered in GO and
canonical pathways of Broad MSigDB, with a P value corrected by the
Bonferroni correction for multiple comparisons (P = 8.1 × 10−6 = 0.05/6,176).
The pathways/biogroups showing significant enrichment of the genes
changed in six other human and mouse conditions were also determined by
pathway enrichment analyses using NextBio. Among the pathways/biogroups with significant enrichment in human burn conditions, those also
showing significant enrichment in each condition were determined. Pathways/biogroups with P values less than the alpha-level corrected by the
Bonferroni correction were considered significantly enriched.
Comparison of Pathways/Biogroups Altered in the Human Diseases and Mouse
Models. Meta-analyses of the human and mouse gene sets using NextBio were
conducted to find the pathways/biogroups commonly changed in the gene
sets. Pathways/biogroups commonly altered across datasets of human diseases and mouse models were determined through a combination of rankbased enrichment statistics (see the following section) and biomedical
ontologies using NextBio. The datasets used in the meta-analysis were as
follows: Human Burn, “leukocytes of patients with severe burns on >20% of
total body surface area vs. healthy controls”; Human Trauma, “white blood
cells of severe blunt trauma patients 28 d after injury vs. healthy subjects”;
Human Sepsis, “whole blood of sepsis patients with community-acquired
infection vs. healthy subjects”; Mouse Burn, “WBC from blood at 7 d after
burn injury vs. burn injury sham”; Mouse Trauma, “WBC from spleen at 3 d
after trauma hemorrhage vs. trauma hemorrhage sham”; Mouse Sepsis,
“blood of C57BL6J mice 4 h after Staphylococcus aureus infection vs. uninfected”; Mouse Infection “blood from 8-wk-old BALB-c mice 1 d after tail
Computing Overlap P Values of Gene Expression Patterns from Different
Datasets. NextBio was used to compare gene expression patterns (signatures)
between experiments from human diseases and mouse models. NextBio was
accessed on December 31, 2013, for Figs. 1–4, unless otherwise noted. NextBio
compares the signatures in publicly available microarray databases with a signature provided by the user using a “Running Fisher” algorithm, as previously
described (15, 18). The overlap P value, that is, the direction of the correlation
between two given gene signature sets, and the P values between subsets of
gene signatures, is calculated with the algorithm. For detailed method for
calculation of overlap P values, see SI Materials and Methods.
ACKNOWLEDGMENTS. This work was supported by Grants-in-Aid for Scientific Research 25116526, 10301780, and 80420397 from the Ministry of Education, Culture, Sports, Science, and Technology of Japan and by grants from
Core Research for Evolutional Science and Technology of the Japan Science
and Technology Agency.
1. Seok J, et al.; Inflammation and Host Response to Injury, Large Scale Collaborative
Research Program (2013) Genomic responses in mouse models poorly mimic human
inflammatory diseases. Proc Natl Acad Sci USA 110(9):3507–3512.
2. Cauwels A, Vandendriessche B, Brouckaert P (2013) Of mice, men, and inflammation.
Proc Natl Acad Sci USA 110(34):E3150.
3. de Souza N (2013) Model organisms: Mouse models challenged. Nat Methods 10(4):
288.
4. Drake AC (2013) Of mice and men: What rodent models don’t tell us. Cell Mol Immunol 10(4):284–285.
5. Leist M, Hartung T (2013) Inflammatory findings on species extrapolations: Humans
are definitely no 70-kg mice. ALTEX 30(2):227–230.
6. Osterburg AR, et al. (2013) Concerns over interspecies transcriptional comparisons in
mice and humans after trauma. Proc Natl Acad Sci USA 110(36):E3370.
7. Remick D (2013) Use of animal models for the study of human disease–a shock society
debate. Shock 40(4):345–346.
8. (2013) Of men, not mice. Nat Med 19(4):379-379.
9. Schaff M (2013) Mouse study findings prove contentious among scientists. Pitt News.
Available at www.pittnews.com/news/article_41cd755a-a081-577a-b752-2c83fc6628e2.
html. Accessed January 17, 2014.
10. Collins DF (2013) Of mice, men, and medicine. NIH Director’s Blog. Available at http://
directorsblog.nih.gov/2013/02/19/of-mice-men-and-medicine/. Accessed January 17,
2014.
11. Kolata G (2013) Health testing on mice is found misleading in some cases. NY Times.
Available at www.nytimes.com/2013/02/12/science/testing-of-some-deadly-diseaseson-mice-mislead-report-says.html. Accessed January 17, 2014.
12. Engber D (2013) Septic shock. Slate. Available at www.slate.com/articles/health_and_
science/the_mouse_trap/2013/02/lab_mouse_models_of_inflammation_for_burns_sepsis_
trauma_animal_testing.html. Accessed January 17, 2014.
13. Cossins D (2013) Do mice make bad models? The Scientist. Available at www.the-scientist.
com/?articles.view/articleNo/34346/title/Do-Mice-Make-Bad-Models-/. Accessed January 17,
2014.
14. Chu E (2013) This is why it’s a mistake to cure mice instead of humans. Popular Science. Available at www.popsci.com/science/article/2013-02/how-scientists-got-lostcuring-mice-instead-humans. Accessed January 17, 2014.
15. Kupershmidt I, et al. (2010) Ontology-based meta-analysis of global collections of
high-throughput public data. PLoS ONE 5(9):e13066.
16. Tseng GC, Ghosh D, Feingold E (2012) Comprehensive literature review and statistical
considerations for microarray meta-analysis. Nucleic Acids Res 40(9):3785–3799.
17. Powell CM, Miyakawa T (2006) Schizophrenia-relevant behavioral testing in rodent
models: a uniquely human disorder? Biol Psychiatry 59(12):1198–1207.
18. Takao K, et al. (2013) Deficiency of schnurri-2, an MHC enhancer binding protein,
induces mild chronic inflammation in the brain and confers molecular, neuronal, and
behavioral phenotypes related to schizophrenia. Neuropsychopharmacology 38(8):
1409–1425.
19. Ohira K, et al. (2013) Synaptosomal-associated protein 25 mutation induces immaturity of the dentate granule cells of adult mice. Mol Brain 6:12.
20. Walton NM, et al. (2012) Detection of an immature dentate gyrus feature in human
schizophrenia/bipolar patients. Transl Psychiatr 2:e135.
21. Hagihara H, Ohira K, Takao K, Miyakawa T (2014) Transcriptomic evidence for immaturity of the prefrontal cortex in patients with schizophrenia. Mol Brain 7:41.
22. Wesche-Soldato DE, Swan RZ, Chung C-S, Ayala A (2007) The apoptotic pathway as
a therapeutic target in sepsis. Curr Drug Targets 8(4):493–500.
23. Marshall JC (2014) Why have clinical trials in sepsis failed? Trends Mol Med 20(4):
195–203.
24. Beutler B, Milsark IW, Cerami AC (1985) Passive immunization against cachectin/tumor
necrosis factor protects mice from lethal effect of endotoxin. Science 229(4716):
869–871.
25. Abraham E, et al.; NORASEPT II Study Group (1998) Double-blind randomised controlled trial of monoclonal antibody to human tumour necrosis factor in treatment of
septic shock. Lancet 351(9107):929–933.
26. Wong HR, et al.; Genomics of Pediatric SIRS/Septic Shock Investigators (2009) Genomic
expression profiling across the pediatric systemic inflammatory response syndrome,
sepsis, and septic shock spectrum. Crit Care Med 37(5):1558–1566.
27. Wong HR (2013) Genome-wide expression profiling in pediatric septic shock. Pediatr
Res 73(4 Pt 2):564–569.
28. Hotchkiss RS, Karl IE (2003) The pathophysiology and treatment of sepsis. N Engl J
Med 348(2):138–150.
29. Oberholzer A, Oberholzer C, Moldawer LL (2000) Cytokine signaling—regulation of
the immune response in normal and critically ill states. Crit Care Med 28(4, Suppl):
N3–N12.
30. Hotchkiss RS, Coopersmith CM, McDunn JE, Ferguson TA (2009) The sepsis seesaw:
Tilting toward immunosuppression. Nat Med 15(5):496–497.
31. Lorente JA, Marshall JC (2005) Neutralization of tumor necrosis factor in preclinical
models of sepsis. Shock 24(Suppl 1):107–119.
32. Levin M, et al. (2000) Recombinant bactericidal/permeability-increasing protein
(rBPI21) as adjunctive treatment for children with severe meningococcal sepsis: A
randomised trial. rBPI21 Meningococcal Sepsis Study Group. Lancet 356(9234):
961–967.
33. Marshall JC, et al. (1995) Multiple organ dysfunction score: A reliable descriptor of
a complex clinical outcome. Crit Care Med 23(10):1638–1652.
6 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1401965111
Takao and Miyakawa