Integrating siRNA and protein–protein interaction
data to identify an expanded insulin signaling network
Zhidong Tu,1Carmen Argmann,1Kenny K. Wong,2Lyndon J. Mitnaul,2
Stephen Edwards,1Iliana C. Sach,1Jun Zhu,1and Eric E. Schadt1,3
1Rosetta Inpharmatics, a wholly owned subsidiary of Merck & Co., Inc., Seattle, Washington 98109, USA;2Department of
Cardiovascular Disease, Merck Research Laboratories, Rahway, New Jersey 07065, USA
Insulin resistance is one of the dominant symptoms of type 2 diabetes (T2D). Although the molecular mechanisms leading
to this resistance are largely unknown, experimental data support that the insulin signaling pathway is impaired in
patients who are insulin resistant. To identify novel components/modulators of the insulin signaling pathway, we
designed siRNAs targeting over 300 genes and tested the effects of knocking down these genes in an insulin-dependent,
anti-lipolysis assay in 3T3-L1 adipocytes. For 126 genes, significant changes in free fatty acid release were observed.
However, due to off-target effects (in addition to other limitations), high-throughput RNAi-based screens in cell-based
systems generate significant amounts of noise. Therefore, to obtain a more reliable set of genes from the siRNA hits in our
screen, we developed and applied a novel network-based approach that elucidates the mechanisms of action for the true
positive siRNA hits. Our analysis results in the identification of a core network underlying the insulin signaling pathway
that is more significantly enriched for genes previously associated with insulinresistance than the set of genes annotated in
the KEGG database as belonging to the insulin signaling pathway. We experimentally validated one of the predictions,
S1pr2, as a novel candidate gene for T2D.
[Supplemental material is available online at www.genome.org.]
Insulin deficiency and resistance are the two major causes of type
2 diabetes (T2D) (Turner et al. 1979). Insulin resistance is tightly
related to the functional state of the insulin signaling pathway
together with several other cellular processes like insulin secretion
and inflammation (Pessin and Saltiel 2000; Clee et al. 2006;
Draznin 2006; Solinas et al. 2007). Previous studies have shown
that knocking out components in the insulin signaling pathway
leads to insulin resistance in mice (Araki et al. 1994; Tamemoto
et al. 1994). While a few key components of the pathway have
been identified (e.g., insulin receptor and insulin receptor sub-
strates), a clear global picture has yet to emerge (Taniguchi et al.
2006). It is of further note that almost all of the T2D genes iden-
tified in the human genome-wide association studies (GWAS) and
extensively replicated appear to be related to beta cell function,
not insulin resistance (Pascoe et al. 2007; Scott et al. 2007).
Therefore, novel approaches are required to enhance our un-
derstanding of the mechanisms of insulin resistance.
The recently developed RNAi technologies provide a way to
interrogate the function of genes individually and in a high-
throughput fashion with respect to biological functions, cellular
processes,or other phenotypesof interest(Mello and Conte2004).
The high-throughput short interfering RNA (siRNA) screening
technologies have gained in popularity over the past several years,
and a plethora of large-scale screens have now been performed in
various organisms, targeting different cellular processes and/or
pathways (Berns et al. 2004; Boutros et al. 2004; Nollen et al. 2004;
Lu et al. 2007). To identify novel components/modulators of the
insulin signaling pathway, we designed siRNAs targeting over 300
genes known or predicted to associate with T2D traits and then
tested the effects of knocking down these genes in an insulin-
dependent anti-lipolysis assay in 3T3-L1 adipocytes. For 126
genes, significant changes in free fatty acid release were observed.
To properly interpret this type of data, a number of concerns
have to be addressed regarding the quality of large-scale siRNA
experiments. Off-target effects resulting from a given siRNA di-
rectly knocking down unintended target genes are now well
established (Kulkarni et al. 2006). Because mismatches and gaps
between small RNA and target RNA sequences can be well toler-
ated, small RNAs have been shown to have potentially hundreds
degree of knockdown of target genes by different siRNAs is typi-
cally highly variable, thus the actual silencing effect on a specific
protein’s activity is hard to estimate. As a result, follow-up studies
have been viewed as critical for filtering out false-positive hits as
well as recovering false-negative hits that result because the extent
of knockdown of the target gene was not significant enough to
achieve efficacy. Further analyses are also required to identify the
underlying mechanisms of the cellular processes of interest. In
particular, analyses at the pathway level are of high interest given
the ultimate aim in large-scale siRNA screening experiments is to
elucidate how genes interact to sense extra/intracellular signals
and to perform complex cellular logic underlying complex bi-
ological processes (Moffat and Sabatini 2006).
To obtain a set of positive siRNA hits with higher confidence
from our insulin resistance siRNA screen, we developed and ap-
plied a novel network method that elucidates the mechanisms of
action for the true positive siRNA hits. This method also leads to
an expanded set of genes predicted to be involved in insulin sig-
naling, beyond those genes represented in the siRNA hit list. This
is achieved by integrating known pathway relationships and
experimentally determined protein–protein interactions to link
true positive siRNA hits to networks of genes related to insulin
signaling. Our analysis results in the identification of a core
network underlying the insulin signaling pathway that is more
E-mail email@example.com; fax (206) 802-6377.
Article published online before print. Article and publication date are at
19:1057–1067 ? 2009 by Cold Spring Harbor Laboratory Press; ISSN 1088-9051/09; www.genome.org
significantly enriched for genes previously associated with insulin
resistance than the set of genes annotated in the Kyoto Encyclo-
pedia of Genesand Genomes (KEGG)databaseas belongingto this
pathway. We experimentally validate one of the predictions,
S1pr2, in this network as a novel candidate gene for T2D.
To obtain a more comprehensive set of insulin pathway compo-
nents/modulators, 313 genes were selected and targeted with
siRNAs in cultured 3T3-L1 adipocytes as shown in Figure 1A,B.
Diabetes- and obesity-related genes were pooled from several dif-
ferent sources: (1) genes supported as causal for diabetes- and
obesity-related traits in published mouse crosses (Schadt et al.
2003; Chen et al. 2008), (2) genes involved in fatty acid beta-
oxidation, based on experiment validations published in the pri-
mary literature, (3) orphan peptidases observed to co-express with
known diabetes and obesity genes, and (4) genes associated with
diabetes and obesity as determined by published data sets and
primary reports in the literature. Genes that give rise to protein
products that have a high druggability potential (Schoenborn
et al. 2006) were identified from this set of diabetes- and obesity-
associated genes, resulting in a set of 313 genes for siRNA screen-
ing. The majority of these 313 candidate genes (177 or 57%) were
supported as causal for diabetes- and obesity-associated traits
based on a previously developed integrative genomics procedure
(Schadt et al. 2005). This procedure integrates genotypic, gene
expression, and clinical trait data to test whether gene expression
and clinical traits linked to common genetic loci are related in
a causal, reactive, or independent fashion. We applied this pro-
cedure to liver and adipose gene expression data and plasma in-
sulin, glucose, and free fatty acid levels monitored in previously
described crosses constructed from the B6, DBA, and C3H strains
(referred to here as the BXD and BXH crosses) (Schadt et al. 2003;
Chen et al. 2008).
We screened each of the 313 genes identified above in an
insulin-dependant anti-lipolysis reporting system in which free
fatty acid (FFA) release was monitored in response to treatment
with siRNAs designed against each of the target genes of interest
(see Methods for more detail). Insulin-dependent FFA release is an
indicator of insulin resistance (Yki-Jarvinen and Taskinen 1988;
Goweretal. 2001).Asasystembecomesmoreinsulinresistant, the
amount of FFA release increases, given that the ability of the sys-
tem to respond to insulin signaling is reduced. Therefore, in this
assay, if reducing the activity of a given gene increases (decreases)
FFA release, we can conclude that the gene has a putative role in
insulin sensitization (resistance). Of the 313 genes screened via
this siRNA assay, FFA release was significantly increased (decreased
insulin sensitivity) or decreased (increased insulin sensitivity) for
126 (;40%) of the genes, based on three repeats for each assay. We
refer to these genes as siRNA (positive)
hits regardless of the direction of the ef-
fect. It is of note that the fraction of
positive hits observed in this study was
much higher than previously reported
screens typically resulted in <10% of the
genes targeted exhibiting a significant
effect (Boutros et al. 2004; Nollen et al.
2004; Lu et al. 2007). The higher hit rate
achieved in our study may reflect the
strong bias toward biological relevance in
the candidate gene selection process.
While a direct comparison to these dif-
ferent studies is not possible given dif-
ferences in biological processes that
were targeted, in the experimental as-
says, in the RNAi designs, and in the
statistical analyses for declaring hits,
a within experiment comparison on
groups of genes selected using different
criteria does suggest that genes identi-
fied using the causality procedure were
more likely to contain relevant pathway
Table S1; Supplemental materials). This
is consistent with previous results we
have achieved applying this procedure
to identify and validate metabolic trait
genes (Mehrabian et al. 2005; Schadt
et al. 2005; Chen et al. 2008).
Some of the siRNA hits we observed
are well known insulin signaling path-
way components or modulators, in-
(representing all four positive controls
used to develop the screen), while the
were selected for screening. Genes from multiple sources were considered and filtered based on
whether their protein products could be targeted by small molecules. (B) Distribution of sources from
which the 313 genes were selected. For example, 177 of 313 genes were selected because they were
supported as causal for diabetes/obesity in an experimental cross population. (C) Distribution of the
ion channel, (4) kinase/protease, and (5) other function. The numbers of genes in each category are pro-
vided after category names, where (+) stands for a positive siRNA hit and (?) denotes a negative siRNA hit.
Selecting genes for the insulin resistance siRNA screen. (A) Global view of how 313 genes
1058 Genome Research
Tu et al.
vast majority are not well supported as genes involved in in-
sulin signaling. Since our assay relied on insulin-dependent anti-
lipolysis as the reporting system, some hits may be more closely
related to lipolysis than to insulin signaling. Pnpla2, a gene re-
cently identified as a key enzyme in triacylglycerol metabolism,
is one such example (Lake et al. 2005; Zechner et al. 2005) of
genes involved in insulin-dependent anti-lipolysis (Kim et al.
2006). Aside from the known insulin pathway genes and lipolysis-
related genes, the siRNA hits are distributed across a number of
categories reflecting different protein characteristics, including
kinases, phosphatases, and G protein–coupled receptors (GPCRs)
The identification of hundreds of genes supported as causal
for diabetes-associated traits from experimental mouse pop-
ulations suggested that large networks of genes were driving
disease-associated traits. The high rate of validation of these pre-
dictions in the siRNA screen further supports that networks,
as opposed to simple, linearly ordered pathways, drive complex
system behavior (Chen et al. 2008). To further increase confidence
in the siRNA results and to provide greater insights into
the mechanisms of the underlying signaling logic, we developed
and applied a novel network-based algorithm (the Pathway
Expansion Analysis, or PEXA) to these data. The primary as-
sumption underlying PEXA is that siRNA hits for a common cel-
lular phenotype are more likely true positives if they interact and
cluster together in the protein–protein interaction (PPI) network.
Instead of taking an unsupervised approach to search for subnet-
works enriched for siRNA hits in the PPI network, an approach
originally proposed for finding differentially expressed gene-
enriched modules (Ideker et al. 2002; Liu et al. 2007), PEXA lev-
erages classic pathway information deposited in KEGG (Kanehisa
et al. 2004) to guide the search in the PPI network. The ultimate
output of this procedure is an interaction
network enriched for genes that are sup-
ported as causal for phenotypic changes
of interest in response to knockdown
Constructing a core network seeded
by siRNA hits
Because the outputs of siRNA screens are
generally noisy, further analyses or fol-
low-up experiments are required to con-
firm whether the observed hits are true
with respect to the intended target. In
addition to analyzing the reliabilities of
individual hits, we were also interested in
obtaining insights into the possible
mechanisms of how these genes poten-
tially interact to perform the complex
logic embedded in the system with re-
spect to insulin signaling. Therefore, we
combined KEGG pathways and PPIs to
increase our confidence in the relevance
of the siRNA hits and to elucidate the
networks in which these genes operate.
In this insulin signaling context the
PEXA method consisted of the following
steps: (1) identifying genes from the
insulin-dependent anti-lipolysis siRNA
screen (referred to here as the hit list), (2)
querying genes in the siRNA hit list against the KEGG pathway
database to identify the initial seeding paths for the networks
underlying the insulin signaling process, (3) expanding the seed-
ing pathways into networks based on PPI data, and (4) pruningthe
expanded network to eliminate components that are not sup-
ported by the siRNA screening results, resulting in a compact
subnetwork that underlies the biological process of interest and
that is enriched for the siRNA hits (Fig. 2).
Seeding the network with KEGG pathways
Since KEGG contains manually curated interactions, it is ideal for
initiating the search process (seeding). A pathway (e.g., insulin
signaling pathway), denoted as Pi, consists of a collection of nodes
Niand edges Ei. Edges can be either directed (such as phosphory-
lation) or undirected (such as PPI), depending on the nature of
the interaction. For the set of positive siRNA hits S, we defined IS;Pi
as the intersection of S and nodes Niin the pathway Pi. If the
numberof elementsin IS;Pi, denotedas jIS;Pij , was $ 2,we searched
for all possible paths in Pithat connected uiand vi, for ui;vi2 IS;Pi
and ui6¼ vi. We call such paths seeding paths, which serve as the
backbone for the network expansion described in the following
step. In this case we stored only the edges and associated nodes,
as opposed to all possible paths. Therefore, the total number of
edges that needs to be stored is bounded by +iEi
compared to the theoretical upper bound on the number of pos-
The seeding paths generated from the siRNA hits and KEGG
pathways are shown in Figure 3A. Several pathways are identified
by the siRNA hit genes, forming a series of disjointed subnet-
works. Not surprisingly, the insulin signaling pathway interacts
with branches of several other KEGG pathways and forms the
largest connected subnetwork. While multiple pathways have
j j, which is small
consists of four steps: (1) perturbing genes of interest using siRNA to identify those that produce the
desired phenotype (referred to here as the hit list); (2) querying through all the pathways in the KEGG
database with the hit list to identify seeding paths; (3) expanding the seeding paths using PPI data to
obtain a more coherent network relating to the biological process of interest (insulin signaling in this
case); and (4) pruning the network obtained in step 3 to enhance the biological coherence of the
network with respect to the biological processes of interest. In the hypothetical networks depicted in
this figure, the orange nodes correspond to genes in the siRNA hit list, while the blue nodes are sup-
ported by the KEGG and/or PPI data as operating in the same part of the network as the hit list genes.
Flow diagram for the PEXA network reconstruction process. This reconstruction process
Integrating siRNA and protein–protein interaction
been pulled in as a result of the seeding process, overall the number
of siRNA hits contained in all seeding paths is only 24 (out of the
126 hits). This limited overlap is mostly due to the limited coverage
of the KEGG database. It is of note that nine of the 187 genes from
the siRNA screen that were tested but that gave no response (neg-
atives in the screen) were also included in the seeding paths, pulled
in as members of the pathway as defined by KEGG, illustrating the
advantage of incorporating known pathway members to help link
the set of positive hits from the siRNA screen.
Iterative expansion of seeding paths using protein–protein interactions
With the seeding paths in hand, we expanded these paths using
thegenome-wide proteininteractionnetwork.Theaimin thisstep
was to include siRNA hits that interact with the seeding paths via
the PPI network. Again, our primary assumption in the expansion
step is that if an siRNA hit has interactions with the seeding paths,
then it is more likely to be a true hit and the pathway to which it
connects is more likely to be relevant to the system under study.
The comprehensive PPI network was assembled from multiple
were screened but that are not in the siRNA hit list, and blue nodes represent genes that were not screened. (A) The siRNA hits (red nodes) serve as seeds
for building up the seeding paths based on pathways represented in the KEGG database. (B) The seeding paths depicted in A are expanded and joined
together using PPI data to form a single network. (C) After pruning we obtain a core network of genes enriched for siRNA hits. Larger sized nodes
represent genes in the KEGG insulin signaling pathway, while the smaller sized yellow nodes represent small molecules in the KEGG database. The red
edges represent interactions between S1pr2 (gold node) and its neighbors. We arbitrarily selected a few nodes and labeled them using a large font size to
indicate the interdependency among the three plotted networks.
Using PEXA to construct the insulin resistance networks. Red nodes represent genes in the siRNA hit list, green nodes represent genes that
Tu et al.
1060 Genome Research
databases (see Methods for more detail). Because the initial gene
set for the siRNA screen was identified based mainly on gene ex-
pression profiles, we did not use the gene expression data to help
differentiate true from false-positive siRNA hits, although coex-
pression networks could be used in a similar fashion to the PPI
data for siRNA screens of genes identified independently of the
expression data. Here, the PPI data were used for a ‘‘stepwise’’ ex-
pansion as follows. If a siRNA hit fell outside of all seeding paths,
but had at least one direct interaction in the PPI network with
a node in a seeding path, then we expanded the seeding paths by
including the siRNA hit gene. The expansion was continued iter-
atively until no extra siRNA hits could be added. By performing
this type of ‘‘stepwise’’ expansion rather than expanding aggres-
sively (e.g., allowing siRNA hits not represented in the network to
connect to seeding paths via non-siRNA hits), a more compact and
presumably more reliable subnetwork was obtained. The ex-
panded network resulting from this process is shown in Figure 3B.
Pruning the expanded network
After the expansion step we observed that the siRNA hits form
a connected network, while other parts of the network contained
no siRNA hits. Clearly, genes that were not tested by the siRNA
screen or genes corresponding to negative siRNA hits could also be
supported as participating in biologicalprocesses of interest if they
were strongly connected to other siRNA hits. In contrast, regions
of the network containing only non-siRNA hits are not supported
by the experimental data and so are considered lower confidence
and lower in priority for experimental follow-up studies. There-
fore, to automatically derive the core module enriched for siRNA
hits, we pruned nodes and their corresponding edges if (1) the
node was a non-siRNA hit gene that had only one connection to
the rest of the network, and it was via an interaction with another
non-siRNA hit gene, or (2) after removing a non-siRNA hit gene
the node belonged to a subnetwork consisting only of non-siRNA
hits (Fig. 4). The full pruning algorithm is described in Supple-
mental Box 1. The pruning procedure is conservative as none of
the siRNA hits can be removed during the process. The final net-
work after pruning was comprised of 202 nodes (185 genes and 17
small molecules), with 79 positive siRNA hits and six negative
siRNA hits (Fig. 3C). Because the network after pruning is con-
nected and compact, it is referred to here as PEXA module. The
PEXA module is not only more significantly enriched for siRNA
positive hits (adjusted P = 0.001), but it is also enriched for genes
whose knockouts lead to insulin resistance and abnormal glucose
tolerance phenotypes (Fisher exact test P = 5.8 3 10?7), supporting
their roles as putative insulin signaling modulators.
Assessing the significance of the PEXA module
Two permutation tests were performed to further validate that (1)
siRNA screen results were at least partially informative, and (2)
PEXA preserved and enhanced the information contained in the
siRNA screen results. If the genes identified as siRNA hits were not
coherent with respect to pathways associated with insulin sig-
naling, then we would not expect them to be more enriched in the
final network than by chance, given that connections in the
network were driven by protein interactions. The null hypothesis
in this case is that the network is not further enriched for siRNA
hit genes beyond what we would expect by chance. To test whether
there was significant enrichment, we empirically estimated the
null distribution by randomly selecting 313 genes from all of the
genes covered by the PPI network and randomly labeled 126 of
them as ‘‘siRNA hits’’ and the rest as ‘‘siRNA non-hits.’’ After ap-
plying PEXA on these gene sets, we counted the numbers of
‘‘siRNA hits’’ and ‘‘siRNA non-hits’’ in the output network for each
set. We performed 1000 permutation tests in total. Because the set
of 313 genes for the permutation test were randomly selected, the
‘‘siRNA hits’’ are less tightly connected in the resulting PEXA
networks. The number of ‘‘siRNA hits’’ in the final network over
the 1000 permutation runs varies significantly, as shown in Sup-
plemental Figure 2A. The P-value for observing equal or larger
numbers of siRNA hits in the final network than we observed in
the network depicted in Figure 3C is 0.003.
Because the original set of 313 genes screened via siRNA was
not randomly selected, we performed a second set of permutations
by resampling from the set of 313 genes that were originally
screenedvia the siRNA assay. We randomlyassigned126 out of the
313 screened genes as hits and then constructed networks using
PEXA. As the 313 genes share coherent biological processes, the
numbers of ‘‘siRNA hits’’ and ‘‘siRNA non-hits’’ in the final net-
works based on 1000 permutations were proportional, as shown in
Supplemental Figure 2B. The results from the siRNA screen based
on the observed data are seen to be dramatically different from all
of the permutation runs. The probability that the observed run
belongs to the null distribution estimated via the permutations is
1.2 3 10?25.
Comparing the derived network with other biological gene sets
To test the performance of PEXA and validate the coherence and
biological relevance of the derived PEXA network, we conducted
several comparisons among different gene sets. First we carried out
a pathway gene set enrichment test for the original 313 screened
genes, the 126 siRNA hits, and the 185 genes in the final PEXA
module. The top 10 enriched KEGG pathways (of the 181 pathway
gene sets detected in total) are listed for these three data sets in
Table 1. As expected, the original genes selected for screening were
not random but biased toward a number of pathways. For exam-
ple, the neuroactive ligand receptors, of which many are G pro-
tein–coupled receptors, were significantly enriched in this set. Of
particular note for the original set of 313 genes selected for
identified for pruning: (A) nodes that are not siRNA hits and that have only
one connection to the network via an interaction with another non-siRNA
hit gene, and (B) nodes that are part of a network component comprised
solely of non-siRNA hit genes and connected to the network via a non-
siRNA hit gene. In the example networks depicted, blue nodes represent
the superficial nodes, orange nodes represent siRNA hits, and green nodes
represent non-siRNA hit nodes incorporated from either the KEGG data-
base or PPI data. The superficial nodes and subnetworks containing no
siRNA hits to which they connect are removed as part of the pruning
PEXA pruning step. Two types of ‘‘superficial’’ nodes were
Integrating siRNA and protein–protein interaction
List of the top 10 most enriched KEGG pathways for 313 screened genes, 126 siRNA hits, and 185 module genes, respectively
313 screened genes
126 siRNA hits
185 PEXA module genes
1.64 3 10?38
2.81 3 10?23
1.29 3 10?41
Calcium signaling pathway
1.48 3 10?12
Calcium signaling pathway
2.66 3 10?11
Insulin signaling pathway
1.92 3 10?37
mTOR signaling pathway
1.91 3 10?8
Insulin signaling pathway
1.91 3 10?7
Natural killer cell mediated
3.90 3 10?29
3.70 3 10?8
Non-small cell lung cancer
2.06 3 10?6
ErbB signaling pathway
6.74 3 10?29
Chronic myeloid leukemia
2.01 3 10?7
MAPK signaling pathway
2.68 3 10?6
Regulation of actin cytoskeleton
7.14 3 10?29
3.53 3 10?7
B cell receptor signaling
8.71 3 10?6
Chronic myeloid leukemia
1.11 3 10?28
Insulin signaling pathway
3.90 3 10?7
1.02 3 10?5
6.68 3 10?27
T cell receptor signaling
4.86 3 10?7
Fc epsilon RI signaling
1.50 3 10?5
1.98 3 10?26
ErbB signaling pathway
7.28 3 10?7
1.73 3 10?5
3.02 3 10?26
Non-small cell lung cancer
7.72 3 10?7
Small cell lung cancer
2.13 3 10?5
Non-small cell lung cancer
4.81 3 10?24
aPWN, the pathway name annotated in the KEGG pathway database.
bPWS, the number of genes in the given pathway set.
cOL, the number of genes overlapping between the pathway set and the indicated gene set (screened gene set, the siRNA hits gene set, and the PEXA module gene set).
dNominal P-values represent the significance of the Fisher exact test statistic under the null hypothesis that the frequency of the indicated KEGG pathway gene set is the same between a reference set of
27,240 genes (genes with protein sequence data in MGI) and the genes comprising the set of screened genes, the set of siRNA hit genes, or the PEXA module.
Tu et al.
1062 Genome Research
screening is that the insulin signaling pathway ranked near the
bottom of the top 10 most enriched pathways. For the genes
reported as siRNA hits, the insulin signaling pathway was among
the most highly ranked pathways, suggesting that the insulin-
dependent anti-lipolysis siRNA screen carried out in this study
indeed favored genes involved in insulin signaling. For the genes
in the PEXA module, the insulin signaling pathway was the sec-
ond most highly ranked pathway, and, in fact, was greater than
fourfold more enriched than the original set of 313 genes selected
for the siRNA screening. The top ranked pathway enriched in the
PEXA module was the focal adhesion pathway, which is of interest
given that free fatty acid and expression of adhesion molecules are
tightly related (Jensen 2006). Although the output of PEXA is
dependent on the input set of genes, a limitation of the partial
genome screening carried out in this study, the method did cor-
rectly prioritize the relevant pathways associated with the bi-
ological process under investigation.
The second comparison we carried out involved comparing
gene sets generated by different computational algorithms, to
analyze the performance of these algorithms against the PEXA
method. The first set was derived by considering siRNA hits that
interacted with each other in the PPI network (referred to here as
the mouse PPI set). The second set was constructed from an in-
termediate step of the PEXA method and consisted of the network
prior to the pruning step. Studying this gene set allowed us to de-
termine whether the pruning step was effective. The third set
we considered was the PEXA generated module. We also considered
the set of siRNA hits and their direct neighbors in the PPI network.
it was effectively useless and excluded it from further comparison.
We tested whether the three gene sets just described were
enriched for genes in the insulin signaling pathway defined in the
KEGG database. The insulin signaling pathway is the most rele-
vant cellular process related to our study, given the siRNA screen
designed for this study was focused on this process. Although this
pathway was provided as input into the PEXA method, it was
providedanonymously alongwith 180other pathways.Therefore,
we considered this a fair comparison given the mouse PPI gene set
was not specifically informed by the insulin signaling pathway
during its construction. As shown in Table 2, the network gener-
ated from PEXA has a much larger overlap (a fourfold increase)
with the insulin signaling pathway than the mouse PPI gene set.
Furthermore, the PEXA module compared to the PEXA network
constructed prior to the pruning step is more significantly
enriched for genes in the insulin signaling pathway. These
enrichments again validate the utility of the PEXA method.
To further establish the power of the PEXA method to lead to
biologically relevant networks, we compared genes in the derived
network with the mouse knockout results extracted from the
combined two sets of genes: (1) knockout models with an insulin
resistance phenotype, and (2) knockouts with abnormal glucose
tolerance phenotypes. Based on these criteria we identified 136
unique genes. We then tested different gene sets related to the
PEXA network we constructed to assess whether the set of 136
genes associated with insulin signaling phenotypes was over-
represented in these different gene sets. The overlap between the
knockout gene set and the original set of siRNA hits was only four
(e.g., Insr and Akt2), which is only marginally significantly
enriched with respect to the 313 genes tested in the siRNA screen
as the background set (Fisher exact test P = 1.1 3 10?2). On the
other hand,theoverlapwiththePEXAmoduleis10 (Cav1,Crebbp,
Prkci, Mapk8, Irs1, Ppargc1a, and the original four genes), an en-
richment that is unlikely to have happened by chance (Fisher
exact test P = 5.8 3 10?7). Given these results, we can conclude
that the PEXA method generated a list of genes significantly more
enriched for genes associated with insulin resistance and glucose
homeostasis than the set of genes that produced positive hits from
the siRNA screen.
Even though the overlap between the network we derived
and the genes causing insulin-related phenotypes is significant,
most of the genes validated as causing insulin-related phenotypes
were not in the PEXA module. One possible explanation is that
these genesmay affect insulin phenotypes viaregulatingthe PEXA
module. To test this we examined adipose gene expression sig-
natures for four gene knockout models that we had access to and
which were validated as impacting diabetes-associated traits, to
assess whether a significant number of genes in the PEXA module
responds to strong perturbations (knockout) of genes that are not
in the module. We defined the adipose gene expression signatures
for Alox5, Lrp5, and Cnr1 mouse knockout models and for the
Mc3r/4r double knockout model as genes that were differentially
expressed (at the 0.01 significance level, see details in Supple-
mental Methods) between the knockout and wild-type mice for
each model. The adipose signatures were then compared with the
KEGG insulin signaling pathway and with our derived network.
Three of the four knockout signatures significantly overlapped
with the KEGG insulin signaling pathway (at the 0.01 significance
level). On the other hand, all four of the knockout signatures
significantly overlapped the PEXA module and the significance
levels were much greater, suggesting that each of these genes may
potentially interact with insulin signaling and glucose homeo-
stasis maintenance (Supplemental Table 2). For each of the
knockout models the MGI database indicated phenotypic effects
on diabetes-associated traits like plasma insulin levels, plasma
glucose levels, and impaired glucose tolerance. For Lrp5 KO mod-
el, the P-value for overlap between the knockout signature and
KEGG insulin signaling pathway was 0.19. However, the overlap
between the knockout signature and PEXA module was 2.1 3
10?6. Therefore, while the knockout phenotype could not be
strongly associated with diabetes based on the KEGG insulin sig-
naling pathway, it was easily predicted based on its impact on the
Identifying and validating S1pr2 as a candidate T2D gene
The gold standard for validating models predicted as causal for
complex phenotypes like insulin signaling is prospective validation.
Because it is not yet feasible to test in vivo all genes supported as
causal in our derived network, we used a number of criteria to
pathway genes (132 genes) with three gene sets described in the
Comparing the overlap of KEGG insulin signaling
Gene set No. of genesNo. of overlapsP-valuea
7 2.86 3 10?9
1.62 3 10?34
1.92 3 10?37
aNominal P-values represent the significance of the Fisher exact test sta-
tistic under the null hypothesis that the frequency of the insulin signaling
the mouse PPI gene set or expanded PEXA module gene set.
Integrating siRNA and protein–protein interaction
prioritize genes for validation: (1) The gene had to reside in the
PEXA module; (2) the gene had to correspond to a positive siRNA
hit; (3) the gene had to be supported as causal for diabetes- and
obesity-related traits in the previously described BXD (Schadt et al.
2003, 2005) and BXH (Chen et al. 2008) crosses; and (4) the gene
knockout model had to exist in Deltabase. Nine genes were iden-
tified afterapplyingthesefilters and,by furtherrestrictingtogenes
belonging to GPCR family (for convenient targeting with small
molecules) and highly expressed in adipose tissue, we obtained
only two genes in the PEXA module, namely, S1pr2 and P2ry1.
P2ry1 expression levels in liver and adipose tissues ranked low
among all tissues, whereas expression levels for S1pr2 were high in
islets, liver, and adipose tissues. Therefore, we selected S1pr2 for
extensive phenotypic validation.
S1pr2 is known to bind sphingosine-1-phosphate, signal via
G protein to elevate intracellular calcium, and to play an impor-
tant role in neuronal excitability (An et al. 2000; MacLennan et al.
2001). Previous studies have shown that S1pr2 transduced S1P-
evoked signaling events relevant to cell proliferation and survival,
including activation of the ERK/MAP kinases (An et al. 2000).
However, S1pr2 has not previously been associated with diabetes
traits. Therefore, S1pr2?/?knockout mice were compared with
wild-type littermate controls for blood insulin, glucose, and FFA
levels. Each group was comprised of nine males and nine females.
Mice were placed on a chow diet after weaning until 11 wk of age
and then switched to a high-fat, Western diet until 21 wk of age, at
which time blood samples were collected after a 4-h fast. As shown
in Figure 5, plasma insulin levels in the male knockout mice were
significantly increased (t-test P = 0.03), consistent with the pre-
diction that this gene may modulate insulin signaling pheno-
types. Although FFA release increased when S1pr2 was knocked
down in 3T3-L1 adipocytes, the plasma FFA levels remained un-
changed in the S1pr2 mouse KO model. This may suggest the ex-
istence of certain compensatory mechanisms operating in the
whole animal that prevent FFA from increasing.
We performed a siRNA screen on 313 genes to test for insulin
signaling phenotypes and found more than one hundred candi-
date genes supported as having putative role in insulin signaling.
To help increase confidence and to enhance the overall coverage
of genes involved in processes associated with insulin signaling,
we developed the pathway expansion analysis (PEXA). This
method leverages the KEGG pathway database and PPI data as
orthogonally generated sources of information that can be used to
identify the signal in noisy siRNA screening data sets. Application
of this method to our siRNA screening data resulted in the iden-
tification of a core module of genes interacting with known in-
sulin signaling pathways. This new view of the insulin signaling
network enhances our previous understanding of this important
pathway and provides a meaningful list of genes to pursue in
follow-up studies. We objectively identified one of the genes in
this network, S1pr2, that was more strongly supported as causal for
insulin resistance phenotypes and experimentally validated this
prediction in a mouse knockout model. As a generalized tool,
PEXA can be applied to other RNAi screening experiments to
help elucidate the underlying mechanisms driving phenotypes of
Compared with PPIs derived from high-throughput tech-
nologies, the KEGG database contains fewer interactions but these
interactions are of higher quality. In addition, genes in the KEGG
database and their interactions are grouped and ordered as path-
ways, where the topologies of such pathways provide extra, im-
portant information compared to the unordered lists of in-
teractions represented in the PPI data. One common analysis
strategy, gene set enrichment analysis (GSEA), tests sets of genes
comprising pathways or other annotated gene sets (Subramanian
et al. 2005) to assess whether they are significantly enriched for an
input set of genes of interest. This type of test is valuable but
provides no direct information on how input genes interact in the
tested pathway. We have demonstrated that the manually curated
biological pathways comprising the KEGG database are extremely
helpful for elucidating complex experimental data. Of course, we
expect that any given researcher may generate their own best prior
knowledge of any number of pathways, in addition to those rep-
resented in KEGG; our method places no restriction on what
pathway sets can be used.
Based on the module we obtained from the PEXA method, it
appears that insulin signaling may be more complicated than
what is presently represented in the KEGG database. However, due
S1pr2?/?, female S1pr2+/+, and female S1pr2?/?. Mice were on a standard chow diet after weaning until 11 wk of age and then were switched to a high-
fat diet until 21 wk of age. Results shown are for blood samples collected at 21 wk of age after a 4-h fast: (A) plasma insulin levels, (B) plasma free fatty acid
levels, and (C) plasma glucose levels.
Phenotypic differences between S1pr2?/?and S1pr2+/+mice. Blood samples were collected from four groups of mice: male S1pr2+/+, male
Tu et al.
1064 Genome Research
to the incompleteness and presumably low quality of the mouse
PPI data, we would expect PEXA to have missed some insulin
signaling modulators or made false-positive calls. Therefore, the
new view our network provides should not be expected to fully
reflect the complete network of genes associated with this bi-
ological process. Further, in practice experimentally determined
interactions normally reflect different levels of experimental sup-
port. As a result, probabilistic modeling of the PPI network could
significantly enhance network accuracy in this case, as we have
shownin othercontexts(Zhuetal.2008).We aimtoconsidersuch
developments in future work to improve PEXA. Nevertheless, the
siRNA screen that generated the raw information upon which the
PEXA algorithm was based and the follow-up network analysis
significantly extend our understanding of the traditional insulin
signaling pathway, and provide us with a new starting point for
Despiteour successwiththe PEXA method,several important
questions still remain. First, the PEXA module we identified covers
only part of the insulin pathway components/modulators, given
our siRNA screen was restricted to 313 genes. Therefore, the PEXA
module represents a lower bound for genes involved in this pro-
cess. In fact, only 10 of the 136 genes represented in the MGI
mutant database that gave rise to phenotypes related to insulin or
glucose tolerance were included in our network (although this
network is still significantly enriched for genes in this knockout
set). To obtain a more global view of this network, a genome-wide
siRNA screen would undoubtedly be helpful. Second, we are still
far away from understanding the complex logic embedded in
mammalian systems that modulates insulin signaling pheno-
types. For example, the network we derived cannot be used to
predict the effect of perturbing a given gene (e.g., will knocking
a given gene down up- or down-regulate FFA release). This type of
information is essential if we are to develop a detailed under-
standing of the system. Developing additional high-throughput
technologies like genome-wide phosphorylation assays as well as
novel methods for incorporating these data with existing data sets
would significantly enable systems biology efforts to construct
models that predict complex system behavior (Olsen et al. 2006;
Dong et al. 2007; Sahin et al. 2007). The availability of these new
high-throughput data types is on the horizon and will enable
efforts to gain much deeper insights into complex living systems
(Yeang et al. 2004; Janes and Yaffe 2006; Ivakhno and Armstrong
2007; Martin et al. 2007; Ourfali et al. 2007; Villen et al. 2007).
Lastly, although we have generated a general picture for the net-
work underlying insulin signaling, it was derived from a mouse
cell line. Even if we assume that the mouse network is similar to
the human network, and even if we ignore the difference between
the in vitro and in vivo networks, specific efforts will be required
to pinpoint the actual causal genes for insulin resistance in human
impact human health.
Insulin-dependent anti-lipolysis based siRNA screen
The anti-lipolysis screen was developed using 3T3-L1 cells that
had been differentiated as described before (Thompson et al.
2004). Small interfering RNAs (siRNA) were obtained from Dhar-
amcon per vendor’s design. For each gene, three unique siRNAs
targeting different regions of the gene were pooled together. The
adipocytes were transfected with 100 nM pooled siRNA for 48 h
while the cells were induced with 1 nM insulin at the same time.
Free fatty acid (FFA) release was measured 4 d after transfection.
FFA release from cells transfected with target-specific siRNAs was
compared with cells transfected with control siRNAs. For each
gene, threereplicatesweremeasuredusingthesame pooledsiRNA.
For one of the replicate, the gene was flagged if FFA release changed
significantly (t-test P < 0.05). A gene was considered a siRNA hit if
two out of three replicates showed same significant effect.
Protein–protein interaction and KEGG pathway data
The set of mouse PPIs was obtained by integrating several public
(BIND, BioGRID, HPRD, MINT, Reactome, DIP, and IntAct) and
commercial (Ingenuity, Proteome, MetaBase, and NetPro) molec-
ular interaction databases. Identifiers for the interacting genes
identified in these different databases were mapped to Entrez
Gene IDs to obtain a unified naming system. Duplicate inter-
actions were removed, although the number of duplicate inter-
actions was small, consistent with previous reports (Mathivanan
et al. 2006). Mechanisms of interaction that were annotated as
proteins involved in regulation (for example, ‘‘activates,’’ ‘‘inhib-
its,’’ and so on) were mapped to directed edges in the network.
Mechanisms of interaction that were annotated as proteins
involved in binding but with no regulatory effect (for example,
‘‘binds,’’ ‘‘covalent binding,’’ ‘‘ppi,’’ and so on) were mapped to
The KEGG pathways were obtained by parsing metabolic and
signaling pathways contained in KEGG’s public collection of
KGML file describes an individual pathway composed of a list of
participating molecules, as well as a list of their relationships. The
relationships were mapped to undirected edges if they were not
regulatory (e.g., ‘‘binding’’), or to directed edges if they were reg-
ulatory (e.g., ‘‘activation’’). Because metabolic pathways can con-
tain reactions (e.g., a substrate binds to an enzyme to produce
a product), reactions were parsed to obtain two edges: (1) an un-
directed binding edge between the substrate and the enzyme, and
(2) a directed edge between the enzyme and the product. Because
KEGG signaling and metabolic pathways are more complete for
human than mouse, we parsed the human KEGG pathways and
then derived the mouse counterparts by mapping the human
genes to the mouse orthologs.
Construction and characterization of S1pr2?/?and control
S1pr2?/?and control mice were obtained and licensed from Na-
tional Institutes of Health and the knockout construction was
reported previously (Kono et al. 2004). S1pr2?/?knockout mice
and wild-type littermate controls were comprised of nine males
and nine females, respectively. Mice were placed on a chow diet
after weaning until 11 wk of age and then switched to a high-fat,
Western diet until 21 wk of age, at which time blood samples were
collected after a 4-h fast.
We thank Sajjad Qureshi and his colleagues for performing the
siRNA screen experiments and the initial statistical analysis on the
An, S., Zheng, Y., and Bleu, T. 2000. Sphingosine 1-phosphate-induced cell
proliferation, survival, and related signaling events mediated by G
protein-coupled receptors Edg3 and S1pr2. J. Biol. Chem. 275: 288–296.
Integrating siRNA and protein–protein interaction
Araki, E., Lipes, M.A., Patti, M.-E., Bruning, J.C., Haag Iii, B., Johnson, R.S.,
and Kahn, C.R. 1994. Alternative pathway of insulin signalling in
mice with targeted disruption of the IRS-1 gene. Nature 372:
Berns, K., Hijmans, E.M., Mullenders, J., Brummelkamp, T.R., Velds, A.,
Heimerikx, M., Kerkhoven, R.M., Madiredjo, M., Nijkamp, W., Weigelt,
B., et al. 2004. A large-scale RNAi screen in human cells identifies new
components of the p53 pathway. Nature 428: 431–437.
Boutros, M., Kiger, A.A., Armknecht, S., Kerr, K., Hild, M., Koch, B., Haas,
S.A., Consortium, H.F.A., Paro, R., and Perrimon, N. 2004. Genome-
wide RNAi analysis of growth and viability in Drosophila cells. Science
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.
Clee, S.M., Yandell, B.S., Schueler, K.M., Rabaglia, M.E., Richards, O.C.,
Raines, S.M., Kabara, E.A., Klass, D.M., Mui, E.T.K., Stapleton, D.S., et al.
2006. Positional cloning of Sorcs1, a type 2 diabetes quantitative trait
locus. Nat. Genet. 38: 688–693.
Dong, M.-Q., Venable, J.D., Au, N., Xu, T., Park, S.K., Cociorva, D., Johnson,
J.R., Dillin, A., and Yates III, J.R. 2007. Quantitative mass spectrometry
identifies insulin signaling targets in C. elegans. Science 317: 660–663.
Draznin, B. 2006. Molecular mechanisms of insulin resistance: Serine
phosphorylation of insulin receptor substrate-1 and increased
expression of p85a: The two sides of a coin. Diabetes 55: 2392–2397.
Eppig, J.T., Bult, C.J., Kadin, J.A., Richardson, J.E., Blake, J.A.,
Anagnostopoulos, A., Baldarelli, R.M., Baya, M., Beal, J.S., Bello, S.M,
et al. 2005. The Mouse Genome Database (MGD): From genes to
mice—A community resource for mouse biology. Nucleic Acids Res. 33:
Gower, B.A., Herd, S.L., and Goran, M.I. 2001. Anti-lipolytic effects of
insulin in African American and white prepubertal boys. Obes. Res. 9:
Ideker, T., Ozier, O., Schwikowski, B., and Siegel, A.F. 2002. Discovering
regulatory and signalling circuits in molecular interaction networks.
Bioinformatics 18: S233–S240.
Ivakhno, S. and Armstrong, J.D. 2007. Non-linear dimensionality reduction
of signaling networks. BMC Syst. Biol. 1: 27. doi: 10.1186/1752-0509-1-
Janes, K.A. and Yaffe, M.B. 2006. Data-driven modelling of signal-
transduction networks. Nat. Rev. Mol. Cell Biol. 7: 820–828.
Jensen, M.D. 2006. Adipose tissue as an endocrine organ: Implications of
its distribution on free fatty acid metabolism. Eur. Heart J. Suppl. 8:
Kanehisa, M., Goto, S., Kawashima, S., Okuno, Y., and Hattori, M. 2004. The
KEGG resource for deciphering the genome. Nucleic Acids Res. 32:
Kanehisa, M., Goto, S., Hattori, M., Aoki-Kinoshita, K.F., Itoh, M.,
Kawashima, S., Katayama, T., Araki, M., and Hirakawa, M. 2006. From
genomics to chemical genomics: New developments in KEGG. Nucleic
Acids Res. 34: D354–D357.
Kim, J.Y., Tillison, K., Lee, J.-H., Rearick, D.A., and Smas, C.M. 2006. The
adipose tissue triglyceride lipase ATGL/PNPLA2 is downregulated by
insulin and TNF-a in 3T3-L1 adipocytes and is a target for
transactivation by PPARg. Am. J. Physiol. Endocrinol. Metab. 291:
Kono, M., Mi, Y., Liu, Y., Sasaki, T., Allende, M.L., Wu, Y.-P., Yamashita, T.,
and Proia, R.L. 2004. The sphingosine-1-phosphate receptors S1P1,
S1P2, and S1P3function coordinately during embryonic angiogenesis. J.
Biol. Chem. 279: 29367–29373.
Kulkarni, M.M., Booker, M., Silver, S.J., Friedman, A., Hong, P., Perrimon,
N., and Mathey-Prevot, B. 2006. Evidence of off-target effects associated
with long dsRNAs in Drosophila melanogaster cell-based assays. Nat.
Methods 3: 833–838.
Lake, A.C., Sun, Y., Li, J.-L., Kim, J.E., Johnson, J.W., Li, D., Revett, T., Shih,
H.H., Liu, W., Paulsen, J.E., et al. 2005. Expression, regulation, and
triglyceride hydrolase activity of Adiponutrin family members. J. Lipid
Res. 46: 2477–2487.
Liu,M., Liberzon, A.,Kong, S.W., Lai, W.R., Park, P.J., Kohane, I.S.,and Kasif,
S. 2007. Network-based analysis of affected biological processes in type
Lu, J., Ruhf, M.-L., Perrimon, N., and Leder, P. 2007. A genome-wide RNA
interference screen identifies putative chromatin regulators essential for
E2F repression. Proc. Natl. Acad. Sci. 104: 9381–9386.
Ma, Y., Creanga, A., Lum, L., and Beachy, P.A. 2006. Prevalence of off-
target effects in Drosophila RNA interference screens. Nature 443:
MacLennan, A.J., Carney, P.R., Zhu, W.J., Chaves, A.H., Garcia, J., Grimes,
J.R., Anderson, K.J., Roper, S.N., and Lee, N. 2001. An essential role for
the H218/AGR16/Edg-5/LPB2 sphingosine 1-phosphate receptor in
neuronal excitability. Eur. J. Neurosci. 14: 203–209.
Martin, M., Reidhaar-Olson, J.F., and Rondinone, C.M. 2007. Genetic
association meets RNA interference: Large-scale genomic screens for
causation and mechanism of complex diseases. Pharmacogenomics 8:
Mathivanan, S., Periaswamy, B., Gandhi, T.K.B., Kandasamy, K., Suresh, S.,
Mohmood, R., Ramachandra, Y.L., and Pandey, A. 2006. An evaluation
of human protein-protein interaction data in the public domain. BMC
Bioinformatics 7: S19. doi: 10.1186/1471-2105-7-S5-S19.
Mehrabian, M., Allayee, H., Stockton, J., Lum, P.Y., Drake, T.A., Castellani,
L.W., Suh, M., Armour, C., Edwards, S., Lamb, J., et al. 2005. Integrating
genotypic and expression data in a segregating mouse population to
identify 5-lipoxygenase as a susceptibility gene for obesity and bone
traits. Nat. Genet. 37: 1224–1233.
Mello, C.C. and Conte, D. 2004. Revealing the world of RNA interference.
Nature 431: 338–342.
Moffat, J. and Sabatini, D.M. 2006. Building mammalian signalling
pathways with RNAi screens. Nat. Rev. Mol. Cell Biol. 7: 177–187.
Nollen, E.A.A., Garcia, S.M., van Haaften, G., Kim, S., Chavez, A.,
Morimoto, R.I., and Plasterk, R.H.A. 2004. From the cover: Genome-
wide RNA interference screen identifies previously undescribed
regulators of polyglutamine aggregation. Proc. Natl. Acad. Sci. 101:
Olsen, J.V., Blagoev, B., Gnad, F., Macek, B., Kumar, C., Mortensen, P., and
Mann, M. 2006. Global, in vivo, and site-specific phosphorylation
dynamics in signaling networks. Cell 127: 635–648.
Ourfali, O., Shlomi, T., Ideker, T., Ruppin, E., and Sharan, R. 2007. SPINE: A
framework for signaling-regulatory pathway inference from cause-effect
experiments. Bioinformatics 23: i359–i366.
Pascoe, L., Tura, A., Patel, S.K., Ibrahim, I.M., Ferrannini, E., Zeggini, E.,
Weedon, M.N., Mari, A., Hattersley, A.T., McCarthy, M.I., et al. 2007.
Common variants of the novel type 2 diabetes genes CDKAL1 and
HHEX/IDE are associated with decreased pancreatic b-cell function.
Diabetes 56: 3101–3104.
Pessin, J.E. and Saltiel, A.R. 2000. Signaling pathways in insulin action:
Molecular targets of insulin resistance. J. Clin. Invest. 106: 165–169.
Sahin, O., Lobke, C., Korf, U., Appelhans, H., Sultmann, H., Poustka, A.,
Wiemann, S., and Arlt, D. 2007. Combinatorial RNAi for quantitative
protein network analysis. Proc. Natl. Acad. Sci. 104: 6579–6584.
Schadt, E.E., Monks, S.A., Drake, T.A., Lusis, A.J., Che, N., Colinayo, V., Ruff,
T.G., Milligan, S.B., Lamb, J.R., Cavet, G., et al. 2003. Genetics of gene
expression surveyed in maize, mouse and man. Nature 422: 297–302.
Schadt, E.E., Lamb, J., Yang, X., Zhu, J., Edwards, S., GuhaThakurta, D.,
Sieberts, S.K., Monks, S., Reitman, M., Zhang, C., et al. 2005. An
integrative genomics approach to infer causal associations between
gene expression and disease. Nat. Genet. 37: 710–717.
Schoenborn, V., Heid, I.M., Vollmert, C., Lingenhel, A., Adams, T.D.,
Hopkins, P.N., Illig, T., Zimmermann, R., Zechner, R., Hunt, S.C., et al.
2006. The ATGL gene is associated with free fatty acids, triglycerides,
and type 2 diabetes. Diabetes 55: 1270–1275.
Scott, L.J., Mohlke, K.L., Bonnycastle, L.L., Willer, C.J., Li, Y., Duren, W.L.,
Erdos, M.R., Stringham, H.M., Chines, P.S., Jackson, A.U., et al. 2007. A
genome-wide association study of type 2 diabetes in Finns detects
multiple susceptibility variants. Science 316: 1341–1345.
Solinas, G., Vilcu, C., Neels, J.G., Bandyopadhyay, G.K., Luo, J.-L., Naugler,
W., Grivennikov, S., Wynshaw-Boris, A., Scadeng, M., Olefsky, J.M.,
et al. 2007. JNK1 in hematopoietically derived cells contributes to diet-
induced inflammation and insulin resistance without affecting obesity.
Cell Metab. 6: 386–397.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L.,
Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S.,
et al. 2005. From the cover: Gene set enrichment analysis: A knowledge-
based approach for interpreting genome-wide expression profiles. Proc.
Natl. Acad. Sci. 102: 15545–15550.
Tamemoto, H., Kadowaki, T., Tobe, K., Yagi, T., Sakura, H., Hayakawa, T.,
Terauchi, Y., Ueki, K., Kaburagi, Y., Satoh, S., et al. 1994. Insulin
resistance and growth retardation in mice lacking insulin receptor
substrate-1. Nature 372: 182–186.
Taniguchi, C.M., Emanuelli, B., and Kahn, C.R. 2006. Critical nodes in
signalling pathways: Insights into insulin action. Nat. Rev. Mol. Cell Biol.
Thompson, G.M., Trainor, D., Biswas, C., LaCerte, C., Berger, J.P., and Kelly,
L.J. 2004. A high-capacity assay for PPARg ligand regulation of
endogenous aP2 expression in 3T3-L1 cells. Anal. Biochem. 330: 21–28.
Turner, R.C., Holman, R.R., Matthews, D., Hockaday, T.D.R., and Peto, J.
1979. Insulin deficiency and insulin resistance interaction in diabetes:
Estimation of their relative contribution by feedback analysis from
basal plasma insulin and glucose concentrations. Metabolism 28: 1086–
Villen, J., Beausoleil, S.A., Gerber, S.A., and Gygi, S.P. 2007. Large-scale
phosphorylation analysis of mouse liver. Proc. Natl. Acad. Sci. 104:
Tu et al.
1066 Genome Research
Yeang, C.-H., Ideker, T., and Jaakkola, T. 2004. Physical network models. J. Download full-text
Comput. Biol. 11: 243–262.
Yki-Jarvinen, H. and Taskinen, M.R. 1988. Interrelationships among
insulin’s antilipolytic and glucoregulatory effects and plasma
triglycerides in nondiabetic and diabetic patients with endogenous
hypertriglyceridemia. Diabetes 37: 1271–1278.
Lipolysis: Pathway under construction. Curr. Opin. Lipidol. 16: 333–340.
Zhu, J., Zhang, B., Smith, E.N., Drees, B., Brem, R.B., Kruglyak, L.,
Bumgarner, R.E., and Schadt, E.E. 2008. Integrating large-scale
functional genomic data to dissect the complexity of yeast regulatory
networks. Nat. Genet. 40: 854–861.
Received October 9, 2008; accepted in revised form February 22, 2009.
Integrating siRNA and protein–protein interaction