Network Analysis of Human In-Stent Restenosis
Euan A. Ashley, MRCP, DPhil; Rossella Ferrara, MD; Jennifer Y. King, BS; Aditya Vailaya, PhD;
Allan Kuchinsky, MS; Xuanmin He, MD, PhD; Blake Byers, BS; Ulrich Gerckens, MD;
Stefan Oblin, MD; Anya Tsalenko, PhD; Angela Soito, BS, JD; Joshua M. Spin, MD, PhD;
Raymond Tabibiazar, MD; Andrew J. Connolly, MD; John B. Simpson, MD, PhD;
Eberhard Grube, MD; Thomas Quertermous, MD
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Background—Recent successes in the treatment of in-stent restenosis (ISR) by drug-eluting stents belie the challenges still
faced in certain lesions and patient groups. We analyzed human coronary atheroma in de novo and restenotic disease
to identify targets of therapy that might avoid these limitations.
Methods and Results—We recruited 89 patients who underwent coronary atherectomy for de novo atherosclerosis (n⫽55)
or in-stent restenosis (ISR) of a bare metal stent (n⫽34). Samples were fixed for histology, and gene expression was
assessed with a dual-dye 22 000 oligonucleotide microarray. Histological analysis revealed significantly greater
cellularity and significantly fewer inflammatory infiltrates and lipid pools in the ISR group. Gene ontology analysis
demonstrated the prominence of cell proliferation programs in ISR and inflammation/immune programs in de novo
restenosis. Network analysis, which combines semantic mining of the published literature with the expression signature
of ISR, revealed gene expression modules suggested as candidates for selective inhibition of restenotic disease. Two
modules are presented in more detail, the procollagen type 1 ␣2 gene and the ADAM17/tumor necrosis factor-␣
converting enzyme gene. We tested our contention that this method is capable of identifying successful targets of
therapy by comparing mean significance scores for networks generated from subsets of the published literature
containing the terms “sirolimus” or “paclitaxel.” In addition, we generated 2 large networks with sirolimus and
paclitaxel at their centers. Both analyses revealed higher mean values for sirolimus, suggesting that this agent has a
broader suppressive action against ISR than paclitaxel.
Conclusions—Comprehensive histological and gene network analysis of human ISR reveals potential targets for directed
abrogation of restenotic disease and recapitulates the results of clinical trials of existing agents. (Circulation. 2006;114:
2644-2654.)
Key Words: atherosclerosis 䡲 paclitaxel 䡲 restenosis 䡲 sirolimus
S
ince Sigwart et al1 first reported the use of a metallic stent
as an adjunct to balloon angioplasty for the treatment of
flow limitation in coronary arteries, the issue of in-stent
restenosis (ISR) has been the focus of attention of interventional cardiologists and vascular biologists alike. Incidence in
published studies ranges between 10% and 60%,2,3 and
despite early attempts with systemic therapy and direct
stenting, only the more recent use of stents coated with
antiproliferative agents has made a significant impact on the
disease.4 Indeed, such is the benefit of these drug-eluting
stents that there has been almost universal adoption, at least in
the United States, for eligible lesions.
Clinical Perspective p 2654
However, although sirolimus- or paclitaxel-eluting stents have
consistently demonstrated lower rates of ISR compared with the
bare metal control in a broad spectrum of lesion types, sufficient
data for specific lesion subsets such as saphenous vein grafts,
bifurcation lesions, and acute myocardial infarctions are still
missing.5 In addition, concerns about the incidence of late stent
thrombosis, related to delayed stent endothelialization, have
tempered enthusiasm.6 Finally, although the occurrence of ISR
within drug-eluting stents is greatly reduced, it remains between
3% and ⬎25% in published trials.6
Received April 29, 2006; revision received September 21, 2006; accepted September 22, 2006.
From the Division of Cardiovascular Medicine, Falk CVRC, Stanford University, Stanford, Calif (E.A.A., R.F., J.Y.K., B.B., J.M.S., A.J.C., T.Q.);
HELIOS Heart Center Siegburg, Siegburg, Germany (U.G., S.O., E.G.); Agilent Laboratories, Palo Alto, Calif (A.V., A.K., A.T.); Foxhollow
Technologies, Redwood City, Calif (X.H., A.S., J.B.S.); and Aviir Inc, Palo Alto, Calif (R.T.).
Correspondence to Dr E.A. Ashley, Stanford University, 300 Pasteur Dr, Stanford, CA 94305 (e-mail [email protected] ); Dr E. Grube, Stanford
University, 300 Pasteur Dr, Stanford, CA 94305 (e-mail [email protected]); or Dr T. Quertermous, Stanford University, 300 Pasteur Dr, Stanford, CA
94305 (e-mail [email protected]).
© 2006 American Heart Association, Inc.
Circulation is available at http://www.circulationaha.org
DOI: 10.1161/CIRCULATIONAHA.106.637025
2644
Ashley et al
Network Analysis of In-Stent Restenosis
2645
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Figure 1. Flow diagram of approach to experiment and analysis. Patients undergo atherectomy, and tissue is either stored for RNA isolation or fixed for histological analysis. The expression profile of de novo atherosclerosis is then subtracted from that of ISR. Independently, biological networks are generated from the literature using text mining. Expression profiles and significance scores of genes are
then overlaid on the networks, and each network is given an overall significance score. Top scoring networks are curated and mined
for biological significance.
For these reasons and others, there is continued interest in
new pharmacotherapeutic agents, suitable for stent-based
delivery to the coronary vasculature, that better address the
needs of this particular vascular bed. One approach lies in a
more sophisticated understanding of restenosis biology. Although classic studies established the role of plateletfacilitated inflammation,7 migration, dedifferentiation, and
proliferation of smooth muscle cells and the elaboration of
extracellular matrix (neointimal hyperplasia),8 more in-depth
characterization has been limited by access to human coronary atheromatous material. Meanwhile, more recent highdimensional genomic and proteomic tools allow not only a
more comprehensive view of the disease at the transcriptional
level but also a perspective not limited by preconceived
speculation as to the process of interest. Moreover, although
light microscopy offers a morphological examination of the
cell types involved at the tissue level, molecular profiling can
offer insight into the ongoing biological program of those
cells.
In the present article, we assess the form and function of
coronary vascular disease through light microscopy and
nearly genome-wide gene expression to offer novel insights
into the process of ISR. Furthermore, we use an original
integrative analysis to identify key components of biological
pathways that may be candidates for future therapeutics.
Finally, we test our approach against results of recently
published clinical trials comparing currently available elution
agents.
Methods
The experimental approach is summarized in Figure 1.
Patient Recruitment and Tissue Harvest
Eighty-nine patients with significant coronary lesions suitable for
atherectomy were recruited at the HELIOS Heart Center Siegburg
(Siegburg, Germany). Standard premedication with oral aspirin,
clopidogrel, and intracoronary heparin was given. Femoral arterial
access was gained with 6F or 8F introducers and a sterile, over-thewire technique. Diagnostic angiography was then performed with
standard catheters, and the coronary anatomy was defined. If a
significant, flow-limiting stenosis was identified in a moderate-sized
vessel, an extra support guidewire was deployed, and longitudinal
atherectomy was performed with the Silverhawk atherectomy catheter (Foxhollow Technologies, Redwood City, Calif). All procedures
were concluded with standard percutaneous coronary transluminal
angiography/stenting at the operator’s discretion. Extracted specimens were divided and stored immediately in both liquid nitrogen
and formalin. Of 89 samples, 56 had RNA of sufficient quality for
hybridization, and 47 arrays were judged to be of sufficient quality
for analysis after rigorous quality control.
The study was approved by the Stanford University Institutional
Review Board. All patients gave written, informed consent.
2646
Circulation
December 12, 2006
TABLE 1.
Patient Characteristics for the 47 Samples Included in the Genomic Analysis
Age, y
Male, n (%)
Unknown
Characteristic
De Novo Restenosis (n⫽32)
ISR (n⫽15)
66⫾11.85
70⫾7.66
21 (84)
9 (64.28)
䡠䡠䡠
0.75
䡠䡠䡠
1 (7.14)
䡠䡠䡠
14 (100)
0.07
P
Ethnicity, n (%)
White
21 (84)
Asian
1 (4)
䡠䡠䡠
䡠䡠䡠
Unknown
3 (12)
䡠䡠䡠
䡠䡠䡠
Hypertension, n (%)
Yes
13 (52)
10 (71.42)
0.12
No
7 (28)
2 (14.28)
䡠䡠䡠
Unknown
5 (20)
2 (14.28)
䡠䡠䡠
Hypercholesterolemia, n (%)
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Yes
15 (60)
12 (85.71)
0.03
No
6 (24)
2 (14.28)
䡠䡠䡠
Unknown
4 (16)
䡠䡠䡠
䡠䡠䡠
Yes
1 (4)
2 (14.28)
0.23
No
11 (44)
8 (57.14)
䡠䡠䡠
Unknown
13 (52)
4 (28.57)
䡠䡠䡠
Diabetes, n (%)
Tobacco, n (%)
Yes
5 (20)
4 (28.57)
0.44
No
20 (80)
10 (71.42)
䡠䡠䡠
Yes
11 (44)
12 (85.71)
0.005
No
4 (16)
䡠䡠䡠
2 (14.28)
䡠䡠䡠
10 (40)
10 (40)
7 (50)
0.34
0
9 (64.28)
䡠䡠䡠
0.53
3 (12)
2 (14.28)
0.65
14 (56)
11 (78.57)
0.07
Previous acute coronary syndrome, n (%)
Unknown
䡠䡠䡠
Drug therapy, n (%)
ACE inhibitors
Angiotensin receptor blockers
␤-Blockers
Ca channel antagonist
Statin
0
15 (60)
Age is given as mean⫾SD. ACE indicates angiotensin-converting enzyme.
Development of Oligonucleotide
Microarray Platform
Direct Labeling and Oligonucleotide
Array Hybridization
A custom microarray was developed by our laboratory in collaboration with Agilent Technologies (Palo Alto, Calif). Its development
has been described in detail previously.9 In brief, a combination of
techniques based on experimental cell culture and data mining was
used to provide a comprehensive catalog of vascular and atherosclerosis-related genes. Probes derived from this list were combined with
those from Agilent human catalog arrays 1A and 1B to create the
final 22 000 –feature microarray.
Our protocol has been described in detail previously.9 In brief, 2 to
5 ␮g RNA for sample and reference was used. A mixture of 80%
HeLa RNA and 20% human umbilical vein endothelial cell RNA
was used for reference because it had previously been shown to
maximize feature fluorescence for this array. After linear amplification, fluorescent-labeled cDNA was reverse transcribed from the
RNA with SuperScript II (Invitrogen, Carlsbad, Calif) and either
cyanine 3-dCTP or cyanine 5-dCTP (Perkin-Elmer/NEN, Boston,
Mass). After purification, hybridization continued for 16 hours at
60°C. Arrays were scanned with the G2565AA Microarray Scanner
System, and features were extracted as previously described. Excluded features were estimated with the K nearest neighbors imputation method of Troyanskaya et al.10
RNA Isolation and Quantification
RNA was isolated using standard techniques based on guanidium
thiocyanate as described previously.9 Extensive quality control
testing of all RNA was carried out. Samples were quantified with the
NanoDrop ND-1000 Spectrophotometer (NanoDrop, Wilmington,
DE), and RNA integrity was assessed with the 2100 Bioanalyzer
System and RNA 6000 Pico LabChip Kit (Agilent Technologies).
Histology
Tissue samples were fixed in 10% neutral buffered formalin,
processed routinely through graded alcohols and clearing reagent,
and embedded in paraffin blocks. Tissue sections (5 ␮m thick) were
Ashley et al
Network Analysis of In-Stent Restenosis
2647
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Figure 2. Histology of atherectomy specimens. Specimens from de novo atherosclerosis and ISR are shown. A, B, Gross specimens
of average and larger size. C. Low-cellularity lesion (Gomori’s trichrome stain, ⫻100). D, High-cellularity lesion (Gomori’s trichrome
stain, ⫻100). E, Example of a foam cell (arrow) (elastic van Gieson stain, ⫻400). F, Lipid deposits (arrow) (Gomori’s trichrome stain,
⫻100). G, Marked inflammatory infiltrate (Gomori’s trichrome stain, ⫻100). H, Rare example of tunica media (elastic van Gieson stain,
⫻100).
cut and stained with hematoxylin and eosin, Gomori’s trichrome,
and elastic van Gieson stain. Sections were classified as de novo
atherosclerosis or ISR and evaluated microscopically. Samples
were graded according to the presence of certain key features: low
cellularity (abundant dense staining of collagen with sparse cells),
high cellularity (rich stellate or spindle cells within a loose
extracellular matrix or limited amount of collagen), lipid deposit
(focal area with clusters of foam cells, cholesterol clefts, pale
granular materials, and acellular debris), and inflammatory focus
(the presence of clusters or aggregates of mononuclear cells,
including monocytes, macrophages, and lymphocytes).
Analytical Methods
Proportions of categorical demographic and histological variables
were compared through the use of Fisher’s exact test.
Our primary difference measure for genomic analysis was the
significance analysis of microarrays.11 With this measure, a significance score is calculated as follows:
y៮ k共i兲⫺y៮ j共i兲/s i⫹s 0,
where y៮ k(i) is the mean expression of gene i in the ISR group, y៮ j(i) is
the mean expression of gene i in the de novo atherosclerosis group,
si is the standard deviation of repeated expression measurements,
and s0 is a small positive constant. The metric is a modified t
statistic (labeled d because of the addition of a small positive
constant to the denominator). Multiple testing is accounted for by
bootstrap permutation. Results from difference analysis are displayed
using heatmap software developed by the authors and freely available to
the academic community (http://quertermous.stanford.edu/heatmap.htm).
Differentially expressed genes were analyzed in the context of
gene ontology (www.geneontology.org) to identify groups with
similar functions or processes. In brief, all genes represented on the
microarray were assigned to corresponding gene ontology terms
with the Biomolecule Naming Service.12 A gene was associated
with a gene ontology term if it was annotated by this term or its
child. For each gene ontology term, we computed a probability
value based on the hypergeometric distribution by comparing the
number of genes annotated by the gene ontology term in a given
list of differentially expressed genes with the expected number of
such genes.
We have previously described a method for identifying highly
connected genes by semantic mining of the published literature.9
We called these genes nexus genes to emphasize their central role in
biological networks and to make a distinction from the hub gene
concept, in which connections are derived from a form of network
analysis centered around correlation of gene expression. In our
method, we use semantic mining of the published literature. An
association network is derived from text mining of Medline abstracts
and association identified between any 2 genes if they appear in the
same sentence as an interaction verb as defined by our user context
file. A series of subnetworks (independent of our experimental data)
is then generated, and the expression values and significance of
genes in our analysis overlaid visually and mathematically on these
networks. An overall significance score for each subnetwork is
calculated by using the mean significance (d) score value, or the
cumulative sum of significance (d) score values, for all members.
Higher scores thus indicate greater significance of the subnetwork.
To test our hypothesis that connectivity analysis is capable of
identifying targets that may lead to more efficacious therapies, we
took advantage of the recent publication of multiple head-to-head
trials of paclitaxel- and sirolimus-eluting stents.4 These clinical trials
demonstrated a small but consistent and significant benefit of
sirolimus over paclitaxel in the inhibition of ISR.13 We posited that
our methods should be capable of predicting this result. Thus, 2
abstract subdatabases were generated by entering the following
queries into Medline: “(rapamycin OR sirolimus) NOT paclitaxel”
TABLE 2. Histological Features of 55 De Novo Restenotic and
34 ISR Lesions
Lesions
De Novo Cases, n (%)
ISR Cases, n (%)
P
Low cellularity
45 (81.8)
9 (26.5)
⬍0.0001
High cellularity
10 (18.2)
25 (73.5)
⬍0.0001
Lipid deposit
14 (25.5)
1 (2.9)
0.001
Inflammatory focus
21 (38.2)
3 (8.8)
0.0004
Thrombus
11 (20)
4 (11.8)
0.25
0.66
Calcification
3 (5.5)
1 (2.9)
Cases, n
55
34
䡠䡠䡠
Cellularity was significantly higher in the ISR group; lipid deposits and
inflammatory foci were more common in the de novo group. There was no
difference in the frequency of thrombus or calcification.
2648
Circulation
December 12, 2006
and “paclitaxel NOT sirolimus NOT rapamycin.” Network analysis
was then carried out within these abstract databases, and the overall
significance of each subnetwork was calculated as detailed above.
The means for sirolimus and paclitaxel groups were compared by
Student’s t test. In a separate analysis, we incorporated the terms
“sirolimus” and “paclitaxel” directly as interacting gene terms to
generate 2 large additional networks. Overall significance also was
compared for these networks.
Pathway interactions are mapped with Cytoscape (version 2.2)14
with the Agilent literature search plug-in (version 2). Using a
separate routine, we overlaid significance levels as colors and
depicted raw data as a short “heat strip” below each node.
The authors had full access to the data and take responsibility for
their integrity. All authors have read and agree to the manuscript as
written.
Results
Patient Population
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Patient characteristics for the 47 samples in the genomic
analysis are detailed in Table 1. Although overall similar,
some demographic variables were significantly different between the 2 populations. Hypercholesterolemia was more
common in the ISR group (15 of 32 versus 12 of 15; P⫽0.03).
This was reflected in greater use of statins (14 of 32 versus 11
of 14; P⫽0.07). In addition, a history of acute coronary
syndrome was much more common in the ISR group (11 of
32 versus 12 of 15; P⫽0.005). The prevalence of known
diabetes was low in our cohort and not different between
groups (1 of 32 versus 2 of 15; P⫽0.23), but this may reflect
the high level of unknowns in this category.
Gross Specimen and Histological Analysis
Lesions were classified as de novo restenosis or ISR by the
operator at the time of catheterization. Samples were weighed
before tissue processing (mean, 1.22⫾1.02 mg; range, 0.1 to
4.77 mg; Figure 2A and 2B). Histological analysis (Figure 2C
through 2H and Table 2) revealed significantly greater
frequency of low-cellularity lesions in the de novo group
(81.8% in de novo versus 26.5% in ISR; P⬍0.0001); there
was a corresponding higher frequency of high-cellularity
lesions within the intima of ISR lesions (73.5% in ISR versus
18.2% in de novo; P⬍0.0001). De novo lesions exhibited
classically recognized features of atherosclerosis to a greater
extent than ISR lesions: lipid deposit, 25.5% in de novo
versus 2.9% in ISR (P⫽0.001); and inflammatory infiltrates,
38.8% in de novo versus 8.8% in ISR( P⫽0.0004). There was
no significant difference in the incidence of calcification or
thrombosis between groups.
Gene Expression Analysis
Comparison analysis using the significance analysis of microarrays revealed highly significant differences between de
novo lesions and ISR lesions. Plotting the false discovery
characteristics (Figure 3A and 3B) reveals a nadir rate of false
discovery at ⬇200 genes (0.77 genes per 207 genes depicted,
⬍1 falsely significant gene, 0.4% false discovery rate). In
addition, at this level of false discovery, all significant genes
had greater expression in the ISR group (Figure 3B and 3C).
Ontology Analysis
Although gene lists are informative, systems analysis lends
structure to large data sets. Patterns of gene expression within
Figure 3. Differential gene expression between de novo restenosis and ISR. A, Graphic representation of false discovery as a
function of genes called significant. The nadir of the graph represents the optimal balance of true positives to false positives.
B, Graphic representation of actual significance (bold black/red
line) and mean permuted significance (thin blue line) for this
data set. The further the former departs from the latter, the
more significant the genes are. The gray dashed line represents
the limit of genes called at a false significance level of 0.77
genes per 207 genes depicted (⬍1 falsely significant gene,
0.4% false discovery rate; red represents significant genes). All
significant genes are higher in the ISR group at this level of false
discovery. C, Heatmap representation of significant genes
shown in B. Samples in columns and rows are genes. Gene
expression is represented as a color, normalized across each
row, with brighter red for higher values and brighter green for
lower values. FDR indicates false discovery rate.
the most upregulated genes in each group (t test) were
assessed for overrepresentation in the context of molecular
function, biological process, and cellular component (Gene
Ontology Consortium). Significant gene ontology terms are
depicted in Figure 4 (all P⬍1.5⫻10⫺5). As shown, the
primary molecular signature differentiating de novo atherosclerosis from ISR is an immune/inflammatory one (Figure
4A and 4B). All significant terms in genes upregulated in de
novo atherosclerosis are children of the “response to stress”
and the “response to external stimulus” parents. In contrast,
the primary molecular signature of ISR is cell growth and
extracellular matrix (Figure 4A and 4B). We found no
evidence of overrepresentation of inflammatory ontologies
compared with de novo atherosclerosis in our data.
Ashley et al
Network Analysis of In-Stent Restenosis
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Figure 3. Continued
2649
2650
Circulation
December 12, 2006
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Figure 4. Ontology analysis of genes differentially expressed between de novo restenosis and ISR. Patterns of gene expression in the
most upregulated genes in each group were assessed for significant overrepresentation in the context of molecular function, biological
process, and cellular component (Gene Ontology Consortium). The size of the nodes is inversely proportional to the probability value of
the term. A, B, Genes upregulated in de novo atherosclerosis show a predominantly immune/inflammatory theme. All significant terms
are children of the “response to stress” and “response to external stimulus” parents (all P⬍1.9⫻10⫺5). C, D, In contrast, the primary
molecular signature of ISR is cell growth and extracellular matrix (all P⬍1.5⫻10⫺5).
Network Analysis
We generated subnetworks using text mining of Medline
abstracts and awarded each an overall significance score
based on the average significance of network members
within our experimental data set (Table 3). Subnetworks
were explored using Cytoscape, and sentences were curated for the most promising candidate networks. Example
subnetworks are reported in Table 3 and illustrated in
Figure 5. The analysis provided a highly enriched source
of candidates, both expected (collagen genes, fibroblast
receptor gene, interleukin-8) and unexpected (ADAM
genes, nm23 protein gene). Figure 5 shows the ADAM17
subnetwork. ADAM17 is the tumor necrosis factor
(TNF)-␣ converting enzyme, which regulates inflammation. Other notable members of the ADAM17 subnetwork
include other disintegrin and metalloproteinase domain
(ADAM) proteins, inducible nitric oxide synthase, tissue
inhibitors of matrix metalloproteinases, and endothelial
growth factor.
Targets for inhibiting ISR include differentially regulated
networks with nexus genes highly upregulated in this tissue.
Several are reported in Table 3, and one is shown in Figure
5B. The procollagen type I ␣2 gene (COL1A2) was as highly
upregulated as any nexus gene in the top 50 subnetworks (d
Ashley et al
Network Analysis of In-Stent Restenosis
2651
TABLE 3. Selected Subnetworks Ranked According to the Mean Significance Score of
the Network
Symbol
Name
Nodes
d Score
Cumulative
Mean
2.48
COL5A2
Collagen type V ␣2
7
4.62
17.38
COL1A1
Collagen type I ␣1
10
4.32
20.45
2.04
ADAM10
A disintegrin and metalloproteinase domain 10
11
⫺2.72
22.22
2.02
MAP4K4
MAP kinase kinase kinase kinase 4
10
3.91
19.66
1.97
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
COL3A1
Collagen type III ␣1
7
4.86
13.29
1.90
COL1A2
Collagen type I ␣2
14
4.97
29.75
1.86
GLI3
GLI-Kruppel family member GLI3
9
2.52
15.67
1.74
GPX1
Glutathione peroxidase 1
13
⫺3.20
22.37
1.72
IL8RA
Interleukin-8 receptor-␣
9
⫺0.38
15.13
1.68
SERPINF1
Serine (or cysteine) proteinase inhibitor, clade F
member 1
13
⫺2.37
21.49
1.65
ADAM17
A disintegrin and metalloproteinase domain 17
(tumor necrosis factor-␣ converting enzyme)
15
⫺2.46
30.78
1.62
NME1
Nonmetastatic cells 1 protein (NM23A)
17
0.83
27.26
1.60
FGFR2
Fibroblast growth factor receptor 2
9
0.09
14.18
1.58
Nodes is the number of nodes in each subnetwork; d score is the significance level of the nexus gene (the higher
the positive number, the more significantly the gene was upregulated in ISR vs de novo; the lower the negative
number, the more significantly the gene was downregulated in ISR vs de novo); cumulative is the summed
significance value for the whole network; mean is cumulative divided by number of nodes. The list is highly enriched
for collagen genes and inflammatory genes such as ADAM proteins and interleukin-8 receptor-␣.
score⫽4.97). The subnetwork contained other collagen genes
(COL1A1) and regulatory elements (Sp1), as well as key
mediators of inflammation: interferon-␥, transforming growth
factor-␤, major histocompatibility class II transactivator,
interleukin-4, fundamental signaling molecules (PI3 kinase,
mitogen-activated protein kinase 8), and matrix metalloproteinase 13.
To test the idea that useful targets of therapy could be
identified by our approach, we took advantage of recently
published head-to-head clinical trials of 2 agents effective
against ISR. Networks were generated using subsets of the
published literature that contained the terms “sirolimus or
rapamycin NOT paclitaxel” or the reverse as noted above.
The paclitaxel search generated 11 136 abstracts; the sirolimus search generated 4854 abstracts. Despite fewer abstracts,
the sirolimus network contained more gene associations. The
sirolimus network consisted of 818 nodes and 3690 associations; the paclitaxel network consisted of 611 nodes and 1862
associations. The overall significance levels of subnetworks
with ⱖ5 nodes are shown in Figure 5C. Empirical testing
suggests that average d scores for subnetworks with ⬍5
nodes are disproportionately affected by 1 or 2 very significant interactions. We hypothesized that whichever of the
agents was more effective should demonstrate a higher mean
significance level among its subnetworks. As shown, the
mean significance level of subnetworks in the sirolimus
group was significantly higher than for the paclitaxel group
(P⫽0.028), suggesting that abstracts that included sirolimus
in their content were likely to contain genes more significantly different between de novo and ISR than those that
included paclitaxel. Furthermore, when individual networks
were generated by incorporating the terms sirolimus and
paclitaxel directly as interacting nodes, the sirolimus network
(518 associations) had an overall significance level (0.85) that
was higher than that of paclitaxel (294 associations; overall
significance level, 0.81). Together, these results support the
idea that sirolimus has a broader efficacy of action in
targeting genetic modules significantly different between de
novo and ISR groups.
Discussion
We present here a comprehensive integrative analysis of ISR,
applying systems tools to make specific predictions regarding
new therapeutic targets. We also test our method against
recently published clinical trials and find that our analysis
supports the previously demonstrated broader efficacy of
sirolimus over paclitaxel in the abrogation of ISR.
Surprisingly few studies have examined the basic cellular
processes of ISR, presumably because tissue is hard to
acquire.15 Many studies have examined circulating markers or
predictors,16 –18 and a small number of studies have begun to
describe allelic variation associated with the restenotic response.19 –21 Indeed, a large National Institutes of Health
study is currently underway to comprehensively assess the
genomics of ISR.22 Identification of genes through the work
described here significantly increases the number of candidate genes that can be investigated through these case-control
genetic epidemiology studies.
In our histological analysis, we confirmed previous findings from cellular microscopy: ISR consists primarily of
smooth muscle cells, seen in much greater number than in de
novo atherosclerosis, and proteoglycan-rich matrix.8,23 Occasional focal collections of inflammatory cells and lipid pools
are seen, but they are more common in de novo disease;
2652
Circulation
December 12, 2006
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Figure 5. Subnets generated using text mining of Medline abstracts were awarded an overall significance score based on the average
significance of network members in our experimental data set. Nodes are individual genes; edges are their connecting lines. Each edge
represents a relationship revealed by semantic mining. Thicker edges imply a stronger relationship (more sentences with associations).
A, The ADAM17 (TNF-␣ converting enzyme) subnetwork. B, The procollagen COL1A2 gene network. Below each node in the network is
a heat strip representation of the raw gene expression data. Brown represents de novo restenosis; blue, ISR. The horizontal center
equates to the median expression value for that gene so that the height of vertical bars above or below the line varies according to
actual expression values. Nodes are colored according to their significance (bright red indicates highly significantly upregulated; bright
green, highly significantly downregulated). C, Mean significance⫾SE for subnetworks generated from subsets of the published literature
containing the terms “sirolimus” or “paclitaxel.”
thrombus and calcification occur at the same frequency in
each group. This description of cellular anatomy complements well the insight offered into cellular function by
expression profiling.
Our technique of using gene expression profiles and
network analysis is designed to identify genetic modules of
interest. One of the most highly differentially regulated
subnetworks in our analysis contained as its nexus ADAM17,
the TNF-␣ converting enzyme. This molecule is believed to
regulate inflammation via cleavage and release of transmembrane proteins from the cell surface, including L-selectin,
TNF-␣, and its receptors TNFR1 and TNFR2. As such,
potentially, both proinflammatory and antiinflammatory effects could result. Release of soluble TNF would enhance
inflammation, whereas release of a TNF receptor ectodomain
might limit the effects of soluble TNF.23 In our analysis,
TNF-␣ converting enzyme itself was highly downregulated
(d score, ⫺2.46), as were many members of the subnetwork,
suggesting a role for this module in normal arterial function
or in de novo atherosclerosis. Although little is known of the
role of ADAM17/TNF-␣ converting enzyme in atherosclerosis, Canault et al24 recently reported the expression of active
enzyme at the surface of macrophages isolated from atherosclerotic lesions of apolipoprotein E⫺/⫺ mice fed a high-fat
diet. Taken together, these findings argue strongly for future
investigation of the TNF system dysregulation in general and
ADAM17 in particular as a target for the interruption of
atherosclerotic or restenotic processes.
Highly upregulated nexus genes in significantly differentially regulated networks are attractive targets for the interruption of ISR. COL1A2 (illustrated in Figure 5) is one such
example and one that suggests potential targets for its own
Ashley et al
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
suppression. For example, decoy Sp1-binding oligonucleotides have been shown to inhibit COL1A2 promoter activity
both in cultured fibroblasts and in vivo in the skin of
transgenic mice.25 Alternatively, interferon-␥ represses collagen in a manner requiring a class II transactivator molecule.26
Both of these molecules are significantly downregulated in
the COL1A2 subnetwork. Thus, a specific hypothesis generated by our data is that targeting of COL1A2, or one of the
regulators shown, would abrogate ISR. Furthermore, we
contend that this directed approach is more elegant than the
global inhibition of cell proliferation induced by sirolimus,
paclitaxel, or brachytherapy and thus is likely to result in
fewer unwanted effects such as slow endothelialization and
late stent thrombosis.
We have previously shown that high-dimensional techniques can identify targets for treatment in heart failure.27,28
To test the idea that such targets could be identified in
vascular disease by our approach, we took advantage of
recently published head-to-head clinical trials of 2 agents
effective against ISR.4,13 Although differences in late lumen
loss or target lesion revascularization in these trials could be
explained by many factors, including those relating to deployment, stent design, polymer design, agent, or time course
of elution, we hypothesized that an important factor was the
agent itself. We found that in networks generated from
subsets of the published literature focused on sirolimus or
paclitaxel, the mean significance level of subnetworks in the
sirolimus group was greater than that for the paclitaxel group.
Thus, our method supports the results of multiple large-scale
clinical trials.
Several limitations of our approach should be noted. First,
tissue derived at 1 time point late in the disease process does
not allow assessment of serial changes or the identification of
early mechanisms in the disease process such as might be
available from animal models.29 This may be why we saw no
inflammatory signal for ISR in our ontology analysis. In
addition, the impossibility of obtaining normal human control
data might limit our ability to identify genes upregulated in
both ISR and atherosclerosis. However, our intention was to
identify the unique signature of ISR, and the best subtraction
profile is clearly atherosclerosis rather than normal. Second,
in network analysis, although publication bias is minimized
by controlling for the size of the network and by combining
semantic mining with experimental data, the former is limited
by the available literature and by our choice of interaction
verbs. The lack of more complex sentence processing means
that a small number of sentences may be included that contain
2 terms and an interaction verb but do not imply a gene
product– gene product relationship (estimates in pathways
curated by hand suggest that this is between 5% and 15%).
However, we believe that the unparalleled scale of analysis
possible with computer processing, the benefit in including
multiple cell types and models via the literature, and the
ability of the technique to account for nontranscriptional
regulation make up for these shortcomings and argue for the
role of this integrative analysis alongside more established
reductionist approaches.
Network Analysis of In-Stent Restenosis
2653
Conclusions
In the present study, we have carried out an extensive
assessment of the biology of human ISR. We confirmed
previously reported findings from cellular microscopy studies
and extended these into the molecular domain with nearly
genome-wide transcriptional profiling. Ontology analysis of
differentially regulated genes revealed a significantly more
inflammatory profile for de novo atherosclerosis lesions than
for ISR lesions. Network analysis confirmed this finding and
identified subnetworks of highly connected genes that we
hypothesize to be promising therapeutic targets. Finally, we
tested our method by forcing it to predict whether sirolimus
or paclitaxel would be the more effective agent in abrogating
ISR and found that it successfully recapitulated the results of
large-scale clinical trials. This is the first implementation of a
network-based analysis of ISR and represents the most
comprehensive survey of the disease process to date.
Acknowledgments
We would like to thank Lutz Buellesfeld for help in data collection
and Eugene Yang for laying the groundwork for the study.
Disclosures
Angela Soito and Drs He and Simpson are employees of Foxhollow
Technologies (Redwood City, Calif). Drs Tsalenko, Vailaya, and
Kuchinsky are employees of Agilent Technologies (Palo Alto, Calif).
Drs Ashley and Quertermous previously held consulting relationships with the Biologics Division of Foxhollow Technologies. Dr
Grube is a stockholder of Foxhollow Technologies. The other
authors report no conflicts.
References
1. Sigwart U, Puel J, Mirkovitch V, Joffre F, Kappenberger L. Intravascular
stents to prevent occlusion and restenosis after transluminal angioplasty.
N Engl J Med. 1987;316:701–706.
2. Fischman DL, Leon MB, Baim DS, Schatz RA, Savage MP, Penn I, Detre
K, Veltri L, Ricci D, Nobuyoshi M. A randomized comparison of
coronary-stent placement and balloon angioplasty in the treatment of
coronary artery disease: Stent Restenosis Study Investigators. N Engl
J Med. 1994;331:496 –501.
3. Serruys PW, de Jaegere P, Kiemeneij F, Macaya C, Rutsch W,
Heyndrickx G, Emanuelsson H, Marco J, Legrand V, Materne P. A
comparison of balloon-expandable-stent implantation with balloon angioplasty in patients with coronary artery disease: Benestent Study Group.
N Engl J Med. 1994;331:489 – 495.
4. Moliterno DJ. Healing Achilles: sirolimus versus paclitaxel. N Engl
J Med. 2005;353:724 –727.
5. Lemos PA, Hoye A, Goedhart D, Arampatzis CA, Saia F, van der Giessen
WJ, McFadden E, Sianos G, Smits PC, Hofma SH, de Feyter PJ, van
Domburg RT, Serruys PW. Clinical, angiographic, and procedural predictors of angiographic restenosis after sirolimus-eluting stent implantation in complex patients: an evaluation from the Rapamycin-Eluting
Stent Evaluated at Rotterdam Cardiology Hospital (RESEARCH) study.
Circulation. 2004;109:1366 –1370.
6. Kuchulakanti PK, Chu WW, Torguson R, Ohlmann P, Rha SW, Clavijo
LC, Kim SW, Bui A, Gevorkian N, Xue Z, Smith K, Fournadjieva J,
Suddath WO, Satler LF, Pichard AD, Kent KM, Waksman R. Correlates
and long-term outcomes of angiographically proven stent thrombosis with
sirolimus- and paclitaxel-eluting stents. Circulation. 2006;113:
1108 –1113.
7. Mitra AK, Agrawal DK. In stent restenosis: bane of the stent era. J Clin
Pathol. 2006;59:232–239.
8. Virmani R, Farb A. Pathology of in-stent restenosis. Curr Opin Lipidol.
1999;10:499 –506.
9. King JY, Ferrara R, Tabibiazar R, Spin JM, Chen MM, Kuchinsky A,
Vailaya A, Kincaid R, Tsalenko A, Deng DX, Connolly A, Zhang P,
Yang E, Watt C, Yakhini Z, Ben-Dor A, Adler A, Bruhn L, Tsao P,
2654
10.
11.
12.
13.
14.
15.
16.
17.
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
18.
19.
20.
21.
Circulation
December 12, 2006
Quertermous T, Ashley EA. Pathway analysis of coronary atherosclerosis. Physiol Genomics. 2005;23:103–118.
Troyanskaya O, Cantor M, Sherlock G, Brown P, Hastie T, Tibshirani R,
Botstein D, Altman RB. Missing value estimation methods for DNA
microarrays. Bioinformatics. 2001;17:520 –525.
Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays
applied to the ionizing radiation response. Proc Natl Acad Sci U S A.
2001;98:5116 –5121.
Kincaid R, Kleusing D, Vailaya A. BNS: An LDAP-Based Biomolecule
Naming Service. Washington DC: OiBC; 2002.
Babapulle MN, Joseph L, Belisle P, Brophy JM, Eisenberg MJ. A hierarchical Bayesian meta-analysis of randomised clinical trials of drugeluting stents. Lancet. 2004;364:583–591.
Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, Amin
N, Schwikowski B, Ideker T. Cytoscape: a software environment for
integrated models of biomolecular interaction networks. Genome Res.
2003;13:2498 –2504.
Welt FG, Rogers C. Inflammation and restenosis in the stent era. Arterioscler Thromb Vasc Biol. 2002;22:1769 –1776.
Kamitani T, Taniguchi T, Miyai N, Kawasaki T, Kawasaki S, Sugihara H.
Association between plasma lipoprotein(a) concentration and restenosis
after stent implantation. Circ J. 2005;69:644 – 649.
Amano T, Matsubara T, Izawa H, Torigoe M, Yoshida T, Hamaguchi Y,
Ishii H, Miura M, Hayashi Y, Ogawa Y, Murohara T. Impact of plasma
aldosterone levels for prediction of in-stent restenosis. Am J Cardiol.
2006;97:785–788.
Breuckmann F, Naber C, Beckert J, Schmermund A, Haude M, Baumgart
D, Erbel R. Postinterventional homocysteine levels: failure as a predictive
biomarker of in-stent restenosis. Int J Cardiol. 2006;108:20 –25.
Petrovic D, Peterlin B. Genetic markers of restenosis after coronary
angioplasty and after stent implantation. Med Sci Monit. 2005;11:
RA127–RA135.
Ribichini F, Steffenino G, Dellavalle A, Matullo G, Colajanni E, Camilla
T, Vado A, Benetton G, Uslenghi E, Piazza A. Plasma activity and
insertion/deletion polymorphism of angiotensin I– converting enzyme: a
major risk factor and a marker of risk for coronary stent restenosis.
Circulation. 1998;97:147–154.
Monraats PS, Rana JS, Zwinderman AH, de Maat MP, Kastelein JP,
Agema WR, Doevendans PA, de Winter RJ, Tio RA, Waltenberger J,
22.
23.
24.
25.
26.
27.
28.
29.
Frants RR, van der Laarse A, van der Wall EE, Jukema JW. ⫺455G/A
polymorphism and preprocedural plasma levels of fibrinogen show no
association with the risk of clinical restenosis in patients with coronary
stent placement. Thromb Haemost. 2005;93:564 –569.
Ganesh SK, Skelding KA, Mehta L, O’Neill K, Joo J, Zheng G, Goldstein
J, Simari R, Billings E, Geller NL, Holmes D, O’Neill WW, Nabel EG.
Rationale and study design of the CardioGene Study: genomics of in-stent
restenosis. Pharmacogenomics. 2004;5:952–1004.
Glover C, Ma X, Chen YX, Miller H, Veinot J, Labinaz M, O’Brien E.
Human in-stent restenosis tissue obtained by means of coronary
atherectomy consists of an abundant proteoglycan matrix with a paucity
of cell proliferation. Am Heart J. 2002;144:702–709.
Canault M, Peiretti F, Kopp F, Bonardo B, Bonzi MF, Coudeyre JC,
Alessi MC, Juhan-Vague I, Nalbone G. The TNF alpha converting
enzyme (TACE/ADAM17) is expressed in the atherosclerotic lesions of
apolipoprotein E-deficient mice: possible contribution to elevated plasma
levels of soluble TNF alpha receptors. Atherosclerosis. 2005.
Verrecchia F, Rossert J, Mauviel A. Blocking sp1 transcription factor
broadly inhibits extracellular matrix gene expression in vitro and in vivo:
implications for the treatment of tissue fibrosis. J Invest Dermatol. 2001;
116:755– 63.
Xu Y, Wang L, Buttice G, Sengupta PK, Smith BD. Major histocompatibility class II transactivator (CIITA) mediates repression of collagen
(COL1A2) transcription by interferon gamma (IFN-gamma). J Biol
Chem. 2004;279:41319 –32.
Chen MM, Ashley EA, Deng DX, Tsalenko A, Deng A, Tabibiazar R,
Ben-Dor A, Fenster B, Yang E, King JY, Fowler M, Robbins R, Johnson
FL, Bruhn L, McDonagh T, Dargie H, Yakhini Z, Tsao PS, Quertermous
T. Novel role for the potent endogenous inotrope apelin in human cardiac
dysfunction. Circulation. 2003;108:1432–1439.
Ashley EA, Powers J, Chen M, Kundu R, Finsterbach T, Caffarelli A,
Deng A, Eichhorn J, Mahajan R, Agrawal R, Greve J, Robbins R,
Patterson AJ, Bernstein D, Quertermous T. The endogenous peptide
apelin potently improves cardiac contractility and reduces cardiac loading
in vivo. Cardiovasc Res. 2005;65:73– 82.
Tabibiazar R, Wagner RA, Ashley EA, King JY, Ferrara R, Spin JM,
Sanan DA, Narasimhan B, Tibshirani R, Tsao PS, Efron B, Quertermous
T. Signature patterns of gene expression in mouse atherosclerosis and
their correlation to human coronary disease. Physiol Genomics. 2005;22:
213–226.
CLINICAL PERSPECTIVE
The present article investigates the biology of in-stent restenosis. Although drug-eluting stents have been highly successful
in abrogating in-stent restenosis, they are not equally efficacious in all lesions and patients. In addition, questions remain
over the incidence of in-stent thrombosis. In this study, we combined light microscopy with a state-of-the-art genomics
tool, the microarray, to assess the gene expression of ⬎20 000 genes in samples of atheroma removed from the coronary
arteries of patients with coronary artery disease. We subtracted the gene expression profile of native lesions from those of
in-stent disease to give the unique signature of in-stent restenosis. Using a new technique that parses the entire biomedical
literature into networks using automated search, we overlaid this gene expression signature onto regulatory networks and
scored each network for the average significance of its members. The highest-scoring networks were curated and tagged
for further investigation. We predict that these networks will make excellent targets for future therapies against in-stent
restenosis. To test the approach, we overlaid our gene expression signature on networks derived from the search terms
“sirolimus” or “paclitaxel.” Our results, showing a mean significance score greater for the sirolimus network, recapitulated
those of major clinical trials that show lower rates of late lumen loss for sirolimus-eluting stents. Future work will
characterize the pathways identified and test their efficacy in abrogating in-stent restenosis.
Network Analysis of Human In-Stent Restenosis
Euan A. Ashley, Rossella Ferrara, Jennifer Y. King, Aditya Vailaya, Allan Kuchinsky, Xuanmin
He, Blake Byers, Ulrich Gerckens, Stefan Oblin, Anya Tsalenko, Angela Soito, Joshua M. Spin,
Raymond Tabibiazar, Andrew J. Connolly, John B. Simpson, Eberhard Grube and Thomas
Quertermous
Downloaded from http://circ.ahajournals.org/ by guest on June 18, 2017
Circulation. 2006;114:2644-2654; originally published online December 4, 2006;
doi: 10.1161/CIRCULATIONAHA.106.637025
Circulation is published by the American Heart Association, 7272 Greenville Avenue, Dallas, TX 75231
Copyright © 2006 American Heart Association, Inc. All rights reserved.
Print ISSN: 0009-7322. Online ISSN: 1524-4539
The online version of this article, along with updated information and services, is located on the
World Wide Web at:
http://circ.ahajournals.org/content/114/24/2644
Permissions: Requests for permissions to reproduce figures, tables, or portions of articles originally published
in Circulation can be obtained via RightsLink, a service of the Copyright Clearance Center, not the Editorial
Office. Once the online version of the published article for which permission is being requested is located,
click Request Permissions in the middle column of the Web page under Services. Further information about
this process is available in the Permissions and Rights Question and Answer document.
Reprints: Information about reprints can be found online at:
http://www.lww.com/reprints
Subscriptions: Information about subscribing to Circulation is online at:
http://circ.ahajournals.org//subscriptions/