Vol. 30 no. 17 2014, pages 2393–2398
BIOINFORMATICS DISCOVERY NOTE doi:10.1093/bioinformatics/btu323
Systems biology Advance Access publication May 7, 2014
Bioinformatics-driven discovery of rational combination for
overcoming EGFR-mutant lung cancer resistance to EGFR therapy
Jihye Kim1
, Vihas T. Vasu2
, Rangnath Mishra2
, Katherine R. Singleton3
, Minjae Yoo1
,
Sonia M. Leach2
, Eveline Farias-Hesson2
, Robert J. Mason2,4
, Jaewoo Kang5
,
Preveen Ramamoorthy2
, Jeffrey A. Kern2,4
, Lynn E. Heasley3
, James H. Finigan2,4,
*,y
and
Aik Choon Tan1,5,
*,y
1
Division of Medical Oncology, Department of Medicine, Translational Bioinformatics and Cancer Systems Biology
Laboratory, University of Colorado Anschutz Medical Campus, 80045 Aurora, 2
Department of Medicine, National Jewish
Health, 80206 Denver, 3
Department of Craniofacial Biology, School of Dental Medicine, 4
Division of Pulmonary Sciences
and Critical Care Medicine, University of Colorado Anschutz Medical Campus, 80045 Aurora, CO, USA and 5
Department
of Computer Science and Engineering, Korea University, Seoul 136-713, Korea
Associate Editor: Inanc Birol
ABSTRACT
Motivation: Non–small-cell lung cancer (NSCLC) is the leading cause
of cancer death in the United States. Targeted tyrosine kinase inhibitors
(TKIs) directed against the epidermal growth factor receptor
(EGFR) have been widely and successfully used in treating NSCLC
patients with activating EGFR mutations. Unfortunately, the duration
of response is short-lived, and all patients eventually relapse by
acquiring resistance mechanisms.
Result: We performed an integrative systems biology approach to
determine essential kinases that drive EGFR-TKI resistance in
cancer cell lines. We used a series of bioinformatics methods to analyze
and integrate the functional genetics screen and RNA-seq data to
identify a set of kinases that are critical in survival and proliferation in
these TKI-resistant lines. By connecting the essential kinases to compounds
using a novel kinase connectivity map (K-Map), we identified
and validated bosutinib as an effective compound that could inhibit
proliferation and induce apoptosis in TKI-resistant lines. A rational
combination of bosutinib and gefitinib showed additive and synergistic
effects in cancer cell lines resistant to EGFR TKI alone.
Conclusions: We have demonstrated a bioinformatics-driven discovery
roadmap for drug repurposing and development in overcoming
resistance in EGFR-mutant NSCLC, which could be generalized to
other cancer types in the era of personalized medicine.
Availability and implementation: K-Map can be accessible at: http://
tanlab.ucdenver.edu/kMap.
Contact: aikchoon.tan@ucdenver.edu or finiganj@njhealth.org
Supplementary information: Supplementary data are available at
Bioinformatics online.
Received on October 13, 2013; revised and accepted on May 1, 2014
1 INTRODUCTION
Genomic medicine has dramatically increased our knowledge
of the molecular changes that underpin disease states.
Understanding alterations in gene expression can identify
proteins and signaling pathways, which might serve as therapeutic
targets. Moreover, recent technologic advances in nextgeneration
sequencing facilitate the rapid assessment of gene
expression changes in specific patients, allowing for individualized
treatment. Although personalized genomics heralds the age
of precision medicine, the comprehensive datasets obtained
through sequencing require sophisticated bioinformatics tools
to identify the critical genes that influence disease. Once these
critical genes are identified, the next challenge is to predict what
drugs would be useful in reversing the disease states.
The connectivity map represents the first attempt to provide a
computational framework to connect genes, drugs and diseases
based on gene expression signatures (Lamb et al., 2006). This
method assumes gene expression changes could be used as a
‘universal language’ to connect distinct biological states (e.g. diseases),
allowing for the successful repurposing of compounds
(Hieronymus et al., 2006; Wei et al., 2006). In short, drugs
known to be effective in one disease can serve as candidates
for use in other diseases marked by similar gene expression
changes. The power of this method has inspired other related
work (Chung et al., 2014; Li et al., 2009; Zhang and Gant,
2008, 2009) with the goal of improving the utility of the
connectivity map in drug repurposing and development.
Non–small-cell lung cancer (NSCLC) serves as an ideal disease
for a connectivity map-based approach. NSCLC accounts for
$85% of lung cancers, and is the leading cause of cancer-related
death in the United States (Siegel et al., 2013) and worldwide.
Comprehensive characterization of cancer genomes have
increased our understanding of cancer biology and moved
NSCLC beyond standard clinico-pathologic classifications and
staging to include molecular characterization based on newly
identified oncogenic drivers, such as the mutant epidermal
growth factor receptor gene (EGFR) (Pao and Chmielecki,
2010). Although therapies directed against EGFR, such as
gefitinib (Hirsch et al., 2003), have revolutionized NSCLC
care, the duration of response is short-lived, and all patients
eventually relapse through acquired resistance mechanisms
*To whom correspondence should be addressed.
yThe authors wish it to be known that, in their opinion, the last two
authors should be regarded as Joint Last Authors.
ß The Author 2014. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com 2393
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom
(Casas-Selves et al., 2012; Engelman and J€anne, 2008; Ohashi
et al., 2013; Ware et al., 2013). Identification of new targeted
therapeutics is a priority, yet the ability to translate genomic
datasets into new drugs has been limited.
We devised an integrative systems biology approach using a
functional screen combined with RNA-seq expression data to
determine essential kinases other than EGFR that drive proliferation
and survival in EGFR-mutant NSCLC. We performed
an initial focused, functional genetic screen using short hairpin
RNAs (shRNAs) that target the kinome ($600 kinases) in the
H1650 lung cancer cell line. This line harbors an EGFR mutation,
yet is intrinsically resistant to EGFR TKIs. We used
Bioinformatics for Next-Generation Sequencing! (BiNGS!)
(Kim and Tan, 2012), a novel bioinformatics pipeline to analyze
and interpret functional genomic screening data by next-generation
sequencing to identify essential kinases that drive the survival
and proliferation signaling pathways of this EGFR-mutant
line. Next, we determined the differentially expressed kinases in
this cancer cell line by next-generation sequencing (RNA-seq).
Integrative analysis was performed to identify kinases that were
essential and dysregulated, and the list of essential and functional
kinases was connected to explore therapeutic opportunities using
the kinase connectivity map (K-Map) (Kim et al., 2013). The
efficacy of the K-map prediction of kinase inhibitors was then
validated in vitro (Fig. 1).
2 MATERIALS AND METHODS
Cell culture. H1650 and H1975 cells were obtained from the University of
Colorado Cancer Center (UCCC) Tissue Culture Core. All cell lines were
routinely cultured in RPMI-1640 growth medium supplemented with
10% fetal bovine serum (Sigma, St. Louis, MO, USA) at 37
C in a
humidified 5% CO2 incubator. Alveolar type II (ATII) cells are primary,
normal cells that were isolated and cultured in an air/liquid interface, as
previously described (Wang et al., 2007).
Kinome essential screen by next-generation sequencing. An essential
screen identifies genes critical for cell survival by knocking out individual
genes in cells (one gene deleted per individual cell). The essential screen is
a high-throughput assessment of gene function as opposed to expression.
A total of 3Â 106
cells of H1650 were transduced with the short-hairpin
loop lentiviral kinome library ($3700 shRNAs targeting $600 kinases)
developed by the RNAi Consortium (TRC 1.0/1.5) and obtained from
the UCCC Functional Genomics Shared Resource Core. Cells were
transduced with the lentiviral shRNA library using conditions to result
in a single shRNA expressed in each infected cell. Cells were cultured and
harvested after 2, 7, 14 and 28 days of transduction. shRNAs from
surviving cells were extracted, reverse transcribed and barcoded for individual
replicates (four replicates per time point). The DNA was purified
and amplified, and Illumina adapters were added. The relative abundance
of the unique shRNA tags was quantified by the Illumina Genome
Analyzer, as previously described (Casas-Selves et al., 2012; Singleton
et al., 2013; Spreafico et al., 2013; Sullivan et al., 2012). Loss of
shRNAs represents essential kinases. See Supplementary Methods for
details.
BiNGS! analysis. We used BiNGS! for analyzing and interpreting the
essential screen (Kim and Tan, 2012). In brief, a preprocessing step filtered
out erroneous and low-quality reads. Filtered reads were mapped
against the shRNA reference library using Bowtie (Langmead et al.,
2009). Output from this step is a PÂ N matrix, where P and N represent
shRNA counts and samples, respectively. We also filtered out shRNAs
where the median raw count in the time-point group 1 is greater than the
maximum raw count in the time-point group 2 if the shRNA is enriched
in the time-point group 1, and vice versa. We then used a negative binomial
to model the count distribution in the sequencing data using edgeR
(Robinson et al., 2010). We computed the q-value of false discovery rate
(FDR) for multiple comparisons for these shRNAs, and carried out a
meta-analysis by combining adjusted P-values for all shRNAs representing
the same gene using weighted Z-transformation (Whitlock, 2005). For
each gene, we computed a P-value P(wZ), and we used this P(wZ) to sort
the kinase list, as previously described (Casas-Selves et al., 2012;
Singleton et al., 2013). We performed pair-wise comparisons for each
of the time points, and we grouped the kinases into three classes using
the following rules based on the P(wZ) obtained from each gene (similar
Fig. 1. Workflow of the experimental and bioinformatics analyses of this study. Blue boxes represent bioinformatics analyses. The K-Map connectivity
results for validation (right)
2394
J.Kim et al.
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom
to the classification rules in Marcotte et al., 2012). In this study, we
considered kinases that, when deleted by shRNAs, induced cell death
in the first time point (day 7) as candidate kinases. These kinases were
‘essential’ for cell survival and they were never recovered in the later time
points (day 14 and day 28). See Supplementary Methods for details.
RNA-seq of kinome and bioinformatics analysis. Transcriptome
libraries were prepared following Applied BioSystems SOLiD Total
RNA-Seq protocol. The libraries were sequenced using the Applied
BioSystems SOLiD 5500 platform, using 75 base pair by 35 base
paired-end reads. Mapping of sequencing reads and quantification of
known RefSeq transcripts were performed using LifeScope v2.1 (ABI).
Expression values for each transcript were calculated as reads per
kilobase of exon model per million mapped reads (RPKM). RPKM is
a method of quantifying gene expression from RNA-seq data by normalizing
for total read length and the number of sequencing reads obtained
within each sample (Mortazavi et al., 2008). See Supplementary Methods
and Supplementary Table S3 for details. To determine the differentially
expressed kinases, we computed a FDR on the P-values obtained by
t-test. We used FDR 5% and fold change 41.25 as thresholds for
determining differentially expressed kinases between cancer and normal
samples.
K-Map analysis. We have recently developed and implemented a
K-Map that systematically connects a kinase profile with a reference
kinase inhibitor database and predicts the most effective inhibitor for a
queried kinase profile (Kim et al., 2013) (Fig. 2). The K-Map is inspired
by the connectivity map concept (Lamb et al., 2006), where the main
assumption of this concept is that gene expression profiles could be
used as ‘universal language’ to connect between biological states, genes
and drugs. The connectivity map has three key components: (i) a reference
database that contains a set of predefined gene expression profiles;
(ii) a query gene signature; and (iii) a pattern matching algorithm or
similarity metric defined between a query gene signature and a reference
gene expression profile to quantify the connection (or similarity) between
the two biological states. Instead of gene expression signatures, we used
the kinase activity profiles as the ‘language’ for connecting kinases with
small molecule kinase inhibitors in K-Map to reveal the interactions of
kinases and inhibitors (Kim et al., 2013). Figure 2 illustrates the concept
of K-Map. Reference Database. We built the reference database based on
two recently published comprehensive analyses of kinase inhibitor
selectivity (Anastassiadis et al., 2011; Davis et al., 2011). The first study
systematically interrogated 178 commercially available inhibitors against
a panel of 300 protein kinases using a radiometric phospho-transfer
method to assess the percent kinase inhibition (IC50) (Anastassiadis
et al., 2011). The second study measured the selectivity and potency of
72 inhibitors against 442 kinases using direct binding affinities between
inhibitors and kinases (Kd) (Davis et al., 2011). These datasets were converted
into rank-ordered lists according to the inhibitors’ potencies
against the kinases and used as the K-Map reference profiles for matching
query kinases. Query Signature. Kinases found to be differentially expressed
and essential were used as the query kinase profile and connected
through the K-Map in this study. Pattern Matching Algorithm. We
implemented the K-Map pattern matching strategy based on the
Kolmogorov–Smirnov (KS) statistics. The KS-test is a non-parametric,
rank-based pattern-matching approach implemented in the connectivity
map (Lamb et al., 2006). The goal of the algorithm is to correlate kinase
inhibitors, based on kinase inhibition profiles in the reference database,
with a given query (i.e. a list of kinases). For every inhibitor in the reference
database, the KS statistic is computed and a ‘connectivity score’ is
defined (see Supplementary Methods for details). K-Map will return a
ranked list of kinase inhibitors that best inhibit the list of queried kinases
based on their ‘connectivity score’. We used K-Map to connect the
differentially expressed and essential kinases with drugs in this study.
Drug Cytotoxicity/Proliferation assays. Bosutinib, gefitinib, sorafenib
and CI-1040 were obtained commercially (LC Laboratories, Woburn,
MA). Cytotoxic/proliferation effects were determined using WST-1
(Roche) assay. Water-soluble tetrazolium salt (WST-1) assay is a colorimetric
assay for the non-radioactive quantification of cellular proliferation,
viability and cytotoxicity. WST-1 is added into cultured cells, and
the reading of absorbance correlates with the number of proliferating
cells. This assay was used to measure the inhibitory effect of drugs
used in this study on cell proliferation. The half maximal inhibitory
concentration (IC50) of individual drugs was determined from the cell
proliferation curves. See Supplementary Methods for details.
Immunoblot assays. Apoptosis markers, cleaved and total caspase-3
protein levels, in cells treated with drugs were measured by immunoblotting.
See Supplementary Methods for details.
Quantifying combination effects. To quantify the combination effects of
drugs used in this study, we used two standard models of additivism
(Borisy et al., 2002). The first approach is the highest single agent
(HSA) model that measures the larger of the effects produced by each
of the combination’s single agents at the same concentrations as in the
mixture. The combination effect can be classified as: (i) additive (if the difference
between combination and HSA is 0); (ii) synergistic (if the difference
between combination and HSA is 40); or (iii) antagonistic (if the
difference between combination and HSA is50). The second approach is
the Bliss additivism model, which predicts the combined response C for
two single compounds, with effects A and B using the following equation:
C = (A+ B)– (AÂ B), where each effect is expressed as fractional inhibition
between 0 (no effect, 0% inhibition) and 1 (maximal effect = 100%
inhibition). By taking the difference between the observed combined
effects and the predicted C, we can classify the combination effect as:
(i) additive (if the difference is 0); (ii) synergistic (if the difference is40); or
(iii) antagonistic (if the difference is50).
3 RESULTS
Essential kinases identified by the BiNGS! analysis. To determine
essential kinases other than EGFR that drive the oncogenic signaling
pathways in NSCLC, we performed a functional kinome
genetic screen using H1650, an NSCLC line with an EGFRactivating
mutation (deletion exon 19), yet for unknown reasons
is relatively resistant to EGFR TKIs. From this kinome essential
Fig. 2. The K-Map concept. Query signature is derived from experimental
study (left). Reference database of K-Map is based on kinase inhibitor
selectivity profiles (center). Pattern-matching algorithm provides a score
for each reference profile based on its enrichment of the query signature
(center). Kinase inhibitors are ranked by the ‘connectivity score’; those at
the top (‘strong connection’) and bottom (‘low connection’) are predicted
to have maximal and minimal efficacy against the query signature (right)
2395
Bioinformatics-driven discovery of drug combinations
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom
screen, we identified 14 candidate kinases (CALM3, CDK1,
CDK6, DDR1, EGFR, EPHA4, GNE, IPMK, MARK3, PBK,
PKN1, ROCK1, RPRD1A and TBK1) as essential for survival
of this cell line (see Supplementary Fig. S1 and Supplementary
Methods for details). Importantly, EGFR was identified from
this functional genetic screen as essential in H1650. To determine
whether these kinases were mutated in this cell line, we queried
the comprehensive curated Catalogue Of Somatic Mutations In
Cancer (COSMIC) database (Forbes et al., 2011). According to
the COSMIC database, besides EGFR, no other known somatic
mutations have been reported in the 13 other essential kinases in
H1650.
Differentially expressed kinases identified by the RNA-seq
analysis. To determine the differentially expressed kinase genes
in H1650, we compared the expressed kinome of H1650 defined
by targeted RNA-seq with four normal, human ATII cells, the
putative cell of origin for lung adenocarcinomas, a histologic
subtype of NSCLC (Xu et al., 2012). From the RNA-seq data,
we focused on the 611 kinases that were the same kinases targeted
in the functional genetic screens. We identified 193 kinases
overexpressed and 131 underexpressed comparing the H1650
data with the normal ATII RNA-seq data (FDR 0.05, fold
change 41.25). Again, EGFR was the top overexpressed gene
in the list, supporting H1650’s dependence on this oncogene.
The heat map of this analysis is illustrated in Supplementary
Figure S2 and Supplementary Table S1.
Integrating a functional kinome genetic screen and RNA-seq
analysis. To refine the panel of identified key kinases driving
the survival and proliferation of H1650 cells, we sought to integrate
the candidate kinase genes identified from the essential
kinome screen with differential kinase transcriptional profiling,
two different yet complementary approaches. Transcriptional
profiling analysis identifies kinases that are differentially expressed
compared with normal lung ATII cells, yet provides no
insight on kinase function. Conversely, the functional genetic
screen allows discovery of kinases required for cell survival, yet
provides no information specific to transformation. Therefore,
integrating these two analyses will provide a list of functionally
essential, potentially transformative kinases in H1650. By comparing
the candidate gene lists obtained from the two approaches
(Fig. 1), seven kinases (CDK6, EGFR, MARK3, PBK, TBK1,
DDR1 and EPHA4) were found to be common to both lists.
Connecting the essential, potentially transformative kinases to
therapeutics using the K-Map. We next asked what compounds
could inhibit these essential and possibly transformative kinases
and serve as a potential therapy to inhibit proliferation of this
cell line. We queried the K-Map using the seven essential and
potentially transformative kinases to connect them to drugs
based on two different kinase activity assays (IC50 and Kd).
The top connection in both assays was staurosporine, a natural
product isolated from the bacterium Streptomyces staurosporeus.
Staurosporine is a potent, general ATP-binding site inhibitor
across kinases, lacking any specificity (Anastassiadis et al.,
2011; Davis et al., 2011). Interestingly, bosutinib, a Src and
Abl dual inhibitor, was also positively connected and ranked
#3 and #29 in the Kd and IC50 assays, respectively (Fig. 1 and
Supplementary Table S2). As expected, the two FDA-approved
EGFR-specific TKIs for treating NSCLC patients, gefitinib and
erlotinib, were not connected by the K-Map as effective therapy
for the queried essential kinases (Fig. 1).
Experimental validation of the compounds identified by K-map
analysis. To test whether compounds identified by the K-Map
could provide better growth inhibition of H1650 cells compared
with the current standard of care with EGFR-specific TKIs, we
selected bosutinib (Fig. 1) as the candidate compound prediction
of the K-map and validated its effect on cell survival. In comparison,
we selected gefitinib, a drug approved for use in EGFRmutant
NSCLC, which was not identified using our strategy and
therefore, we predicted, would be less efficacious in H1650 cells
(Fig. 1). To further test our algorithm and K-Map, we identified
sorafenib (connectivity scores of 0.234 and 0.226 in both kinase
assays, Fig. 1) as a compound which was predicted to be
ineffective against the H1650 cell line.
To evaluate the sensitivity of bosutinib, gefitinib and sorafenib,
H1650 NSCLC cells were exposed to increasing concentrations
of these individual drugs and assessed for proliferation. As
depicted in Figure 3A, bosutinib had the lowest IC50 among the
three tested compounds. Conversely, gefitinib and sorafenib had
higher IC50 values (IC50410 mM), indicating both drugs were
less effective in inhibiting the growth of H1650 cells. As an additional
negative control, we validated CI-1040, a compound that
has a connectivity score of 0 (minimal effect and ranked #238 of
250 drugs) and was the least effective of the four drugs tested
(IC50 $34 mM) in H1650 (Supplementary Fig. S3).
To estimate the statistical significance of the connection, we
used the connection testing proposed by Zhang and Gant (2008,
2009), where the two-tailed P-value associated with the observed
connection score is the number of times the connection score
obtained by a random gene signature with the same number of
Fig. 3. Experimental validation of bosutinib in H1650 and H1975.
Proliferation curves for bosutinib, gefitinib and sorafenib in (A) H1650
and (B) H1975. Combination of bosutinib and genfitinib in (C) H1650
and (D) H1975. Bosutinib and the combination induce apoptosis in
(E) H1650 and (F) H1975 cell lines
2396
J.Kim et al.
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom
genes when queried to the database. Given a reference database
and a query signature S with m kinases (here, m = 7), the K-Map
connectivity score is KSS, and the two-tailed P-value is estimated
as: p = Prob {jKSrj ! jKSSj}, where KSr is the connection score
obtained from the random query signature with m kinases. For
this particular study, we performed 500 permutations, and computed
the P-value of the bosutinib connection as p = 0.052 (see
Supplementary Methods for details). As demonstrated by the
experimental validation, this indicates that the connection is
not a false positive identified by the method.
Generalizing the K-map prediction to acquired EGFR-TKI
resistance. Acquired resistance to EGFR TKIs after therapy is
universal in lung cancer patients limiting their usefulness as a
treatment of NSCLC. Drugs that can overcome this resistance
would dramatically impact lung cancer care. To explore whether
bosutinib can be extrapolated as an effective therapy for an
acquired mutation conferring resistance to EGFR TKIs, we
tested H1975, a NSCLC line that harbors EGFR T790M, a
gatekeeper mutation that is commonly acquired during treatment
with EGFR TKIs and renders standard EGFR TKIs ineffective.
As illustrated in Figure 3B, the IC50 value of bosutinib on
H1975 was $3.7 mM. As a negative control, we also determined
the IC50 value of gefitinib, sorafenib and CI-1040 for this
cell line as $12, $17 and $31mM, respectively (Fig. 3B and
Supplementary Fig. S3). Thus, bosutinib is effective in inhibiting
the proliferation of a NSCLC line with an acquired EGFR
resistance mutation.
Rational combination of bosutinib and gefitinib shows synergistic
effects in EGFR-mutant NSCLC cells. We further hypothesized
that the addition of bosutinib to gefitinib would be
synergistic in EGFR-mutant NSCLC cells resistant to single
agent EGFR-TKI treatment. Indeed, the combination of bosutinib
and gefitinib demonstrated additive or synergistic effects
in both EGFR-mutant NSCLC lines (Fig. 3C and D).
Combination therapy was significantly improved in H1650 and
H1975 (P50.05, Welch two sample t-test) in inhibiting cell proliferation
when compared with individual drug alone (Fig. 3C
and D). Based on the additivism models, in H1650 combination
therapy was synergistic (HSA= 0.34 and Bliss = 0.10), whereas
in H1975 was additive (HSA = 0.19 and Bliss = –0.04)
(see Supplementary Methods for details). Immunoblotting of
H1650 and H1975 revealed an increase of cleaved caspase-3
(Casp3), an apoptotic marker, in the bosutinib and combination
treated cells as compared with gefitinib alone and vehicle as early
as 48 h after treatment (Fig. 3E and F).
4 DISCUSSION
Functional genetic screens have the potential to identify genes
essential for cancer cell survival and proliferation, providing a
‘functional’ map of human cancer. Complementing the functional
genetic screen with comprehensive genomics studies such
as transcriptional profiling could reveal the ‘vulnerable targets’
for targeted treatment of cancer cells. Here, we performed a systematic
and unbiased approach to determine essential kinases
driving survival mechanisms of EGFR-mutant NSCLC lines resistant
to EGFR TKIs. Using a series of novel bioinformatics
analyses, specifically connecting the essential kinases with small
molecules based on inhibition activities, we have identified that
bosutinib effectively inhibits the essential kinases in H1650
resulting in cell death. We validated bosutinib in H1650 (intrinsic
EGFR-TKI resistance) and H1975 (acquired gatekeeper T790M
mutation) and demonstrated that this compound inhibited cell
proliferation and induced apoptosis in these cancer cell lines that
are resistant to EGFR-specific TKIs better than standard care.
Our strategy combines high-throughput genetic screens with a
new computational technique (BiNGS!) to rapidly identify kinases
that are essential for cancer cell survival. As described earlier,
the H1650 cell line is driven by an activating EGFR
mutation; therefore, it was expected that EGFR would be one
of the seven essential kinases found in the overlapping gene lists.
Other differentially expressed and essential kinases identified in
this study have been previously reported to play an important
role in cancer cells. For example, MARK3 has been shown to
play a role in regulating cell cycle progression in cancer cells (Sha
et al., 2007). EPHA4 was recently identified as an inhibitor of cell
migration and invasion in lung cancer (Saintigny et al., 2012),
supporting our finding that the expression of this gene is lower
than that found in normal ATII cells, yet functionally important
in driving lung oncogenesis. DDR1 was identified as an essential
kinase across three different tumor types from a recent largescale
functional genetic screen (Marcotte et al., 2012). Despite
these data, prior to our analyses there was no method for identifying
these seven kinases as the functional vulnerabilities of
H1650, which could be targeted for therapeutic intervention.
By connecting the H1650 essential kinase profile through the
K-Map to existing kinase inhibitors, we could repurpose bosutinib
to treat EGFR-mutant NSCLC cell line resistant to EGFR
TKIs. Bosutinib is a Src/Abl dual TKI, which has recently been
approved by the FDA to treat CML patients. Our data suggest
that by using the connectivity map concept, we could ‘connect’
essential kinases to therapeutics, facilitating the translation from
in silico discovery to clinical trials. Recently, bosutinib has been
tested in a Phase I clinical trial of advanced solid tumors (Daud
et al., 2012). Among the 16 NSCLC patients treated with the
optimal dose of bosutinib, seven of the patients observed tumor
shrinkage and nine patients had stable disease. As concluded by
the authors (Daud et al., 2012), bosutinib might provide benefit
in combination with other drugs to result in better treatment
responses. In this study, we provide a biological rationale for
the effectiveness of bosutinib in NSCLC. Moreover, we confirmed
that the combination of bosutinib with gefitinib has additive
or synergistic effects in two gefitinib-resistant NSCLC cell
lines. Future experiments for the combination of bosutinib and
gefitinib are warranted to investigate this bioinformatics-driven
discovery in more cell lines and in animal models.
In summary, we have demonstrated a proof-of-concept, bioinformatics-driven
discovery roadmap for drug repurposing and
development in cancer research, which could be generalized to
other diseases in the era of personalized and precision medicine.
ACKNOWLEDGEMENTS
The authors would like to acknowledge the University of
Colorado Lung Cancer SPORE Career Development Program
and the Program for the Evaluation of Targeted Therapy
(PETT) for their useful comments. We also like to thank the
2397
Bioinformatics-driven discovery of drug combinations
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom
comments and suggestions from the three reviewers that have
helped to improve the presentation of this manuscript.
Funding: This research is partly supported by the Cancer League
of Colorado (to J.K. and A.C.T.), Institutional Start-Up Fund
(to A.C.T.), Korea University Global Professorship Program
(to A.C.T.), the Department of Defense Award W81XWH-11-
1-0527 (to A.C.T.), National Institute of Health (NIH)
P50CA58187 (to L.H., J.H.F. and A.C.T.), R01CA127105 (to
L.H.), R01HL111674 (to J.H.F.), R01HL106112 (to R.J.M.),
P30CA046934, Flight Attendants Medical Research Initiative
113038 (to J.H.F.) and National Jewish Health Feil Family
Foundation Translational Research Award (to J.K.).
Conflict of Interest: none declared.
REFERENCES
Anastassiadis,T. et al. (2011) Comprehensive assay of kinase catalytic activity
reveals features of kinase inhibitor selectivity. Nat. Biotechnol., 29, 1039–1045.
Borisy,A.A. et al. (2002) Systematic discovery of multicomponent therapeutics.
Proc. Natl Acad. Sci. USA, 100, 7977–7982.
Casas-Selves,M. et al. (2012) Tankyrase and the canonical Wnt pathway protect
lung cancer cells from EGFR inhibition. Cancer Res., 72, 4154–4164.
Chung,F.H. et al. (2014) Functional module connectivity map (FMCM): a framework
for searching repurposed drug compounds for systems treatment of cancer
and an application to colorectal adenocarcinoma. PLoS One, 9, e86299.
Daud,A.I. et al. (2012) Phase I study of bosutinib, a Src/Abl tyrosine kinase inhibitor,
administered to patients with advanced solid tumors. Clin. Cancer Res., 18,
1092–1100.
Davis,M.I. et al. (2011) Comprehensive analysis of kinase inhibitor selectivity. Nat.
Biotechnol., 29, 1046–1051.
Engelman,J.A. and J€anne,P.A. (2008) Mechanisms of acquired resistance to epidermal
growth factor receptor tyrosine kinase inhibitors in non-small cell lung
cancer. Clin. Cancer Res., 14, 2895–2899.
Forbes,S.A. et al. (2011) COSMIC: mining complete cancer genomes in the catalogue
of somatic mutations in cancer. Nucleic Acids Res., 39, D945–D950.
Hieronymus,H. et al. (2006) Gene expression signature-based chemical genomic
prediction identifies a novel class of HSP90 pathway modulators. Cancer Cell,
10, 321–330.
Hirsch,F.R. et al. (2003) Epidermal growth factor receptor in non-small-cell lung
carcinomas: correlation between gene copy number and protein expression and
impact on prognosis. J. Clin. Oncol., 21, 3798–3807.
Kim,J. and Tan,A.C. (2012) BiNGS!SL-seq: a bioinformatics pipeline for the analysis
and interpretation of deep sequencing genome-wide synthetic lethal screen.
Methods Mol. Biol., 802, 389–398.
Kim,J. et al. (2013) K-map: connecting kinases with therapeutics for drug repurposing
and development. Hum. Genomics, 7, 20.
Lamb,J. et al. (2006) The connectivity map: using gene-expression signatures to
connect small molecules, genes and disease. Science, 131, 1929–1935.
Langmead,B. et al. (2009) Ultrafast and memory efficient alignment of short DNA
sequences to the human genome. Genome Biol., 10, R25.
Li,J. et al. (2009) Building disease-specific drug-protein connectivity maps from
molecular interaction networks and PubMed abstracts. PLoS Comput. Biol.,
5, e1000450.
Marcotte,R. et al. (2012) Essential gene profiles in breast, pancreatic, and ovarian
cancer cells. Cancer Discov., 2, 172–189.
Mortazavi,A. et al. (2008) Mapping and quantifying mammalian transcriptomes by
RNA-Seq. Nat. Methods, 5, 621–628.
Ohashi,K. et al. (2013) Epidermal growth factor receptor tyrosine kinase inhibitorresistant
disease. J. Clin. Oncol., 31, 1070–1080.
Pao,W. and Chmielecki,J. (2010) Rational, biologically based treatment of EGFRmutant
non-small-cell lung cancer. Nat. Rev. Cancer, 10, 760–774.
Robinson,M.D. et al. (2010) edgeR: a Bioconductor package for differential expression
analysis of digital gene expression data. Bioinformatics, 26, 139–140.
Saintigny,P. et al. (2012) Global evaluation of Eph receptors and ephrins in lung
adenocarcinomas identifies EphA4 as an inhibitor of cell migration and invasion.
Mol. Cancer Ther., 11, 2021–2032.
Sha,S.K. et al. (2007) Cell cycle phenotype-based optimization of G2-abrogating
peptides yields CBP501 with a unique mechanism of action at the G2 checkpoint.
Mol. Cancer Ther., 6, 147–153.
Siegel,R. et al. (2013) Cancer statistics, 2013. CA Cancer J. Clin., 63, 11–30.
Singleton,K.R. et al. (2013) A receptor tyrosine kinase network composed of fibroblast
growth factor receptors, epidermal growth factor receptor, v-erb-b2
erythroblastic leukemia viral oncogene homolog 2, and hepatocyte growth
factor receptor drives growth and survival of head and neck squamous carcinoma
cell lines. Mol. Pharmacol., 83, 882–893.
Spreafico,A. et al. (2013) Rational combination of a MEK inhibitor, selumetinib,
and the Wnt/calcium pathway modulator, cyclosporin A, in preclinical models
of colorectal cancer. Clin. Cancer Res., 19, 4149–4162.
Sullivan,K.D. et al. (2012) ATM and MET kinases are synthetic lethal with nongenotoxic
activation of p53. Nat. Chem. Biol., 8, 646–654.
Wang,J. et al. (2007) Differentiated human alveolar epithelial cells and reversibility
of their phenotype in vitro. Am. J. Respir. Cell Mol. Biol., 36, 661–668.
Ware,K. et al. (2013) A mechanism of resistance to gefitinib mediated by cellular
reprogramming and the acquisition of an FGF2-FGFR1 autocrine growth loop.
Oncogenesis, 2, e39.
Wei,G. et al. (2006) Gene expression-based chemical genomics identifies rapamycin
as a modulator of MCL1 and glucocorticoid resistance. Cancer Cell, 10,
331–342.
Whitlock,M.C. (2005) Combining probability from independent tests: the weighted
Z-method is superior to Fisher’s approach. J. Evol. Biol., 18, 1368–1373.
Xu,X. et al. (2012) Evidence of type II cells as cells of origin of K-Ras-induced distal
lung adenocarcinoma. Proc. Natl Acad. Sci. USA, 109, 4910–4915.
Zhang,S.-D. and Gant,T.W. (2008) A simple and robust method for connecting
small-molecule drugs using gene-expression signatures. BMC
Bioinformatics, 9, 258.
Zhang,S.-D. and Gant,T.W. (2009) sscMap: an extensible Java application for connecting
small-molecule drugs using gene-expression signatures. BMC
Bioinformatics, 10, 236.
2398
J.Kim et al.
atMasarykuniversityonOctober2,2015http://bioinformatics.oxfordjournals.org/Downloadedfrom