In silico identification of potential targets and drugs for non‐small cell lung cancer

Abstract

Lung cancer is one of the leading causes of death in both the USA and Taiwan, and it is thought that the cause of cancer could be because of the gain of function of an oncoprotein or the loss of function of a tumour suppressor protein. Consequently, these proteins are potential targets for drugs. In this study, differentially expressed genes are identified, via an expression dataset generated from lung adenocarcinoma tumour and adjacent non‐tumour tissues. This study has integrated many complementary resources, that is, microarray, protein‐protein interaction and protein complex. After constructing the lung cancer protein‐protein interaction network (PPIN), the authors performed graph theory analysis of PPIN. Highly dense modules are identified, which are potential cancer‐associated protein complexes. Up‐ and down‐regulated communities were used as queries to perform functional enrichment analysis. Enriched biological processes and pathways are determined. These sets of up‐ and down‐regulated genes were submitted to the Connectivity Map web resource to identify potential drugs. The authors' findings suggested that eight drugs from DrugBank and three drugs from NCBI can potentially reverse certain up‐ and down‐regulated genes' expression. In conclusion, this study provides a systematic strategy to discover potential drugs and target genes for lung cancer.

1 Introduction

Lung cancer is the leading cause of death in both the USA [1] and Taiwan [2]. According to the World Health Organization classification, lung cancer can be divided into two major classes: small cell lung cancer (SCLC) and non‐SCLC (NSCLC). NSCLC accounts for > 85% of all lung cancer cases, and adenocarcinoma is the most common subtype. However, it is very difficult to make the connections from cancer through its gene or protein expression to drug discovery. With the advancement of microarray technology, accumulation of protein–protein interactions (PPIs) and human cells treated with many drugs, it is now possible that the drug discovery process can be accelerated.

It is known that many proteins are associated with human diseases, although very often their precise functional role in disease pathogenesis remains unclear. A strategy to gain a better understanding into the interaction and function of these proteins is to make use of the PPI data, and construct a set of interaction rules for disease‐associated proteins. The recent availability of PPI data has made it possible to study human disease at a systemic level. The key challenge is the identification of potential drug targets in a disease‐associated PPI network. To address these questions, we first collected lung cancer associated genes and hypothesised that the PPI network, derived from the gene signature, could be analysed topologically to prioritise potential targets. We further performed gene set enrichment analysis and pathway analysis, and then made use of drug–gene interaction databases and the Connectivity Map (cMap) to find potential drugs for the treatment of nasopharyngeal cancer [3]. It is conjectured that a small drug molecule may potentially reverse the disease signature if the molecule‐induced signature is significantly negatively correlated with the disease‐induced signature found in the cMap [3]. In fact, potential new treatments for cancers have been successfully identified via the cMap, including acute leukaemia, colon cancer, hepatocellular carcinoma, neuroblastoma, NSCLC and renal cell carcinoma [4–6]. Both up‐ and down‐expressed genes are potential therapeutic targets; therefore identification of potential drugs to treat lung cancer by using an ‘in silico’ screening approach followed by empirical validation might be easier and faster than traditional drug discovery.

2 Methods

This study proposes an ‘in silico’ strategy to narrow down the search for lung cancer genes for target identification and drug discovery; the workflow of this study is shown in Fig. 1.

Fig. 1.

Work flow of this study

2.1 Input dataset

Microarray data for lung cancer was downloaded from [4]the experiment ID E‐TABM‐15, which consisted of 41 samples from a cohort of 18 patients with cancerous and non‐cancerous lung tissue. We conducted a two‐pair test (normal as well as cancer tissues are taken from the same patient); therefore only 36 samples were used. The age distribution of the 18 patients was two age 40–49 years, three age 50–59 years, three age 60–69 years, nine age 70–79 years and one age 80–89 years.

2.2 Microarray data analysis

Microarray technology allows for high‐throughput screening and analysing tens of thousands of genes at the same time. Some genes are activated or inhibited, and some are differentially expressed genes (DEGs) which, because of certain regulatory factors, results in changes in gene expression levels by a few times, ten times or more. Given sets of microarray data, one can identify DEGs among a large number of gene expressions, and understand the mechanism of lung cancer formation induced by these DEGs.

There are many microarray data analysis methods, such as using the concept of false discovery rate to screen for significant genes [7], using analysis of variance to explore the impact of microarray gene expression values within a single factor [8] and clustering analysis. Among the many statistical methods, significance analysis of microarray (SAM) [9, 10], empirical Bayes analysis of microarrays (EBAM) [11] and empirical Bayes statistics (eBayes) [12] are the three commonly employed approaches to screen DEGs. The publicly available microarray data analysis package ‘Bioconductor’ [13, 14] was adopted to perform such calculations.

The statistical method eBayes was chosen in this study because it was found that eBayes, SAM and EBAM achieve a similar level of cancer gene prediction accuracy [15]. The selected DEGs were divided into two groups, an up‐regulated group (up probes in Fig. 1) and a down‐regulated group (down probes), according to the gene expression fold change (FC) values.

Among the DEGs, genes were classified as either up‐ or down‐regulated genes with log₂ FC less than or greater than zero, respectively. Any gene expression level with FC < 5.64 (log₂ 50), was reset to 5.64 in order to facilitate the cMap search.

2.3 Cluster analysis

In a PPI network, a densely connected area is referred to as a cluster, which is a functional module. Nodes having high degree of connection are defined as hubs and are more likely to be essential. The members of a cluster are usually involved in similar biological processes, and protein complexes can be identified through the clustering of a network [16, 17]. It is suggested that a protein complex is a biologically functional module composed of subunits performing similar functions [18]. Given two proteins A and B have a PPI, if both of A and B are obtained from the eBayes prediction as up‐regulated, then the PPI among A and B are the so‐called up PPI. Communities constructed from up PPI are called up‐regulated communities.

To investigate the functional modules in which potential lung cancer‐related proteins are involved, a set of highly confident ‘human’ PPIs were input into the CFinder software [19] to perform an analysis based on the clique percolation clustering approach [20]. A 3‐community was set as k being equal to three (complete subgraphs of size k). Any two k ‐communities are adjacent if they share k −1 common nodes. A k ‐community (k ≥ 4) is constructed by merging all possible adjacent (k −1) communities. Communities with up‐regulated PPI are called up‐regulated communities.

In this study, we compared the k ‐community results with known protein complexes obtained from MIPS [21] in order to identify experimentally determined cancer‐related protein modules. Subunits from k ‐community are compared with 1818 protein complexes by the Jaccard index (JI), which is an index used to quantify the similarity between two sets. Hence, given two modules A and B, the JI is given by

$$\text{JI}(A,B)=\frac{|A\cap B|}{|A\cup B|}$$ (1)

where |A ∩ B | and |A ∪ B | denote the cardinality of A ∩ B and A ∪ B, respectively. It is noted that JI lies between 0 and 1.

2.4 Gene set enrichment analysis

DAVID [22] is a web‐based resource which provides batch annotation and gene ontology (GO) [23] term enrichment analysis to highlight the most relevant GO terms associated with a given gene list. The ConsensusPathDB (CPDB) [24] tool provides gene set analysis and metabolite set analysis. DAVID tool is based on the Fisher's exact test, whereas the CPDB tool is based on the Wilcoxon test. To find the enriched pathways of our lung cancer gene signature, we performed an over‐representation pathway analysis by using both DAVID and CPDB. Under the threshold of a P ‐value of < 0.005, enriched pathways from the over‐representation analysis include up‐ and down‐regulated k ‐communities obtained from CFinder analysis. Significant pathway results were ranked according to the P ‐value. Thus, enriched GO terms for these two protein groups were obtained. We used both tools in this stage for cross‐verification. For DAVID and CPDB, as the default setting, enriched biological processes with a P ‐value of < 0.01 were examined in this work.

2.5 Potential target genes and drugs discovery

Both of the up‐ and down‐regulated communities derived from the CFinder tool were used to query the cMap database, where potential drugs with P ‐value < 0.05 are retained.

To identify target genes, the Food and Drug Administration‐approved drugs and the chemical–protein links data from STITCH [25] were merged. The Gene Name Service was then used to translate the protein ID to its corresponding human gene organisation (HUGO)‐approved gene symbol and Entrez gene ID. Drugs obtained from the cMap output were mapped, and finally identified with known drug targets in the cancer up‐ or down‐regulated PPI network.

The idea of drug repositioning is a rather popular approach in the pharmaceutical industry. This approach intends to identify new uses for existing drugs, and it has achieved certain successes [26]. Furthermore, this approach has the potential to accelerate the development time for drugs and reduce side‐effects as well. There are many works on identifying repositioned drugs which is based on various methods: graph‐based inference method [27, 28], microarray expression method [29] and differential expressed correlation method [30].

2.6 MTT cell viability test

Cell viability was determined using an MTT assay. NSCLC cell lines A549, A14 and CL1‐5 were seeded in a 96‐well microplate for 16–20 h and treated with drugs. After drug treatment for 72 h, 50 µl MTT (3‐[4, 5‐dimethylthiazol‐2‐yl]‐2, 5‐diphenyltetrazolium bromide) solution [2 mg/ml in phosphate‐buffered saline (PBS)] per well was added and incubated at 37°C. Two hours later, dimethyl sulfoxide was added and the absorbance at 570 nm was detected by using enzyme‐linked immunosorbent assay reader (Infinite^® M1000, TECAN, Switzerland). The untreated groups were considered as 100% viability.

2.7 Clonogenic assay

Four NSCLC cell lines A549, H460, H23, H1299‐wild and H1299‐L858R were seeded in six‐well plates with 400 cells per well for 10 days. Each well contained 2 ml Rosewell Park Memorial Institute (RPMI) medium as the cultured condition and the tested drugs were added 24 h after seeding of the cells. The medium and drugs were changed once on day 4. After the treatments, cells were washed with PBS, and the colonies were fixed with fix solution (acetic acid: methanol = 1:3) and stained with 0.5% crystal violet in methanol. After removing the crystal violet carefully and rinsing with tap water, the colonies were counted manually.

3 Results

3.1 Microarray data analysis

In this study, the microarray source data, E‐TABM‐15, comprised 22 283 genes. Robust multi‐array average was used for gene expression normalisation.

The eBayes analysis was subsequently conducted on the previous results. DEGs were predicted by adjusted P ‐value of 0.005. After eliminating unlabelled and repeated genes, 2291 genes remained. According to their log FC values, 952 and 1339 genes are classified as the up‐ and down‐group, respectively. By integrating these genes with the BioGrid [31] PPI data, a list of binary interactions among DEGs was determined for the up‐ and down‐groups.

It is known that cancer is a heterogeneous disease. There is concern that changing to a different dataset may lead to totally different DEGs. Here, we repeated the DEGs analysis using a different dataset, that is, GSE‐7670, obtained from NCBI. Experiment GSE‐7670 comprises of tissue took from adjacent normal‐tumour lung tissue from 19 patients. The age distribution of the 19 patients was three age 40–49 years, eight age 50–59 years, four age 60–69 years and four age 70–79 years. Again a two‐pair test was conducted, a total of 2558 DEGs were found, in which 1558 DEGs are the same as given by E‐TABM‐15, a 68% (i.e. 1558 of 2291) overlap. This result suggested that the analysis is rather robust, it may be because of fact that normal‐tumour lung tissue was sampled, which could possibly eliminate heterogeneity because of different individuals.

3.2 Cluster analysis

Based on the assumption that genes which do not highly interact with other genes are viewed to be less important, they were removed before the subsequent analysis. Hence, by CFinder, any gene which did not belong to a k ‐community was excluded. We also counted the number of k ‐community in the lung cancer PPI network, and there was no community with k larger than five. Table 1 states the number of k ‐community identified by CFinder. A total of 331 and 939 genes belong to the 86 up‐ and 482 down‐regulated k ‐communities, respectively.

Table 1.

Statistics of k ‐community search results by CFinder

k	Up‐group	Down‐group
3	74	407
4	12	71
5	null	4
total	86	482

To examine the reliability of these modules, these communities were compared with the protein complex data, that is, CORUM in MIPS, the corresponding JI values were computed to evaluate the similarity level. Those protein complexes whose subunits were completely different from any k ‐community modules were excluded from further discussion. The results of the identified communities are summarised in Table 2.

Table 2.

Total number of k ‐community identified by CFinder and their corresponding JI values

k	Up‐group	JI, (%)	Down‐group	JI, (%)
3	73/192 = 0.38	5%–100%	302/466 = 0.65	3%–75%
4	12/17 = 0.70	20%–100%	65/123 = 0.53	13%–50%

The number before the slash (/) denotes the number of k ‐communities retrieved by CFinder, whereas the first number after the slash denotes the corresponding number of protein complexes found in CORUM.

Among the 86 (482) communities identified by CFinder only 85 (371), that is, 99% (77%), had non‐zero JI values. The up‐group data seems to have a higher JI value than the down‐group; for instance, certain 3‐community and 4‐community groups achieved a 100% JI value. On the other hand, a few down‐group communities did not correspond to real protein complexes. Nevertheless, the present data indicated that interaction dense regions represent protein complexes in most of the cases. Also, from Table 2, it was found that JI interval decreases as k increases; for instance, the interval decreases from 72% (75 − 3%) to 23% (40 − 17%) for the down‐group as k increases from three to five. Moreover, the high end of the JI value decreases as k increases. The distributions of the JI values for up‐ and down‐regulated groups are summarised in Figs. 2 a and b, respectively.

Fig. 2.

Distributions of the JI values for up‐ and down‐regulated groups

a JI value distribution of up‐regulated group

b JI value distribution of down‐regulated group

Fig. 3 depicts the results of four 5‐community down‐groups extracted by CFinder; it was found that the four down‐groups were exact cliques.

Fig. 3.

Results of four 5‐community down‐groups extracted by CFinder

We further asked which protein complexes had more lung cancer associated genes. Eight complexes, comprised of communities with at least three cancer‐associated genes were found, among them the MCM2‐MCM4‐MCM6‐MCM7 complex comprised the highest number of common DEGs, that is, four. The results are reported in the second column of Table 3. The up‐regulated proteins are involved in the first seven protein complexes (Table 3), whereas the down‐group proteins are involved in the SNX complex. These results are consistent with the hypothesis, where protein complex's subunits are co‐expressed.

Table 3.

Associated genes and enriched GO biological processes for lung cancer‐associated protein complexes

Complex name	Number of lung cancer‐associated genes	Go description	Genes
STAGA complex		DNA topological change	KAT2A, SF3B3, TAF10
		transcriptional activator activity
		protein amino acid acetylation
		protein amino acid deacetylation
		response to DNA damage stimulus
		chromosome organisation and biogenesis
H2AX complex I	3	DNA repair	NPM1 (O, T), PARP1 (T), H2AFX
		DNA topological change
		cell cycle checkpoint
		RNA elongation
		response to DNA damage stimulus
		chromosome organisation and biogenesis
FEN1‐9‐1‐1 complex	3	DNA catabolism	FEN1, HUS1, RAD1
		RNA catabolism
		DNA repair
		cell cycle checkpoint
		response to DNA damage stimulus
PCNA‐RFC2‐5 complex	3	DNA replication	PCNA, RFC4, RFC3
chromosomal passenger complex CPC	3	mitotic cell cycle	AURKB (T), BIRC5(T), CDCA8
		chromosome segregation
MDC1‐H2AFX‐TP53BP1 complex	3	DNA repair	H2AFX, MDC1(T), TP53BP1 (T)
		cell cycle checkpoint
		DNA binding
		response to DNA damage stimulus
MCM2‐MCM4‐MCM6‐MCM7 complex	4	phosphate metabolism	MCM2, MCM4, MCM6, MCM7
		DNA replication
		ATP binding
SNX complex	3	protein transport	LEPR, SNX1, SNX2
		protein transporter activity
		receptor mediated endocytosis
		enzyme‐linked receptor protein signalling pathway

Letters O and T in the last column denote oncoprotein and tumour suppressor protein, respectively.

Enriched GO biological processes for the complexes are also summarised in Table 3. The STAGA complex is a co‐activator required for the transcription of a subset of RNA polymerase II‐dependent genes. The H2AX complex I is involved in DNA repair, cell cycle checkpoint and DNA damage. The FEN1‐9‐1‐1 complex is also involved in DNA repair, cell cycle checkpoint and DNA damage. Another protein, the PCNA‐RFC2‐5 complex is associated with DNA replication. Chromosomal passenger complex CPC is involved in cell cycle and chromosome segregation. The MDC1‐H2AFX‐TP53BP1 complex performs similar functions to the H2AX complex I and FEN1‐9‐1‐1 complex. The MCM2‐MCM4‐MCM6‐MCM7 complex is involved in phosphate metabolism, DNA replication and ATP binding. Finally, the SNX complex is mainly associated with protein transportation.

Furthermore, it is interesting to note that six tumour suppressor proteins are found in three up‐regulated complexes, that is, H2AX complex, CPC complex and MDC1‐H2AFX‐TP53BP1 complex. For instance, PARP1, a DNA repair protein and E2F1 co‐activator, was highly expressed at the mRNA and protein levels in SCLC [32]. AURKB transcripts are frequently over‐represented in lung tumour [33]. BIRC5 (survivin) is a member of the inhibitor of apoptosis (IAP) gene family, which encodes negative regulatory proteins that prevent apoptotic cell death. Pervious work has shown that BIRC5 protein expression in tumour samples is often elevated in squamous cell lung cancer [34, 35]. MDC1 and TP53BP1 are critical components of the DNA damage response machinery that protects genome integrity and guards against cancer. These two proteins are well recognised in tumorigenesis, such as in lung cancer formation [36, 37]. NPM1 is known to be involved in several processes including regulation of the ARF/p53 pathway. Mutations in this gene are associated with acute myeloid leukaemia [38] and colon cancer [39]. Its role in lung cancer is still unknown [40].

As we mentioned above, a total of 86 up‐ and 482 down‐regulated clusters of k ‐community were obtained for the up‐ and down‐groups, respectively. Only genes belong to the communities identified by CFinder were selected for the next stage of the investigation; hence, 88 up‐ and 381 down‐regulated genes were chosen.

3.3 Enriched biological pathways

The functional annotation of community was given by implementing DAVID and CPDB. Enriched terms of biological processes and pathways were retrieved as the output. For the DAVID tool, the top five pathways were sorted by their P ‐value according to REACTOME [41] and KEGG [42], and are listed in Tables 4 and 5, respectively, where the ‘Count’ column denotes the number of overlapped genes in the filtered community genes and the corresponding pathway. The cell cycle pathway ranked among the top three both in REACTOME and KEGG, respectively.

Table 4.

Summary of top five pathways returned by REACTOME using DAVID

Term	Count	%	P value
cell cycle checkpoints	28	6	1.40 × 10−10
cell cycle, mitotic	44	9.4	3.20 × 10−8
DNA replication	18	3.8	4.60 × 10−5
signalling by NGF	24	5.1	2.60 × 10−4
botulinum neurotoxicity	7	1.5	4.40 × 10−4

Table 5.

Summary of top five pathways returned by KEGG using DAVID

Term	Count	%	P value
focal adhesion	35	7.5	1.10 × 10−9
cell cycle	27	5.8	1.20 × 10−9
pathways in cancer	42	9	1.80 × 10−7
regulation of actin cytoskeleton	32	6.8	2.90 × 10−7
endocytosis	26	5.5	1.30 × 10−5

In contrast, by using the CPDB software, the top five most significant pathways returned by REACTOME and KEGG are listed in Tables 6 and 7, respectively, with P < 0.01. It is known that PAK‐2p34 and SNARE are associated with apoptosis and vesicle formation, respectively.

Table 6.

Summary of the top five pathways returned by REACTOME using CPDB

Pathway name	Set size	Candidates	% of overlap	P value
stimulation of the cell death response by PAK‐2p34	2	2	100%	0.00161
localisation of the PINCH‐ILK‐PARVIN complex to focal adhesions	4	3	75%	0.000249
phosphorylation of Emi1	6	4	66.7%	3.59 × 10−5
regulation of PLK1 activity at G2/M transition	6	4	66.7%	3.59 × 10−5
phosphorylation of proteins involved in the G2/M transition by Cyclin A:Cdc2 complexes	3	2	66.7%	0.00469

Table 7.

Summary of top five pathways returned by KEGG using CPDB

Pathway name	Size	Candidates	% of overlap	P value
SNARE interactions in vesicular transport	36	8	22.2%	7.05 × 10−5
DNA replication	36	8	22.2%	7.05 × 10−5
cell cycle	124	27	21.8%	3.38 × 10−13
colorectal cancer	62	13	21.0%	7.99 × 10−7
chronic myeloid leukaemia	73	15	20.5%	1.50 × 10−7

Again, the cell cycle pathway ranked among the top three both in REACTOME and KEGG by using CPDB. In other words, analysis using DAVID and CPDB are in good agreement. To be specific on DNA and cancer‐related pathways, a total of 32 and 294 pathways were obtained from DAVID and CPDB, respectively, among them, 21 (5 pathways from DAVID and 16 pathways from CPDB) and 18 (2 pathways from DAVID and 16 pathways from CPDB) pathways are, respectively, related to DNA biological process and cancer were retrieved. Table 8 summarises the pathway classification statistics. Relative to DAVID, CPDB returns a greater number of pathways.

Table 8.

Results of statistics for the number of biological pathways which are related to DNA and cancer

DataBase	DAVID			CPDB
	Pathways	Cancer	DNA	Pathways	Cancer	DNA
REACTOME	7	0	1	223	7	15
KEGG	25	5	1	71	9	1
total	32	5	2	294	16	16

Table 9 lists the derived DNA pathways sorted by P ‐value. Most of these pathways are related to DNA replication and repair. Both the DAVID and CPDB analysis predicted DNA replication as a potentially significant pathway.

Table 9.

Results of biological pathways related to DNA according to CPDB and DAVID analysis

CPDB
Pathway name	P value	Pathway source
DNA replication	7.59 × 10−10	REACTOME
synthesis of DNA	1.35 × 10−9	REACTOME
regulation of DNA replication	6.84 × 10−9	REACTOME
DNA replication pre‐initiation	1.27 × 10−6	REACTOME
DNA strand elongation	2.26 × 10−6	REACTOME
unwinding of DNA	3.83 × 10−5	REACTOME
DNA replication	7.05 × 10−5	KEGG
DNA repair	0.00018	REACTOME
DAVID
Pathway name	P value	Pathway source
DNA replication	4.60 × 10−5	REACTOME
DNA replication	2.60 × 10−3	KEGG

The DNA pathways returned by CPDB can be further subdivided based on their annotations. In Table 10, we show the results of gene‐associated DNA pathways, where the majority are related to the p53 gene.

Table 10.

Gene‐associated DNA pathways

CPDB
Pathway name	P value	Pathway source
p53‐dependent G1 DNA damage response	0.000218	REACTOME
p53‐dependent G1/S DNA damage checkpoint	0.000218	REACTOME
E2F‐mediated regulation of DNA replication	0.00152	REACTOME
p53‐independent DNA damage response	0.00449	REACTOME
p53‐independent G1/S DNA damage checkpoint	0.00449	REACTOME

Tablee 11 shows the cell‐cycle‐checkpoint‐related DNA pathways.

Table 11.

Cell cycle‐associated DNA pathways

CPDB
Pathway name	P value	Pathway source
G2/M DNA damage checkpoint	1.12 × 10−5	REACTOME
G1/S DNA damage checkpoints	4.94 × 10−5	REACTOME
G2/M DNA replication checkpoint	0.00913	REACTOME

Tables 12 – 14 list the pathways related to cancer, which are classified according to signal transduction, cancer types, signal transduction and genes. From Table 12, we observed that some of these pathways are related to the cell signalling transmitting system such as the NOTCH series [43]. In addition, the combined inhibition of PI3 K (Table 12) and Mitogen‐activated protein kinases (MAPK) signalling pathway is regarded as an effective therapeutic treatment strategy [44–46]. From Table 13, other cancer type annotations are also reported, such as bladder, colorectal, pancreatic, SCLC cancers etc., this suggested that the identified proteins are rather prevalent in carcinogenesis.

Table 12.

Signal transduction‐associated lung cancer pathways with P < 0.01

CPDB
Pathway name	P value	Pathway source
signalling by NOTCH1 HD domain mutants in cancer	2.42 × 10−6	REACTOME
signalling by NOTCH1 HD + PEST domain mutants in cancer	2.42 × 10−6	REACTOME
signalling by NOTCH1 PEST domain mutants in cancer	2.42 × 10−6	REACTOME
signalling by NOTCH1 in cancer	2.42 × 10−6	REACTOME
transcriptional misregulation in cancer	3.46 × 10−5	KEGG
signalling by EGFR in cancer	0.000301	REACTOME
constitutive PI3 K/AKT signalling in cancer	0.00795	REACTOME

Table 14.

Gene‐associated lung cancer pathways with P < 0.01

CPDB
Pathway name	P value	Pathway source
FBXW7 mutants and NOTCH1 in cancer	2.42 × 10−6	REACTOME

Table 13.

Results of cancer pathways returned by CPDB and DAVID according to KEGG with P < 0.01

CPDB
Pathway name	P value
proteoglycans in cancer	2.01 × 10−11
pathways in cancer	6.23 × 10−11
colorectal cancer	7.99 × 10−7
pancreatic cancer	1.70 × 10−6
prostate cancer	5.05 × 10−5
small cell lung cancer	0.000153
bladder cancer	0.00373
endometrial cancer	0.00446
DAVID
Pathway name	P value
pathways in cancer	1.80 × 10−7
pancreatic cancer	3.80 × 10−4
colorectal cancer	1.60 × 10−3
small cell lung cancer	1.60 × 10−3
prostate cancer	2.60 × 10−3

Table 14 shows the result of gene‐associated cancer pathway, in which recurrent mutations in FBXW7 are found in squamous cell lung cancers [47].

3.4 Retrieving target genes for lung cancer

Both the up‐ and down‐groups extracted from CFinder in Section 3.2 were analysed by cMap. Under the constraint of a P ‐value < 0.005, a total of 30 drugs were retrieved, and the results ranked by P ‐value in Table 15.

Table 15.

Summary of the potential 30 drugs returned by cMap

Rank	cMap drug name	Mean	N	Enrichment	P value
1	vorinostat	−0.744	12	−0.861	0
2	trichostatin A	−0.693	182	−0.818	0
3	prochlorperazine	−0.463	16	−0.622	0
4	thioridazine	−0.482	20	−0.606	0
5	trifluoperazine	−0.365	16	−0.578	0
6	fluphenazine	−0.453	18	−0.575	0.00002
7	MS‐275	−0.899	2	−0.998	0.00004
8	Prestwick‐675	0.54	4	0.904	0.0001
9	Scriptaid	−0.755	3	−0.963	0.00016
10	metergoline	−0.648	4	−0.905	0.00016
11	ciclopirox	−0.636	4	−0.897	0.00022
12	perphenazine	−0.626	5	−0.848	0.00024
13	15‐delta prostaglandin J2	−0.423	15	−0.506	0.00042
14	desipramine	−0.637	4	−0.881	0.00048
15	piperacetazine	−0.584	4	−0.88	0.00052
16	niclosamide	−0.497	5	−0.81	0.00056
17	perhexiline	−0.679	4	−0.87	0.00058
18	bambuterol	0.295	4	0.854	0.0006
19	econazole	−0.544	4	−0.861	0.00068
20	loperamide	−0.332	6	−0.73	0.00073
21	Anisomycin	−0.516	4	−0.846	0.00101
22	pyrvinium	−0.452	6	−0.713	0.00119
23	AH‐6809	0.594	2	0.972	0.00125
24	resveratrol	−0.358	9	−0.598	0.00132
25	mefloquine	−0.343	5	−0.757	0.00154
26	alimemazine	−0.496	4	−0.824	0.00181
27	azapropazone	0.34	3	0.888	0.0027
28	ivermectin	−0.373	5	−0.715	0.00411
29	rifabutin	−0.628	3	−0.872	0.00415
30	(−)‐MK‐801	−0.297	4	−0.779	0.00491

Only P < 0.005 were retained.

Bold faced indicated cMap drug names validated by MTT or clonogenic assays.

In the 30 potential drugs from Table 15, there were 11 drugs tested by MTT or clonogenic assays and validated as effective, including trichostatin A, prochlorperazine, thioridazine, trifluoperazine, ciclopirox, perphenazine, 15‐delta prostaglandin J2, niclosamide, Anisomycin, pyrvinium and mefloquine.

We submitted the selected 30 drugs to DrugBank and NCBI to search for their target genes. A total of 76 and 101 target genes were derived from DrugBank and NCBI, respectively. After eliminating the repeated genes, we obtained 46 and 39 target genes from DrugBank and NCBI, respectively.

Secondly, DrugBank and NCBI queries return three up‐regulated genes – MEN1, CSNK2A1, HSP90B1 and four down‐regulated genes – NR3C1, TNNC1, ADRB2 and CALM1. Of the selected 30 drugs, 8 drugs (trifluoperazine, fluphenazine, perphenazine, desipramine, bambuterol, loperamide, resveratrol and rifabutin) from DrugBank and 3 drugs (trifluoperazine, metergoline and mefloquine) from NCBI target at least one of the above seven genes; therefore it is concluded that those are potential drug targets. Among these retrieved drugs, trifluoperazine was found in both DrugBank and NCBI but its target genes are different. The results of drugs and targets obtained from cMap are summarised in Table 16, which are the potential drugs and therapeutic targets for future lung cancer clinical trials.

Table 16.

Results of 11 drugs and 7 target genes predicted by cMap

Rank	Drug name	Target gene
DrugBank
5	trifluoperazine	DRD2	DRD1IP	ADRA1A	CALM1	TNNC1	S100A4
6	fluphenazine	DRD2	DRD1	CALM1
12	perphenazine	DRD2	DRD1	CALM1
14	desipramine	SLC6A2	SLC6A4	HTR2A	ADRB2	ADRB1	SMPD1	SNF
		HRH1	ADRA1A	CHRM1	CHRM2	CHRM3	CHRM4	CHRM5
18	bambuterol	ADRB2
20	loperamide	OPRM1	OPRD1	OPRK1	CACNA1A	POMC	CALM1
24	resveratrol	NQO2	CSNK2A1	PTGS1	PTGS2
29	rifabutin	rpoA	rpoB	rpoC	HSP90AA1	HSP90B1
NCBI
5	trifluoperazine	NR3C1	TSHR	THRB	CYP2C19
10	metergoline	THRB	MEN1	MAPT	GLA
25	mefloquine	PMP22	MEN1	HTT	GMNN

Boldface and underlined italics denote up‐ and down‐regulated DEGs obtained from microarray study.

It is supposed that a small molecule (drug) may potentially reverse the disease signature if the molecule‐induced signature is significantly negatively correlated with the disease‐induced signature in cMap. In other words, the drugs extracted from cMap are of up or down connectivity, up/down connectivity implies that a drug may activate/suppress the expression of certain genes. In Table 16, only bambuterol (ranked 18th) is of up connectivity, the others are of down connectivity. It is also worth noting that both bambuterol and desipramine (ranked 14th) target ADRB2, which implies that the two drugs induced target expression in an opposite manner. Therefore bambuterol and desipramine cannot be used simultaneously in cancer treatment since gene expression may be affected in opposite ways.

Lung cancer cells were treated with potential drugs for 72 h, followed by detection of cell viability by MTT assay. Table 17 summarised the IC50 values for the 11 drugs. It is known that trifluoperazine and perphenazine target dopamine receptors. Moreover, trifluoperazine could inhibit cancer stem cells growth [48] and has better IC50 in a clonogenic assay than MTT assay. Since a clonogenic assay reflects a more inhibitory effect on stemness function, this result supports the previous study [49] that phenothiazine‐like anti‐psychotics may target cancer stem cells.

Table 17.

Results of IC50 for the 11 drugs predicted by cMap

Rank	Drug name	MTT	Clonogenic
		IC50, μM	IC50, μM
DrugBank
5	trifluoperazine	> 10	< 3.3
6	fluphenazine	> 5	—
12	perphenazine	> 10	< 10
14	desipramine	—	—
18	bambuterol	—	—
20	loperamide	—	—
24	resveratrol	—	—
29	rifabutin	—	—
NCBI
5	trifluoperazine	> 10	< 3.3
10	metergoline	—	—
25	mefloquine	> 5	< 10

3.5 Target genes‐related PPI

In this section, we focus on the seven target genes derived from Section 3.4 and their adjacent genes in PPI. Table 18 lists the F ‐rank, degree and degree rank for each target genes, where F ‐rank represents the rank (of 331up‐genes and 939 down‐genes) of logFC value and D ‐rank denotes the rank of graph degree. It can be seen that most of the target genes ranked low in F ‐rank.

Table 18.

Results of F ‐rank, logFC, degree and the degree rank of the seven predicted target genes

Up gene	F ‐rank	logFC	Degree	D ‐rank	Down gene	F ‐rank	Log FC	Degree	D ‐rank
HSP90B1	108	0.68	5	49	NR3C1	447	−0.75	22	4
MEN1	278	0.38	4	66	CALM1	716	−1.16	21	5
CSNK2A1	289	0.36	6	38	ADRB2	847	−1.62	4	132
					TNNC1	930	−2.81	2	291

Fig. 4 depicts the PPI network of up‐regulated target genes by using Cytoscape [50]. The up‐regulated target gene MEN1 directly interacts with TP53, which has the highest degree of PPI and it is a well‐known tumour suppressor gene. Also, this gene is one of the most frequently mutated genes in lung adenocarcinoma [51, 52], such relationship could further highlight the possibility of targeting p53‐associated pathways [53].

Fig. 4.

Up‐regulated target genes (circles) PPI partners (squares)

Fig. 5 represents the PPI network of down‐regulated target genes. Three out of the four down‐regulated target genes directly interact with the UBC gene, which has the highest degree of PPI. If drugs could indirectly affect down‐stream PPI of the UBC gene, the therapeutic role of this gene is worth further exploration.

Fig. 5.

Down‐regulated target genes (circles) PPI partners (squares)

4 Conclusions

In this study, the ‘Bioconductor’ package was adopted to identify DEGs for lung cancer microarray data. A two‐pair test was conducted to minimise heterogeneity among individuals. Both the up and down DEGs are identified. It is supposed that they are potential therapeutic targets. By integrating the DEG results with PPI data, it was found that DEGs can be classified into up‐ and down‐regulated PPI communities. Lung cancer‐related protein complexes are identified, suggesting these complexes can potentially play an oncogenic or tumour suppressor role in cancer. Our findings suggest that the three up‐regulated genes – MEN1, CSNK2A1, HSP90B1, and four down‐regulated genes – NR3C1, TNNC1, ADRB2 and CALM1 are the potential drug targets. Furthermore, the eight drugs derived from DrugBank and the three derived from NCBI are the potential lung cancer therapeutic drugs.

In summary, we have developed a pipeline to infer therapeutic drugs for disease treatment by integrating microarray, PPI and the cMap resources. The analysis starts from DEGs identification, and then PPI data are employed to construct dense PPI modules. Given the modules, protein complexes identification was performed. Up‐ and down‐regulated communities were used as queries to perform functional enrichment analysis. Over‐represented or enriched biological processes and pathways are determined. The results of our drug findings are supported by IC50 experimental data.

It is expected that the approach developed in the current work should be of value for future studies into understanding the molecular mechanism of lung cancer formation and identifying the therapeutic drug targets. In conclusion, this study provided a systematic strategy for the identification of DEGs, cancer‐related complexes, drug targets and potential drugs for lung cancer; this systematic strategy could be applied to other types of cancer.