Identification of subsets of actionable genetic alterations in KRAS-mutant lung cancers using association rule mining

Abstract

Background

Lung cancer is the leading cause of cancer-related death in both men and women. KRAS mutations occur in ~ 25% of patients with lung cancer, and the presence of these mutations is associated with a poor prognosis. Unfortunately, efforts to directly target KRAS or its associated downstream MAPK or PI3K/AKT/mTOR pathways have seen little or no benefits. Here, I hypothesize that KRAS-mutant tumors do not respond to KRAS pathway therapies due to the co-occurrence of other activated cell survival pathways and/or mechanisms.

Methods and results

To identify other potentially activated cell survival pathways in KRAS-mutant tumors, I performed association rule mining on somatic mutations in 725 metastatic lung cancer patient samples. I identified 67 additional genes that were mutated in at least 10% of the samples with KRAS mutations. This gene list was enriched with genes involved in the MAPK, AKT and STAT3 pathways, as well as in cell-cell adhesion, DNA repair, chromatin remodeling and the Wnt/β-catenin pathway. I also identified 160 overlapping subsets of three or more genes that code for oncogenic or tumor suppressive proteins that were mutated in at least 10% of the KRAS-mutant tumors.

Conclusions

I identified several genes that are co-mutated in primary KRAS-mutant lung cancer samples. I also identified subpopulations of KRAS-mutant lung cancers based on sets of genes that were co-mutated. Pre-clinical models that capture these subsets of KRAS-mutant tumors may enhance our understanding of lung cancer development and, in addition, facilitate the design of personalized treatment strategies for lung cancer patients carrying KRAS mutations.

Electronic supplementary material

The online version of this article (10.1007/s13402-018-0377-5) contains supplementary material, which is available to authorized users.

Introduction

Advanced lung cancer is the most commonly diagnosed, and the leading cause of all cancer deaths worldwide [1–3]. Morphologically, lung cancer can be broken down into two major categories: non-small cell lung carcinoma (NSCLC), which makes up 85% of all the cases, and small cell lung carcinoma (SCLC) [2]. Data accumulated over the last 15 years show that lung cancer is not a single disease but a collection of genetically varied neoplasms with distinct molecular features and clinical outcomes. Known oncogenic drivers of lung cancer include the receptor tyrosine kinases (RTKs) EGFR, ALK, RET, MET, DDR2, ROS1 and FGFR1, and their downstream effectors KRAS, PIK3CA and BRAF [4]. Tumors with EGFR alterations are targetable by EGFR tyrosine kinase inhibitors (TKIs) such as erlotinib, gefitinib and afatinib [5–7], whereas tumors with ALK and ROS1 alterations can be targeted with crizotinib and ceritinib [8]. Progress has also been made in therapeutically targeting the other RTKs mentioned [4, 9], but for tumors containing KRAS alterations there are currently no FDA approved drugs available. KRAS mutations are present in 20 to 25% of all lung cancers and are associated with a poor prognosis [10, 11]. KRAS signals primarily through the MAPK and the PI3K/AKT/mTOR pathways [12–14]. Efforts to directly target KRAS or its downstream effectors have seen little or no benefit [15]. The reason for this limited success may be the development of adaptive response mechanisms or the co-occurrence of mutations in other genes that activate different cell survival pathways in these tumors.

Here, I employed association rule mining, a type of unsupervised machine learning, to identify genes whose alterations co-occur with KRAS mutations in advanced lung cancers. The building blocks of association rules are the items that may appear in any given transaction [16]. I used the Apriori association rule algorithm [17], which states that all subsets of a frequent itemset must also be frequent. For example, for the rule {mutation X} → {mutation Y}, then the items {mutation X}, {mutation Y} and {mutation X and mutation Y} must all be frequent. The strength of a rule is determined by three statistical measures, i.e., support, confidence and lift statistics. The support of a mutation (or mutation set) is a measure of how frequent it occurs in the tumor samples. The support for one or more mutations in a gene X in N tumor samples is given by

$$\mathit{\text{Support}}\left(\mathit{\text{mutation}}X\right)=\frac{\mathit{\text{the number of samples with mutations in gene}}X}{N}$$

The confidence is the measure of the accuracy of a rule. For a given rule {mutation X} → {mutation Y}, the confidence is defined as

$$\mathit{\text{Confidene}}\left(\mathit{\text{mutation}}X\to \mathit{\text{mutation}}Y\right)=\frac{\mathit{\text{Support}}\left(\mathit{\text{mutation}}X,\mathit{\text{mutation}}Y\right)}{\mathit{\text{Support}}\left(\mathit{\text{mutation}}X\right)}$$

The confidence provides a measure of the proportion of the samples in which a mutation in gene X coincides with a mutation in gene Y. The lift statistic measures how much more likely a mutation or a set of mutations occur in a sample relative to its typical rate of occurrence, given that another mutation or set of mutations is present in the sample. The lift for a given rule {mutation X} → {mutation Y} is defined by

$$\mathit{\text{Lift}}\left(\mathit{\text{mutation}}X\to \mathit{\text{mutation}}Y\right)=\frac{\mathit{\text{Confidence}}\left(\mathit{\text{mutation}}X,\mathit{\text{mutation}}Y\right)}{\mathit{\text{Support}}\left(\mathit{\text{mutation}}Y\right)}$$

For a hypothetical example, {TP53, BRACA1} → {CD43}, support = 0.2, confidence = 0.80, lift = 3, count = 125, this rule can be interpreted as a tumor that has mutations in both TP53 and BRCA1 will also have a mutation in CD43. This rule covers 20% (support = 0.2) of all the tumor samples, and the rule is correct in 80% (confidence = 0.80) of tumors having TP53 and BRCA1 mutations. A lift of 3 implies that samples with mutations in both TP53 and BRCA1 are three times more likely to have a mutation in CD43 than a typical cancer sample.

Here, the Apriori machine learning algorithm is applied to a dataset containing somatic mutations from 725 advanced lung cancer patient samples to identify gene subsets that are also altered in tumors with KRAS mutations.

Methods

Data acquisition

The data used in this study were obtained from the CBioportal website on October 6th, 2017. The data represent the findings from a study on the mutational landscape of advanced cancers from more than 10,000 patients by Zehir et al. [18]. After extraction of all the information on mutations (deletions, insertions, substitutions and gene rearrangements) from all metastatic lung cancer samples, the columns of interest were merged by sample ID. The resulting file containing the genes with alterations was converted into a transactional data format (supplementary Table 1). The genes coding for the receptor tyrosine phosphatases (PTPRT, PTPRD, and PTPRS) were converted to PTPRs and the B melanoma antigen precursors (BAGE 2, 3, 4 and 5) to BAGE.

Apriori algorithm

Association rule was performed using the Apriori algorithm that was developed by Agrawal and Srikant [17]. The “arules” and “aruleviz” packages were used in the R programming language to execute the Apriori algorithm and for visualization of the resulting rules, respectively. For a typical analysis, I used the codes:

cancer_rules <− apriori (data, parameter = list (support = 0.01, confidence = α, minlen = β, maxlen = μ)).

A support of 0.01 was used in all analyses. This threshold will include only genes that were mutated in at least 1% of the samples. More than 600 genes were above this threshold. The values for the confidence, the minimum (minlen) and the maximum (maxlen) number of the members in a rule were selected based on the question that was addressed.

Graph analysis

Rules or sub-rules generated using the Apriori algorithm in R were saved as Gephi files and manipulated in Gephi 0.9.2 to generate directed graphs as shown in Figs. 1d and 4a. For example, the rule.

Fig. 1. KRAS-mutant lung cancer is associated with alterations in genes that regulate diverse cellular processes.

a Scatter plot showing the distribution of all the rules generated for the genes that were mutated in all the metastatic lung cancer samples included. b Subset of the rules containing KRAS mutations. c Associations between KRAS mutations and mutations in other tumor drivers. The lift (likelihood) for KRAS-mutant tumors to also have mutations in EGFR, MET or PI3K3CA is less than one would expect by chance. Fisher’s exact test (F.E.T) was used to determine the significance of the rules. (*p-value < 0.05). d Directed graph showing associations between KRAS and 67 genes. The query was limited to a one-to-one association with KRAS on the LHS to identify the genes that were also mutated in KRAS-mutant samples. A confidence of 0.1 was used to include only the genes that were altered in at least 10% of the KRAS-mutant samples. The sub-rules generated in R were imported into Gephi and manipulated to represent the nodes with distinct colors based on their degree of connectivity. e The 67 genes that were also altered in the KRAS-mutant tumor samples were used as input in the REACTOME pathway analysis tool. e A subset of the pathways that were enriched in the 67-member gene list

Fig. 4. Analyses of rules with 3 or more members. .

a All rules were generated that contained mutant KRAS, as indicated in Fig. 1b. Next, the rules were subdivided to include only rules with at least three genes, with KRAS on the LHS, and that were present in at least 10% of the KRAS-mutant samples. The resulting 160 rules were imported as a graph file into Gephi. This file was manipulated to show the notes of different colors and font sizes based on the number of edges coming into the nodes (in-degree). Note that KRAS had an in-degree of zero, since it was restricted to the LHS. Mutations in PTPRs (PTPRT, PTPRD and PTPRS) were the most connected with 32 in-degrees. b and c show tabular representations of the rules that converge to the PTPRs and STK11 nodes. Fisher’s exact test (F.E.T) was used to determine the significance of the rules. *p-value < 0.05, ** p-value < 0.01 and *** p-value < 0.001

cancer_rules <− apriori (data, parameter = list (support = 0.01, confidence = 0.5, minlen = 2, maxlen = 10)),

can be saved as a Gephi file using the code:

saveAsGraph((cancer_rules), file = “cancer_rules.graphml”).

In Gephi, graphs were manipulated to represent nodes with different colors or labels with different font sizes based on the directionality and magnitude of the degree of connectivity of the nodes.

Fisher’s exact test

The association rule mining technique is vulnerable to the discovery of large numbers of spurious patterns [19, 20]. The Fisher’s exact test was used to determine the significance of the rules [19]. The test considers all the possible rules that can be obtained from the dataset. A p-value of less than 0.05 was considered statistically significant.

Bioinformatics

The REACTOME pathway analysis site was used to determine the pathways that were enriched in the gene sets obtained from association rule analyses.

Data Availability

The data supporting the findings in this study were obtained from the CBioportal site [18]. Supplementary files and R codes for all the analyses performed are included.

Results

Distribution of mutations in known oncogenic drivers in metastatic lung cancers

The dataset used in this study contained 725 metastatic lung cancer patient samples (93% NSCLC and 7% SCLC) [18]. Consistent with previous studies [21], TP53 was found to be mutated in 63% of all metastatic lung cancer samples. ROS1 and ALK1 mutations were present in 17 and 15% of the samples, respectively. EGFR mutations were present in 29% of the samples, and DDR2, RET, MET and FGFR1 were found to be mutated in 8, 12, 9 and 6% of the samples, respectively. The RTK facilitators PIK3CA, BRAF and KRAS were found to be mutated in 8, 16 and 23% of the samples, respectively. The samples that carried KRAS mutations were almost exclusively NSCLC, i.e., only 0.6% (1 tumor sample) was SCLC. All the tumors that carried EGFR mutations were NSCLC.

Analyses of rules statistics

Using a support of 1% and a confidence of 10%, 187,168 rules were identified for the genes that were mutated in all the samples. KRAS mutations were found to be associated with 5724 of these rules. The supports for the 187,168 rules were in the range of 1–25% of the data (Fig. 1a). The supports for rules containing KRAS mutations were predominantly between 1 and 5%. The maximum support for a rule that had a KRAS mutation was 12% (Fig. 1b). The confidence for all the rules and the KRAS-mutant sub-rules were evenly distributed between 10 and 100% (Fig. 1a, b).

KRAS-mutant lung cancers are less likely to have EGFR, MET or PIK3CA mutations

Next, the co-occurrence of KRAS mutations and alterations in other oncogenic drivers in the metastatic lung cancer samples were examined. Consistent with previous reports [22], EGFR and KRAS mutations were found to be almost mutually exclusive, i.e., only 3% (support = 0.03, p-value = 1.00) of the samples carried mutations in both oncogenes (Fig. 1c and Table 1). The likelihood for a sample with a KRAS mutation to also have an EGFR mutation was two-fold less (lift = 0.46) than one would expect in a typical metastatic lung cancer sample (Fig. 1c). Similarly, the likelihoods for mutations in MET (found in 1% of the KRAS-mutant samples) or PIK3CA (found in 1% of the KRAS-mutant samples) to occur in KRAS-mutant samples were less than one would expect by chance. The likelihoods that samples with a KRAS mutation also have mutations in ALK, ROS1, FGFR1, DDR2 or BRAF were all close to 1 (Fig. 1c). Mutations in RET were more likely to be found in KRAS-mutant tumor samples (p-value = 0.043, Fig. 1c).

Table 1.

Genes that are altered in at least 10% of KRAS-mutant metastatic lung cancers: Support (sup) represents the proportion of all metastatic lung cancers covered by the rule. Confidence (conf. or accuracy) represents the proportion of the KRAS-mutant tumors covered by the rule. Lift is the likelihood of having a mutation in the gene on the right side of the rule given that the sample has a KRAS mutation. Fisher’s exact test (F.E.T) was used to determine the significance of the rules. *p-value < 0.05, ** p-value < 0.01 and *** p-value < 0.001

KRAS mutations are associated with alterations in 67 other genes

To identify potential cell survival pathways that might also be activated in KRAS-mutant tumors, the rules generated from the Apriori algorithm for genes associated with KRAS mutations were queried. By doing so, 67 genes were identified that were mutated in at least 10% of the KRAS-mutant tumors (Fig. 1d and Table 1). Using a Fisher’s exact test, 35 out of the 67 associations were found to be statistically significant (p-value < 0.05, Table 1). The 67 genes are known to be involved in diverse biological pathways that play important roles in cancer cell survival (Fig. 1e). Genes involved in RTK signaling and its downstream MAPK, PI3K/AKT/mTOR and JAK/STAT pathways were predominantly altered (Figs. 1e and 2a). This list was also enriched with genes coding for proteins involved in chromatin remodeling and transcription regulation by TP53 (Figs. 1e and 2c). Several FDA approved and investigational drugs are available that can be used to target at least 45% of the proteins encoded by the 67 genes (Supplementary Table 2, Fig. 3).

Fig. 2. Schematic representation of altered genes encoding proteins in the MAPK, mTOR and STAT3 pathways, as well as in the oxidative stress response and the DNA repair pathways, in KRAS-mutant lung cancers. .

Red denotes oncogenic proteins and blue denotes tumor suppressive proteins. a Members of the MAPK, PI3K/AKT/mTOR and STAT3 pathways, (b) KEAP1-mediated oxidative stress responses and (c) TP53/ATM/ATR-mediated cell cycle arrests

Fig. 3. Drugs targeting proteins of which the encoding genes are altered in KRAS-mutant tumors. .

Using the proteins encoded by the 67 mutated genes from Fig. 1d as input, the DrugBank database (www.drugbank.ca) was queried. The query was limited to FDA approved and investigational drugs that may directly target proteins (a) belonging to the RTK family, (b) acting in the mTOR pathway, (c) acting in the MAPK and STAT3 pathways and (d) acting in chromatin remodeling and DNA repair. A detailed list and the DrugBank IDs are presented in Supplementary Table S2

RTK mutations in samples with KRAS mutations

In addition to the known oncogenic RTK drivers discussed above, several other RTKs were identified that were altered in KRAS-mutant tumors, including members of the EGFR family and the EPH, NTRK and VEGF receptor families. ERBB4, a member of the EGFR family, was found to be altered in 17% of the samples with KRAS mutations (Fig. 2a). Unlike EGFR, the likelihood of having a KRAS-mutant tumor with a mutation in ERBB4 was higher than one would expect by chance in an advanced lung cancer sample (lift = 1.35 and p-value < 0.05, Table 1). KRAS-mutant samples also exhibited alterations in several members of the VEGF receptor family including the FLT1 (16%, p-value = 0.055), FLT3 (10%, p-value < 0.05), FLT4 (19%, p-value < 0.001) and KDR (15%, p-value < 0.05) receptors (Fig. 2a, Table 1). NTRK1 and NTRK3 that code for the TrkA and TrkC receptors were found to be mutated in 10% and 25% of the KRAS-mutant samples, respectively (Fig. 2a, Table 1). Four members of the ephrin receptor family, EPHA3 (22%), EPHA5 (22%), EPHA7 (12%) and EPHB1 (11%), were also found to be altered in KRAS-mutant tumors (Table 1). The platelet-derived growth factor receptors PDGFRA and PDGFRB were found to be altered in 14 and 11% of the KRAS-mutant samples, respectively (Fig. 2a and Table 1). Mutations in MET and KRAS were found to be almost mutually exclusive (Fig. 1c), whereas the MET ligand HGF was found to be mutated in 20% of the KRAS-mutant samples (p-value = 0.054, Table 1). In summary, 18 genes coding for RTKs were identified that were also altered in KRAS-mutant tumors. Thirteen of these RTKs are direct targets of currently available FDA approved or investigational drugs (Fig. 3a and Supplementary Table S2).

Mutations in genes involved in the MAPK and PI3K/AKT/mTOR pathways

KRAS is an effector of RTK signaling, and KRAS mutations can lead to constitutive activation of its downstream cell survival MAPK pathway, irrespective of receptor activation [23, 24] (Fig. 2a). Consequently, RTK inhibitors have been shown to be futile in the presence of activating KRAS mutations [23, 24]. BRAF is a serine-threonine kinase that acts directly downstream of KRAS in the MAPK pathway (Fig. 2a). Constitutively active BRAF mutations can also drive tumor progression by activating MEK, irrespective of RTK and/or KRAS mutation statuses [25]. In colorectal cancer, it has been shown that mutations in KRAS and BRAF are almost mutually exclusive [26, 27]. However, in the metastatic lung cancer dataset used in this study, BRAF was mutated in 16% of the samples with KRAS mutations (Fig. 1c and Table 1). The likelihood of having a KRAS-mutant tumor with a BRAF alteration was 0.95 (p-value = 0.671, Table 1). Activating mutations in the Gαs subunits of heterotrimeric G-proteins can also promote cancer cell proliferation via the MAPK pathway [28]. GNAS, which codes for the Gαs subunit, was found to be mutated in 13% of the KRAS-mutant samples (Fig. 2a). The tumor suppressor NF1 was found to be altered in 11% of the KRAS-mutant tumors (Fig. 2a and Table 1). NF1 activates the GTPase activity of KRAS [29]. However, loss of NF1 is not redundant in KRAS-induced myeloproliferative disorders, as alterations in both genes have been found to result in disease with a shorter latency period [30]. These observations suggest that NF1 may have additional tumor suppressive functions, and that the presence of a concurrent mutation in KRAS may lead to a more aggressive disease. In summary, three mutant genes (BRAF, GNAS and NF1) were identified downstream of receptor activation in the MAPK pathway. BRAF and GNAS are direct targets of currently available FDA approved and investigational drugs (Fig. 3c and Supplementary Table S2.)

RTKs and KRAS can also activate the PI3K/AKT/mTOR pathway [14]. This pathway is commonly activated and genetically altered in lung cancer [31, 32]. Here, alterations of several components of the PI3K/AKT/mTOR pathway were found to be present in KRAS-mutant tumors, i.e., PIK3CG, PIK3C2G, TSC2 and mTOR were mutated in 12, 10, 13 and 13% of the tumors with KRAS mutations, respectively (Fig. 2a and Table 1). Consistent with previous studies [32], mutations in individual components of this pathway were found to be almost mutually exclusive (data not shown). Other members of the mTOR pathway, INPP4B and STK11, were found to be altered in 19 and 35% of KRAS-mutant tumors (Fig. 2a and Table 1). PI(3,4)P2 is essential for AKT activation. INPP4B inactivates AKT by dephosphorylating PI(3,4)P2 [33]. Also, alterations in INPP4B can drive tumorigenesis via activation of serum and glucocorticoid-regulated kinase 3 (SGK3) [34]. STK11 is a serine/threonine kinase that functions as a tumor suppressor by regulating TP53-dependent apoptosis pathways [35], as well as by activating TSC1/2 in the PI3K/AKT/mTOR pathway [36, 37] (Fig. 2a). STK11 is commonly mutated in lung cancer [22], and consistent with prior studies [22, 38], STK11 mutations were found to be two times more likely to occur in KRAS-mutant tumors compared to KRAS wild-type tumors (Table 1). On the contrary, STK11 mutations were found to be three times less likely to occur in EGFR-mutant tumors compared to EGFR wild-type tumors. Only 6% of the EGFR-mutant tumors exhibited alterations in STK11 (data not shown). These findings are consistent with the observation that STK11 and KRAS mutations occur predominantly in smokers, while EGFR mutations are more prevalent in tumors from patients without a smoking history [22]. STK11 also seems to have additional effects in KRAS-mutant tumors. Mutations in STK11 have been shown to suppress immune surveillance in KRAS-mutant lung cancers, and to serve as predictors of de novo resistance to immune checkpoint blockade [39, 40]. In summary, at least six mutant genes downstream of RTK activation in the PI3K/AKT/mTOR pathway were identified. At least five of these mutated genes code for proteins that are direct targets of currently available FDA approved and investigational drugs (Fig. 3b and Supplementary Table S2).

Mutations in protein receptor tyrosine phosphatases and STAT3 pathway genes

The receptor-type protein tyrosine phosphatase PTPRT, PTPRD and PTPRS encoding genes were found to be mutated in more than 50% of the KRAS-mutant samples (Fig. 2a and Table 1). Mutations in these tumor suppressors were found to be two times more likely to occur in samples with KRAS mutations compared to tumors with wild-type KRAS (p-value < 0.001, Table 1). On the other hand, these PTP mutations were found to be less likely to be present in samples with EGFR alterations (data not shown). Consistent with previous studies [41, 42], PTPRT was found to be mutated in 17% of all metastatic lung cancer samples, while PTPRD and PTPRS were found to be mutated in 18 and 7% of the samples, respectively. At least one of these PTPs was mutated in 33% of all samples (data not shown). The role of PTPs in lung cancer is yet to be unraveled. However, phospho-STAT3 is a substrate for these phosphatases [43, 44] and frequent loss of function mutations in PTPRT and PTPRD have been linked to hyperphosphorylation of STAT3 in head and neck squamous cell carcinomas (HNSCCs) and in gliomas [45–47]. STAT3 has been found to be hyperactive in some lung cancers [48]. This hyperactivity may be due to enhanced signaling from RTKs, cytokines or JAK kinases (altered in 11% of KRAS-mutant tumors) or by inactivating PTP mutations [45, 49, 50].

Mutations in antioxidant stress response pathway genes

Another tumor suppressor gene that exhibited a high rate of genetic alterations in KRAS-mutant samples was KEAP1 (Fig. 2 and Table 1). Concurrent mutations in KRAS and KEAP1 in lung cancers have been reported to be associated with a more aggressive disease [51]. The KEAP1 gene codes for Kelch-like ECH associated protein 1, which is a negative modulator of NFE2L2, a master transcriptional regulator of anti-oxidant responses. Loss of KEAP1 has been shown to alter the response of KRAS-mutant lung cancer cells to MAPK inhibitors. Krall et al. [52] showed that loss of KEAP1 may abolish increases in reactive oxygen species (ROS) that are associated with MAPK inhibition. They also showed that deletion of KEAP1 may change cell metabolism and enhance the survival and proliferation of lung cancer cells in the absence of MAPK signaling. In another study, Romero et al. [53] found that lung cancer cells with concurrent mutations in KRAS and KEAP1 were dependent on increases in glutaminolysis and, as such, more sensitive to glutaminase inhibition. Here, it was found that KEAP1 was 1.7 times more likely to be mutated in KRAS-mutant compared to KRAS-wildtype tumors, and that 30% of the KRAS-mutant tumor samples carried mutations in KEAP1 (p-value < 0.001 and Table 1). Conversely, it was found that EGFR-mutant tumors were more than two-fold less likely to carry a mutation in KEAP1 (data not shown).

Mutations in DNA replication stress response genes

Replication stress is an attribute of many malignant cells [54]. ATR mediates stress responses by arresting cell cycle progression while stabilizing and repairing the replication fork. ATM and TP53 maintain genome integrity when ATR is defective by arresting cell cycle progression and allowing DNA repair [55, 56]. Here, TP53, ATM and ATR were found to be altered in 51, 22 and 10% of the tumor samples with KRAS mutations, respectively (Table 1 and Fig. 2c). The rate of occurrence of TP53 mutations was less than one would expect by chance in KRAS-mutant tumors (Table 1). ATM and ATR mutations were found to be almost mutually exclusive in KRAS-mutant tumors, and none of the samples exhibited concurrent KRAS, ATM, ATR and TP53 mutations. However, {ATM, KRAS, TP53} mutant and {ATR, KRAS, TP53} mutant subsets were found to be present in 9.5 and 6% of the KRAS-mutant samples, respectively (data not shown). KRAS-mutant samples also exhibited alterations in SETD2 (16%), FANCA (10%) and MRE11 (11%), which encode proteins that regulate DNA mismatch and double-strand break repair. In summary, six genes that are involved in DNA repair and replication stress response mechanisms were found to be mutated in at least 10% of the KRAS-mutant samples.

Mutations in transcription factor and chromatin remodeling genes

KRAS-mutant tumors were also found to carry alterations in the SMARCA (19%), ARIDIA (13%), PBRM1 (10%), PRDM9 (10%), CREBBP (13%), TBX3 (11%), ATRX (14%), MLL2 (14%), MLL3 (14%), TERT (13%), ZFHX3 (13%) and MED12 (16%) genes, which code for transcription factors and proteins involved in chromatin remodeling (Table 1). MED12 promotes tumorigenesis at least in part by regulating the Wnt/β-catenin pathway [57]. Mutations were also found in genes that are important for the formation of the β-catenin/TCF transactivating complex including MLL2, MLL3, CREBBP, TERT and SMARCA. MLL2 and MLL3 are targets of Entacapone, a methyltransferase inhibitor that is used to treat Parkinson’s disease. (Fig. 3d and Supplementary Table S2). SMARCA, ARIDIA and PBRM1 are part of the SWI/SNF chromatin remodeling complex. The genes that code for the subunits of the SWI/SNF complex are altered in more than 20% of all human cancers. Deficiency in one or more of these subunits, including ARIDIA, have been found to be associated with enhanced tumor growth and invasiveness [58]. Taken together, 12 genes coding for transcription factors and chromatin modifying proteins were found to be altered in at least 10% of the KRAS-mutant tumors.

Mutations in E3 ubiquitin ligase genes

Another group of genes that was found to be altered in KRAS-mutant samples codes for E3 ubiquitin ligases. PARK2, TRAF7 and RFWD2 were found to be mutated in 12, 11 and 10% of the samples with KRAS mutations (Table 1). E3 ubiquitin ligases mediate the degradation of many proteins that are important for cancer development [59]. PARK2 deletion or loss of function has been shown to promote inflammation, genomic instability and lung cancer progression [60, 61]. TRAF and RFWD2 can ubiquitinate and promote TP53 degradation, leading to decreased TP53-dependent transcription and, subsequently, apoptosis [62]. In KRAS-mutant tumor samples, mutations in individual E3 ubiquitin ligases were found to be mutually exclusive (data not shown).

Mutations in cadherins and β-catenin genes

Cadherins and catenins are known to control cell-cell adhesion, and deregulation of their expression has been found to be associated with increased metastasis and a poor prognosis of several cancers, including lung cancer [63, 64]. Here, CDH12 and the atypical cadherin FAT1 were found to be altered in 13 and 15% of the KRAS-mutant tumors, respectively (Table 1). Mutations in FAT1 and CDH12 were also found to be mutually exclusive in KRAS-mutant tumors. CTNNA2, the gene coding for α2-catenin, was found to be mutated in 34% of the KRAS-mutant samples (Table 1). KRAS-mutant tumors were two-fold more likely to have a mutation in CTNNA2 compared to KRAS wild-type tumors (Table 1). Mutations in CTNNA2 were less likely to occur in EGFR mutant samples, i.e., only 10% of the EGFR mutant samples exhibited alterations in CTNNA2, and half of these samples carried concurrent CTNNA2, EGFR and KRAS mutations (data not shown). The effects of CTNNA2 dysfunction on lung cancer have not been addressed yet, but frequent mutations in CTNNA2 have been shown to promote the migration and invasion of HNSCC cells [65]. The {KRAS, CTNNA2, FAT1} and {KRAS, CTNNA2, CDH12} mutation subsets were found to be present in 5 and 8% of the KRAS-mutant tumors, respectively. In summary, it was found that cadherins and catenins were collectively altered in at least 40% of the KRAS-mutant tumors.

Analyses of rules involving three or more genes

To understand the complexity of KRAS-mutant tumor subpoptulations, the rules that were generated for the subtypes of cancers with mutations in three or more genes were queried. By doing so, mutations in KRAS were found to be associated with 5724 rules (Fig. 1b). When the query to rules was limited to at least three genes that were present in at least 10% of the KRAS-mutant tumors, 160 sub-rules were generated (Fig. 4a and Supplementary Table S3). Most (79%) of these rules had mutations in three genes. Many of these rules were associated with mutations in PTPRs (PTPRT, PTPRD and PTPRS), TP53, CTNNA2 and the pseudogene LOC1750 (Fig. 4a). A high degree of “overlap” was noted between the rules. For example, all 17 rules with CTNNA2 mutations on the right-hand side of the rule were similar to 95% of the rules with LOC1750 mutations on the right-hand side (Supplementary Table S3).

Mutations in PTPRs can lead to constitutive activation of STAT3, which has been associated with lung cancer [45–47]. PTPRs were found to be collectively altered in at least 50% of the KRAS-mutant tumors (Table 1). A concurrent mutation in KRAS may lead to a tumor exhibiting hyperactive MAPK, AKT and/or STAT3 signaling [23, 24]. The subsets of tumors with KRAS and PTPR mutations were subsequently assessed to identify other pathways that might be defective. This effort resulted in {KRAS, PTPRT, SMARCA4}, {KRAS, PTPRT, MED12} and {KRAS, PTPRT, MLL2} mutant subsets in 11.9, 11.3 and 10.7% of the KRAS-mutant tumors, respectively (Fig. 4b). In addition to hyperactivity of the MAPK, AKT and STAT3 pathways, additional mutations in SMARCA4, MED12 or MLL2 may hint at the Wnt/β-catenin pathway and transcriptional (de) regulation by the SWI/SNF complex [57, 58]. Other subsets of tumors with KRAS and PTPRT mutations were {KRAS, KEAP1, PTPRT} and {CTNNA2, KRAS, PTPRT} occurring in 14 and 21.4% of the KRAS-mutant tumors, respectively. Concurrent mutations in KRAS, KEAP1 and PTPRT can lead to tumors with hyperactive MAPK, AKT and STAT3 pathways and defective anti-oxidant response mechanisms [52, 53], while mutations in CTNNA2, KRAS and PTPRT could lead to tumors with defective cell-cell adhesion in addition to constitutive activation of the MAPK, AKT and STAT3 pathways [65].

STK11 alterations have been reported to be associated with a poor prognosis in lung cancer [38]. There are, however, conflicting reports on the effects of alterations in STK11 in KRAS-mutant tumors. Some researchers found that concurrent STK11 and KRAS mutations did not affect the aggressiveness of the disease [51, 66], whereas others found that STK11 can potentiate tumor progression in KRAS-mutant tumors in mouse models [67] and that mutated STK11 may be associated with a more aggressive disease [39]. This discrepancy may be due to the presence of other alterations in subpopulations of KRAS^mut/STK11^mut lung cancers. Here, about 66% of the KRAS^mut/STK11^mut tumors were found to also exhibit alterations in KEAP1, which regulates oxidative stress responses [52, 53]. In addition, it was found that STK11^mut/PTPRs^mut and STK11^mut/CTNNA2^mut genotypes were present in about 35 and 33% of the tumors with KRAS mutations, respectively (Fig. 4c). Such combinations of mutations may affect disease biology and, thus, lead to different clinical outcomes. The design of pre-clinical models that capture such combinations may enhance our understanding of the etiology of lung cancer.

Discussion

Lung cancer is characterized by the presence of mutations in genes coding for oncogenic and tumor suppressive proteins [22]. For patients with KRAS mutations, there are currently no FDA approved treatment options. Efforts aimed at targeting KRAS downstream effectors have seen little or no benefits. A probable reason for the lack of success in targeting these effectors may be the presence/activation of other cell survival pathways and mechanisms. Using association rule mining, 67 genes were identified that were found to be mutated in at least 10% of the samples with KRAS mutations. More than 30 of the proteins encoded by these genes turned out to be direct targets of FDA approved or investigational drugs. Using Fisher’s exact test, more than 50% of the associations between KRAS and these genes were found to be statistically significant. Genes whose mutations were previously associated with KRAS-mutant tumors were found to be clinically relevant. For example, lung cancer patients with a smoking history are known to be more likely to have mutations in KRAS, STK11 and KEAP1 [22]. As expected, mutations in STK11 or KEAP1 were more likely to be found in tumors with KRAS mutations (p-value < 0.001). Also, tumors that exhibited mutations in KRAS and KEAP1 were four times more likely to also have a mutation in STK11. On the other hand, EGFR alterations which are known to occur mainly in non-smokers [22] were found to be less likely to occur in KRAS-mutant tumors (p-value = 1.00).

In addition, several novel associations between alterations in KRAS and other genes were found. Genes coding for protein receptor tyrosine phosphatases, which dephosphorylate and inactivate STAT3 [43, 44], were more likely found to be mutated in KRAS mutant tumors. Previously, somatic mutations or promoter hypermethylation of PTPRT and PTPRD have been associated with increased STAT3 activity in HNSCCs and gliomas [45–47, 68]. The high rate of occurrence of PTP alterations in KRAS-mutant lung cancers may explain the STAT3 hyperactivity observed in some NSCLCs [48]. Cancer cells harboring loss of function mutations in PTPRT and PTPRD are highly sensitive to STAT3 inhibition [45]. These findings suggest that PTP mutations may serve as biomarkers for KRAS-mutant lung cancer patients that may benefit from STAT3 inhibitors.

MAPK is a downstream effector of KRAS, and drugs targeting the MAPK pathway have been shown to be ineffective in KRAS-mutant lung cancers. Inhibition of MEK in, for example, epithelial-like and mesenchymal-like KRAS-mutant tumors has been shown to be ineffective, which may be attributed to a bypass of KRAS/BRAF/MEF by ERBB3 and FGFR1 receptors to activate MAPK. In patient-derived xenografts containing mesenchymal-like cells, simultaneous treatment with BRAF and FGFR inhibitors has been found to result in significant reductions in tumor growth [69, 70]. In another study, inhibition of the MAPK pathway in KRAS-mutant tumors that harbor loss of function KEAP1 mutations were found to result in cells that were heavily dependent on glutaminolysis for their survival. Here, KEAP1 mutations were found to be 1.7 times more likely to occur in KRAS-mutant tumors. Romero et al. [53] showed that concomitant blockade of MAPK and glutaminolysis significantly abrogated the proliferation of KRAS-mutant lung cancer cells. These examples indicate that targeting KRAS downstream effectors and their co-occurring cell survival mechanisms in KRAS-mutant lung cancers may serve as promising approaches to ensure better clinical outcomes.

In summary, association rule mining was employed to identify subpopulations of metastatic KRAS-mutant lung cancer patients based on combinations of somatic mutations in genes coding for oncogenic and/or tumor suppressive proteins. The design of pre-clinical models that capture these subpopulations may lead to the development of novel effective combination therapies and evasion of potential resistance mechanisms in KRAS-mutant metastatic lung cancers.

Compliance with ethical standards

Ethics approval and consent of participants

Not applicable.

Consent for publication

Not applicable.

Competing interests

None declared.