24h-gene variation effect of combined bevacizumab/erlotinib in advanced non-squamous non-small cell lung cancer using exon array blood profiling

The gene expression data set originated from a translational substudy of the phase II SAKK 19/05 trial from the Swiss Group for Clinical Cancer Research (ClinicalTrials.gov: NCT00354549). In the original study, 103 unselected patients with advanced non-squamous NSCLC were enrolled, among which 101 were evaluable. The experimental design of the trial SAKK 19/05 is summarized in Fig. 1. Patients were treated using the combined targeted therapy BE until disease progression or unacceptable toxicity. At progression, standard platinum-based chemotherapy (CT) was used. The primary endpoint of this trial was disease stabilization 12 weeks after initiation of therapy. Secondary endpoints included tumor shrinkage at 12 weeks (TS12), time to progression under BE (TTPBE), time to progression under CT (TTPCT) and OS. Further detailed information about this trial can be found in previous publications [11, 17, 18]. As part of a translational substudy, blood samples were taken at baseline and 24h after initiation of treatment for gene expression analysis in a subset of 49 patients. The current study was approved by the ethics committee of the canton St. Gallen (EKSG 06/012).

RNA from whole blood samples was extracted and quality checked. Six pairs of samples had to be excluded from the analysis due to low quality, whereas RNA extracts provided sufficient quality for microarray hybridization in 43 out of 49 pairs of sample. Messenger RNAs were hybridized on Affymetrix Human Exon 1.0 ST arrays (Affymetrix, Santa Clara, CA, USA) following standard recommendations from the manufacturer. This microarray platform measures genome-wide exon-level expression in over 1.4 million probe sets, and allows the investigation of genomic variations both at the gene and at the exon level. For the sake of this analysis, 439,778 exonic probe sets (within 38,900 genes) were kept in the analysis, after filtering out intronic, intergenic and unreliable probe sets (according to the nomenclature defined in the R package annmap [19]). Raw data (Affymetrix CEL files) have been deposited in NCBI’s Gene Expression Omnibus (GEO), and are accessible through GEO Series accession number GSE61676. The exon level probe sets were pre-processed, quality checked and normalized using the RMA procedure (including background correction, quantile normalization and median-polish summarization) as implemented in the R package oligo [20, 21].

The current experimental design includes a repeated measurement of genome-wide exon-level expressions in 43 patients at two different time points (baseline and 24h after initiation of therapy). The exon expression data are stored in a pair of fully matched tables. Dually constrained correspondence analysis (DCCA)—an extension of the multivariate technique correspondence analysis—was used to investigate the 24h-change in exon expression levels [16]. DCCA uses observation- and variable-wise linear constraints in order to take into account complex experimental designs including within-patient repeated measurements and variables grouped into hierarchical levels (within-gene exonic structures). The theoretical scheme of DCCA as applied to our experimental design is summarized in Fig. 2. The mathematical underpinnings of DCCA are further described in Additional file 1. In the current work, the following parametrization of DCCA was used: (1) the two fully matched tables of exon-level expression intensities at baseline and 24h were stacked observation-wise; (2) an observation-wise between-time constraint (baseline vs. 24h) was applied after partialling out the within-patient effect; (3) a variable-wise constraint indicating the within-gene exonic structure was applied.

Generalized linear (mixed effects) models were used in order to test the predictive value of the identified biomarkers. Binary endpoints were tested using logistic regression, time to event data were modeled using Cox proportional hazards regression, and continuous variables were modeled using (multiple) linear regression. Mixed effects modeling was used when testing associations in the frame of the within-patient repeated measurement design. The adjustment for patients’ EGFR mutational status was done by including the mutational status as covariate in the predictive models. The significance level used for the discovery of the putative predictive markers was set to 0.001. Gene signatures combining the information of the best predictive candidates were built using the metagene approach [22]. In this approach, the linear combination of several genes is calculated as follows: (1) The matrix of normalized gene expression intensities for all patients at a given time point (rows) and for all candidate genes (columns) is analyzed using unscaled principal component analysis (PCA); (2) The row coordinates (scores) on the first PCA axis summarizing the largest amount of variance are extracted (metagene score); (3) Based on the median of the metagene score, a binary score is created to discriminate between low versus high-risk patients.

All analyses were implemented using the R statistical software [23] including the extension package ade4 [24], as well as dedicated packages from the Bioconductor project [25] such as the package oligo for microarray preprocessing [21] and the annotation package annmap (Homo sapiens database version 86). Functions available in ade4 (including dudi.coa, wca, bca, pcaiv and pcaivortho) can theoretically be used to carry out DCCA. However, for reasons of computational efficiency due to the extensive size of exon array datasets, the DCCA algorithm was substantially optimized and a new R extension package dcca is available (see Additional file 1). Hypothesis testing for the identification of predictive biomarkers was carried out using the following R packages: lme4, coxme and multcomp.

Gene set enrichment analysis was done by interrogating the molecular knowledge databases Kyoto Encyclopedia of Genes and Genomes (KEGG) [26] and WikiPathways [1] using the functional annotation web-service WebGestalt [27]. Enrichment analyses were based on the list of 100 most dysregulated genes (as identified by DCCA), as well as the lists of genes which were significantly predicting patient’s outcome (each of the investigated endpoints). The significance of the enrichment was obtained using hypergeometric tests. Validation was carried out using the lung data set from the Kaplan–Meier Plotter (KMplotter) web tool [28]. KMplotter is a manually curated database including gene expression level information about more than 50,000 Affymetrix probe set IDs together with associated clinical information. The prognostic value of single or multiple genes can be assessed with regard to relapse free and overall survival. Another independent lung cancer dataset was used for external validation. This gene expression microarray dataset includes 85 lung adenocarcinoma tumor samples and is part of the program “Carte d’Identité des Tumeurs” (CIT) from the french national cancer league [29]. Samples were profiled using the Affymetrix Human Genome U133 Plus 2.0 Array and raw data are available in NCBI’s Gene Expression Omnibus through GEO Series accession number GSE30219. Furthermore, the results were discussed in the light of available literature findings.

The characteristics of the 43 patients are reported in Table 1. Patients were late stage (91% stage IV/9% stage IIIb) non-squamous NSCLC. Five out of 43 patients had demonstrable EGFR mutations: one on exon 18 (E709A-G719S), three on exon 19 (Del L747-G749, Del E746-A750 and R748-S752) and one on exon 21 (L858R). The median age was 61 years old (IQR 54–66) and the sex ratio was 0.44 (19 males/24 females). Disease stabilization at 12 weeks was reached in 53% of patients. The median tumor shrinkage at 12 weeks was 15.8%. The median overall survival was 11.1 (95% CI 10.1–17.9) months. The median time-to-progression under BE was 4.0 (95% CI 2.8–6.0) months, whereas the median time-to-progression under CT was 2.6 (95% CI 1.7–5.7) months.

The 100 genes mostly dysregulated (54 up-regulated vs. 46 down-regulated) by the 24h effect of BE are summarized in Table  2. The genes were involved in all aspects of tumor biology. This includes genes involved in mitosis and cell cycle processes such as the cancer susceptibility candidate 5 (CASC5) encoding for a protein influencing the spindle assembly checkpoint during eukaryotic cell cycle; the centromere-associated protein E (CENPE) which accumulates and play a key stabilizing role during mitosis; the protein furry homolog (FRY) which plays a crucial role in the structural integrity of mitotic centrosomes; kinetochore associated 1 (KNTC1) encoding for a protein ensuring proper chromosome segregation during cell division; Phospholipase D1 (PLD1) involved in cancer progression [30] and in the regulation of mitosis in relationship with the “Ras signaling pathway” and “Pathways in cancer”. Other dysregulated genes were involved in energy-dependent metabolisms (ATPase/GTPase). This includes genes from the ATPase Family (ATAD2) known to be related to gastric cancer network, and which may play an important role in cell proliferation and cell cycle progression of breast cancer cells; Guanylate Binding Protein 4 (GBP4) related to GTPase activity and associated with the interferon signaling pathway; dedicator of cytokinesis 10 (DOCK10) acting on GTPase and related to hemostasis and regulation of cell division cycle 42 (CDC42) activity. Mechanisms of cell migration are also controlled by a series of dysregulated genes such as ADAM Metallopeptidase Domain 19 (ADAM19) regulating cell migration, cell adhesion and cell-cell/cell-matrix interactions, supposed to play an important role in pathological processes including cancer; BMX non-receptor tyrosine kinase (BMX) encoding for a protein implicated in several signal transduction pathways regulating tumorigenicity of cancer cells; Epidermal Growth Factor Receptor Pathway Substrate 8 (EPS8) encoding for a protein having functions in part of the EGFR pathway and being related to Tyrosine Kinases/Adaptors and Development FGFR signaling pathways; Fms-Related Tyrosine Kinase 3 (FLT3) encoding for a class III receptor tyrosine kinase regulating hematopoiesis, and whose action is related to apotosis, proliferation and differentiation processes; Interleukin 1 Receptor, Type II (IL1R2) controling many cellular functions including proliferation, differentiation, and cell survival/apoptosis; Matrix Metallopeptidase 9 (MMP9) involved in tissue/matrix remodeling, playing a central role in cell proliferation, migration, differentiation, showing an altered expression in a number of different human cancers with poor prognosis. Apoptosis was regulated through the action of various genes including baculoviral IAP repeat containing 3 (BIRC3) encoding for an inhibitor of apoptosis protein acting on killing tumor cells; death-associated protein kinase 2 (DAPK2) whose overexpression was shown to induce cell apoptosis; Retinoblastoma-Like 1 (RBL1) encoding for a tumor suppressor protein involved in cell cycle regulation; insulin-like growth factor 1 receptor (IGF1R) encoding for a growth factor with tyrosine kinase activity, having an anti-apoptotic effect and being highly overexpressed in most malignant tissues [31].

Based on the 100 best candidates identified by DCCA, three main pathways (according to the KEGG functional annotation database) were significantly altered by the 24h effect of the combined therapy BE: Hematopoietic cell lineage (KEGG pathway hsa04640; \(p=0.0094\)); ABC transporters (KEGG pathway hsa02010; \(p=0.0085\)); Pathways in cancer (KEGG pathway hsa05200; \(p=0.0204\)).

Figure 3 displays the gene expression levels of each genes (baseline and 24h) within these dysregulated pathways. Figure  4 shows the detailed KEGG’s “Pathways in Cancer” highlighting the down-regulated genes within the sub-pathways “Cytokine–cytokine receptor information” and “MAPK signaling pathways”. The down-regulation of these sub-pathways contributes to the limitation of cell proliferation.

These pathway dysregulations are well in line with the expected effects of both erlotinib (“Pathways in cancer”; “ABC transporters”) and bevacizumab (“Hematopoietic cell lineage”). “Pathways in cancer” is a generic pathway, including genes involved in various aspects of tumorigenesis such as phenomena of proliferation, invasion, resistance and apoptosis. Several diseases are inter-related to this pathway, including non-small cell lung cancer. Most genes from the “Pathways in cancer” (FLT3, IGF1R, DAPK2, PLD1 and MMP9) are inhibited due to the action of BE resulting in an anti-proliferative and pro-apoptotic action of the combined therapy, with the notable exception of the up-regulation of the apoptosis inhibitor BIRC3. The combined action of BE results in the inhibition of all genes that belong to the “Hematopoietic cell lineage” (FLT3, IL1R1, IL1R2, MME). Hematopoietic stem cells play important roles for angiogenesis [32]. The down-regulation of genes within the pathway “Hematopoietic cell lineage” may be related to the specific anti-angiogenic action of bevacizumab [33, 34]. On the other hand, the dysregulation of genes that belong to the pathway “ABC transporters” is probably associated with the energy-related mechanisms of action of erlotinib [35, 36].

Gene set enrichment analysis based on the WikiPathways database provides information on additional cancer- and energy-related activated pathways including “gastric cancer network 2” (WP2363, \(p=0.010\)), “IL1 megakaryocytes in obesity” (WP2865, \(p=0.006\)), “apoptosis modulation and signaling” (WP1772, \(P=0.011\)) and “gastric cancer network 1” (WP2361, \(p=0.009\)).

The predictive value of the genes that were mainly dysregulated due to the 24h effect of BE was investigated. The following endpoints were considered: disease stabilization at 12 weeks, tumor shrinkage at 12 weeks, time to progression under BE, time to progression under chemotherapy, and overall survival. The magnitude of the 24h change in expression of the 100 most dysregulated genes was not significantly associated with any of the investigated endpoints, after adjustment for the patients mutational status.

Table 3 summarizes the predictive value of the blood gene expression at baseline and 24h after initiation of BE. Putative markers at baseline that predicted patient’s overall survival included Cancer susceptibility candidate 1 (CASC1). At baseline, 142 genes were identified as putative predictive markers of tumor shrinkage at 12 weeks. Among those genes, there was a significant enrichment of the KEGG signaling pathways “Phagosome” (hsa04145) and “Protein digestion and absorption” (hsa04974).

Putative predictive markers from blood gene expression at 24h after initiation of treatment of the TTPBE included four genes enriched in the pathway “Pathways in cancer”. These genes were the E2F transcription factor 1 (E2F1), RAD51 recombinase (RAD51), the junction plakoglobin (JUP), and the microphthalmia-associated transcription factor (MITF). E2F1 plays a critical role in the control of cell cycle and acts as a tumor suppressor. E2F1 was found to be associated with phenomena of resistance of targeted therapy in breast cancer [37]. There was also a significant enrichment in the signaling pathway “Pathways in cancer” among the predictors of TS12 at 24h: genes frizzled class receptor 5 (FZD5) and epidermal growth factor (EGF). EGF encodes for a protein playing an important role in the cell growth, proliferation and differentiation. It binds with high affinity epidermal growth factor receptor. Its dysregulation has been associated with cancer progression [38]. Other pathways associated with 24h predictive markers of TTPBE included the cancer-related pathways “Melanoma”, “Pancreatic cancer”, “PPAR signaling cancer”, as well as the metabolism-related pathways “arginine and proline metabolism” and “Pyrimidine metabolism”.

All putative predictive markers of TTPBE at 24h were combined into a 91-gene metagene. Patients could be significantly classified into low-risk versus high-risk according to their median metagene score (HR 4.93 [95% CI 2.34 to 10.39], log-rank test: \(p < 0.001\)) (Fig. 5, left panel). The median TTPBE were 2.46 (95% CI 1.54–3.22) months vs. 6.87 (95% CI 4.14–13.31) months in the high-risk and low-risk populations, respectively. This finding was successfully validated using the KMplotter web tool (Fig. 5, central panel) and the external CIT validation dataset (Fig. 5, right panel). The predictive value of the metagene remained significant after adjusting for the patient’s mutational status (Cox proportional hazards regression after adjustment for the mutational status: HR 2.63 [95% CI 1.87–3.70], \(p < 0.001\)). An illustration of selection of responders based on the metagene score is provided in the Results section of the Additioanl file 1.