Gene Expression of the EGF System-a Prognostic Model in Non-Small Cell Lung Cancer Patients Without Activating EGFR Mutations.

OBJECTIVES: Contradicting results have been demonstrated for the expression of the epidermal growth factor receptor (EGFR) as a prognostic marker in non-small cell lung cancer (NSCLC). The complexity of the EGF system with four interacting receptors and more than a dozen activating ligands is a likely explanation. The aim of this study is to demonstrate that the combined network of receptors and ligands from the EGF system is a prognostic marker.
MATERIAL AND METHODS: Gene expression of the receptors EGFR, HER2, HER3, HER4, and the ligands AREG, HB-EGF, EPI, TGF-α, and EGF was measured by quantitative polymerase chain reaction in tumor samples from 100 NSCLC patients without EGFR activating mutations. Results were dichotomized into high or low levels of target expression. Coexpression of the ligands and receptors was observed, and a score was developed based on the summed effect of receptors and ligands. Akaike's information criteria selected the optimal score. Results were correlated with age, sex, stage, histology, performance status, and overall survival.
RESULTS: Patients were randomly split 1:1 to create test and validation cohorts. In multivariate analyses, the only individual prognostic marker was EPI (hazard ratio [HR] 0.38 [0.20-0.72], P = .003). The optimal score in the test cohort was validated as a marker of inferior survival in the validation cohort and by bootstrapping. Multivariate analysis confirmed the combined score as a prognostic marker of inferior survival (HR 3.75 [2.17-6.47], P < .00001).
CONCLUSION: Our study has developed a model that takes the complexity of the EGF system into account and shows that this model is a strong prognostic marker in NSCLC patients.

Despite advances in the treatment, non–small cell lung cancer (NSCLC) remains the leading cause of cancer-related death in the world [1]. In particular, the overall prognosis is poor for the metastatic stages, with a median overall survival (OS) of only 8 to 10 months. Even in the early nonmetastatic stages, the 5-year survival rate is as low as 50% [2], [3]. Prognostic markers are needed to stratify patients with different risk outcome. Several biomarkers have been evaluated in NSCLC, but only a few have proven to be clinically relevant. An activating mutation in the epidermal growth factor receptor (EGFR) is both a well-described predictive marker of benefit of EGFR-targeted tyrosine kinase inhibitors but also a debated prognostic marker of better OS [4], [5], [6], [7], [8], [9]. As EGFR expression has been associated with OS in head and neck, colorectal, and esophagus cancer [10], [11], [12], attention has been directed toward the use of EGFR expression as a prognostic marker in NSCLC, but contradicting results have been demonstrated [13], [14], [15], [16]. The EGF system is complex, and the effect of ligand-receptor interaction depends on a variety of different factors, which provides a plausible explanation for the divergence observed between studies that only evaluate EGFR expression in general. EGFR is one out of four related receptors from the EGF system and is capable of forming homodimers or heterodimers with one of the three other receptors when activated by a ligand. Several ligands from the EGF system such as amphiregulin (AREG), epidermal growth factor (EGF), and transforming growth factor–α (TGF-α) only activate EGFR, whereas some have the ability to activate several combinations of the four EGF receptors like heparin-binding epidermal growth factor (HB-EGF), epiregulin (EPI), and betacellulin (BCL). Most knowledge on the role of the ligands in NSCLC is from in vitro studies or from smaller clinical studies. In vitro studies have suggested that the biological effect of EGFR activation is dependent on the specific activating ligand as well as the dimerization partner [17]. Yet, no clinical studies have evaluated the effect of the network of receptors and ligands influencing EGFR in NSCLC. Furthermore, the majority of the clinical studies exploring EGFR expression are based on immunohistochemistry which is a semiquantitative method with a great risk of interobserver variability. Quantitative gene expression analyses provide a more accurate measure and are therefore more suitable for studies comparing expression levels. Prospectively, we have collected fresh tumor samples from patients suspected of lung cancer. Accordingly, the aim of this study is to evaluate the gene expression of the network of receptors and ligands of the EGF system affecting EGFR as a prognostic markers in NSCLC.

Materials and Methods

Patients and Tumor Samples

In this study, 1093 patients referred to the Department of Pulmonary Medicine, Aarhus University Hospital, Denmark, for diagnostic workup of lung cancer were included consecutively from April 2011 until January 2013. Patients with NSCLC without mutations in EGFR defined the NSCLC group used in this study. Patients with other types of cancer than NSCLC were excluded. To establish normal values of gene expression, a reference group was made of patients without cancer, as they had comparable clinical features (age, smoking history, clinical symptoms, and anamnestic symptoms leading to lung cancer suspicion). Patient selection is depicted in Figure 1. Clinicopathological characteristics were collected at time of inclusion. To minimize the risk of positive overestimation when establishing a score, the NSCLC group was randomly split 1:1 into a test and a validation cohort matched on histology, stage, WHO performance status (PS), and sex.

Figure 1

Patient selection.

All patients gave informed written consent before inclusion, and the study was approved by the Central Denmark Region Committees on Biomedical Research Ethics (M-20100246).

Tumor samples were collected at time of diagnosis with the Ultra-micro sampling technique, allowing an aliquot of the diagnostic sample to be used for gene expression studies [18]. Samples were stored in RNAprotect at − 80°C until total RNA extraction.

RNA Extraction and cDNA Preparation

Total RNA was extracted from all tissue samples using the RNeasy micro kit from Qiagen, (Qiagen, Hilden, Germany; www.qiagen.com) according to the instructions of the manufacturer. Purified RNA was diluted in 14 μl of RNase-free H2O. cDNA was synthesized from 2 μl of total RNA in a 20-μl reaction, following manufacturer’s instructions (Applied Biosystems Inc., CA).

Quantitative Polymerase Chain Reaction (PCR)

Expression of EGFR, HER2, HER3, HER4, AREG, HB-EGF, EPI, TGF-α, EGF, and the reference gene β-actin was quantified using the LightCycler 480 instrument. The PCR was performed using SYBR Green I Master Mix Buffer except HER2 that was analyzed by hydrolysis probes (SYBR Green RT-PCR Reagents Kit, Roche, Mannheim, Germany) following manufacturer’s instructions. Primer sequences, concentrations, and annealing temperature are shown in Table S1. All primers were purchased from Eurofins MWG Synthesis GmbH, Ebersberg, Germany. In an earlier study, we found that, in these samples, β-actin was the most stable household gene [18]. Therefore, each target gene was normalized to the household gene β-actin and used for further analysis.

Cut Points

A cut point was determined in the reference group for each target gene. The cut point was defined by the mean + 2 SD for the individual target gene, and expression levels above the cut points were defined as elevated.

The EGF Score Model

To evaluate the EGF system as a network, a score based on the summed effect of each receptor and ligand was established. Each ligand and receptor were given the value of + 1 or − 1 based on the individual markers’ hazard ratio (HR) in the univariate analysis in the test cohort [19]. If the HR was greater than 1, the value given was + 1, and if the HR was less than 1, the value was − 1. The sum of all markers created the score. The score was then confirmed by bootstrapping (1000 times) to create bias-corrected confidence intervals (BCIs). Redundancy of one or several markers was tested to optimize the score. To select the score with the best model fit to explain the data, Akaike’s information criteria (AIC) were calculated in the test cohort for every combination excluding one or several of the markers. When the difference of the AIC values was larger than 2, the score with the lowest AIC was selected. If the difference of the AIC was less than 2, the score with the least target genes was selected [20]. Based on the calculated AIC, the score with the best model fit in the test cohort was chosen and validated in the validation cohort. By bootstrapping (1000 times), the optimal score was validated. After validation, a multivariate analysis was carried out in the entire NSCLC group taking clinicopathological characteristics into account.

Statistical Analysis

Associations between clinical characteristics were tested by Person χ2 or t test. Survival was calculated from date of diagnosis until date of death or last day of follow-up (May 1, 2015). OS was defined as death of any cause or last date of follow-up. OS was estimated by the Kaplan-Meier method and compared by the log-rank test. Uni- and multivariate HRs were estimated by Cox proportional hazard analyses. HR is presented with a 95% CI. All P values are two-sided with a 5% significance level. The proportional hazard assumption was assessed graphically by plotting Schoenfeld residuals and log-log plots. Bootstrapping was used for validation of estimated HR. Bias in the bootstrap-derived calculations was used as an indicator of robustness of estimates. To identify effect modification, interaction terms were established and added to the regression model. Backward selection was used to identify the multivariate model with best fit. Variables were stepwise removed until only significant variables remained (P < .10). Model fit was evaluated by AIC. Statistical analyses were carried out in STATA version 13.

Results

Patient Characteristics

Clinical and pathological characteristics are detailed in Table 1. One hundred patients with NSCLC without EGFR mutations and 111 patients without cancer were included. From the NSCLC patients, a test cohort (n = 50) was randomly generated with a similar distribution of age, stage, PS, and histologic subtype as the total cohort. In the reference cohort, the mean age was lower (60.7) than in the cancer cohorts (67.5) and the distribution of never smokers versus current or former was significantly different. Furthermore, in the reference cohort, more samples originated from lymph nodes compared with the primary tumor than in the cancer cohorts.

Table 1

Clinical Characteristics

	Test Cohort n = 50	Validation Cohort n = 50	P Values*, †	Reference Cohort n = 111 (282 Samples)	P Values*, ‡
Agemean (min-max)	66.4(37-86)	68.7(50-86)	.21	60.7(21-84)	< .001
Sex n (%)
Female	30 (60)	20 (40)	.05	51 (46)	.51
Male	20 (40)	30 (60)		60 (54)
Smoking
Never	2	1	.23	28	< .001
Former	26	32		59
Current	22	17		24
PS n (%)
0	21 (42)	15 (30)	.24
1	17 (34)	27 (54)
2	10 (20)	7 (14)
3	2 (4)	1 (3)
Stage
I & II	11 (22)	9 (18)	.44
III & IV	39 (78)	41 (82)
Histology
Squamous	15 (30)	12 (24)	.50
Nonsquamous	35 (70)	38 (76)
Sample origin
Tumor	18 (36)	11	.12	34	.02
Lymph node	32 (64)	39		248

P values calculated by Person χ2 or t test.

Test cohort versus validation cohort.

Reference cohort versus test and validation cohorts.

The EGF System

Reference group: 282 samples each representing individual lesions from 111 patients were examined, and a cut point for each target gene was calculated as described above. Because this part of the study was carried out to characterize the background expression of the target genes, all available samples from each patient were included. Cut points are shown in Table 2.

Table 2

Characteristics of the EGF System

		Univariate			Multivariate*			Score Value	Nonsquamous (n = 73)		Squamous (n = 27)
									Stage I & II(n = 12)	Stage III & IV(n = 61)N (%)	Stage I & II(n = 7)N (%)	Stage III & IV(n = 20)N (%)
Biomarker	Cut Point	HR	CI	P Value	HR	CI	P Value	Initial Score	N† (%)	N† (%)	N† (%)	N† (%)
HER1	0.20	0.98	0.48-2.00	.754	1.06	0.65-1.72	.826	− 1	1 (8)	24 (39)	3 (42)	5 (25)
HER2	6.40	1.25	0.38-4.09	.707	1.97	0.77-4.99	.155	+ 1	0	4 (7)	1 (14)	0
HER3	0.12	0.91	0.57-1.46	.710	1.01	0.50-2.07	.959	− 1	0	9 (15)	1 (14)	2 (10)
AR	1.03	2.03	0.93-4.44	.077	1.34	0.81-2.22	.254	+ 1	0	19 (31)	0	6 (30)
EPI	0.25	0.56	0.31-1.02	.060	0.38	0.20-0.72	.003	− 1	1 (8)	16 (26)	0	2 (10)
HB-EGF	1.20	1.92	1.11-3.31	.018	1.86	0.96-3.59	.065	+ 1	0	8 (13)	2 (29)	9 (45)
TGFα	0.20	0.29	0.07-1.20	.086	0.59	0.27-1.32	.203	− 1	1 (8)	8 (13)	1 (14)	2 (10)
EGF	0.35	0.22	0.05-0.91	.037	0.99	0.55-1.78	.991	− 1	1 (8)	14 (23)	3 (43)	2 (10)
The initial score‡		2.31	1.26-4.56	.008	2.17	1.11-4.22	.02
The optimized score‡		3.70	2.26-8.08	.00001	2.62	1.49-4.61	.001		0 (0)	12 (20)	1 (14)	8 (40)

Multivariate results are adjusted for age, sex, PS, stage, and histology.

Number of patients with upregulated biomarker.

Results are presented with bias-corrected interval.

NSCLC group: Level and distribution of each target gene are shown in Figure 2, Figure S1 and in Table 2. For the receptors EGFR and HER3 and the ligands, elevated levels were observed in both nonsquamous and squamous cell type and predominantly in the advanced stages. For EPI and HB-EGF, the distribution between histological types was not equal. EPI was elevated in 7% of patients with squamous cell histology (2 of 27) as opposed to 23% (17 of 73) in patients with nonsquamous histology. For HB-EGF, the corresponding frequencies were 29% (8 of 27) in squamous cell and 15% (11 of 73) in nonsquamous cell. HER4 was hardly measurable in the reference group, and only one patient had elevated HER4 mRNA. Therefore, HER4 was not included in further analyses. HER2 was found in 72 of 100 samples, and yet, only 5 patients had samples with elevated expression compared with the reference group.

Figure 2

HER1/β-actin levels. Dashed line indicates the cutoff value defined in the reference group. Patients are split by stage: I and II versus III and IV.

Figure S1

Expression levels of HER2/β-actin, HER3/β-actin, AREG/β-actin, HB-EGF/β-actin, EPI/β-actin, TGF-α/β-actin, and EGF/β-actin. Dashed lines indicate the cutoff value. Patients are split by stage: I and II versus III and IV.

Survival Analyses

In a univariate Cox proportional hazards model, gene expression of all target genes was analyzed for the potential of being an individual prognostic marker (Table 2). This demonstrated that overall EGFR gene expression was not a prognostic marker (HR 0.98 [0.48-2.00]; P = .75) as opposed to EGF (HR 0.35 [0.05-0.91]; P = .037) and HB-EGF (HR 1.2 [1.11-3.31]; P = .018). However, neither of the associations was significant when tested in a multivariate analysis taking known prognostic clinicopathological characteristics into account. For EPI, the opposite was found. In a multivariate analysis, high gene expression levels of EPI were found to be a marker of better prognosis (HR 0.38 [0.20-0.72]; P = .003). None of the remaining receptors were significant prognostic markers when analyzed individually.

Score Development

Values creating the initial score are listed in Table 2. The initial score was tested and validated by bootstrapping demonstrating that the score could predict an inferior OS (HR 2.31, BCI [1.26-4.56]). To optimize the initial score, it was tested if one or several of the markers were excessive. Therefore, different combinations were tested in the test cohort. Based on the AIC, the score with the optimal model fit only included the ligands AREG, HBEGF, EPI, and TGF-α. In the test cohort as well as in the validation cohort, the optimized score was a prognostic marker of inferior OS (test: HR 6.17, 95% CI [2.60-14.64], P > .001, BCI [1.94-18.89]; validation: HR 2.41, 95% CI [1.23-4.71], P = .01, BCI [1.92-18.61]) (illustrated in Figure 3). In the entire NSCLC group, the optimized score was tested in a multivariate analysis adjusting for the clinical parameters age, sex, PS, stage, and histology. The optimized score was a significant prognostic marker of inferior outcome (HR 2.93 [1.64-5.23], P < .0001, BCI [1.68-4.97]) as seen in Table 3. Median survival for patients with a high score was 4.7 months (95% CI: 2.9-8.3) versus 15.9 months (95% CI: 11.5-22.0) for patients with a low score.

Figure 3

Kaplan-Meier Survival curves for OS in the test cohort (A) and the entire cohort (B). Differences between groups were calculated using the log-rank test.

Table 3

Multivariate Cox Proportional Hazard

	HR	95% CI	P Value	HR*	95% CI*	P Value*	BCI
EGF score	3.52	2.00-6.20	< .001	3.75	2.17-6.47	< .0001	2.09-6.45
Age: ≤ 67 vs > 67	1.28	0.80-2.04	.304
Sex: female vs male	0.85	0.54-1.32	.462
PS: 0, 1 vs 2, 3	1.89	1.06-3.37	.030	1.97	1.14-6.47	.014	1.02-4.01
Stage: I & II vs III & IV	4.16	1.92-9.03	< .001	3.63	1.71-7.68	.001	1.47-8.19
Histology: squamous vs nonsquamous	0.72	0.42-1.24	.24

Results from the multivariate analysis including the optimized EGF score.

Significant results using backward selection. Removal criterion: P < .10.

Discussion

The impact of EGFR expression on OS in NSCLC is controversial. Studies evaluating EGFR expression as a prognostic marker have presented contradicting conclusions [13], [14], [15], [16]. In a meta-analysis by Nakamura et al., expression of the EGFR was not a prognostic marker in NSCLC [13] as opposed to what has been found in other cancers [10], [11], [12]. In vitro studies have indicated that it is not EGFR in itself but the summed effect of the EGF system that leads to tumor growth [17]. In certain studies, coexpression of the other receptors from the EGF system has been evaluated as a prognostic factor, though without convincing results when known clinical confounders are taken into consideration [21], [22], [23]. Expression of the complete network of ligands and receptors interacting with EGFR has not been evaluated as a prognostic factor in NSCLC.

In concordance with the complexity of the EGF system, we have evaluated the gene expression of each receptor and ligand in 100 NSCLC patients EGFR wild type in a model. We found that this model predicts the outcome of the individual patient, whereas the individual receptor or ligand does not. Our findings support the hypothesis that it is not the individual receptor or ligand but the network that is reflected in the OS. Gene score models have been used in several settings to sum the effect of several markers that may be coexpressed but differ in impact on a given outcome [19], [24], [25]. We show by internal validation that our model is a strong prognostic marker in NSCLC patients even when the known clinical characteristics are taken into account. We also demonstrate that it is the increased expression of the ligands AREG, HB-EGF, EPI, and TGF-α and not the receptors that has the primary impact on the summed effect of the system because removing the effect of the receptors increases the fit of the model. To our knowledge, this has not previously been demonstrated in clinical samples. When establishing the score, we chose an arbitrary approach giving each marker the value of + 1 or − 1 and not the actual coefficient as is done elsewhere [19]. Because of to the small sample size, the actual coefficient for the individual marker is not reliable, especially when splitting into test cohorts of only 50. Optimizing the estimate would lead to a risk of overfitting, and therefore, we chose − 1 or + 1 as coefficients. By doing so, we accept the risk of not being able to incorporate the magnitude of each factor of the EGF system, but in contrast, we do not run the risk of assigning inaccurate coefficients.

A major challenge in explorative gene expression studies is to establish normal values of expression. We identified 111 patients affected by symptoms leading to lung cancer suspicion but where the diagnosis was excluded based on a biopsy from a relevant area of the lung or lymph node. Because of to the consecutive nature of the study, the reference group is younger and there are more never smokers in this group. Despite these differences, we believe the reference group to be valid because it reflects the everyday clinical setting. From the reference group, cut points were defined, resulting in a test where expression above the cut points could be interpreted as truly elevated. Subsequently, we evaluated the prognostic value of the receptors EGFR, HER2, and HER3 and the ligands AREG, HB-EGF, EPI, TGF-α, and EGF. We found that only EPI was a prognostic marker of longer OS, whereas none of the receptors or the remaining ligands were significant predictors of outcome in a multivariate analysis taking the clinicopathological characteristics into account. This contradicts other studies [26], [27]. In a study by Sunaga et al., it was shown that high EPI expression could predict inferior outcome especially in KRAS-mutated patients [27]. It is known that 25% to 30% of NSCLC patients are expected to have a mutation in KRAS. The KRAS mutation status of our patients is unknown, which might explain the contradictory results. Our ability to find a correlation could also be explained by the definition of the cut points. We have used biopsies from patients entering the clinic with symptoms resembling lung cancer to define the cut points, whereas other studies exploring the role of EPI have used either the mean or median value of gene expression [27], [28] or predefined percentages of cancer cells expressing EPI[24], [25], [29]. By using our approach, we address the fact that the EGF system is not exclusively cancer related but also involved in other biological processes likely to take place in normal tissue (i.e., inflammation, etc.). Furthermore, the samples used in this study were collected with the intent to perform gene expression studies compared with studies where mRNA was extracted from archived formalin-fixed, paraffin-embedded samples.

Despite being unique in this regard, our study faced some limitations. Firstly, the cohort used in this study was consecutively collected, but stages I and II were underrepresented (19 patients), just as only 27 of the patients had squamous histology which was less than expected. We found no effect modification between the score and histology (data not shown), indicating that the expression of ligands had an equal effect in the squamous cell type as well as the nonsquamous EGFR wild-type patient. Secondly, though affecting EGFR, the receptor HER4 and the ligand BTC were not included in the final score model. HER4 was measurable only in 1 of 50 patients in the test cohort; it was not deemed relevant. In the literature, only a few studies have evaluated BTC in clinical samples, and because of limited tissue amount, we chose not to include it in this study. Our findings seem promising, but they need to be validated in an independent cohort. A better understanding of the role of EGFR in NSCLC patients without activating mutations in the EGFR can help generate new strategies for treatment development and perhaps improve the OS for NSCLC patients.

Conclusion

It is well established that activation of EGFR is dependent on ligands and other receptors of the EGF system. Therefore, to evaluate EGFR as a prognostic marker, it is necessary to take these surrounding ligands and receptors into account. In this study, we show by internal validation that a score combining factors from the EGF system can be developed and that the primary impact on OS is increased expression of the ligands and not the receptors. We have demonstrated a strong prognostic marker that can identify a group of patients where different treatment strategies may be needed.

The following are the Supplementary data related to this article.