Prediction of Lymph Node Metastasis by Analysis of Gene Expression Profiles in Primary Lung Adenocarcinomas

Abstract

Purpose

Lymph node status is a strong predictor of outcome for lung cancer patients. Recently, several reports have hinted that gene expression profiles of primary tumor may be able to predict node status. The goals of this study were to determine if microarray data could be used to accurately classify patients with regard to pathologic lymph node status, and to determine if this analysis could identify patients at risk for occult disease and worse survival.

Experimental Design

Two previously published lung adenocarcinoma microarray data sets were reanalyzed. Patients were separated into two groups based on pathologic lymph node positive (pN+) or negative (pN0) status, and prediction analysis of microarray (PAM) was used for training and validation to classify nodal status. Overall survival analysis was performed based on PAM classifications.

Results

In the training phase, a 318-gene set gave classification accuracy of 88.4% when compared with pathology. Survival was significantly worse in PAM-positive compared with PAM-negative patients overall (P < 0.0001) and also when confined to pN0 patients only (P = 0.0037). In the validation set, classification accuracy was again 94.1% in the pN+ patients but only 21.2% in the pN0 patients. However, among the pN0 patients, recurrence rates and overall survival were significantly worse in the PAM-positive compared with PAM-negative patients (P = 0.0258 and 0.0507).

Conclusions

Analysis of gene expression profiles from primary tumor may predict lymph node status but frequently misclassifies pN0 patients as node positive. Recurrence rates and overall survival are worse in these “misclassified” patients, implying that they may in fact have occult disease spread.

Non–small cell lung cancer is the leading cause of cancer death in the United States and overall 5-year survival is ∼14% (1). Even in stage I (lymph node negative) patients, recurrence rates range from 25% to 50% and result in poor (50-70%) 5-year survival in this group (2). Current staging methods are unable to identify those patients who will ultimately have poor outcome, and we are therefore in need of a clinically useful approach to better stratify patients with respect to the risk of recurrence and survival. One much studied approach is to enhance the detection of tumor spread to lymph nodes using either immunohistochemistry for tumor cells or molecular detection of cancer-related RNA. A review of this literature indicates that occult lymph node metastases (micro-metastases) are indeed clinically relevant in non–small cell lung cancer and confer worse prognosis (3). Recently, however, Beer et al. reported that microarray analysis of gene expression in lung adenocarcinomas can also predict outcome in stage I patients (4). In this study, survival of the low-risk stage I patients was almost 90%, whereas in the high-risk group, survival was similar to that expected for pathologic N1 (pN1) lung cancer patients (42%; ref. 5). Also interesting was the fact that the stage III patients did not show a significant survival difference between the high-risk and low-risk groups. Whereas this may be attributable to the small sample size, an alternative hypothesis is that the microarray data identifies patients at high risk for occult lymph node disease in the stage I group and that this is irrelevant in the stage III patients who already have overtly positive lymph nodes. Although this hypothesis is very tenuous based solely on the data from Beer et al., two recent articles lend some support (6, 7). In the first article by Ramaswamy et al., the authors analyzed gene expression in primary adenocarcinoma samples and compared this with gene expression in metastatic tumors derived from unmatched primary adenocarcinomas. Using this data, a 128-gene signature was identified that distinguished reasonably well between primary tumors and metastases. When applied to 62 lung adenocarcinomas, the authors found that survival was significantly worse in patients whose tumors bore the metastasis-associated gene expression signature (P = 0.009). The authors, however, did not report on correlation of the microarray data with pathologic nodal status in these 62 patients. From their data, Ramaswamy et al. concluded that some primary tumors are preconfigured to metastasize, and that this propensity is detectable at the time of initial diagnosis. In the second study, Huang et al. analyzed breast cancer samples and identified separate gene sets that seem to predict lymph node metastasis and overall patient outcome. Specifically relating to lymph node status, the authors identified a “metagene” signature that predicted pathologic nodal involvement with ∼90% accuracy. From these articles, it seems that while metastatic potential and patient outcome are clearly interrelated events, they may be the result of distinct biological processes, the signatures for which can be identified by analysis of the primary tumor.

The goals of this study were to reanalyze publicly available gene expression data from lung adenocarcinoma samples and identify gene expression patterns that correlated with lymph node metastases. These gene expression patterns were then applied to pathologically node-negative patients to determine if the primary tumor microarray data could identify patients at risk for occult disease and worse survival.

Materials and Methods

Microarray data

Analysis was done on two independent, publicly available lung adenocarcinoma microarray data sets. Data set 1 was previously published by Beer et al. (4) from the University of Michigan, and both gene expression data and patient outcome information were downloaded from the web site http://dot.ped.med.umich.edu:2000/ourimage/pub/Lung/index.html. This data set was originally obtained using Affymetrix HuFL arrays and included a total of 86 adenocarcinomas consisting of 69 pN0 and 17 pN+ (either pN1 or pN2) patients. The second data set was that previously published by Bhattacharjee et al. (8) from Harvard University, and was used as a test data set in Beer's study. The Harvard data was downloaded from the web site http://dot.ped.med.umich.edu:2000/ourimage/pub/Lung/index.html. This set consisted of data from 84 lung adenocarcinoma samples and was obtained using Affymetrix HG-U95A arrays. Of the 84 patients in this data set, 52 were pN0, 17 were pN+, and the remaining 15 were listed as having unknown lymph node status. These patients were excluded from our analysis, leaving 69 patients with definitive lymph node staging by pathology.

Data compilation and preanalysis

There were 7,129 Affymetrix probe sets (genes) in the first data set and 12,625 probe sets in the second set. Data for NULL gene name (EST) were excluded and a total of 5,377 genes were subsequently identified as being common to both data sets (Supplementary data). Expression values for the same gene from multiple probe sets were average of values generated from all probe sets. To assess data quality and to identify any differences between data sets or between pN0 and pN+ patients, a variety of data visualization and data quality measurements were employed. These included data quality graphs (mean versus variance; multiplicative-additive plots; correlograms) and measures (confounding index; among-array coefficient of variation; mean group correlations). All methods were done using the University of Pittsburgh Cancer Institute web application for microarray analysis (caGEDA; cancer gene expression data analyzer), located at http://bioinformatics.upmc.edu/GE2/GEDA.html (9).

Data analysis and prediction of lymph node status

Prediction of lymph node status was done in a training set (Michigan data) and test set (Harvard data) approach using prediction analysis of microarray (PAM) software available free from the url http://www-stat.stanford.edu/~tibs/PAM. This software uses a variation of shrunken-centroid classification with an automated gene selection step integrated into the algorithm. During 10-fold cross-validation, a shrinkage factor Δ is varied in search of optimal classifier performance. The value that returns the lowest classification error with the fewest genes is favored as optimal. The variable Δ is a soft-thresholding variable that is reported to usually produce more reliable estimates of the true means (10). The Δ value obtained from training on the entire data set from the Michigan study was subsequently used for predicting the lymph node status of each sample in the Harvard data set. Survival comparisons between the predicted classes were done using Kaplan-Meier survival estimates and the log-rank test for determining survival differences.

Results

Evaluation of data set quality and compatibility

To evaluate the quality of the expression data from the 5,377 overlapping genes, confounding index (see footnote to Table 1), among-array coefficient of variation, and mean group correlation coefficients were calculated for each data set using caGEDA (Table 1). This analysis indicated overall high data quality, and no further normalization was done. To determine comparability of data from pN0 and pN+ patients, box-whisker plots were generated. No apparent difference in the mean or distribution of total expression of the 5,377 genes was identified between pN+ and pN0 groups within each data set (Fig. 1A and B). However, box-whisker plots indicated large differences in the global expression distributions between the two data sets (Fig. 1C). Rather than attempting to combine the two data sets by performing additional normalization, we instead treated the two data sets independently using one as a training set and the other as a validation or test set.

Table 1.

Data quality measurement between pN0 and pN+ groups in both data sets

Measurement	Michigan data set	Harvard data set
Confounding index*	1.007	0.997
Among-array coefficient of variation	0.034	0.028
Mean group correlation (r2)	0.984	0.994

^*A confounding index for global expression compares the average within-group correlation with the average between-group correlation. It should be as close to 1.0 as possible. Values higher than 1.0 may indicate incidental confounding factors in the experimental design. For equation and more details, see ref. 9.

Fig. 1.

Box-and-whisker plots with sample names on the X axis and expression values on the Y axis. A, data from pN+ ( ) and pN0 ( ) groups in Michigan data set. B, data from pN+ and pN0 groups in Harvard data set. C, data from both data sets plotted together; , Harvard; , Michigan.

Prediction of lymph node status in the training set

Analysis of the training data set using PAM identified a group of 318 genes that appeared to correlate with pathologic lymph node status (all 318 genes are listed in Supplementary data). The top 50 of these genes are listed in Table 2. Because the output from the PAM software is in the form of continuous probability values, comparison with pathology was facilitated by transforming the PAM output into binary data (lymph node positive or negative) using P = 0.5 as the cutoff. When compared with pathologic nodal status, lymph node classification by PAM was correct in 76 of 86 (88.4%) cases overall. PAM correctly classified 16 of 17 (94.1%) pN+ samples and 60 of 69 (87.0%) pN0 samples in this set. Furthermore, on individual samples, the classification probability reported from PAM was 100% for 73 samples, 80% to 99% for 11 samples, and 50% to 80% for only 2 samples (Fig. 2A).

Table 2.

Top 50 genes associated with nodal metastases from PAM classification

Rank in PAM output	Gene symbol	Gene name	Function
Overexpressed genes
1	COL5A2	collagen, typeV, α2	Extracellular matrix
4	COL1A1	collagen, type I, α1
9	COL3A1	collagen, type III, α1
14	COL1A2	collagen, type I, α2
18	TGFBI	transforming growth factor β induced, 68 kDa
19	OSF-2	periostin, osteoblast-specific factor
10	LGALS1	lectin, galactoside binding, soluble, 1	Cytoskeletal/cell adhesion
22	SPARC	SPARC-like 1
27	CSPG2	chondroitin sulfate proteoglycan 2
37	VIM	vimentin
38	TAGLN	transgelin
45	KRT15	keratin 15
46	FN1	fibronectin1
5	RPLP1	ribosomal protein, large, P1	Protein synthesis/degradation
6	RPS23	ribosomal protein S23
23	RPL37	ribosomal protein L37
29	RPS26	ribosomal protein S26
20	CSTB	cystatin B
33	CTSL	cathepsin L
13	S100A10	S100 calcium binding protein A10	Calcium
17	S100A11	S100 calcium binding protein A11
35	S100A4	S100 calcium binding protein A4
39	S100A2	S100 calcium binding protein A2
15	HSPA8	heat shock 70 kDa protein 8	Chaperones
28	CEBPB	CCAAT/enhancer binding protein, β	Transcription/translation
43	SUI1	putative translation initiation factor
3	GAPD	glyceraldehyde-3-phosphate dehydrogenase	Enzymes, cellular metabolism
11	GPI	glucose phosphate isomerase
25	PGK1	phosphoglycerate kinase 1
26	TPI1	triosephosphate isomerase 1
36	LDHB	lactate dehydrogenase B
7	H2AFZ	H2A histone family, member Z	Other
34	HNRPA2B1	heterogeneous nuclear ribonucleoprotein A2/B1
42	PLP2	proteolipid protein 2
49	RAB31	RAB31, member RAS oncogene family
Underexpressed genes
2	HLA-B	MHC, class I, B	Other
8	OAZ1	ornithine decarboxylase antizyme 1
12	JUND	jun D proto-oncogene
16	USP11	ubiquitin specific protease 11
21	KIAA0220	KIAA0220-like protein
24	RPS8	ribosomal protein S8
30	PTGDS	prostaglandin D2 synthase 21kDa
31	CST3	cystatin C
32	MGP	matrix Gla protein
40	SLU7	step II splicing factor SLU7
41	CIRBP	cold inducible RNA binding protein
44	ARHA	ras homolog gene family, member A
47	HU-K5	monoglyceride lipase
48	CAP	adenylate cyclase – associated protein 1
50	SFTPD	surfactant, pulmonary-associated protein D

Fig. 2.

Classification probability by PAM with 10-fold cross-validation for training set (A) and for test set (B). Original sample names and pathologic lymph node status are labeled on the top of each graph. PAM-classified nodal status is indicated by ◆, (pN0); , (pN+) with the Y axis indicating the probability associated with the classification of each patient.

Prediction of lymph node status in the validation set

The 318-gene classification algorithm generated on the training set was next applied to the data from Harvard as a validation test. In this data set, the classification accuracy for PAM in pN+ cases remained high (94.1%; 16 of 17) but accuracy in the pN0 group was only 21.2% (11 of 52). Overall, classification accuracy for PAM in the validation set was 39.1% (27 of 69) when compared with pathology. The classification probabilities for individual samples were similar to those in the test set with 100% probability for 54 samples, 80% to 99% for 11 samples, and 50% to 80% for 4 samples (Fig. 2B).

Correlation of prediction analysis of microarray nodal classification with recurrence and survival

In the training data set, Kaplan-Meier survival estimates showed that survival was significantly worse in PAM-positive cases than in PAM-negative cases (P < 0.0001) for all 86 patients. Furthermore, when the survival analysis was restricted to the pN0 patients only, we found that survival was worse in 9 PAM-positive cases compared with 60 PAM-negative cases (P = 0.0037; Fig. 3A). Thus, PAM successfully identified a subgroup (13%) of patients with worse outcome within the pN0 patients based on the microarray nodal status classification.

Fig. 3.

Kaplan-Meier survival curves for pathologic node-negative patients classified as positive or negative by PAM analysis. Michigan data set (A) and Harvard data set (B).

In the validation data set, survival analysis of all 69 patients showed that PAM-negative and PAM-positive patients again differed significantly in overall survival (P = 0.0258) although the stratification was not as strong as in the training set. In the 52 pN0 patients only, disease recurrence was significantly higher (19 of 41, 46.3%) in PAM-positive patients than in PAM-negative patients (1 of 11, 9%; Table 3) and survival analysis showed that survival was worse in 41 PAM-positive patients compared with 11 PAM-negative patients (P = 0.0507; Fig. 3B).

Table 3.

Patients relapse/metastasis status in Harvard data set

Node status				Site of relapse/metastasis*
Pathology	No. patients	PAM predicted	No. patients	Lymph node	Lung only	Other site	Total (%)
Negative	52	N0	11	0	1	0	1 (9.0)
		N+	41	5	2	12	19 (46.3) †
Positive	17	N0	1	0	0	0	0
		N+	16	3	3	6	12 (75.0)

^*Data was from http://research.dfci.harvard.edu/meyersonlab/lungca/.

^†Significantly higher than PAM predicted N0 patients within pathologic negative group (P = 0.024).

Discussion

Treatment options for patients with non–small cell lung cancer are largely dictated by the status of local and regional lymph nodes. Node-negative patients are typically treated with surgery alone whereas patients with lymph node involvement will frequently receive adjuvant or neoadjuvant chemotherapy, as well as surgery, in the belief that their tumor has spread and that surgery alone is less likely to result in a cure. Identification of lymph node metastases therefore plays a critical part in the current management of non–small cell lung cancer patients. Unfortunately, however, current methods for clinical and pathologic staging of lymph nodes are not very sensitive and many patients are understaged. As a result, many studies have been undertaken in attempts to improve the sensitivity of lymph node metastasis detection using either immunohistochemistry or molecular methods such as reverse transcription-PCR (3). In both cases, these assays directly identify metastatic disease in the lymph nodes.

More recently, analysis of gene expression in the primary tumor has been shown to correlate with patient outcome (4). Thus, it is possible that surgical removal and pathologic evaluation of lymph nodes may not be necessary to accurately predict a patients' survival probability. However, because survival is linked to effective treatment, and effective treatment is currently determined by nodal stage, it is unclear whether predicting patient outcome alone will be sufficient to avoid the need for accurate lymph node assessment. With this in mind, other recent studies have also shown that it may be possible to predict metastatic potential (6) and specifically lymph node metastases (7, 11) using analysis of gene expression in the primary tumor. For example, Huang et al. used microarrays to determine gene expression in 89 primary breast tumors and then used complex statistical approaches to identify patterns and interactions between genes that correlated with patient outcome. The authors identified patterns of gene expression that were associated with either lymph node involvement or survival, indicating that these two end points represent distinct biological processes, involving different gene sets. Interestingly, the genes that correlated with survival were the ‘classic’ oncogenes associated with cell cycle, growth, cell signaling, etc., whereas the genes associated with lymph node involvement tended to be chemokines, chemokine receptors, and IFN-associated genes. This is very similar to the data reported by Ramaswamy et al. and raises the hypothesis that the ability of a tumor to metastasize is dependent on the host response to the tumor, and that this is what can actually be detected by analysis of the primary tumor. Thus, it is possible that analysis of the primary tumor may provide information on both the probability of survival and lymph node metastasis. Understanding the genomic differences between these closely related end points may prove useful for determining the appropriate treatment of non-small cell lung cancer patients.

In the current study, we specifically aimed to determine whether analysis of gene expression in primary non–small cell lung cancer could predict lymph node status and, if so, predict outcome in lymph node–negative patients, many of whom are presumed to have occult lymph node metastases (3). For our study, we used pathologic lymph node status as the training classifier, and used microarray analysis software (PAM) to identify gene patterns that specifically correlate with pathologic nodal status. PAM identified 318 significant genes that were associated with nodal metastases. Of the top 50 genes (Table 2), 35 are overexpressed in tumors with lymph node metastases and the remaining 15 genes are underexpressed. These genes are involved in a wide spectrum of biological functions including extracellular matrix, protein synthesis/degradation, cell adhesion and structure, and regulation of transcription/translation. Interestingly, among the 35 overexpressed genes, four genes are from the collagen gene family and four are from the S100 calcium binding protein family. The up-regulation of collagen genes in tumors with nodal metastasis is consistent with observations that interactions between tumor cells and the surrounding stroma are critical for tumor cell invasion (12–14). High levels of collagen gene expression have also been observed in advanced gastric and ovarian cancers using cDNA microarrays and serial analysis of gene expression, and in colorectal cancer using differential display reverse transcription-PCR (15–17). Overexpression of S100A4 and S100A2 has also been associated with metastasis and survival in patients with early-stage breast cancer (18) and non–small cell lung cancer (19). Of the underexpressed genes, down-regulation of the HLA-B gene may allow tumor cells to escape the T-cell response and this is a potential mechanism for tumor cells to escape immune surveillance (20–22). Down-regulation of the another gene, MGP, has also been seen in colon cancer (23) and has been associated with metastasis in prostate adenocarcinoma (24). Therefore, capture of these genes as the top genes associated with nodal metastases by PAM classification probably reflects the biological pathways contributing to the nodal metastases phenotype.

One clear weakness in our study is that many of the pN0 patients may actually have occult nodal involvement and, therefore, our gold standard for the training classifier is not necessarily accurate in pathologically node-negative cases. However, it should be 100% accurate in the case of node-positive tumors. Indeed, in both the training and validation data sets, we found extremely high (94%) concordance between microarray analysis and pathologic node status for tumors with pathologically positive lymph nodes. In node-negative tumors, however, microarray analysis of the training set data predicted that 9 of 69 (13%) pN0 tumors were actually node positive. This finding agrees quite well with estimates on the prevalence of occult lymph node disease in non–small cell lung cancer (3) and survival of these nine patients was significantly worse than for the remaining 60 pN0 patients (P = 0.0037). In the validation set, however, an unexpectedly high number (41 of 52, 79%) of pN0 tumors were classified as positive by PAM. Clearly, this is too high to be explained by the presence of occult disease alone and, therefore, we considered the following alternative possibilities: (1) Average expression values from the Harvard data set were much higher than that from Michigan data. This might cause prediction bias when the training algorithm generated with the Michigan data set was applied to the Harvard data. To make the differences smaller between the two data sets, we normalized the data using Z transformation but did not find any improvement in the performance of training and testing. (2) PAM was overtrained with the training data set. To test this hypothesis, a simulated data set that included 5,377 values (randomly assigned from −10 to 2,000) in 86 samples with two classes (randomly assigned with 1 or 2) was generated and applied to PAM for classification training. The overall classification error with cross-validation was 40% to 50% and classification probabilities for individual samples were from 40% to 60% (Supplementary Fig. 1). This shows that PAM classified the simulated data randomly (as it should) and indicates that the method does not overtrain for classification. (3) The population of patients between the two data sets was significantly different in terms of staging and/or outcome. Comparison of patient survival between the two data sets was analyzed and the results showed that patients from the Harvard data set had a significantly worse survival than the patients from the Michigan data set (P = 0.0245). Interestingly, this difference was completely due to differences in survival among the pN0 patients in the two data sets (P = 0.058), whereas there was no difference in survival of the pN+ patients between the two data sets (P = 0.9499; Supplementary Fig. 2).

When we analyzed survival in PAM-negative and PAM-positive patients within Harvard pN0 patients, 46.3% of PAM-positive patients suffered disease recurrence compared with only 9% in the PAM-negative patients. Thus, we believe that the poor outcome in the pN0 Harvard patients reflects understaging (possibly occult nodal disease) in a higher than normal percentage of patients, and to some degree this is responsible for the high number of patients called node positive in the microarray analysis. Furthermore, when we used the Harvard data set for training, 70% of the pathology-positive and 38% of the pathology-negative patients were incorrectly classified by PAM. Overall, cross-validated classification accuracy when training with the Harvard data set was only 54%. Again, we hypothesize that this inability to attain good classification accuracy (even within the training set itself) may be a result of a high frequency of occult lymph node disease in this patient set. Despite this, however, 22 patients who did not suffer disease recurrence were still called node positive in our analysis (when training with the Michigan data set), implying that many patients who were probably truly node negative were misclassified in the test set. This may be a result of differences between the two microarray data sets or may reflect the need to refine the analysis tools used for prediction. Another possibility is that some of the misclassified tumors do actually have the potential to metastasize but have not yet done so. Presumably, metastasis requires not only metastatic ability but also time, and it is therefore possible that the actual detection of lymph node disease (overt or occult) will prove to be a more powerful predictor of outcome than the primary tumor gene expression. Conversely, however, one could envision a scenario where tumor cells have spread to lymph nodes (or distant sites) but are unable to divide, grow, and/or evade the host immune system. This may be encoded in the genetics of the primary tumor in which case the gene expression data may be the more powerful predictor.

In summary, our analysis of gene expression in primary tumors correlated very highly in pathologically node-positive patients. Furthermore, recurrence rates and overall survival were worse in the pN0 patients who were “misclassified” as positive by microarray analysis, and it seems likely that the analysis may be identifying patients whose tumors have already metastasized or have the potential to do so. However, one problem with this interpretation of the data is that we do not have actual data on occult lymph node disease (detected by immunohistochemistry or reverse transcription-PCR) in pN0 patients. Instead, we are using recurrence and overall survival as surrogates for occult disease and this end point may reflect not only lymph node spread but also hematogenous spread and clinically occult metastases to other organs. Furthermore, because lymph node metastasis and survival are closely correlated, it is possible that despite training based on lymph node status, our analysis is actually identifying genes that correlate with overall outcome and not specifically with the ability to metastasize. Thus, we realize that our explanation for the unexpectedly high false-positive results from microarray analysis may in fact be somewhat circular. This issue can only be resolved in a situation where tumor gene expression, accurate clinical and surgical staging, and presence of occult nodal disease are determined in the same set of patients.