Gene expression variation to predict 10-year survival in lymph-node-negative breast cancer

Since the prevalence of breast cancer among women is very high (one out of eight American women is affected in their life-time [1]) the economical burden for the treatment is considerable, as well as the suffering it causes. After surgery the majority of breast cancer cases are at the present over-treated since adequate diagnostic markers are not currently available. Therefore it is important to find better markers to correctly distinguish between the high-risk patients that need additional treatment and the low-risk patients where further treatment after surgery will have no positive effect and might actually harm the patient. At present, the most relevant marker in breast cancer diagnostics is the lymph-node status of the patient. Although lymph-node negative breast cancer patients have a better survival rate than patients with metastasis positive lymph nodes, around 20% will succumb to their cancer in less than 15 years [2]. In this study, gene expression microarrays were used to find a set of genes whose expression profiles can predict clinical outcome in lymph-node negative breast cancer patients. We also wanted to evaluate our results on data sets derived from similar studies done previously on tumours from lymph-node negative patients [3, 4].

Expression microarray analysis was performed on 46 lymph-node negative breast cancer tumours divided into two equally sized groups, where 23 tumours were from 10-year survivors and 23 tumours were from patients that died within 10 years after diagnosis. Previously, 43 of these tumours were included in a CGH study (Comparative Genomic Hybridisation) that searched for chromosomal alterations differing between deceased patients and 10-year survivors [16]. The aim of this study was to search for specific genes differentially expressed between the two groups and test the prognostic potential of these genes on other tumour sets. Previous studies have used expression microarrays to find sets of genes whose expression could predict clinical outcome in breast cancer or candidate genes that could be correlated to prognostic features [3, 4, 17‐32]. Some of these studies focused on node-negative tumours [3, 4, 32]. We wanted to further expand the current knowledge of the genetic events associated with clinical outcome in lymph-node negative breast cancer using tumours from a relatively homogenous population, and compare our results with previous findings, particularly focusing on the studies reported by van't Veer et al. and Wang et al. [3, 4], since both these studies developed gene-lists superior to current methods in classifying node-negative breast tumours.

A reporter list consisting of 51 genes with significant differential expression between the two groups (p < 0.001) was assembled. Hierarchical clustering using these genes resulted in two clusters, separating the two survival groups with only four survivor tumours misclassified (figure 1). Only two tumours were incorrectly classified in the cross-validation using the VFI classification algorithm (accuracy rate: 96%). The correlation-based classifier showed 89% accuracy and 100% NPV (figure 2a). We considered the NPV to be particularly important as it reflects the patient's probability to survive ten years after diagnosis if it is classified into the "good prognosis group". No patient in the "good prognosis group" died within ten years from diagnosis, which is visualized in the Kaplan-Meier curves (figure 4). Among the 18 patients that were classified in the "good prognosis group", which would not benefit from further treatment, 13 had been post-surgically treated by radio therapy, chemo therapy, hormonal treatment or a combinatory treatment (additional data file 1. Clinical data of each patient.). On the other hand, in the "bad prognosis group" 16 of the patients did not receive post-surgical treatment which they should have according to this study (of these, three were incorrectly classified 10 year survivors). Consequently, based on the results generated from the classifier using the 51 genes, the adjuvant treatment among many patients in this particular material would be reconsidered. Furthermore, the tumours within each group show similar patterns of gene expression in the 51 selected genes and these specific genes are relevant for predicting clinical outcome in our tumour material.

In order to test this gene-list further, we analysed its predictive potential in two independent data sets [3, 4]. Van't Veer et al. identified a set of 70 genes to classify 78 node-negative breast tumours from young women into poor and good prognostic groups with an accuracy of 83% (81% when the classification threshold were calibrated so that less than 10% of the metastasis patients were classified in the "good prognosis group") [3]. Another study by van de Vijver et al. confirmed the relevance of the 70-gene classifier on 295 tumours, of which 151 were from node-negative patients [33]. This classifier was designed to not classify tumours from deceased patients into the "good prognosis" group, and few of the incorrectly classified tumours are from patients where the disease recurred in both studies. Recently, a mini-microarray customised based on this 70-gene classifier was tested on 162 of the lymph node-negative patients used in the previous two studies with good prognostic correlation [34]. In another study of node-negative breast cancer by Wang et al., a 76-gene signature that correlated to disease-free 5-year survival was developed using samples from 115 patients [4]. This set of genes showed an accuracy of 63% (93% sensitivity and 48% NPV) when tested on 171 additional tumours. These two main studies (van't Veer et al. and Wang et al.) have focused on detecting patients with good prognosis where adjuvant chemotherapy is not required. Interestingly, none of the genes used in these two studies were found in our gene list, even though some genes seems to be involved in the same pathways. Ein-Dor et al. suggest that since many genes are similarly correlated to breast cancer survival, several lists of genes from the same data-set would be equally predictive [35]. These lists could be rather trustworthy prognostic tools, but the specific genes in the lists are not necessarily of importance for survival if considered individually.

When classifying van't Veer's tumours based on our gene-list, the correlation-based classifier showed 74% accuracy and 85% NPV. The corresponding percentage of correctly classified tumours in van't Veer's study was 83% (88% NPV). Overall, the average time of disease-free survival of the five patients that were misclassified into the "good prognosis group" was noticeably higher than for the patients that were correctly classified into the "bad prognosis group", although the difference fell short of being statistically significant (41 months versus 28 months, p = 0,11). Moreover, 74% accuracy, especially with high NPV, is a good result (figure 2b). Using the VFI method, 67% of the tumours were classified correctly. The results would probably have been even better if all genes in the gene list could have been found in the material. The attempt to classify Wang's data using our genes did not provide a good correlation. This may be explained by the fact that only 23 of our 51 genes could be found in their material. Furthermore, not even Wang et al. themselves found particularly high accuracy when classifying their tumours (63%). Still they could specify a sub-group of 56 patients where 93% were free of metastasis within five years from diagnosis, which we could not do with our gene set. In our gene set, four genes were of special interest since they differed significantly in expression between the groups with (p < 0.05) in van't Veer's data as well; GGH, PIR, TAF5L and FAAH. Of these, only FAAH was significant in Wang's data. This gene has earlier only been correlated to multiple drug addiction [36]. In summary, our gene set had high accuracy when classifying our own material and was relatively competent in classifying the samples in van't Veer's study, but did not show high accuracy on Wang's tumours.

A correlation-based classifier showed 72% accuracy in predicting 10-year survival (figure 3a) and the VFI classification of our tumours using van't Veer's genes showed 67% accuracy. Hereby, the results both with the VFI classifier and the correlation-based classifier were moderate, considering that the NPV was only 73%. Using Wang's genes, the VFI classifier showed 70% accuracy which is moderate, while the correlation-based classifier worked poorly, showing only 61% accuracy and as many as twelve deceased patients were classified in the "good prognosis group" (figure 3b). In general, classifying a material is naturally more accurate when using a set of genes selected for that specific material. Moreover, the genes in our list seem to be more accurate in predicting 10-year survival, whereas van't Veer's and Wang's genes are likely to better predict metastasis within five years.

Many of the 51 genes with significantly different expression between the two survival groups have previously been implicated in cancer. The genes BCAT1, GGH and SERPINB9 have been correlated to clinical outcome in other types of cancer, such as colorectal cancer, neuroendocrine cancer, large cell lymphoma and melanoma [37‐40]. The genes SALL4 and TM4SF5 were expressed to a higher extent in the tumours from deceased patients in our material and have been reported to be up-regulated in other cancers, and may thereby represent putative oncogenes [41, 42]. NEIL1, a gene involved in DNA repair [43], showed lower expression levels in the tumours from deceased patients in our material and has been reported as down-regulated in gastric cancer [44]. The GIT2 gene was higher expressed in the tumours from deceased patents and may be implicated in the transformation of epithelial cells to cancer cells as well as inducing cell motility and invasion [45]. The genes CCNB1IP1, CUL7 and E2F2 are involved in cell cycle control and cell growth [46‐48] and have expression levels in our study that promotes cell growth in the tumours from deceased patients. In our previous CGH-study of the same tumours, seven chromosomal regions were altered significantly more in the tumours from deceased patients than the tumours from 10-year survivors (4q, 5q, 6q, 12q, 17p, 18p and Xq) [16]. Five of the selected genes in the microarray study were located in these regions (C4orf26, CCDC99, SPATA5, TM4SF5 and TRPV2) and might be of special interest since they revealed significance in both studies.

There are disadvantages using this type of selected material for gene expression studies, the tumours in this study come from a relatively homogenous population in Sweden which might make the results less applicable for breast cancer tumours in general. Furthermore, the number of tumours investigated in this study is low, further studies using larger sets of tumours are needed to verify the significance of the 51 gene list. Also, the tumours have been frozen for a long period of time which might affect the quality of the RNA yielding less reliable results.

It can be concluded that the list of 51 genes we identified (table 2) could predict clinical outcome in our material with great certainty. They were competent in van't Veer's as well but not in Wang's material, probably due to the low number of genes found in the material. In classifying our material, our gene set clearly worked best, but the genes found by van't Veer et al. and Wang et al. had some prognostic potential as well. The gene set found by Wang et al. had the lowest impact on our material, particularly considering the low NPV (59%) in the correlation based classification. Overall, our gene set worked similarly well in classifying van't Veer's material as their gene set on our material, slightly better considering the NPV.

The list of 51 genes might contain specific genes interesting for clinical outcome in breast cancer as well as being a good prognostic gene set. Additional studies using larger sets of tumours are needed to define the significance of these genes during the genesis of lymph-node-negative breast tumours.