Predicting a local recurrence after breast-conserving therapy by gene expression profiling

This analysis is based on previously reported gene expression profiles of tumors from a series of 295 stage I and II breast cancer patients treated at the Netherlands Cancer Institute between 1984 and 1995 [23]. All patients were under 53 years of age at the time of diagnosis. For this study all the patients from this series who received BCT (n = 161) were selected. BCT consisted of a wide local excision and axillary lymph node dissection followed by whole breast irradiation (median dose 50 Gy, range 50 to 54 Gy; the use of 6, 8 or 18 MV photons depending on breast diameter); 144 patients received a boost to a median dose of 15 Gy, 98 patients received a low boost (14 to 18 Gy), and 46 patients received a higher boost (20 to 26 Gy). The boost technique was delivered using ¹⁹²I (iridium) implantation (88 patients), electrons (31 patients) or photons (25 patients). Pathological margins were assessed as free of tumor in 134 patients, focal involvement by invasive carcinoma only in 3 patients, focal involvement by ductal carcinoma in situ (DCIS) only in 5 patients, and focal involvement by both DCIS and invasive carcinoma in 4 patients; in 8 patients there was more than focal involvement of the margins by DCIS, or the degree of margin involvement could not be determined with certainty. For four patients there was invasive carcinoma at the margin, but the extent of involvement could not be determined with certainty, and for three patients this was the case for both invasive carcinoma and a DCIS component. Of all 27 patients with any extent of involved margins, 2 did not receive a boost, 6 received a low boost and 19 a high boost of radiotherapy. Follow-up assessment was at least yearly, including mammography. Seventeen patients developed a local recurrence; the resection margins of the primary tumor had been free of any tumor for 15 of these 17 patients. One patient who later developed a local recurrence had focal involvement of the resection margins with invasive carcinoma, and one patient had focal involvement with DCIS. The 10-year local-recurrence-free percentage was 85% (median follow-up for all patients was 7.7 years, and 8.5 years for alive patients).

On the basis of the microarray identification number, patients were alternately assigned to the training set (that is, developing a predictor of local recurrence) or to the validation set. This was done separately for locally recurring and locally tumor-free patients, thereby creating almost equal-sized sets with balancing for local recurrence.

Patient characteristics for both the training and validation series are equally divided and are shown in Table 1. In Table 2 characteristics are listed for patients with and without a local recurrence. A full clinical data sheet is available online as Additional file 1, on [28] and on the paper's home page [29].

RNA isolation, labeling of complementary RNA, and hybridization to 25,000-element oliogonucleotide microarrays, and measurement of expression ratios were performed as described previously [23].

The microarray data for analysis of the 70-gene prognosis profile were obtained with the use of the same 25,000-element oligonucleotide array; the correlation coefficient of the expression of these 70 genes with the good prognosis profile was calculated as reported previously [22]. The two other gene lists used are derived from Stanford cDNA arrays [30]. The correspondence between genes represented by Stanford cDNA microarrays and Rosetta/NKI oligonucleotide microarrays were established by using Unigene identifiers (build 158, release date 18 January 2003 for the wound signature, as this was the original mapping carried out for the validation of the wound signature [26]). The primary hypoxia analysis was performed at a later time. We used build 172, release date 17 July 2004 [27]. Because of the relatively small number of genes, probes on the Rosetta/NKI array that were mapped to the same unigene cluster were averaged.

The entire expression database for all 295 patients is publicly available online [28], as are the files containing the expression ratios for the specific experiments (70 genes (Additional file 2), hypoxia signature (Additional file 3) and wound signature (Additional file 4) [29].

The first method we used for predicting local recurrence is PAM (prediction analysis for microarrays): class prediction as described by Tibshirani and colleagues [31]. This method shifts the mean expression level of each gene (possibly after transformation) for each class (for example local recurrence) towards the overall mean expression level for all classes by a fixed standardized difference (shrunken centroids). The class for which the shrunken centroid most closely reaches the observed expression pattern of a certain patient by using a Pearson correlation is then the predicted class for that patient. For a given shrunken centroid only those genes for which the shrunken means still differ from the overall mean will contribute to the distance between centroids and any individual tumor sample's expression pattern. The standardized difference, and thereby the number of relevant genes, is chosen by minimizing the prediction error using 10-fold balanced, leave-10%-out cross-validation within the training set. The same method is then used to predict classes for new samples (validation set).

Subramanian and colleagues [32] have described a method that analyzes whether or not predefined gene sets (representing biological pathways) are coordinately and significantly upregulated or downregulated between samples that represent different biological or clinical entities [32]. The program corrects for multiple testing. We applied this method to our data set and analyzed gene set C2 ('functional gene set'). The software and gene set are available for download [33]. C2 represents 504 gene sets. Standard parameters were used for the analysis. The threshold for false discovery rate is 25% before a gene set is called significantly upregulated or downregulated.

The following previously established gene expression profiles were used to find a classifier for local recurrence after BCT:

First, these gene expression profiles were used in their original form and tested for their value in predicting local recurrence after BCT in the complete set of 161 patients.

Second, we optimized these profiles by using the clinical data on local recurrence.

Using the training data set, we calculated for each of the three profiles the average pattern of expression of the genes that comprised each profile, for patients who developed a local recurrence (local recurrence profile or local recurrence centroid). For all patients in the training and validation set we then calculated the non-centralized Pearson correlation to each of the three average patterns of expression (local recurrence profile or local recurrence centroid). A correlation value between -1 and +1 is thereby derived for all patients and indicates the similarity in gene expression between an individual patient (tumor) and the average gene expression of tumors from patients in the training set who developed a local recurrence.

For each of the three signatures a plot of these correlation values (see Figure 1 for the core serum response genes) is drawn for the training set and shows the difference in expression between patients who developed a local recurrence and those who did not. On the basis of the difference in the distribution of these correlation values, we selected the maximal threshold, misclassifying only one out of nine local recurrences in the training set. Patients with a correlation value above the threshold were classified as high-risk, and patients below the threshold as low-risk. Except for the gene selection process, this is the same procedure as followed by van't Veer and colleagues [21]. The local recurrence profile and its associated threshold value were subsequently validated on the validation set. The methods for training and validation are depicted schematically in Figure 2.

All patients who developed a local recurrence were considered; 16 patients developed a local recurrence as a first event and in 1 patient local recurrence occurred concurrently with distant metastases. Local-recurrence-free percentages were calculated with the Kaplan–Meier method (log-rank test for comparison). All Kaplan–Meier curves are available as supplementary data in Additional file 5. Comparison of categorical variables between groups was performed with a χ² test. For the multivariable analysis we used a Cox regression model. All p values are two-sided. All correlations were calculated by the non-centralized Pearson method. For analysis we used Winstat^® for Excel (R. Fitch Software, Staufen, Germany), SPSS 13.0^® (SPSS Inc., Chicago, IL, USA) and Microsoft Excel^® (Microsoft Corporation) were used. The PAM analysis was performed with version 1.2 (November 2003) [31, 35].

Additional material and methods are available described in Additional file 9.

Nine out of 81 patients in the training set had developed a local recurrence at a median interval of 6.2 years (range 1.0 to 10.8 years); the median follow-up time for the patients in the entire training set was 7.9 years. Six out of nine local recurrence patients had received a boost (five patients received 15 Gy, one patient 25 Gy).

In the validation set, 8 out of 80 patients had developed a local recurrence at a median interval of 3.72 years (range 1.2 to 8.4 years); the median follow-up time for all patients in the validation set was 7.6 years. Seven out of these eight patients had received a boost (three patients received 15 Gy, four patients 25 Gy).

We used the PAM algorithm with the training data set to search for gene expression profiles predictive of local recurrence, starting with a gene list that consisted of the top 5,000 most variably expressed genes with technically well-measured expression, represented by microarrays (out of 25,000). The performance of this approach, as evaluated by cross-validation, seemed to be poor: only 4 out of 9 local recurrences and 50 of the 72 non-local recurrences could be predicted accurately in the training set at this threshold (using a profile based on 684 genes). The use of fewer genes resulted in more accurate prediction for non-local-recurrence patients, but less accuracy for local-recurrence patients. In the validation series only 1 out of 8 local recurrences and 65 out of 72 non-local recurrences were predicted correctly (PAM analysis is provided in Additional files 6, 7, 8).

We used the GSEA (gene set enrichment analysis) software to analyze whether or not gene sets defined on the basis of biological processes or pathways were significantly correlated with local recurrence. We first ran the analysis on the full probe set (more than 23,000). We than ran the analysis on the same selection of genes used for the PAM analysis (the top 5,000). Neither the full probe set nor the top 5,000 probe set revealed any gene set to be significantly upregulated or downregulated in samples from tumors that recurred locally, after correction for multiple testing.

We have previously shown that three distinct gene expression signatures can each be used to subdivide breast carcinomas into two groups that are associated, respectively, with good and poor distant-metastasis-free and overall survival: a '70-gene' prognosis profile [22, 23], a wound-response signature [26, 30] and a hypoxia-response signature [27]. We tested whether these three signatures were also associated with a difference in the risk of local recurrence. The 70-gene prognosis profile was originally trained to predict metastasis within 5 years ('poor prognosis' group) or no metastasis within 5 years ('good prognosis' group); in the current series of 161 patients, 97 patients had tumors with a poor prognosis profile and 64 patients had tumors with a good prognosis profile.

The wound-response signature was derived from an in vitro model for wound healing and can be used to classify tumors as having an 'activated wound-response signature' or a 'quiescent wound-response signature' [26, 30]. In this series of tumors, 380 (out of 459) genes from the wound-response signature were identified; 65 tumors had an activated wound-response signature and 96 tumors had a quiescent wound-response signature.

The hypoxia-response profile is a gene expression classifier derived from an ex vivo model in which epithelial cells are exposed to hypoxia [27]. In this series of tumors, 123 (out of 168) genes from the hypoxia-response signature were identified. On the basis of the expression of these hypoxia-associated genes, the breast carcinomas in this study could be assigned to 36 cases with a hypoxia-response signature and 125 with a non-hypoxia signature.

Because the supervised approach using all genes (class prediction; PAM) failed and the previously identified gene expression signatures could not convincingly predict the risk of a local recurrence, we used genes from the original profiles and constructed a new classifier based on the relationship of their expression patterns to clinical outcome observed in our study (local recurrence or no local recurrence).

Each of these three signatures was originally defined for purposes other than the prediction of local recurrence, but we hypothesized that the optimization of each signature's pattern threshold by training on local recurrence data might improve their performance for this new purpose.

Applying the method described in the Materials and methods section, a correlation value of 0.2364 was calculated as a cutoff point for the 70-gene expression profile as a predictor of local recurrence, resulting in one misclassified patient with local recurrence. The recurrence risks at 10 years were 7% and 28% for the low-risk and high-risk groups, respectively.

The wound-response signature was first identified as a pattern of changes in gene expression induced by exposing fibroblasts to serum. A threshold value of 0.3179 for the correlation with the wound-response signature (as derived from the 380 available genes) resulted in one missed local recurrence in the training set. With the use of the threshold value for classification, 30 patients (37%) were classified as high-risk, and 51 (63%) as low-risk. In the training set, the 10-year recurrence risk was 6% for the 'low-risk' group and 35% for the 'high-risk' group. With this threshold, in the validation set 26 patients (33%) were classified as high-risk, and 54 (67%) as low-risk. The 10-year recurrence risk in the validation set was 5% for the low-risk group, in contrast with 29% for the high-risk group (p = 0.0008; log-rank). The predictive performance of the wound-response local recurrence signature on the validation set, using the threshold derived from the training set, is shown in Figure 2. The sensitivity of this classifier for predicting local recurrence in the validation set was 88% (7 out of 8) with a specificity of 74% (53 out of 72). The hazard ratio was 15 (95% confidence interval (CI) 1.8 to 123; log-rank 0.01). An overview of the sensitivity and specificity for the three supervised approaches is shown in Table 3.

To assess the independent prognostic value of the supervised wound-response signature, it was tested in a Cox regression model that also included known clinico-pathological risk factors for local recurrence, using the independent validation data set. The factors that independently predicted a local recurrence in multivariable analysis in the EORTC 'boost versus no boost' trial were tested together with the local recurrence classifiers [17]. In the presence of age, diameter and boost treatment, all the gene expression classifiers were tested. Only the wound-response local recurrence signature (average local recurrence centroid) added significant prognostic information to the classical risk factors in predicting a local recurrence (hazard ratio 16 (95% CI 1.9 to 125); p = 0.01). Further results of the multivariable analysis are shown in Table 4.