Gene expression profiling in primary breast cancer distinguishes patients developing local recurrence after breast-conservation surgery, with or without postoperative radiotherapy

Postoperative radiotherapy with a median absorbed dose of 50 Gy (range 48 to 54 Gy) was given in 24 to 27 fractions in one series to the remaining breast parenchyma. Adjuvant systemic therapy was given to 16 patients (Tables 1 and 2).

Histological grade was re-evaluated according to Elston and Ellis [27]. ER and progesterone receptor content were analyzed routinely after primary operation with enzyme immunoassay according to kit instructions (Abbott Laboratories, Diagnostics Division, Chicago, IL, USA) and expressed as femtomoles per milligram of cytosol protein. Receptor values greater than or equal to 25 fmol/mg protein were considered positive.

RNA was extracted from freshly frozen invasive breast tumors as previously described [28]. The RNA quality was assessed using an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA), and the RNA integrity number (RIN) method [29] was used to validate the RNA quality. Twenty-one samples were excluded due to RIN values of below 6. Labeling and hybridization were performed as previously described [28]. By means of Human Genome Oligo Set Version 2.1 (containing 21,329 70-mer probes) and Human Genome Oligo Set Version 2.1 Upgrade 27 (containing 5,462 70-mer probes), oligonucleotide arrays were produced by the Swegene DNA Microarray Resource Centre, Lund University [30]. In inclusion step 1, probes were printed in duplicate, creating 55 K slides, and in inclusion step 2, in single, creating 27 K slides.

Wilcoxon rank sum tests, Sammon maps, and support vector machine (SVM) classifications [31] were performed with the statistical language R [32] using the libraries MASS (Sammon) and e1071 (SVM). For the SVM, only genes with no missing values were used. For the LR⁺RT⁺ versus LR^-RT⁺/LR^-RT^- groups, the numbers of genes with no missing values were 9,128, 13,362, and 8,834 for the ER⁺, ER^-, and combined ER groups, respectively. For the LR⁺RT^- versus LR^-RT^- groups, they were 11,209, 13,547, and 10,658, respectively. For the wound-response genes, the corresponding numbers were 93, 120, and 91 for the LR⁺RT⁺ versus LR^-RT⁺/LR^-RT^- groups and 105, 122, and 99 for the LR⁺RT^- versus LR^-RT^- groups. Leave-one-out cross-validation was used. When a sample was left out, the SVM was trained on the remaining samples, and the distance to the maximal margin hyperplane (the decision value) was calculated for the left-out sample. A linear kernel was used and the cost of constraints violation (C constant) was fixed to one. No parameter tuning was performed even if the use of another layer of cross-validation might have improved the results. The goal of this study was to prove that gene expression profiles can distinguish the groups, not to find the optimal classifier. Actually, the optimal classifier does not even need to be an SVM. We also minimized potential suspicions about information leak by restricting the parameters of the SVM to the default values of the R function svm. A receiver operating characteristic (ROC) curve and area were calculated using the decision values. The expected average value of the ROC area is 0.5 if there is no discrimination between the groups. Due to random variations, ROC areas above 0.5 are often obtained even when there is no discrimination between the groups. To distinguish a real discrimination between the groups from the case of no discrimination, a p-value was calculated. A small p-value makes it unlikely that the ROC area can be reconciled with the case of no discrimination. The p-value was calculated by a permutation test. The local recurrence labels were shuffled randomly 1,000 times and the ROC areas were found for the corresponding classifications. The P value was calculated as the fraction of the 1,000 permutations that had an ROC area larger than the real one. If the P value was zero, the random ROC areas were fitted to a normal distribution and the area under the tail above the real ROC area was used as the P value. The P value of the ROC area for the case of a fixed test set (that is, no cross-validation) was calculated by a permutation test of the labels in the test set. Odds ratios, confidence intervals of odds ratios, and P values of odds ratios were calculated with the R function Fisher test, which uses the conditional maximum likelihood estimator.

The Gene Ontology (GO) [33] OBO (open biomedical ontologies) file of 14 November 2006 was used. Gene annotation was performed using ACID (Array Clone Information Database), which is a publicly available web application that provides GO categories for genes [34]. A total of 6,841 GO categories belonging to 'Cellular component', 'Biological process', or 'Molecular function' had at least one gene in common with our data. The genes were ranked according to their Wilcoxon rank sum P value between LR⁺RT⁺ and LR^-RT⁺/LR^-RT^- groups in the ER⁺ group. A Wilcoxon rank sum test was performed for each GO category to test for over-representation of genes toward the top of the ranked gene list using Catmap [35].

To identify this group of patients, we compared LR⁺RT⁺ versus LR^-RT⁺/LR^-RT^- groups. There was an association between the LR⁺RT⁺ group and ER^- status (odds ratio 6.8, 95% confidence interval 2.0 to 24; P = 0.0007) (Table 3). In this analysis, only the patients from the first inclusion were used, as the second inclusion was made on ER and local recurrence status. Age can also distinguish the LR⁺RT⁺ group. Histological grade was marginally able to separate the LR⁺RT⁺ group from the LR^-RT⁺/LR^-RT^- group, whereas tumor size was not (Table 3).

After filtering the microarray data on low-quality spots and missing values, 26,824 reporters remained, representing 16,895 unique genes. From a Sammon map of the gene expression profiles of the 100 ER⁺ patients from both inclusions, it was evident that the LR⁺RT⁺ and LR^-RT⁺/LR^-RT^- groups were well separated even without gene selection (Figure 2). For supervised classification, an SVM was used. The areas under the receiver operating curve (ROC areas) were 0.91 (P = 9 × 10^-6) within the ER⁺ group, 0.74 (P = 0.08) within the ER^- group (Figure 3), and 0.83 (P = 9 × 10^-5) within the combined ER⁺/ER^- group (data not shown). The ER⁺ group was by far the larger group, which could explain the superior performance of the ER⁺ group compared with the ER^- one. Also, the classification performance was deteriorated by combining ER⁺ and ER^- in one SVM; it was preferable to use distinct SVMs for the two subpopulations (Figure 3). For the ER⁺ group, at 90% sensitivity (18 of 20 LR⁺RT⁺ correctly classified), the specificity was 87.5% (70 of 80 LR^-RT⁺/LR^-RT^- correctly classified), and at 90% specificity (72 of 80 LR^-RT⁺/LR^-RT^- correctly classified), the sensitivity was 80% (16 of 20 LR⁺RT⁺ correctly classified).

As age is a risk factor for local recurrence, we investigated whether the gene expression profiling has classification ability beyond that of age in the ER⁺ subgroup. We constructed a training set from the 77 patients who were either older than 50 years and in the LR^-RT⁺/LR^-RT^- group or younger than 50 years and in the LR⁺RT⁺ group. The test set consisted of the remaining 23 patients (for example, those who were either younger than 50 years and in the LR^-RT⁺/LR^-RT^- group or older than 50 years and in the LR⁺RT⁺ group). The point is that the test set chosen contains patients who behave exactly opposite to the usual connection between age and local recurrence. Applying an SVM, we obtained an ROC area of 0.88 (P = 0.002). Furthermore, we checked the influence of health care regions by using the 68 samples from the South and South-East health care regions as a training set and the 32 samples from the West or Stockholm health care regions as a test set. The specific split into health care regions was done to get a reasonable amount of samples in LR⁺RT⁺ and LR^-RT⁺/LR^-RT^- groups in both the training and the test set. No optimizations were performed in this regard. The ROC area of 0.87 (P = 0.002) shows that the classifier indeed works across health care regions.

The wound-response signature genes, also known as the core serum response genes [21], were shown to have the ability of partially predicting local recurrence [20]. We mapped the wound-response signature to our microarrays and performed an SVM classification using only this signature. The ROC areas were 0.75 (P = 0.007) within the ER⁺ group, 0.75 (P = 0.08) within the ER^- group, and 0.61 (P = 0.10) within the combined ER⁺/ER^- group.

A Wilcoxon rank sum test between LR⁺RT⁺ and LR^-RT⁺/LR^-RT^- groups within the ER⁺ subgroup was performed for all 26,824 reporters. A clear over-representation of genes with small P values was found; for example, there are 5,237 reporters with a P value below 0.001 corresponding to a Benjamini-Hochberg [36] false discovery rate of 0.005. A heatmap of the 81 genes with a known gene name, no missing values, and a Wilcoxon rank sum test P value between the LR⁺RT⁺ and LR^-RT⁺/LR^-RT^- groups of below 10^-6 is shown in Figure 4. A GO analysis was performed using Catmap [35]. A total of 6,841 GO categories belonging to 'Cellular component', 'Biological process', or 'Molecular function' were included. At a false discovery rate of 0.05, only the four categories of cytosolic ribosome (sensu Eukaryota), eukaryotic 43S preinitiation complex, eukaryotic 48S initiation complex, and cytosolic small ribosomal subunit (sensu Eukaryota) were significant.

To identify this group of patients, we analyzed LR⁺RT^- versus LR^-RT^- groups. ER^- status was weakly correlated with the LR⁺RT^- group (odds ratio 2.5, 95% confidence interval 0.6 to 11; P = 0.21) (Table 4). Young age was correlated with local recurrence in the ER⁺ group (Table 4). Neither histological grade nor tumor size had the power to separate the two groups.

An SVM gene expression classifier yielded an ROC area of 0.66 (P = 0.04) within the combined ER⁺/ER^- group. The ER^- and ER⁺ subgroups were too small to give a significant result on their own, even though there was a tendency of discriminative power within the ER+ subgroup. (ROC area = 0.62; P = 0.14). For the wound-response signature, the ROC areas were 0.64 (P = 0.10) in the ER⁺ group, 0.69 (P = 0.27) in the ER^- group, and 0.68 (P = 0.03) in the combined group.