An 8-gene qRT-PCR-based gene expression score that has prognostic value in early breast cancer

A key aspect of the management of women with early breast cancer is the selection of adjuvant therapy, which is guided by the use of prognostic factors included in the guidelines[1, 2]. However, the use of these criteria leads to unnecessary treatment in many women. Gene expression profiles may improve prognostic and predictive information in breast cancer patients, so that adjuvant chemotherapy is given only to those with the higher risk. A number of such profiles are currently available and two of them, the 70-Gene Signature (MammaPrint™) and the Recurrence Score (OncoType™) are being evaluated in phase III studies [3, 4].

The major drawbacks to the widespread use of gene expression profiling include reservations regarding their cost/effectiveness ratio and their lack of widespread availability. For some profiles, the need for fresh-frozen biological samples adds to these drawbacks. We have previously demonstrated that 1) quantitative reverse-transcriptase polymerase chain reaction (qRT-PCR) can be used to assess the genes included in the 70-Gene Signature and 2) commercially available probes can be used on formalin-fixed, paraffin-embedded (FFPE) samples to determine the 70-Gene Signature, the Recurrence Score and the Two-Gene Ratio [5, 6].

In the present study, we sought to identify a gene expression profile based on the genes included in the 70-Gene Signature, the Recurrence Score and the Two-Gene Ratio. Our hypotheses were as follows: 1) commercially available assays can be used to quantify gene expression by qRT-PCR using RNA extracted from FFPE tissues and 2) it is possible to find a prognostic gene expression profile using a reduced number of genes.

We considered 736 patients for inclusion in the study and a total of 153 patients were finally included. Additional file 2 Table S1 includes raw data of gene expression, whereas Additional file 3 Table S2 checks the REMARK recommendations along the manuscript. Table 1 describes the clinical features of and therapies received by the 153 patients. The median age was 58 years and the median follow-up was 91 months. Thirty-four patients (22%) had a distant relapse, of whom 17 died and 7 were lost to follow-up after the relapse. Among 119 patients who did not have a distant relapse, four had a local/regional recurrence that was successfully treated with surgery.

We first selected 53 genes that exhibited highly correlated expression levels in FF and FFPE samples. Seventeen genes were subsequently filtered based on their P-values related to DMFS. A model was built using these 17 genes (data not shown). Models with 10, 8, 7 or even fewer genes also demonstrated a good separation between groups at a low and high risk of distant recurrence. We selected a score based on 8 genes because it demonstrated the best performance. The 8-gene Score was calculated for each sample using reference-normalized expression measurements based on the following equation:

8-gene Score = 0.1936*DTL + 0.2176*ECT2 + 0.0454*MTDH + 0.1329*PRC1 + 0.0556*RFC4 - 0.1913*SCUBE2 - 0.0443*STK32B - 0.1182*ZNF533.

BRB ArrayTools was used to define a cutoff point for risk stratification. We assessed cut-off points leaving from 10% to 90% of patients in the low risk group increasing by 10% each time. Best performance was obtained when 60% of the patients were allocated to the low-risk group. Patients with a total score <2.86 were assigned to the low risk of recurrence group and patients with a score ≥2.86 constituted the high risk of recurrence group. The P-value of the log-rank test used to calculate recurrence risk between risk groups based on 1000 permutations was 0.018.

DMFS at five years was 97.7% for patients in the low-risk group and 60.6% for patients in the high-risk group (HR: 20.4, CI 95%: 6.2 - 67.5; p < 0.001) (Figure 1A). DMFS at five years decreased continuously as the 8-Gene Score increased (Figure 2). Overall survival at five years was also calculated, and was found to be 98.9% for the low-risk group and 86.6% for the high-risk group (HR: 7.496, CI 95%: 2.4 - 23.4; p < 0.001). A univariate analysis evaluating the effects of pathological factors on DMFS and OS is presented in Additional file 4 Table S3.

A sub-analysis was performed to evaluate the influence of lymph node status on DMFS (Figures 1B and 1C). In patients with lymph node involvement, the 8-gene Score included half of each group (low- vs. high-risk), and DMFS at five years was 93.3% vs. 39.5%, respectively. In women without lymph node involvement, the five-year DMFS rates were 100% vs. 75.7%, respectively. We also evaluated the performance in the population of patients receiving just endocrine therapy and no chemotherapy (figure 1D).

We used Cramer's V statistic to asses the concordance between other gene profiles [3, 4] and the 8-gene Score. The correlation between the 8-gene Score and the 70-Gene Signature was 0.65 and the correlation was 0.58 for the Recurrence Score and 0.30 for the Two-Gene Index. These results indicate that there is a strong correlation between the 8-gene Score and both the 70-Gene Signature and the Recurrence Score. To assess the discriminative capability of each prognostic profile at five years, Harrell's bias-corrected concordance index was calculated. The calculated values were as follows: 8-gene Score = 0.81, Recurrence Score = 0.73, 70-Gene Signature = 0.70, and Two-Gene Index = 0.59.

We then applied the 8-gene Score to an online database from the Dutch Cancer Institute (NKI) that has been previously used to compare several gene profiles[16]. The 8-gene Score identified significant differences in DMFS for the entire group of 295 patients (Figure 3), as well as for the N+, N-, and ER + subgroups based on their score (Additional file 5 table S4). In comparison with the 70-Gene Signature, the 8-gene Score categorized more patients as being low-risk, and DMFS was a bit lower for all groups. However, if the cutoff value was modified to include the same number of patients in the low-risk category as in the 70-Gene Signature, then the results for DMFS were virtually identical.

We subsequently analyzed the performance of the 8-gene Score in three additional data sets: SWE (GSE1456)[20], UPP (GSE4922)[21], and LOI (GSE6532)[22]. Although the use of other external databases does not constitute a formal validation, it may provide insight about the performance of the gene set. A similar approach has recently been reported in bladder carcinoma [25]. In all data sets, the 8-Gene Score was able to identify significant differences in DMFS (figure 3). In all cases, the 8-gene Score assigned more than 55% of the patients to the low-risk group. It predicted a DMFS rate of >90% in the LOI and SWE data sets, and >80% in the UPP data set.

In the present study, we propose a qRT-PCR-derived prognostic score based on the expression levels of eight genes (the 8-Gene Score) for use in women with early breast cancer. The score can be determined using FFPE samples. It identified patients at high risk of recurrence in our series of patients as well as in four independent databases available online.

In comparison with women included in the other online databases, all of our patients received tamoxifen and two thirds were also treated with chemotherapy. The fact that the 8-gene Score worked well in these databases reflects its strength as a prognostic tool that can be used in large populations. Our gene expression profile was developed using hormonal receptor-positive tumors and, although it was derived from a signature validated in both positive and negative tumors [4], it should not be used for women with hormone receptor-negative tumors.

The genes included in the 8-gene Score are also part of the 70-Gene Signature, but the former should not be considered a scaled down version of the latter, because the 8-Gene Score is applicable to FFPE samples and can be obtained using commercially available probes. Models with fewer than 8 genes performed similarly to the 8-gene Score, but the 8-gene Score was finally selected because it was more stable than smaller models and, at the same time, did not contain redundant information.

DTL, ECT2, MTDH, PRC1 and RFC4 have all been previously implicated in breast cancer; SCUBE2 and STK32B deletions have been found to be related to mental deficits in humans; and ZNF533 has been found to be associated with the Hedgehog signaling pathway in Zebrafish [26–35]. Of note, MTDH activation by 8q22 genomic gain promotes chemoresistance and metastasis of poor-prognosis breast cancer [36]. The eight genes themselves could play an important role in the prognosis of breast cancer. Alternatively, their expression levels could vary in response to changes in the expression of more influential genes. This means that rather than identifying the ultimate genetic cause of a cancer, gene expression profiles may instead provide information about the molecular consequences of critical mutations. In our population, the 8-gene Score identified groups of patients whose DMFS rates differed, both in the whole patient population and in the node positive and node negative subgroups (figure 1). A sub-analysis of patients receiving endocrine therapy and no chemotherapy showed the possible relevance of the 8-gene profile for decision of chemotherapy.

According to the results from the NKI series [16], the 8-gene Score performed as well as the 70-Gene Signature (Additional file 5 table S4). The DMFS and OS rate values were higher using the 70-Gene Signature, probably due to the different number of patients assigned to low- versus high-risk groups in each profile, as well as to the fact that many women in the NKI series did not receive adjuvant therapy [37].

The 8-gene Score did not perform as well in other data sets as it did in ours, probably because of substantial differences in the use of adjuvant therapy. For instance, many women with estrogen receptor-positive tumors did not receive adjuvant hormonal therapy in the UPP series, and most did not even receive chemotherapy although over one third of these patients had positive lymph nodes[21]. The series by Loi et al. included women with hormone receptor-positive tumors who were treated with tamoxifen, but chemotherapy was not used in any patient [22]. Our profile was generated in a population of women who had a very good prognosis and received optimal adjuvant therapy, and therefore the cutoff values identified using this cohort may be a bit too optimistic for "sub-optimally" treated populations. This suggests that the 8-gene Score may provide not only prognostic but also predictive information. However, because our profile works as a continuous variable, the cutoff value can be changed to become less restrictive depending on the patient group to which it is applied.

Our study has some limitations. Whereas the 70-Gene Signature and the Recurrence Score have undergone extensive investigation and clinical use, the 8-gene Score is still undergoing initial development. Our retrospective series included a limited number of patients and the treatment they received was not uniform. Furthermore, without an independent validation series, we cannot rule out overestimation of the prognostic value of the 8-gene Score. The application of our score to external data sets may temper this limitation, but it is not a formal validation. Without such validation, the score cannot facilitate treatment choices, as it neither distinguishes between therapy benefit nor identifies a group of patients that do not need any therapy. The use of different populations in the discovery phase may certainly yield different profiles, but our point was to demonstrate that a small subset of informative genes could provide a significantly prognostic score.

The 8-gene Score correlated very well with two currently available gene profiles: the 70-Gene Signature and the Recurrence Score. The discrimination capacity (Harrell's C index) was high for all three profiles. The Two-Gene Index did not perform similarly to the 8-Gene Score in our cohort of patients, but this does not demonstrate inferiority. On the other hand, the performance of this index has recently been improved with the incorporation of five additional genes related to the tumor grade [38].

We feel that the 8-gene Score has three main advantages. First, it can be used in FFPE samples with commercial probes. Second, considering the modular nature of this platform, it would be easy and inexpensive to add further genes that might provide additional information, as has been done with the Two-Gene Index [38]. Next generation profiles could include both prognostic and predictive information, for instance. Finally, qRT-PCR using TLDAs can be easily implemented in the clinical setting.

We have described an 8-gene prognostic Score that can be used in FFPE samples from patients with early hormone receptor-positive breast cancer. The main interest of this study is that it opens the door to the use of gene expression profiling based on small sets of genes in local laboratories.