Identifying the patients who would and those who would not benefit from taxanes is crucial to moving away from the "one shoe fits all" strategy. Here, we have analyzed the expression of 34 selected proteins in a subset of 1,099 patients included in the PACS01 trial.
We show that a Ki67-positive status is not only independently associated with shorter DFS, but also with the benefit of a docetaxel addition in women treated with adjuvant anthracycline-based chemotherapy. Ki67, expressed during the cell cycle, is a well-established cell proliferation marker. Its expression in breast cancer correlates with poor prognosis [35, 36] and higher response to chemotherapy. In the neo-adjuvant setting, correlation between Ki67 positivity and response to taxanes, either as monotherapy  or in association with anthracyclines , has been reported, although the relationship was not observed in other small series [39, 40]. We found a prognostic correlation much more important in docetaxel-free patients (P <0.001, log-rank test) than in docetaxel-treated patients (P = 0.048), in relation to the interaction observed with docetaxel benefit. Such interaction has been previously studied in randomized trials of adjuvant chemotherapy. Higher efficiency of the CMF regimen vs. no chemotherapy (P = 0.16 for interaction) was reported in Ki67-positive ER-positive patients treated in the NSABP-20 trial , whereas Ki67 labeling index was not predictive of better response to adjuvant chemotherapy in endocrine-responsive tumors . Bartlett et al. analyzed data of the UK NEAT/BR9601 trial, which showed benefit for the addition of anthracyclines to CMF regimen, and did not detect any interaction with anthracycline benefit for Ki67 status . The benefit of docetaxel (P = 0.11 for interaction) in Ki67-positive ER-positive patients enrolled in the PACS01 trial has been reported . Our present analysis, applied to all available PACS01 samples regardless of the ER status, showed a significant interaction, with an adjusted HR for the risk of relapse in Ki67-positive patients (HR = 0.51) equal to that previously reported for ER-positive patients only, and superior to that observed in the unselected whole population (HR = 0.81). By contrast, in an analysis of the BCIRG001 trial , Ki67 did not present any interaction with docetaxel benefit. In this study, the positivity threshold was the median value of the tested population, whereas we used the more classical 20% cut-off . We did not find in the literature any pre-clinical or clinical data in the neo-adjuvant or the metastatic setting regarding Ki67 status and the response to or the benefit associated with docetaxel specifically. Of course, our observations will require validation before application in routine. But already, they clearly suggest that patients with Ki67-positive breast cancer potentially derive a high benefit from adjuvant docetaxel (adjusted HR = 0.51), and it might be a candidate for intensifying taxane delivery (six cycles and/or a dose-dense scheme). By contrast, we did not observe any additional benefit for docetaxel in patients with Ki67-negative tumors (adjusted HR = 1.10), who represent more than two-thirds of the tested patients. A potential application might be to omit docetaxel for these patients and go back to six cycles of FEC. However, because that would increase the risk of cardiotoxicity and leukemia, it is reasonable to think that the delivery of three FEC three docetaxel cycles would remain "as good" as six FEC cycles and worthwhile giving. Whether Ki67 is mechanistically involved in the response to docetaxel remains to be demonstrated, but even if it is not, it may be a marker of a phenotype more sensitive to docetaxel. One may suppose that docetaxel, which exerts its cytotoxic effects in the G2/M phase of the cycle by inhibiting the microtubule disassembly (antimitotic effect), is more active on rapidly proliferating cells (high Ki67).
We did not find any interaction among the 33 other markers and docetaxel benefit. In fact, the relapse risk was decreased by docetaxel addition in nearly all subgroups defined according to these markers, but without significant difference for benefit between the marker-positive and marker-negative subgroups. Univariate analysis showed higher reduction of risk in ER-negative patients than in ER-positive ones, but the interaction was not significant. Similar observation was reported in the pooled analysis of two randomized trials , and the analyses of BCIRG001 , NSABP-B28 , CALGB9344 , and TACT  trials. Today, the predictive value of ER status for response to taxanes is not demonstrated when considering ER alone, and additional markers are required for identifying subgroups of ER-positive and ER-negative who most benefit from these drugs. Our previous  and present data suggest that Ki67 is a potential candidate. This predictive value is also debated for HER2. We found a significant interaction with HER2 status in univariate analysis, with more benefit of docetaxel addition in HER2-positive patients (HR for relapse: 0.46 vs 0.84), but the interaction lost significance in multivariate analysis. Analysis of two other adjuvant trials reported a benefit from the addition of taxanes in HER2-positive patients [9, 28], with significant interaction in one , but analyses of other trials yielded conflicting results [16, 18, 32, 46]. Regarding the other markers, we observed a trend towards a benefit from docetaxel in patients whose tumor was positive for CK14, Angiogenin, and β-Catenin, with a statistically borderline interaction, which deserves further analysis in larger series.
More complex molecular combinations could prove more informative than single markers for predicting docetaxel benefit [9, 29–31, 47]. Here, we have defined molecular subtypes according to the status of ER, PR, HER2 and Ki67 proteins. These subtypes displayed expected histo-clinical features. Notably, five-year DFS was relatively good in the luminal A subtype, poor in HER2-overexpressing and the triple-negative subtypes, and intermediate in the luminal B subtype, in close agreement with survival rates observed in the BCIRG001  and GEICAM9906 trials , even if the definition of luminal B and HER2-overexpressing subtypes was a little different. Clearly, the subtype that did not benefit from docetaxel in our series was luminal A, as reported in the BCIRG001 trial , and for ER-positive, HER2-negative tumors (assimilated to luminal A) in the CALGB9344  and TACT  trials. By contrast, luminal A tumors unexpectedly benefited from paclitaxel in the GEICAM9906 trial . The reasons for discrepancy between this latter observation and the former ones are unclear, and longer follow-up is required. The benefit we observed in the triple-negative tumors (adjusted HR = 0.88) was not different from the benefit observed in the whole-population, and interaction was not significant. Two subtypes benefited from docetaxel addition: luminal B and HER2-overexpressing. In luminal B patients, the absolute benefit in five-year DFS was 19% (DFS: 64% with FEC vs 83% with FEC-D); docetaxel reduced the risk of relapse by 53% after adjustment for histo-clinical variables, and the interaction was significant. Similarly, luminal B BCIRG001 patients showed a 44% reduction of relapse risk with docetaxel . This was not confirmed in the GEICAM9906 sub-study , which, however, used the same definition as the BCIRG001 trial (HR-positive, and either Ki67-positive or HER2-positive). In these two studies, the positivity cut-off was similar for ER and PR (1%), but different for Ki67 (10% in the GEICAM9906, 13% in the BCIRG001 study). From literature data and our comparison of luminal A vs luminal B tumors (Additional file 6, Table S5), some biological features of luminal B vs luminal A tumors might speculatively explain this differential benefit from docetaxel: not only the higher proliferation rate (Ki67), but also the lesser expression of BCL2 and TAU whose expression has been associated with taxane resistance [48, 49], and the higher expression of P53, whose mutations have been associated with resistance to DNA-damaging agents such as anthracyclines , and relative sensitivity to taxanes . However, none of these markers, except Ki67, was associated with docetaxel benefit in our univariate analysis (Additional file 5, Table S4). Finally, docetaxel led to a 34% adjusted reduction of the relapse risk (HR = 0.66) in the HER2-overexpressing subtype, with a borderline interaction (P = 0.14). No adjuvant trastuzumab was given in our series, but it is likely that trastuzumab would not decrease the observed docetaxel benefit. Significant interaction between HER2, treatment and outcome was found in the CALGB9344 trial, independently of ER status . Higher benefit of taxanes was also reported in HER2-positive, HR-negative BCIRG001  and TACT  patients, but not in the GEICAM9906 trial . Because the current IHC definition of luminal B and HER2-overexpressing subtypes is not consensual, we repeated analyses using definitions used by others in the BCIRG001  and GEICAM9906 trials : luminal B tumors were defined as HR-positive/HER2-negative/Ki67-positive or HR-positive/HER2-positive (N = 206, 21%), and HER2-overexpressing tumors as HR-negative/HER2-positive (N = 86, 9%). Luminal A and triple-negative subtypes were not changed. As shown in Additional file 9, Table S7, unadjusted and adjusted HR for relapse were very similar to previous analysis regarding the benefit of docetaxel per subtype, with more benefit in the luminal B and HER2-overexpressing subtypes, and no benefit in the luminal A subtype..
Our study presents several strengths (randomized prospective trial, high number of samples representative of the whole trial population and of tested proteins, including novel markers), but, like retrospective subset biomarker studies reported in adjuvant trials, suffers also from several limitations: a relatively small number of analyzed markers when compared with high-throughput profiling of frozen samples, a relatively limited proportion of available samples (55%), and limitations intrinsic to unplanned analyses. The relatively low number of events and the relatively small benefit of the experimental arm (HR = 0.81) make the study not powered enough to detect small interactions between markers and docetaxel benefit. The PACS01 trial, like other trials, was not designed to detect the benefit of taxanes in patient subgroups defined by markers. Conversely, unplanned analyses confer a risk of false positive results. For both reasons, meta-analysis of first-generation taxane trials incorporating molecular data, ideally centrally generated, is warranted in the context of international collaborations (planned future EBCTCG meta-analysis), as well as validation in ongoing second-generation taxane trials. For example, a validation study is planned to test the predictive value of Ki67 in the PACS04 trial, which compared three cycles of FEC three of docetaxel vs. six of ED (Epirubicine Docetaxel).
Other limitations are methodological (IHC) and conceptual (breast cancer heterogeneity). Breast cancer is heterogeneous. Given the extent of differences between the four molecular subtypes (luminal A and B, HER2-overexpressing, and triple-negative), and because a signal relevant in a given subtype may be diluted and undetectable in the whole population, and conversely a signal relevant in the whole population may not be detected in a given subtype, another promising approach is to redefine prognostic and predictive markers in each subtype . Results of univariate prognostic analyses are detailed in Additional file 10, Table S8, showing, for example, that the negativity of BCL2, EGFR, and TAU were significant prognosticators in the luminal A subtype, but not in the other subtypes, or that TACC2-positivity had an unfavorable prognostic value in the HER2-overexpressing subtype, but favorable in the triple-negative subtype. However, given the relatively small size of each subtype in our series, this deserves to be reassessed in a larger series and in multivariate analysis.
Regarding IHC, it is known to be non-quantitative, and it is difficult to know the most biologically relevant cut-offs for prognostic and predictive analyses. In the present study, all cut-offs were predefined before statistical analyses. For ER, PR and HER2, we used the Saint-Gallen cut-offs . For the other proteins, the challenge is more substantial. For Ki67, important guidelines are under development [42, 54–57]. A meta-analysis of 46 studies published in 2007 reported many different Ki67 cut-offs, ranging from 3.5% to 34% . The 2011 Saint-Gallen cut-off has been fixed at 14%, based on comparison with gene array data (PAM50) as a prognostic factor . However, as stated in the recommendations , the "optimal cut-points in Ki67 labeling index for prediction of efficacy of endocrine or cytotoxic therapy may vary". Here, to remain consistent with our previous PACS01 sub-study , we used a 20% cut-off. For the other 30 tested proteins, and in the absence of consensual guidelines, we used a 1% cut-off. We are aware that the chosen cut-offs may not be optimal in terms of prognostic and/or predictive values. We have thus planned the analysis of different cut-offs for each protein, which will require identification and validation in independent series. In the present series, the prognostic value of Ki67 and its value predictive for docetaxel benefit remained significant in multivariate analysis when we applied a 15% cut-off (data not shown), regardless of the cut-off for ER and PR: 1% (P = 0.028 and P for interaction = 0.033 respectively) or 10% (P = 0.022 and P for interaction = 0.031 respectively). By contrast, the favorable impact of the luminal B subtype for the docetaxel benefit lost its significance (P-value for interaction = 0.208 in multivariate analysis) when we used a 15% cut-off for Ki67 (Additional file 11, Table S9). Another analytic approach is to study the prognostic and predictive values of continuous IHC variables. As a preliminary approach, we used quartiles of IHC measurements to categorize our population (percent of stained tumor cells for the data generated on standard slides and quick score QS for the TMA data; Additional file 12, Table S10). We then assessed for each protein the magnitude of the prognostic and treatment effects as a function of such categorization. The results of multivariate analyses (Additional file 13, Table S11) showed a differential DFS according to the value of Ki67 and PR, with a maximum difference for the highest quartile (effect unfavorable for Ki67 and favorable for PR). There was also a trend towards a differential benefit of docetaxel versus no docetaxel according to the value of Ki67 and Aurora A, with a maximum benefit for the highest quartile for both proliferation markers (Additional file 14, Table S12). Complementary analyses, such as STEPP analysis (subpopulation treatment effect pattern plot) , are warranted.