Assessment of an RNA interference screen-derived mitotic and ceramide pathway metagene as a predictor of response to neoadjuvant paclitaxel for primary triple-negative breast cancer: a retrospective analysis of ﬁ ve clinical trials

Background Addition of taxanes to preoperative chemotherapy in breast cancer increases the proportion of patients who have a pathological complete response (pCR). However, a substantial proportion of patients do not respond, and the prognosis is particularly poor for patients with oestrogen-receptor (ER)/progesterone-receptor (PR)/human epidermal growth factor receptor 2 (HER2; ERBB2)-negative (triple-negative) disease who do not achieve a pCR. Reliable identiﬁ cation of such patients is the ﬁ rst step in determining who might beneﬁ t from alternative treatment regimens in clinical trials. We previously identiﬁ ed genes involved in mitosis or ceramide metabolism that inﬂ uenced sensitivity to paclitaxel, with an RNA interference (RNAi) screen in three cancer cell lines, including a triple-negative breast-cancer cell line. Here, we assess these genes as a predictor of pCR to paclitaxel combination chemotherapy in triple-negative breast cancer. 0·64 [0·43–0·81]). In multivariate logistic regression, the metagene was associated with pCR (OR 19·92, 2·62–151·57; p=0·0039) with paclitaxel-containing chemotherapy. Interpretation The paclitaxel response metagene shows promise as a paclitaxel-speciﬁ c predictor of pCR in patients with triple-negative breast cancer. The metagene is suitable for development into a reverse transcription-PCR assay, for which clinically relevant thresholds could be established in randomised clinical trials. These results highlight the potential for functional genomics to accelerate development of drug-speciﬁ c predictive biomarkers without the need for training clinical trial cohorts.


Introduction
Despite the use of modern cytotoxic agents, the proportion of patients who achieve a complete pathological response (pCR) to preoperative chemotherapy remains low, at 15-25% for all breast-cancer histopathological subtypes. 1 Rates of pCR in sporadic oestrogen-receptor (ER)/ progesterone-receptor (PR)/human epidermal growth factor receptor 2 (HER2; ERBB2)-negative (triple-negative) breast cancer range from 12% for taxane monotherapy to 45% with combination neoadjuvant chemotherapy regimens. [2][3][4] Patients who achieve a pCR after chemotherapy have excellent disease-free and overall survival. 1 The outcome for patients who do not achieve a pCR varies, and the prognosis for patients with triple-negative cancers and residual disease after preoperative chemotherapy is particularly poor. 2 Better understanding of breast-cancer biology is likely to expand the list of potentially eff ective chemotherapeutic agents in the neoadjuvant setting, and will help identify tailored chemotherapy schedules for distinct patient cohorts based on tumour molecular characterisation. If it can be reliably established that patients resistant to one type of therapy are sensitive to a diff erent agent, then robust predictors of chemotherapy response will have an essential role in selecting the optimum treatment in the neoadjuvant setting. The identifi cation of patients with disease that is resistant to conventional chemotherapy combinations is the fi rst step in this process.
Several predictive biomarkers have been discovered using associative-learning strategies. Predictive geneexpression signatures derived from an associative analysis approach are susceptible to chance associations which lead to overestimation of true clinical accuracy; therefore, two separate trial cohorts are required to train and validate the predictive signature. 5,6 Associative strategies developed from genomics signatures that are predictive of drug response in cell lines might circumvent the need for training and validation of trial cohorts. 7 Quantifying distinct biological processes within gene-expression datasets, instead of a gene-by-gene based associative analysis, may further avoid these problems and accelerate biomarker development. 8 However, potentially relevant biological processes or functional modules must be identifi ed in advance for such an analysis to be possible.
RNA interference (RNAi) functional screening may be an applicable method to identify biological processes relevant to drug response in cancer medicine. To test this hypothesis, we revisited results of our previous study 9 of an RNAi drugresistance screen across three cancer cell lines, including a triple-negative breast-cancer cell line, MDA-MB-231. In this screen, we identifi ed two distinct gene sets regulating sensitivity to paclitaxel. 9 The fi rst set of genes is involved in mitosis and the mitotic spindle assembly checkpoint (SAC), and the second set is involved in metabolism of the proapoptotic lipid, ceramide. The involvement of both of these gene sets is consistent with the current biological understanding of the mechanism of action of paclitaxel. An activated SAC orchestrates a paclitaxel-induced mitotic arrest, and in our RNAi screen, silencing several genes implicated in SAC control impaired the accumulation of cells in mitosis and subsequent cell death in response to paclitaxel. Identifi cation of ceramide pathway genes as regulators of paclitaxel sensitivity 9 is consistent with published evidence that overexpression of glucosylceramide synthase (UGCG) promotes resistance to paclitaxel and repression promotes paclitaxel sensitivity. [10][11][12] We created a metagene using established methodology 8,13 to quantify the activity of these two biological pathways identifi ed by our RNAi screen, and tested its paclitaxelpredictive value in patients with triple-negative breast cancer who were treated in clinical trials with either a paclitaxel-containing combination regimen, T-FAC (paclitaxel followed by fl uorouracil, doxorubicin, and cyclophosphamide) or by regimens without paclitaxel.

Patients and procedures
Gene-expression and treatment-response data were retrieved from six cohorts in fi ve neoadjuvant clinical trials, referred to in this study as MDA1, MDA/MAQC-II, TOP, EORTC FEC , EORTC TET , and DFCI. In all trials, pCR was determined at the time of surgery (no evidence of residual invasive cancer in the breast or lymph nodes at time of   Validated datasets from our previously published RNAi screen were used for the derivation of the paclitaxel response metagene (webappendix pp [10][11][12]. 9 The RNAi screen focused on kinases and ceramide-pathway genes. In the kinase screen, we identifi ed ten mitosis-associated genes that infl uenced sensitivity to paclitaxel in three diff erent cell lines. In addition, we identifi ed three genes from the ceramide pathway (COL4A3BP, GBA1, GBA3) that together with the published gene UGCG, 10-12 encode proximal regulators of ceramide metabolism and infl uence paclitaxel sensitivity. Of these 14 genes (fi gure 1), six mitotic genes and three ceramide genes could be assayed on the Aff ymetrix HGU133A platform, and these were combined into the mitotic and the ceramide gene sets. The number of genes within each gene set was further reduced to the mitotic module (four genes) and the ceramide module (two genes) by only including genes that were signifi cantly correlated with each other across four independent breast cancer datasets. [16][17][18][19] The expression of the genes within each module was compressed by taking the mean. Since the two modules were expected to predict sensitivity in the opposite direction (high mitotic expression=sensitivity, high ceramide expression=resistance), the paclitaxel response metagene was defi ned as the mitotic module minus the ceramide module. As a comparison, 10 000 random combinations of nine-gene sets were subjected to the same correlative approach, across four clinical datasets, and the mean expression of signifi cantly correlated genes were tested for their ability to predict pCR in T-FAC-treated cohorts. We also did a literature search and found 24 genes reported as aff ecting taxane sensitivity when either overexpressed or repressed (webappendix p 6). These genes were subjected to the same correlation analysis as the paclitaxel response metagene, and the reported direction of expression relating to paclitaxel sensitivity was included as weights (minus 1 if the gene induced sensitivity when repressed, 1 if the gene induced sensitivity when overexpressed). The genomic grade index and the stroma signatures were calculated as described previously. 8,20 Statistical analysis A binomial test was used to test for enrichment in genes that predicted response to treatment with paclitaxelcontaining neoadjuvant chemotherapy, among the genes that substantially aff ected paclitaxel sensitivity across three cell lines. The signifi cance of association between module or metagene scores and pCR was estimated with the one-sided Wilcoxon signed-rank test and plotted with receiver operating characteristics curves. We used a one-sided Wilcoxon test because we had a prior expectation about the direction of association, based on the RNAi screening results. 9 The eff ect size was estimated with AUC and logistic regression. Since the MDA1 and MDA/MAQC-II trials were done by the same investigators, at the same site, with identical geneexpression platforms, we combined the two T-FAC (B) Proximal regulators of the proapoptotic lipid, ceramide, that alter paclitaxel sensitivity through conversion to sphingomyelin (via COL4A3BP ceramide transporter) or glucosylceramide (UGCG glucosylceramide synthase) and conversion of glucosylceramide to ceramide (GBA1 and GBA3 beta-glucosidase). Genes for which repression promotes paclitaxel sensitivity are shown in green and for which repression promotes paclitaxel resistance are shown in red. RNAi=RNA interference.

Glucosylceramide
Cell survival

Role of the funding source
The funding sources and sponsors of the trials had no role in the design of the study; collection, analysis, or interpretation of the data; or writing of this report. All authors had access to the raw data. The corresponding author had full access to all data and had the fi nal responsibility to submit for publication.

Results
To develop an analytical framework based on experimentally established functional links instead of associative correlations, we revisited our previously published functional genomic RNAi screen that identifi ed genes which either promote resistance or sensitivity to paclitaxel. 9 Based on this screen and previously published data, we identifi ed two gene modules of tightly correlated genes: a four-gene mitotic module where higher expression predicted sensitivity, and a two-gene ceramide module where higher expression predicted resistance. We then used the mean expression of the genes within each module as a single predictive value; an approach previously shown to improve reliability across many tumour samples. 8,13 Finally, we combined the mitotic and ceramide modules into a single functionally derived paclitaxelresponse metagene by subtracting the mean expression of the genes in the ceramide module from the mean expression of the genes in the mitotic module (fi gure 1). Therefore, this summary measure refl ects the diff erence in the mean expression of the two modules and is predicted to correlate with paclitaxel sensitivity.
Measurements of gene expression in the cohorts treated with neoadjuvant paclitaxel were consistent with in-vitro functional observations; repression of genes in the mitotic module was associated with resistance to T-FAC therapy, and repression of genes in the ceramide module was associated with sensitivity to T-FAC. Moreover, genes whose suppression increased paclitaxel resistance across all three cancer cell lines of diff erent tissue origin were enriched in a list of genes that are diff erentially expressed in patients who were sensitive or resistant to paclitaxel (a binomial test addressing the probability that four of six  . These data suggest that the concordance of the functional genomic and clinical trial genomics datasets is unlikely to result from a chance association. For multivariate analysis, we combined the two T-FACtreated MDA1 and MDA/MAQC-II cohorts to increase statistical power. We found that the paclitaxel response metagene was the covariate most signifi cantly associated with pCR (p=0·0039; odds ratio 19·92; 95% CI 2·62-151·57) in T-FAC-treated patients with triple-negative breast cancer, more than nodal status, T stage, tumour grade, and Ki67 (table 3). The paclitaxel response metagene also did better than the genomic grade index and the stroma signature in both univariate and multivariate analysis of the combined T-FAC clinical trials, when all patients were considered and in the triple-negative cohorts (table 4). 8,20 These data support the conclusion that the paclitaxel response metagene derived from an RNAi screen has predictive power in two paclitaxel clinical trial cohorts that had no role in the discovery of the genes included in the metagene. The response metagene was not signifi cantly associated with recurrence-free survival in two untreated triple-negative breast cancer cohorts, indicating the predictive rather than prognostic power of the metagene (webappendix p 4). 16,18 To further validate the functional importance of the genes contained within the metagene and limit the possibility that the correlation step would artifi cially enrich for genes predictive of pCR, we tested the ability to predict for pCR of 10 000 random nine-gene sets subjected to the same correlative approach. None of the 10 000 combinations predicted for pCR better than the paclitaxel response metagene (p<0·0001). We addressed the relevance of the RNAi screening process to select genes for inclusion in the metagene by performing the same analysis with 24 genes reported to be associated with paclitaxel or docetaxel resistance. The literature metagene did not show any predictive value in the MDA1 or MDA/MAQC-II cohorts. In a further analysis, we eliminated the correlation step from the derivation of the paclitaxel response metagene. Although the nine-gene set (fi gure 1) selected before the expression correlation step was still predictive of pCR with T-FAC, this gene set did not do as well as the paclitaxel response metagene. These results suggest that unbiased selection of correlated genes with consistent phenotypes across an RNAi screen improves the performance of the predictive metagene (table 2).
To assess the paclitaxel specifi city of the paclitaxel response metagene, we assessed its predictive power in four cohorts that did not receive paclitaxel (table 1, webappendix p 5): the EORTC 10994 FEC trial cohort, EORTC 10994 TET trial cohort, the TOP epirubicin trial cohort, and the triple-negative DFCI cisplatin-treated cohort. The paclitaxel response metagene did not predict response in the triple-negative subtype (p>0·05, fi gure 2 C-F). Notably, the paclitaxel response metagene did not predict a signifi cant response to the TET regimen. The triple-negative tumours analysed were somewhat homogeneous (refl ecting the neoadjuvant setting of these trials), with similar nodal status, T stage, and were higher grade tumours (webappendix p 5), indicating that tumour heterogeneity is unlikely to account for this result.  Using a meta-analysis, we combined the odds ratios of the paclitaxel response metagene to pCR in the two paclitaxel-treated cohorts and in the four non-paclitaxel treated cohorts. We found that the summary odds ratio of the paclitaxel treated cohorts was 5·65 (95% CI 1·67-19·11; p=0·0053), whereas the summary odds ratio of the non-paclitaxel treated cohorts was 0·87 (95% CI 0·44 to 1·67; p=0·67), consistent with improved predictive power of the paclitaxel response metagene in paclitaxel treated cohorts (fi gure 3).
Finally, we combined paclitaxel-treated with nonpaclitaxel-treated triple-negative cohorts and did logisticregression analysis using a mixed eff ects model, considering paclitaxel treatment, binary metagene status, and their interaction. We observed a signifi cant interaction term between paclitaxel treatment and binary metagene status (OR 5·9, 95% CI 1·61-23·18; p=0·0089), indicating that the paclitaxel response metagene has paclitaxel-specifi c predictive power.

Discussion
This study supports the use of high-throughput RNAi functional genomics screening to accelerate discovery of predictive biomarkers in cancer medicine. By fi ltering for common paclitaxel resistance pathways through RNAi screening across three cell lines of diff erent tumour origin, and selecting genes which correlate across independent cohorts, we derived a paclitaxel response metagene that is predictive of T-FAC response in two clinical trial datasets, but not in cohorts treated without paclitaxel. Our results show the usefulness of this approach to identify drugspecifi c response predictors. Data reported here support a model whereby expression of genes that regulate mitotic arrest and chromosomal stability, mediated through spindle assembly checkpoint signalling, and genes that infl uence ceramide conversion to sphingomyelin or glucosylceramide, are associated with altered response to T-FAC therapy in vivo (fi gure 4). 9,21 Since the metagene was derived from an RNAi screen in a triple-negative breast-cancer cell line, we investigated the power of the metagene to predict response to paclitaxel in clinical trials that include patients with this histological subtype. The patient demographics in the four nonpaclitaxel trials included in this study are typical of the neoadjuvant setting and similar to MDA1 and MDA/MAQC-II, with a bias towards node positive, higher grade (G2-G3) and larger tumours (T2 and above), indicating that tumour heterogeneity across the trial cohorts is unlikely to contribute to the observed results. However, although the paclitaxel response metagene performs better than the genomic grade index and the stromal signature in multivariate analysis, both of which have predictive value in patients treated with T-FAC, 8,20 the AUCs reported in the MDA1 cohort result in false negatives and denial of active therapy in two of 13 (15%) of the responding patients, to spare suboptimal treatment in ten of 14 patients with resistant disease. A false negative proportion of one in 13 (8%) would be possible by sacrifi cing specifi city, resulting in the sparing of suboptimal therapy in seven of 14 patients with resistant disease. A false negative proportion of 8% would compare with the false negative proportion of ER analysis in breast cancer (Sotiriou C, unpublished data). We plan to assess whether the    performance of the paclitaxel response metagene approach might be improved through gene selection from genomewide RNAi screening approaches targeting more than 21 000 genes across multiple cell lines, by contrast with the 829 genes assessed in this study. Sources of random and systematic error should be considered when interpreting these data. Notably, the triple-negative breast-cancer cohorts are likely to be molecularly heterogeneous. Although no patients enrolled in the T-FAC trials were known to have germline BRCA mutations, the same DNA repair-pathway mechanisms may be disrupted in sporadic breast cancers that have also been implicated in taxane resistance in vitro. 22,23 Also, it should be noted that the T-FAC clinical trial datasets were acquired from fi ne needle aspirations whereas the nonpaclitaxel datasets were acquired from core biopsies. RNA yield and expression profi ling are similar using both techniques, 24 but we cannot exclude the possibility that the enrichment of stromal elements in the core biopsy datasets contribute to the lack of predictive power of the paclitaxel response metagene. There is heterogeneity between the clinical studies including the timing and measurement of pCR by diff erent pathological centres. The fi ve trials examined here varied in chemotherapy exposure from 12 to 24 weeks, and two trials used monotherapy schedules that might aff ect the proportion of patients achieving a pCR.
A metagene derived from reports of genes implicated in taxane resistance did not show T-FAC predictive power.
Furthermore, none of 10 000 random nine-gene sets performed better than the metagene. While these results support the RNAi approach to biomarker discovery and argue against the role of chance in these fi ndings, the modest cohort sizes and the heterogeneity of the nonpaclitaxel trials require replication in larger prospective studies to confi rm the relevance of this method with a clinically applicable gene-expression assay. Experience with the Oncotype DX assay has shown that RNA-based expression measurements (real-time PCR) from paraffi nfi xed tumour material can inform clinical decision making. A similar assay to assess the expression of the paclitaxel response metagene could be developed that, we estimate, would cost less than €30 per patient. With this assay, exact thresholds of mitotic and ceramide module expression should be defi ned retrospectively with tumours from the T-FAC trial before testing the defi ned threshold in a prospective trial. A randomised clinical trial comparing a paclitaxel with a non-paclitaxel regimen will be required to formally support the paclitaxel-specifi city of the metagene and the relevance of RNAi to the biomarker discovery process. The usefulness of this approach in routine clinical practice should then be assessed, since patients enrolled in clinical trials may not accurately refl ect the demographics and clinical stage of patients diagnosed with primary breast cancer.
Notably, the paclitaxel response metagene was not predictive of pCR in the EORTC TET cohort (docetaxel then epirubicin and docetaxel), which may be explained by the non-overlapping pathways of drug resistance to the two taxanes. The metagene was derived from a screen to identify mediators of paclitaxel not docetaxel resistance across three cell lines, and preclinical data has shown that docetaxel binds to β-tubulin with greater affi nity than paclitaxel and has increased interphase (G1/S/G2) cell-cycle activity, mediating cell death through the induction of BCL2 phosphorylation. 25,26 Finally, 19-31% of patients respond to paclitaxel having progressed on or after docetaxel, 27,28 supporting the divergent drugresistance mechanisms of these two taxanes.
We cannot be certain that the paclitaxel response metagene is paclitaxel specifi c, despite the test for interaction, since silencing of COL4A3BP and UGCG has been shown to promote doxorubicin sensitisation. 9,29 Consistent with this hypothesis, we note that the paclitaxel response metagene is weakly predictive of pCR in the EORTC FEC dataset across all patients (AUC 0·62; 95% CI 0·49-0·72; p=0·025). The ceramide module is likely the main contributor, since COL4A3BP expression alone predicts for pCR in all patients (0·59; 0·48-0·71; p=0·061); this would support the role of the ceramide pathway in the regulation of multidrug sensitivity in vivo.
In summary, we used in-vitro functional genomics analyses to guide the development of a metagene to predict paclitaxel response in patients with breast cancer. Although we cannot conclude that our approach is better than associative predictive strategies, the latter strategy Paclitaxel stabilises microtubules (green) and causes mitotic arrest, due to spindle assembly checkpoint signalling (red) generated by unattached kinetochores (yellow). BUB1B, TTK, and AURKB all have roles in checkpoint signalling, which maintains the activity of CDC2 to keep cells arrested in mitosis. AURKB may maintain checkpoint activation through detachment of syntelic attachments (purple rings and magnifi ed section) which may be promoted during paclitaxel exposure. Following a mitotic arrest, sensitive tumour cells undergo apoptosis. The mechanisms for this are currently unclear. Ceramide is a proapoptotic lipid that may contribute to paclitaxel-induced cell death following a mitotic arrest. Decreased levels of the mitotic module may attenuate a mitotic arrest, and consequently, cell death. Conversely, decreased levels of the ceramide transporter, COL4A3BP (CERT), and glucosylceramide synthase, UGCG, may lead to an increase in the ceramide pool and potentiate paclitaxel-induced cytotoxicity.

Checkpoint signalling
Taxane treatment requires training of two large cohorts and validation of such signatures. The functional genomics approach used in this study could be an effi cient method to accelerate biomarker development for experimental therapies in single-cohort early phase clinical trials, where stratifi cation of response according to tumour expression of a functional metagene could be considered.