Plasma miRNA Levels for Predicting Therapeutic Response to Neoadjuvant Treatment in HER2-positive Breast Cancer: Results from the NeoALTTO Trial

The HER2 gene is amplified and/or overexpressed in approximately 20% of breast cancer and has been historically associated with aggressive disease and poor prognosis (1). The combination of trastuzumab with either lapatinib or pertuzumab dramatically improved the clinical outcome of HER2-positive (HER2⁺) early and metastatic breast cancer patients (2–6). Nevertheless, dual anti-HER2 therapies do not exert the same effect in all patients and may be unnecessary in those who already benefit from single-agent trastuzumab. Tumor/host factors accounting for anti-HER2 sensitivity/resistance include the PI3K signaling (7), the hormone receptors cross-talk (8), and the mechanisms of cell-cycle control (9). Although the prognostic and predictive values of baseline PAM50, PI3K mutations, and tumor-infiltrating lymphocytes (TIL) have been addressed in several clinical trials (10–12), HER2 status in primary tumor still remains the only biomarker used in clinical practice (13, 14). miRNAs are a class of non-coding 15–27 nucleotide-long single-stranded RNAs that modulate gene expression posttranscriptionally by base-pairing to target messenger RNAs (15, 16). miRNAs play essential roles in cancer proliferation/differentiation, invasion, angiogenesis, apoptosis, and contribute to the development of innate/adaptive immune responses (17–23). miR-205, miR-21, and miR-210 are aberrantly expressed in HER2⁺ breast cancer cell lines and are associated with resistance to anti-HER2 treatment (24, 25). Notably, miRNAs are also detectable in different bio-fluids including blood (26–28). Although the precise mechanisms of miRNA release to extracellular environment remain unknown, circulating (ct)-miRNAs act as messengers by modulating certain pathways in the recipient cells and provide clinically relevant information in different tumor types (29, 30). In this study, we profiled ct-miRNAs in plasma samples obtained from patients enrolled in the NeoALTTO trial (2) before and during treatment to identify those associated with pathologic complete response (pCR) and to assess their robustness on long-term outcome.

Details on the NeoALTTO study (Breast International Group 1-06) and its results have been published elsewhere (2). According to REMARK guidelines to describe the characteristics of the study patients (31), NeoALTTO was a multicenter randomized phase III study in which HER2⁺ breast cancer patients with primary tumor >2 cm were randomly assigned to either lapatinib (n = 154), trastuzumab (n = 149), or their combination (n = 152) for 6 weeks, followed by 12 weeks with additional paclitaxel. Surgery was performed within 4 weeks from the last paclitaxel dose. After surgery, all patients received fluorouracil, epirubicin, and cyclophosphamide for three cycles and continued the same anti-HER2–targeted therapy received prior to surgery to complete 52 weeks of treatment. The primary endpoint was pCR (i.e., absence of invasive tumor cells in the breast; ref. 32). The secondary endpoint event-free survival (EFS) was defined as the time from randomization to first event (33). All patients enrolled were asked to sign the main study consent form, which included a nonspecific clause for use of blood samples collected at baseline (T0), after 2 weeks (T1), prior to surgery (T2), and eventually at the time of relapse, for future biomarker research. The study complied with the Declaration of Helsinki. The Internal Review and Ethics Boards of Fondazione IRCCS Istituto Nazionale dei Tumori (Milan, Italy) approved the miRNA study protocol.

Plasma samples stored at the central biobank of Vall d’Hebron University Hospital (Barcelona, Spain) were shipped to Fondazione IRCCS Istituto Nazionale dei Tumori. Patients (n = 429, Consort diagram; Fig. 1) with at least one plasma sample at T0 or T1 were considered suitable for the aim of the study, which was to identify early predictors of pCR. To control for data overfitting, we resorted to a training-testing approach by splitting the data in a training set (for model building) and a testing set (for model validation) to provide an estimate of the prediction error in order to evaluate the performance of miRNA signatures (34). We used computer-generated random numbers to assign 183 of these patients to the training set, and 246 to the testing set.

Procedures for plasma preparation and RNA isolation are detailed in the Supplementary Material. Reverse transcription and ct-miRNA profile was performed at Exiqon A/S, using the miRCURY LNA Universal RT microRNA PCR system according to the Exiqon manufacturer's instructions, which was characterized by good performance and robustness (35–37). On the training samples, 752 miRNAs were profiled using microRNA Ready-to-Use PCR, Human panel I+II. Individual miRNAs were then assessed using custom PCR panels (miRCURY LNA Universal RT Pick-&-Mix microRNA PCR panel from Exiqon) in the testing samples. Amplification was performed using the LightCycler 480 Real-Time PCR System (Roche) and ExiLENT SYBR Green master mix, in 384-well PCR plates. Negative controls were run in parallel.

The amplification curves were analyzed using the Roche LC software for determination of quantification cycle (Cq) values. Within each treatment arm (i.e., trastuzumab, lapatinib, and their combination) and timing of blood collection (i.e., T0 and T1), the following workflow was applied: the relative quantity (RQ) of each miRNA was calculated using the comparative threshold cycle method following the formula 2⁁- DCq where DCq = Cq miRNA - Cq reference (38). The Cq reference was computed by averaging the Cq values of the reference miRNAs identified by an updated version of the NqA algorithm (39). The RQ of each miRNA, expressed in logarithmic scale (log₂ RQ = −DCq), was then analyzed in univariate manner to identify those miRNAs statistically associated with pCR. In this selection step, we considered as potentially relevant only those miRNAs detected in at least 10 patients with pCR and 10 without pCR. Specifically, the miRNAs considered according to this criterion were 171 and 199 for lapatinib at T0 and T1, 220 and 223 for trastuzumab at T0 and T1, 276 and 279 for lapatinib + trastuzumab at T0 and T1, respectively. The miRNAs selected by the univariate analysis were combined by resorting to a multivariate logistic regression approach to identify miRNA signatures (40). Finally, to confirm the predictive capability of the identified signatures, the selected relevant miRNAs, together with the related set of references miRNAs, were evaluated in the testing sample. The area under the ROC curve (AUC) and its 95% confidence interval (CI) were calculated to estimate each signature predictive capability. REMARK recommendations (31) have been followed to describe assay methods and study design.

For the statistical analysis, we considered background filtered (BF) data as supplied by Exiqon (i.e., for assays that do not yield any signal on the negative control, the upper limit of detection was set to Cq = 37; otherwise, it was set to 3 Cq lower than the Cq value of a negative control). The BF Cq values were then processed with SAS and R softwares to obtain log₂ (RQ). In the training set, to identify differentially expressed miRNAs according to pCR, the univariate analysis was performed by resorting to the nonparametric Kruskal–Wallis test and to the logistic regression model. The relationship between each miRNA and pCR probability was investigated by resorting to a regression model based on restricted cubic splines. The procedure for the selection of miRNAs to be included in the initial multivariate models is detailed in the Supplementary Material. When number of events per variables (EPV) was <10, the selected miRNAs were combined by using a logistic regression model with least absolute shrinkage and selection operator (LASSO) penalty to identify a final model (i.e., miRNA signature). After outlier detection, the final model was fitted to the testing data by applying the same regression coefficients obtained in the training set (i.e., model validation) and by reestimating the regression coefficients (i.e., model confirmation) of the miRNAs included in the signature (41). In the latter case, according to the required EPV, a penalized logistic regression model was used whenever appropriate. The AUC and its 95% CIs were calculated to estimate the predictive capability of each model. Furthermore, in both the training and testing sets, each of the available clinicopathologic variables was added to the signature in a multivariate logistic regression model together with the corresponding first order interaction term. For each signature with a lower limit of the 95% CI of the AUC > 0.50 in the testing set (i.e., confirmed signature) an optimal cut-off value was identified by maximizing the Youden index from the ROC curve obtained by fitting a logistic regression model with the miRNAs included in the signature on the total patient series. Consequently, on the basis of the model's predicted probabilities, patients were classified as signature-positive or signature-negative. Finally, the prognostic role of each dichotomized signature was investigated in terms of EFS using a Cox regression model in uni- and multivariate fashion. The patterns of EFS were estimated using the Kaplan–Meier method and compared using log-rank test.

Among the 455 patients of the NeoALTTO trial (2), 429 (94.3%) had plasma samples available for the current study. Reasons for excluding 26 patients are reported in Fig. 1. The demographics and clinicopathologic characteristics of patients in the training (n = 183) and in the testing (n = 246) sets are outlined in Table 1. No difference was observed between training and testing sets, which were superimposable with the whole NeoALTTO patient population.

Among the 183 patients randomized in the training set, plasma of enough quantity for ct-miRNA high-throughput analysis was available in 170 and 177 samples at baseline (T0) and after 2 weeks (T1), respectively (REMARK flow diagram, Supplementary Fig. S1). ct-miRNAs underwent univariate analysis in 167 and 176 cases at T0 and T1, respectively. In the six considered scenarios, one for each treatment arm (i.e., lapatinib, trastuzumab, and their combination) and timing of blood drawn (i.e., T0 and T1), 51 ct-miRNAs were found to be differentially expressed in univariate analysis in patients achieving or not the pCR. For all these ct-miRNAs, a linear relationship between the pCR probability and their expression was found to be appropriate. Three ct-miRNAs were present in more than one scenario: miR-100-5p (in lapatinib at T0 and T1 and in the combination arm at T1), miR-377-3p (in lapatinib at T1 and in the combination arm at T0), and miR-493-5p (in the combination arm at T0 and T1). According to the fold change computed by considering the median expression value in plasma of all the patients (attaining or not pCR), a total of 38 miRNAs were upregulated and 12 downregulated. Only miR-100-5p showed a discrepant behavior resulting upregulated in patients treated with lapatinib and downregulated in patients treated with lapatinib plus trastuzumab. After excluding 17 miRNAs since undetectable in more than 15% of cases or hemolysis-related, 34 ct-miRNAs were included in the multivariate analysis and are listed in Table 2, with the number of available samples, and the identification of those entered the final models (one for each scenario). It is worth mentioning that they are detectable in almost all the samples (median value, 99%; range 68–100), as reported in Supplementary Table S1. The list of the 17 ct-miRNAs not considered in the initial multivariate is reported in Supplementary Table S2. With regard to the final LASSO regression models (defining miRNA signatures associated with pCR in the training set), statistically significant AUC values were obtained for each scenario (Supplementary Table S3; Fig. 2). Three of the identified signatures proved to be treatment and time-specific as shown by the AUC values of each signature evaluated in the other scenarios (Supplementary Table S4).

Among the 246 patients randomized in the testing set, plasma of enough quantity for ct-miRNA real-time PCR analysis was available in 224 and 228 samples at T0 and T1, respectively (Supplementary Fig. S1). Reference and signature-related ct-miRNAs were detectable and allowed the confirmation analysis in 157 and 172 cases at T0 and T1, respectively. The predictive capability of 4 of the 6 signatures identified in the training set was confirmed in the testing set (Supplementary Table S3; Fig. 2) as following: at T0 and T1 lapatinib AUC (95% CI), 0.86 (0.73–0.98) and 0.71 (0.55–0.86), respectively; at T1 trastuzumab 0.81 (0.70–0.92), and lapatinib plus trastuzumab 0.67 (0.51–0.83).

In both training and testing sets, multivariate logistic regression analyses showed that, when adjusted for each single clinicopathologic variables, including age, tumor size, nodal involvement, and estrogen receptor status (ER), the association between the confirmed signatures and pCR remained statistically significant (alpha level of 0.10). For explorative purpose, in Fig. 3, the 95% CIs of AUC computed within the categories of each clinicopathologic variables are depicted.

As a secondary endpoint, we examined whether the confirmed ct-miRNA signatures were associated with long-term outcome in the overall series (i.e., combining training and testing sets) to have sufficient numbers of events for each signature. By considering the 338 patients (137 for lapatinib, 106 for trastuzumab, and 95 for the combination) in the training and testing sets, the median follow-up was 6.7 years (IQR, 5.8–6.8), and 96 events were observed. Supplementary Figure S2 reports the EFS curves of the confirmed signatures dichotomized as described in the Statistical analysis section. The log-rank test and the HRs reported in Supplementary Table S5 indicated that none of the signatures were associated with EFS. In addition, we assessed the prognostic role of each ct-miRNA included in the confirmed signatures by resorting to a univariate Cox regression model and considering its detectable values regardless of the presence of the other miRNAs of the signature. Only ct-miR-140-5p, belonging to the signature of T1 trastuzumab, was significantly associated with EFS (n = 127; HR = 0.73; 95% CI, 0.59–0.91) when analyzed as a continuous variable (Supplementary Fig. S3A), with an increased probability of EFS by increasing the ct-miRNA expression level. Supplementary Figure S3B reports the EFS Kaplan–Meier curve (HR = 0.43; CI, 0.22–0.84; log-rank test: 0.014) for ct-miR-140-5p dichotomized according to its median value. Of note, ct-miR-140-5p maintained its prognostic significance when adjusted for ER status and pCR in a multivariate Cox regression model (continuous HR = 0.73; 95% CI, 0.59–0.91 and dichotomized HR = 0.39; 95% CI, 0.20–0.76; Supplementary Tables S6 and S7).

This study represents one of the largest retrospective analyses of prospectively collected blood samples within a randomized clinical trial evaluating ct-miRNA as a novel biomarker associated with pCR and outcome in patients with HER2-positive breast cancer. The clinical value of ct-miRNAs emerged also in other previous neoadjuvant studies. Jung and colleagues showed that ct-miR-210 levels were associated with response to neoadjuvant trastuzumab-based chemotherapy in 29 HER2⁺ breast cancer patients (42). Nevertheless, Müller and colleagues showed no predictive value of ct-miR210 in the Geparquinto study population (n = 127), which was instead found to be a negative prognostic factor (25). More recently, miR-4734 and miR-150-5p have been associated with prognosis in early breast cancer patients treated with adjuvant trastuzumab (43). These conflicting results can be ascribed to moderately sized patient cohorts, heterogeneity of disease stage or treatment type, different sample source (plasma or serum), profiling platforms, normalization approaches, and limited number of ct-miRNAs analyzed (44–47). Our ct-miRNA signatures were built on high-throughput analysis of samples collected within a controlled randomized study and were developed through a statistical strategy based on penalized regression model, which controls overfitting and is based on the principle of parsimony (40, 44). Over a total of 429 patients, miRNA profile was obtained in 306 (71%) and 328 (76%) cases at baseline and week 2, respectively. These attrition rates are acceptable taken into consideration clinical (i.e., multicentric scenario), technical (i.e., high-throughput analyses) and analytic (i.e., stringent criteria to identify and validate predictive signatures) factors. Several preclinical studies showed that the biological functions of most of the miRNAs in our validated signatures are related to important tumor-related processes such as proliferation, apoptosis, migration, and invasion (48–52). Tumor-suppressing functions has been mainly described for miR-376c-3p, miR-874-3p, miR-320c, miR-144-3p, miR-362-3p, miR-140-5p, miR-328-3p, miR-145-5p, miR-34-5p, and miR-98-5p, whereas oncogenic proprieties have been ascribed to miR-197-3p (Supplementary Table S8). Interestingly, in our study, the predictive value of basal and week two ct-miRNA signatures is mostly treatment specific. Trastuzumab and lapatinib exert their antitumor effects by different mechanisms, and their combination may act differently from the sum of the antitumor effect of each single agent (53–57). Consistent with this, and considering that this is the first time that ct-miRNAs are evaluated in patients treated with both agents either alone or in combination in a large patient population, the current findings may indicate that ct-miRNA expression during treatment can serve as a new tool for assessing how anticancer agents act as single agents and in combination, and may provide a basis for optimizing combinatorial treatment. In fact, differences in miRNAs expression profile could help not only to elucidate the distinct nature of cellular responses provoked by these agents and the mechanisms of action of dual HER2 blockade, but also to identify new targets for therapy when such mechanisms fail to work. In our study, none of the four miRNA signatures, which were able to discriminate between patients with and without pCR, were associated with EFS. NeoALTTO was not originally planned to evaluate the differences between the treatment arms in terms of EFS (33, 58), and our findings could be weakened by low statistical power. Notwithstanding, ct-miR-140-5p after 2 weeks of trastuzumab is associated with prognosis. Because of the role of miR-140-5p in the trastuzumab scenario, we performed gene enrichment analysis on a list of possible miR-140-5p gene targets by considering as putative target genes those predicted by different miRNA-target prediction tools. Enrichments of pathways related to breast cancer, apoptosis process, signal transduction involved in cell-cycle control were observed. In addition, miR-140-5p putative target genes exhibited enrichment of those involved in Wnt signaling pathway. Of note, basal miR-140-5p levels are associated with neither pCR nor EFS. A possible explanation for this divergence may be provided, if we assign to ct-miR 140-5p two independent roles in the two different specific situations, at baseline and during treatment. Baseline miR-140-5p is influenced by an untreated tumor burden and does not necessarily imply a direct relationship with response to the treatment intended to be used. The relationship with EFS in this case may thus indicate a pure prognostic value. Changes in ct-miRNA-140-5p, on the other hand, is the consequence of the delivered HER2 targeted agent and thus reflects an entirely different situation where tumor biology, treatment responsiveness, and the host interplay balance in the patient may be the primary mediators of miR-140-5p into the circulation, thus indicating a possible predictive value. In effect, miR-140-5p is certainly expressed by tumor cells, as indicated by TCGA data (http://cancergenome.nih.gov/), but its circulating levels might be the result of the contribution of different cell types, including immune components, known to be relevant for trastuzumab activity. As a matter of fact, the combined consideration of bioinformatic tools and literature search showed miR-140-5p expression pattern in normal blood cells, with an apparent enrichment in granulocytes (Supplementary Fig. S4). Thus, rather than evaluating absolute levels of single miRNA in patients with different genetic backgrounds, comorbidities, and lifestyle, our findings suggest that it is more informative to evaluate the expression patterns of miRNAs due to therapy and during the course of drug/tumor and host interplay/disease.

In conclusion, the results obtained in the current study give promising evidence for future analyses using ct-miRNAs to examine the response to anti-HER2 agents. Although confirmatory studies on independent case series are needed to validate and to evaluate the generalizability of our signature(s) before drawing more definite conclusion, the data presented here may have direct implications for future clinical trials, because miRNA analyses in plasma may be a promising strategy for predicting response to trastuzumab as monotherapy and are exploitable to guide deescalating therapy.

S. Di Cosimo reports receiving speakers bureau honoraria from Novartis Pharma. E. de Azambuja is a consultant/advisory board member for and reports receiving commercial research support from Roche. M.A. Izquierdo has ownership interests (including patents) at Novartis. J. Huober reports receiving speakers bureau honoraria from and is a consultant/advisory board member for Novartis and Roche, and reports receiving commercial research grants from Novartis. J. Baselga is an employee of AstraZeneca, Foghorn, Varian Medical Systems, Bristol-Myers Squibb, Grail, Aura, and Infinity Pharmaceuticals, has ownership interests (including patents) at PMV Pharma, Tango, Venthera, Seragon, Juno, and Northern Biologies, is a consultant/advisory board member for Eli Lilly and Novartis, and reports receiving commercial research support from Roche. F.G. de Braud reports receiving speakers bureau honoraria from BMS, Eli Lilly, Roche, Amgen, AstraZeneca, Gentili, Fondazione Menarini, Novartis, MSD, Ignyta, Bayer, Noema S.r.l., ACCMED, Dephaforum S.r.l., Nadirex, Roche, Biotechspert Lts., PriME Oncology, Pfizer, and is a consultant/advisory board member for BMS, TizianaLife Sciences, Celgene, Novartis, Servier, Phanm Research Associated, Daiichi Sankyo, Ignyta, Amgen, Pfizer, Roche, Teogarms, and Pierre Fabre. No potential conflicts of interest were disclosed by the other authors.

GlaxoSmithKline participated in the collection of biological samples but had no role in the design and conduct of this substudy; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. Novartis reviewed and approved the final version of the manuscript.

Conception and design: S. Di Cosimo, V. Appierto, S. Pizzamiglio, P. Tiberio, P. Verderio, M.G. Daidone

Development of methodology: S. Di Cosimo, S. Pizzamiglio, P. Verderio

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S. Di Cosimo, E. de Azambuja, J. Huober, J. Baselga, M. Piccart, M.G. Daidone

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Di Cosimo, V. Appierto, S. Pizzamiglio, P. Tiberio, G. Apolone, P. Verderio, M.G. Daidone

Writing, review, and/or revision of the manuscript: S. Di Cosimo, V. Appierto, S. Pizzamiglio, P. Tiberio, M.V. Iorio, F. Hilbers, E. de Azambuja, L. de la Peña, M. Izquierdo, J. Huober, J. Baselga, M. Piccart, F.G. de Braud, G. Apolone, P. Verderio, M.G. Daidone

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S. Di Cosimo, S. Pizzamiglio, P. Verderio, M.G. Daidone

Study supervision: S. Di Cosimo, P. Verderio, M.G. Daidone

This substudy of the NeoALTTO trial was supported by the Italian Association for Cancer Research (AIRC: MFAG 14361 and IG 20774 to SDC, IG 16900 to MGD) and by the Progetto Bandiera “NANOMAX” (Consiglio Nazionale delle Ricerche). Special thanks should be given to the Scientific Directorate of the Fondazione IRCCS Istituto Nazionale dei Tumori that allocated to this study part of the Italian Ministry of Health funds obtained through an Italian law that allows taxpayers to allocate the “5 × 1000” share of their payments to research. We thank Dr. Sandra Romero-Cordoba for the in silico analysis of miR-140-5p in blood cells. The NeoALTTO study was sponsored by GlaxoSmithKline; lapatinib is an asset of Novartis AG as of March 2, 2015.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.