- Email: [email protected]
Construction of novel immune-related signature for prediction of pathological complete response to neoadjuvant chemotherapy in human breast cancer
original articles 11.
12.
13.
14.
15.
16. Wilke H, Stahl M. Colorectal cancer and metastasectomy: treatment of advanced disease. Ann Oncol 2000; 11(Suppl 3): 45–49. 17. Cersosimo RJ. Lung cancer: a review. Am J Health Syst Pharm 2002; 59(7): 611–642. 18. Gibson L, Lawrence D, Dawson C et al. Aromatase inhibitors for treatment of advanced breast cancer in postmenopausal women. Cochrane Database Syst Rev 2009; 7(4): CD003370. 19. Kopetz S, Chang GJ, Overman MJ et al. Improved survival in metastatic colorectal cancer is associated with adoption of hepatic resection and improved chemotherapy. J Clin Oncol 2009; 27(22): 3677–3683. 20. Dawood S, Broglio K, Gonzalez-Angulo AM et al. Trends in survival over the past two decades among white and black patients with newly diagnosed stage IV breast cancer. J Clin Oncol 2008; 26(30): 4891–4898. 21. Voogd AC, van Oost FJ, Rutgers EJ et al. Long-term prognosis of patients with local recurrence after conservative surgery and radiotherapy for early breast cancer. Eur J Cancer 2005; 41(17): 2637–2644.
Annals of Oncology 25: 100–106, 2014 doi:10.1093/annonc/mdt427
Construction of novel immune-related signature for prediction of pathological complete response to neoadjuvant chemotherapy in human breast cancer Y. Sota1, Y. Naoi1, R. Tsunashima1, N. Kagara1, K. Shimazu1, N. Maruyama1, A. Shimomura1, M. Shimoda1, K. Kishi2, Y. Baba2, S. J. Kim1 & S. Noguchi1* 1
Department of Breast and Endocrine Surgery, Osaka University Graduate School of Medicine, Osaka; 2LS Business Unit, Sysmex Corporation, Kobe, Japan
Received 5 June 2013; revised 3 September 2013; accepted 4 September 2013
Background: The aim of this study was to construct a novel prediction model for the pathological complete response ( pCR) to neoadjuvant chemotherapy (NAC) using immune-related gene expression data.
Patients and methods: DNA microarray data were used to perform a gene expression analysis of tumor samples obtained before NAC from 117 primary breast cancer patients. The samples were randomly divided into the training (n = 58) and the internal validation (n = 59) sets that were used to construct the prediction model for pCR. The model was further validated using an external validation set consisting of 901 patients treated with NAC from six public datasets. Results: The training set was used to construct an immune-related 23-gene signature for NAC (IRSN-23) that is capable of classifying the patients as either genomically predicted responders (Gp-R) or non-responders (Gp-NR). IRSN-23 was first validated using an internal validation set, and the results showed that the pCR rate for Gp-R was significantly higher than that obtained for Gp-NR (38 versus 0%, P = 1.04E−04). The model was then tested using an external validation set, and this analysis showed that the pCR rate for Gp-R was also significantly higher (40 versus 11%, P = 4.98E−23). IRSN23 predicted pCR regardless of the intrinsic subtypes (PAM50) and chemotherapeutic regimens, and a multivariate analysis showed that IRSN-23 was the most important predictor of pCR (odds ratio = 4.6; 95% confidence interval = 2.7–7.7; P = 8.25E−09). Conclusion: The novel prediction model (IRSN-23) constructed with immune-related genes can predict pCR independently of the intrinsic subtypes and chemotherapeutic regimens. Key words: breast cancer, immune signature, intrinsic subtype, neoadjuvant chemotherapy, pathological complete response *Correspondence to: Prof S. Noguchi, Department of Breast and Endocrine Surgery, Osaka University Graduate School of Medicine, 2-2-E10 Yamadaoka, Suita-shi, Osaka 565-0871, Japan. Tel: +81-6-6879-3772; Fax: +81-6-6879-3779; E-mail: noguchi@ onsurg.med.osaka-u.ac.jp
© The Author 2013. Published by Oxford University Press on behalf of the European Society for Medical Oncology. All rights reserved. For permissions, please email: [email protected].
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
analysis of the Munich Cancer Registry. Breast Cancer Res Treat 2011; 128(3): 795–805. Pal SK, Dehaven M, Nelson RA et al. Impact of modern chemotherapy on the survival of women presenting with de novo metastatic breast cancer. BMC Cancer 2012; 12: 435. Dawood S, Haaland B, Albarracin CT et al. Is the proportion of patients with synchronous stage IV breast cancer surviving >2 years increasing over time? J Clin Oncol 2013; 31(Suppl. 12): Abstr 524. Chan S, Friedrichs K, Noel D et al. Prospective randomized trial of docetaxel versus doxorubicin in patients with metastatic breast cancer. J Clin Oncol 1999; 17(8): 2341–2354. Nabholtz JM, Buzdar A, Pollak M. Anastrozole is superior to tamoxifen as first-line therapy for advanced breast cancer in postmenopausal women: results of a North American multicenter randomized trial. J Clin Oncol 2000; 18(22): 3758–3767. Slamon DJ, Leyland-Jones B, Shak S et al. Use of chemotherapy plus a monoclonal antibody against HER2 for metastatic breast cancer that overexpresses HER2. N Engl J Med 2001; 344(11): 783–792.
Annals of Oncology
Annals of Oncology
introduction
patients and methods
previously [7]. The presence of tumor cells in the biopsy samples was estimated through the histological confirmation of their presence in the adjacent tumor biopsy samples, regardless of the percentage of tumor cells in the histological section.
histological evaluation of the response to chemotherapy The surgical specimens were cut into 5-mm slices, and hematoxylin and eosin-stained 3-μm sections were prepared to determine the presence or the absence of tumor cells. The complete disappearance of invasive tumor cells in the breast and negative lymph node were defined as pCR, regardless of the presence or the absence of ductal carcinoma in situ in the breast.
immunohistological examination The ER and Ki67 levels in the tumor biopsy samples were determined immunohistochemically as described previously [7]. The cut-off values for ER and Ki67 were 10 and 20%, respectively. The HER2 amplification was performed through fluorescence in situ hybridization (FISH) using the PathVysion HER-2 DNA Probe Kit (Vysis, Abbott Molecular Inc., Chicago, IL). A tumor was classified as HER2-amplified if the FISH ratio was ≥2.0. FOXP3 and CD8 were determined immunohistochemically as described previously (the 117 patients analyzed in the present study were included in a previous study involving FOXP3 and CD8 [8]).
public datasets of NAC-treated patients The patients who fulfilled the following inclusion criteria were selected from the public datasets (GEO, http://www.ncbi.nlm.nih.gov/geo/): (i) gene expression data were obtained using Affymetrix HG-U133; (ii) NAC consists of anthracycline-based regimens with or without antimicrotubule agents and does not include trastuzumab or hormonal therapy; (iii) information on the pathologically documented response is available and the definition of pCR is the same as described above and (iv) the quality of the gene expression data is sufficiently high. A total of 901 patients who fulfilled the above-mentioned criteria were selected from the six datasets (GSE16446 [9], GSE20194 [10], GSE20271 [11], GSE22093 [12], GSE23988 [4] and GSE41998 [13]) (Table 1) and served as the external validation set.
patients and tumor samples In this study, 117 breast cancer patients (stages II and III) were retrospectively recruited. These patients had been treated with NAC consisting of paclitaxel (Bristol-Myers Squibb Co., New York, USA) 80 mg/m2 weekly for 12 cycles followed by combined 5-fluorouracil (Kyowa Hakko Kirin Co., Tokyo, Japan) 500 mg/m2, epirubicin (Nippon Kayaku Co., Tokyo, Japan) 75 mg/m2 and cyclophosphamide (Shionogi & Co., Osaka, Japan) 500 mg/ m2 every 3 weeks for four cycles (P-FEC) at Osaka University Hospital (OUH) between 2002 and 2010. The patients treated with neoadjuvant trastuzumab or hormonal therapy were excluded from the study. Before NAC, all of the patients underwent a tumor biopsy with a vacuum-assisted corebiopsy instrument (Mammotome 8G; HH Ethicon Endosurgery/Johnson and Johnson Company, Langhorne, PA) under ultrasonographic guidance for histological examination and gene expression analysis. The tumor samples for histological examination were fixed in 10% buffered formaldehyde, and the tumor samples for gene expression analysis were snap-frozen in liquid nitrogen and stored at −80°C until use. Informed consent for the study was obtained from all the patients before the tumor biopsy.
RNA extraction and DNA microarray analysis The RNA extraction and DNA microarray analysis (Human Genome U133 plus 2.0 Array; Affymetrix, Santa Clara, CA) were conducted as described
Volume 25 | No. 1 | January 2014
construction of a prediction model for pCR First, the 934 probes that are classified as ‘immune response’ in the Gene Ontology Biological Process from the HG-U133A annotation file (version na32; http://www.affymetrix.com/) were selected. The 117 patients treated at OUH were randomly divided into the training set (n = 58) and the internal validation set (n = 59) using the outcome-stratified randomization method (Table 1) such that the training set included double the number of pCR cases in the validation set in order to enhance the statistical power for the detection of the differentially expressed genes between the pCR and the non-pCR cases in the training set [14]. All remaining processes were executed by using the training set only. The expression of the 30 of 934 probes was found to be significantly (P < 0.010) different between the pCR and the non-pCR cases through Welch’s t-test. Supervised analysis using DLDA (Diagonal Linear Discriminant Analysis) was employed for the construction of the prediction model using these probes. The sequential forward filtering method, which assesses the leave-one-out cross-validation, was employed to optimize the prediction model, and we found that the DLDA model comprising 23 probes (19 genes; supplementary Table S1, available at Annals of Oncology online) exhibited the highest accuracy in the training set (supplementary Figure S1, available at Annals of Oncology online) and was therefore adopted as the immune-related 23-gene signature for NAC (IRSN-23).
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Recently, an increasing number of breast cancer patients have been treated with neoadjuvant chemotherapy (NAC) to improve not only the operability of locally advanced breast cancer but also the feasibility of breast-conserving surgery for patients with a relatively large tumor. In addition, NAC yields important prognostic information, i.e. a pathological complete response ( pCR) after NAC is strongly correlated with excellent prognosis [1]. Currently, sequential chemotherapy with taxane and anthracycline-based regimens is most commonly used in NAC in clinical practice, but the pCR rate for these regimens is not very high (in the range of 10–30%) [2]. It therefore appears important to develop a predictive factor for pCR that could aid decision-making process concerning the indication for NAC. Various pathological factors, such as negative estrogen receptor (ER), high-Ki67 index, high histological grade (HG) and positive human epidermal growth factor receptor 2 (HER2), have been repeatedly shown to be significantly associated with pCR. In addition, several models for the prediction of pCR have been developed using gene expression analysis with DNA microarrays or other techniques [3]. However, these predictive factors or models are not sufficiently powerful to be incorporated into clinical practice. Recently, Iwamoto et al. [4], who performed a pathway analysis based on gene expression data from DNA microarrays, identified the importance of immune-related pathways in the prediction of pCR. More recently, Denkert et al. [5] reported that tumorinfiltrating lymphocytes (TILs) are associated with a favorable response to NAC, and Schmidt et al. [6] found that an immunerelated signature, which is represented by an IGKC-positive lymphocyte, is important for the prediction of pCR. The aim of our study was therefore to construct a novel prediction model for NAC that focuses on immune-related genes.
original articles
original articles
Annals of Oncology
Table 1. Clinicopathological characteristics of the patients analyzed (training set, internal validation set, and external validation sets) Characteristic Training Internal External validation validation All trials GSE16446 A AT AT (n = 901) (n = 114) (n = 58) (n = 59)
GSE20271 A AT (n = 87) (n = 91)
GSE20194 AT (n = 221)
GSE23988 AT (n = 61)
GSE41998 AT AIx (n = 117) (n = 128)
27 31
30 29
507 394
69 45
48 34
52 35
48 43
105 116
35 26
71 46
79 49
2 46 9 1 0
3 42 9 5 0
59 506 222 113 1
16 79 5 14 0
3 44 21 14 0
5 37 18 26 1
8 39 19 25 0
22 131 34 34 0
2 19 40 0 0
2 77 38 0 0
1 80 47 0 0
17 41 0 0 0
15 44 0 0 0
229 297 80 44 251
52 57 3 2 0
33 32 10 2 5
28 33 22 3 1
31 38 16 6 0
65 105 24 27 0
20 32 5 4 0
0 0 0 0 117
0 0 0 0 128
22 36
22 37
494 407
114 0
41 41
38 49
42 49
83 138
30 31
72 45
74 54
41 17 0
41 18 0
761 114 26
59 29 26
82 0 0
77 10 0
75 16 0
186 35 0
61 0 0
108 9 0
113 15 0
8 37 13 0
9 42 8 0
34 218 349 300
2 20 87 5
3 25 37 17
5 31 36 15
10 30 36 15
13 93 115 0
1 19 38 3
0 0 0 117
0 0 0 128
18 40
9 50
198 703
16 98
24 58
7 80
19 72
45 176
20 41
34 83
33 95
AT, anthracycline plus taxane-based sequential neoadjuvant chemotherapy; A, anthracycline-based neoadjuvant chemotherapy; AIx, anthracycline plus ixabepilone-based sequential neoadjuvant chemotherapy; cT, clinical tumor size; cN, clinical nodal status; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; HG, histological grade; pCR, pathologic complete response.
intrinsic subtype and genomic grade index The 901 samples in the external dataset were classified into luminal A, luminal B, HER2-enriched, basal-like and normal breast-like subtypes by PAM50 [15]. The normal breast-like tumor subtype was excluded from further analysis because this subtype contains only a small amount of tumor cells [16]. The genomic grade index (GGI) was calculated according to the method described by Sotiriou et al. [17].
(logistic regression model). All of the statistical analyses were two-sided, and P-values of <0.050 were considered to be statistically significant. For the hierarchical cluster analysis, Pearson’s correlation coefficients and the averaging method were used, and the results are shown as heat maps.
results
statistics
prediction of pCR by IRSN-23 in the training and internal validation sets
The gene expression data obtained from DNA microarrays were analyzed with the statistical software R (version 2.15.0; http://www.r-project.org/, accessed 2 April 2012) and Bioconductor (http://www.bioconductor.org/, accessed 2 April 2012). The SPSS software package (version 11.0.1; SPSS, Inc., Chicago, IL) was used for the following analyses, i.e. the association among the various parameters (Chi-square test and Fisher’s exact test) and the univariate and multivariate analyses of the association of the various parameters with pCR
Using IRSN-23, the patients were classified as either genomically predicted responders (Gp-R), who are likely to achieve pCR, or genomically predicted non-responders (Gp-NR), who are unlikely to achieve pCR. In the training set, the Gp-R showed a significantly higher pCR rate than the Gp-NR (62 versus 6%, P = 7.12E −06). The accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of IRSN-23
| Sota et al.
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Age ≤50 >50 cT T0–1 T2 T3 T4 Unknown cN N0 N1 N2 N3 Unknown ER Negative Positive HER2 Negative Positive Unknown HG 1 2 3 Unknown pCR Yes No
GSE22093 A (n = 82)
original articles
Annals of Oncology
A Dataset
Gp-R (n) pCR Total
GSE 16446 GSE 20194 GSE 20271 GSE 22093 GSE 23988 GSE 41998
11 35 14 18 12 44
38 77 54 36 22 112
Gp-NR (n) pCR Total 5 10 12 6 8 23
pCR rate (%) Gp-R vs. Gp-NR
76 144 124 46 39 133
29 vs. 45 vs. 26 vs. 50 vs. 55 vs. 39 vs.
7 7 10 13 21 17
Fixed effect model Random effects model
P value
5.8 11.2 3.3 6.7 4.7 3.1
1.8 – 18.2 5.1 – 24.5 1.4 – 7.7 2.3 – 19.6 1.5 – 14.6 1.7 – 5.6
1.19E-03 1.25E-11 4.78E-03 2.62E-04 6.55E-03 1.19E-04
4.8 5.0
3.4 – 6.8 3.2 – 7.9
<0.0001 <0.0001
0.2
80% pCR rate (%)
95%CI
0.5 1 2 Odds ratio
5
20
P = 4.98E-23
60% 40%
11%
40% 20% 0% Gp-R (n = 339)
Gp-NR (n = 562)
Figure 1. Evaluation of IRSN-23 with the external validation set. The predictive value of IRSN-23 for pCR was evaluated using six external validation datasets (n = 901). The odds ratios for pCR are shown in the forest plot according to each dataset (A), and the pCR rates for the Gp-R and Gp-NR in the pooled analysis of the six datasets are shown in (B). Gp-R, genomically predicted responders; Gp-NR, genomically predicted non-responders.
were 79, 89, 75, 62 and 94%, respectively. The IRSN-23 model was then applied to the internal validation set, and the Gp-R showed a significantly higher pCR rate than the Gp-NR (38 versus 0%, P = 1.04E−04). In this set, the accuracy, sensitivity, specificity, PPV and NPV of IRSN-23 were 75, 100, 70, 38 and 100%, respectively. The heat maps were generated through the hierarchical cluster analysis using the genes included in IRSN-23 in the training and internal validation sets (supplementary Figure S2, available at Annals of Oncology online).
relationships between IRSN-23 and clinic pathological parameters The Gp-R were significantly more likely to be more than 50 years of age (P = 4.66E−04), lymph node metastasis-positive (P = 0.034), ER-negative (P = 4.54E−09), HG3 (P = 0.049) and Ki67-positive (P = 0.009) and to exhibit CD8-positive T-cell infiltration (P = 0.024) and FOXP3-positive T-cell infiltration (P = 8.27E−05) compared with the Gp-NR (supplementary Table S2, available at Annals of Oncology online).
external validation of IRSN-23 using public datasets IRSN-23 was further evaluated using six external datasets (II validation set), which included 901 patients treated with NAC. Of these patients, 339 were classified as Gp-R, and 562 were classified as Gp-NR. The pCR rate for the Gp-R was significantly higher than that obtained for the Gp-NR (40 versus 11%; P = 4.98E−23), as shown in Figure 1. The accuracy, sensitivity, specificity, PPV and NPV of IRSN-23 were 70, 68, 70, 40 and
Volume 25 | No. 1 | January 2014
89%, respectively. The multivariate analysis showed that IRSN23 was the most significant [odds ratio (OR) = 4.6; 95% confidence interval (CI) = 2.7–7.7; P = 8.25E−09] predictor of pCR, followed by HER2 (P = 0.012), ER (P = 0.018) and HG (P = 0.049) (Table 2).
prediction of pCR by IRSN-23 according to the intrinsic subtype In the external validation set, the Gp-R showed a significantly higher pCR rate than the Gp-NR in the ER(+)/HER2(−), ER (−)/HER2(+) and ER(−)/HER2(−) subtypes, and a similar tendency was also observed in the ER(+)/HER2(+) subtype (Figure 2A). In each intrinsic subtype determined by PAM50, the Gp-R showed a significantly higher pCR rate than the GpNR (Figure 2B).
prediction of pCR by IRSN-23 according to the NAC regimen In the external validation set, the predictability of pCR by IRSN23 was evaluated with respect to the different NAC regimens, which include epirubicin (n = 114), FAC (5-fluorouracil, doxorubicin and cyclophosphamide) or FEC (5-fluorouracil, epirubicin and cyclophosphamide) (n = 169), sequential anthracycline and paclitaxel (n = 429), sequential anthracycline and docetaxel (n = 61) and sequential anthracycline and ixabepilone (n = 128). As shown in Figure 2C, the Gp-R showed a significantly higher pCR rate than the Gp-NR for each regimen.
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
B 100%
OR
original articles
Annals of Oncology
Table 2. Univariate and multivariate analyses of various predictive factors for pCR using the external validation datasets (n = 901) Characteristics
pCR
Univariate OR
95% CI
P-valuea
Multivariate OR 95% CI
P-valueb
394 507
75 123
1 1.4
0.9–1.9
0.060
335 565 1
71 127 0
1 1.1
0.8–1.5
0.653
403 217 281
80 32 86
1 0.7
0.5–1.1
0.115
407 494 0
42 156 0
1 4.0
2.8–5.8
1.72E−14
2.0
1.1–3.5
0.018
761 114 26
161 35 2
1 1.7
1.1–2.6
0.023
2.1
1.2–3.8
0.012
252 348 301
21 96 81
1 4.2
2.5–6.9
4.23E−09
1.8
1.1–3.2
0.049
562 339
64 134
1 5.1
3.6–7.1
4.98E−23
4.6
2.7–7.7
8.25E−09
pCR, pathological complete response; OR, odds ratio; CI, confidence interval; cT, clinical tumor size; cN, clinical nodal status; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; HG, histological grade; Gp-R, genomically predicted responders; Gp-NR, genomically predicted non-responders. a Chi-square test. Unknown data were not included in the analysis. b Logistic regression analysis. Unknown data were not included in the analysis.
combination of IRSN-23 and the GGI Because the GGI proved to be a important predictor of pCR for ER-positive tumors (P = 6.02E−08) but not for ER-negative tumors (P = 0.071), an analysis combining IRSN-23 and the GGI was performed for the ER-positive tumors only (supplementary Figure S3, available at Annals of Oncology online). The Gp-R with a high GGI showed the highest pCR rate (31%), and the Gp-NR with a low GGI showed the lowest pCR rate (5%). The Gp-R with a low GGI (20%) and the Gp-NR with a high GGI (15%) showed intermediate pCR rates.
discussion Because it is well known that TILs play a important role in the response to chemotherapy for breast cancer [5, 18], it is assumed that the immune-related signature is also associated with the response to chemotherapy [6, 19, 20]. In the current study, we analyzed all of the immune-related genes to select those that differentially expressed in pCR and non-pCR and
| Sota et al.
then used these selected genes to construct the IRSN-23 model using the training set, the diagnostic accuracy of which could be validated with the internal validation set. The IRSN-23 captures the true characteristics of TILs, as was verified through its significant association with the infiltration of CD8-positive T-cells (cytotoxic T-cells) and FOXP3-positive T-cells (regulatory T-cells, T-regs), both of which have been demonstrated to be significantly associated with pCR [8]. It is well known that non-biological experimental variations or batch effects are commonly observed across multiple batches of DNA microarray gene expression experiments, which often renders the analysis of the combined data from the multiple batches difficult [21]. Thus, our prediction model was initially constructed using our data (n = 117), which were obtained under the same experimental conditions to avoid any batch effects, rather than using the public datasets, although these included a much larger number of patients. The external validation of the IRSN-23 model was then performed using six public datasets (n = 901). The IRSN-23 model was able to classify the patients as either Gp-R or Gp-
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Age >50 ≤50 cT T3, T4 T0–1, T2 Unknown cN N1, N2, N3 N0 Unknown ER Positive Negative Unknown HER2 Negative Positive Unknown HG 1, 2 3 Unknown IRSN-23 Gp-NR Gp-R
n
original articles
Annals of Oncology
A ER and HER2
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
ER (+) ER (+) ER (–) ER (–)
12 2 19 99
26 2 12 24
HER2 (–) HER2 (+) HER2 (+) HER2 (–)
B
45 5 35 243
330 27 47 143
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
Luminal A Luminal B HER2-enriched Basal-like
4 10 26 90
19 15 4 13
14 25 52 230
266 124 37 68
27 vs. 40 vs. 54 vs. 41 vs.
8 7 26 17
pCR rate (%) Gp-R vs. Gp-NR 29 vs. 40 vs. 50 vs. 39 vs.
7 12 11 19
OR
95%CI
P value
4.3 8.3 3.5 3.4
2.0 – 9.2 0.8 – 82.9 1.4 – 8.8 2.1 – 5.7
8.93E-05 1.05E-01 7.91E-03 1.07E-06
OR
95%CI
P value
5.2 4.8 8.3 2.7
1.5 – 18.2 1.9 – 12.7 2.6 – 26.6 1.4 – 5.3
4.43E-03 6.59E-04 1.16E-04 2.30E-03
0.2
2 0.5 1 Odds ratio
5
20
0.2
2 0.5 1 Odds ratio
5
20
0.2
2 0.5 1 Odds ratio
5
20
C Chemotherapeutic Regimen
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
Epirubicin FAC or FEC A-Paclitaxel A-Docetaxel A-Ixabepilone
11 22 66 12 23
5 9 32 8 10
38 62 156 22 61
76 107 273 39 67
pCR rate (%) Gp-R vs. Gp-NR
OR
95%CI
P value
29 vs. 7 36 vs. 8 42 vs. 12 55 vs. 21 38 vs. 15
5.8 6.0 5.5 4.7 3.5
1.8 – 18.2 2.5 – 14.1 3.4 – 9.0 1.5 – 14.6 1.5 – 8.1
1.19E-03 1.17E-05 3.90E-13 6.55E-03 3.25E-03
Figure 2. Analysis of IRSN-23 using the external validation set according to the intrinsic subtypes and chemotherapeutic regimens. The odds ratios for pCR are shown in the forest plot according to the ER and HER2 status (A), the intrinsic subtypes determined by PAM50 (B), and chemotherapeutic regimens (C). FAC, 5-fluorouracil, doxorubicin and cyclophosphamide; FEC, 5-fluorouracil, epirubicin and cyclophosphamide; A-Paclitaxel, sequential anthracycline and paclitaxel; A-Docetaxel, sequential anthracycline and docetaxel; A-Ixabepilone, sequential anthracycline and ixabepilone.
NR with pCR rates of 40 and 11%, respectively (P = 4.98E −23). The multivariate analysis showed that IRSN-23 was a highly significant (P = 8.25E−09) predictor of pCR and far superior to HER2 (P = 0.012), ER (P = 0.018) and HG (P = 0.049). Among the immune-related signatures that have been found for the prediction of the response to NAC to date, IRSN-23 appears to have the lowest P-value and the highest OR [6, 19]. Because it is well established that the intrinsic subtypes determined by PAM50 are significantly associated with pCR [22], we also evaluated IRSN-23 for each intrinsic subtype and were able to show that IRSN-23 can predict pCR for each intrinsic subtype. It is noteworthy that IRSN-23 was able to predict pCR even for the luminal A subtype, even though no prediction model has yet been able to do so [19]. In addition, we were able to show that IRSN-23 can predict pCR regardless of the chemotherapeutic regimen. These results appear to indicate that IRSN-23 has a better predictive capability for pCR than the previously reported immune-related signatures and suggest that antitumor immunity is involved
Volume 25 | No. 1 | January 2014
in the response to chemotherapy, regardless of the intrinsic subtypes or chemotherapeutic regimens. IRSN-23 includes CXCL9, which is significantly up-regulated in Gp-R. This observation is consistent with the report by Denkert et al. [5], who found that CXCL9 expression is significantly associated with TILs and the response to NAC. IRSN-23 also includes IDO1, which inhibits the T-cell response by increasing the number of T-regs [23]. Wainwright et al. [24] reported that IDO deficiency reduces T-regs recruitment and enhances T-cell mediated immunity against brain tumors. The association of the intra-tumoral infiltration of T-regs with a good response to chemotherapy for breast cancer has also been reported by us and other investigators [8, 18]. The clinical value of IRSN-23 appears to be limited because its diagnostic accuracy is not sufficiently high. We thus attempted to enhance its accuracy by combining it with another signature. Because the tumor cell proliferation status and the antitumor immunity status are considered important determinants for chemosensitivity, we speculated that the combination of an immune signature with a tumor cell proliferation signature
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Intrinsic subtype
pCR rate (%) Gp-R vs. Gp-NR
original articles 7.
8.
9. 10.
11.
12.
13.
14.
funding This work was supported by Grants-in-Aid from the Knowledge Cluster Initiative of the Ministry of Education, Culture, Sports, Science and Technology of Japan and by a cooperative research fund from Sysmex Corporation.
15. 16.
17.
disclosure S.N. and S.J.K. received honoraria from Sysmex Corporation. K.K. and Y.B. are employees of Sysmex Corporation, and their role was to provide technical assistance in the statistical analysis of the DNA microarray gene expression data. All of the remaining authors have no conflicts of interest to declare.
18.
19.
20.
references 1. von Minckwitz G, Untch M, Blohmer JU et al. Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol 2012; 30: 1796–1804. 2. Sachelarie I, Grossbard ML, Chadha M et al. Primary systemic therapy of breast cancer. Oncologist 2006; 11: 574–589. 3. Fuksa L, Micuda S, Grim J et al. Predictive biomarkers in breast cancer: their value in neoadjuvant chemotherapy. Cancer Invest 2012; 30: 663–678. 4. Iwamoto T, Bianchini G, Booser D et al. Gene pathways associated with prognosis and chemotherapy sensitivity in molecular subtypes of breast cancer. J Natl Cancer Inst 2011; 103: 264–272. 5. Denkert C, Loibl S, Noske A et al. Tumor-associated lymphocytes as an independent predictor of response to neoadjuvant chemotherapy in breast cancer. J Clin Oncol 2010; 28: 105–113. 6. Schmidt M, Hellwig B, Hammad S et al. A comprehensive analysis of human gene expression profiles identifies stromal immunoglobulin kappa C as a compatible
| Sota et al.
21. 22.
23.
24.
25.
prognostic marker in human solid tumors. Clin Cancer Res 2012; 18: 2695–2703. Naoi Y, Kishi K, Tanei T et al. Prediction of pathologic complete response to sequential paclitaxel and 5-fluorouracil/epirubicin/cyclophosphamide therapy using a 70-gene classifier for breast cancers. Cancer 2011; 117: 3682–3690. Oda N, Shimazu K, Naoi Y et al. Intratumoral regulatory T cells as an independent predictive factor for pathological complete response to neoadjuvant paclitaxel followed by 5-FU/epirubicin/cyclophosphamide in breast cancer patients. Breast Cancer Res Treat 2012; 136: 107–116. Desmedt C, Di Leo A, de Azambuja E et al. Multifactorial approach to predicting resistance to anthracyclines. J Clin Oncol 2011; 29: 1578–1586. Popovici V, Chen W, Gallas BG et al. Effect of training-sample size and classification difficulty on the accuracy of genomic predictors. Breast Cancer Res 2010; 12: R5. Tabchy A, Valero V, Vidaurre T et al. Evaluation of a 30-gene paclitaxel, fluorouracil, doxorubicin, and cyclophosphamide chemotherapy response predictor in a multicenter randomized trial in breast cancer. Clin Cancer Res 2010; 16: 5351–5361. Hatzis C, Pusztai L, Valero V et al. A genomic predictor of response and survival following taxane-anthracycline chemotherapy for invasive breast cancer. JAMA 2011; 305: 1873–1881. Horak CE, Pusztai L, Xing G et al. Biomarker analysis of neoadjuvant doxorubicin/ cyclophosphamide followed by ixabepilone or paclitaxel in early-stage breast cancer. Clin Cancer Res 2013; 19: 1587–1595. Lerman J. Study design in clinical research: sample size estimation and power analysis. Can J Anaesth 1996; 43: 184–191. Parker JS, Mullins M, Cheang MC et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol 2009; 27: 1160–1167. Elloumi F, Hu Z, Li Y et al. Systematic bias in genomic classification due to contaminating non-neoplastic tissue in breast tumor samples. BMC Med Genomics 2011; 4: 54. Sotiriou C, Wirapati P, Loi S et al. Gene expression profiling in breast cancer: understanding the molecular basis of histologic grade to improve prognosis. J Natl Cancer Inst 2006; 98: 262–272. Ladoire S, Arnould L, Apetoh L et al. Pathologic complete response to neoadjuvant chemotherapy of breast carcinoma is associated with the disappearance of tumorinfiltrating foxp3+ regulatory T cells. Clin Cancer Res 2008; 14: 2413–2420. Ignatiadis M, Singhal SK, Desmedt C et al. Gene modules and response to neoadjuvant chemotherapy in breast cancer subtypes: a pooled analysis. J Clin Oncol 2012; 30: 1996–2004. West NR, Milne K, Truong PT et al. Tumor-infiltrating lymphocytes predict response to anthracycline-based chemotherapy in estrogen receptor-negative breast cancer. Breast Cancer Res 2011; 13: R126. Leek JT, Scharpf RB, Bravo HC et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet 2010; 11: 733–739. Prat A, Parker JS, Fan C et al. PAM50 Assay and the three-gene model for identifying the major and clinically relevant molecular subtypes of breast cancer. Breast Cancer Res Treat 2012; 135: 301–306. Viguier M, Lemaitre F, Verola O et al. Foxp3 expressing CD4+CD25(high) regulatory T cells are overrepresented in human metastatic melanoma lymph nodes and inhibit the function of infiltrating T cells. J Immunol 2004; 173: 1444–1453. Wainwright DA, Balyasnikova IV, Chang AL et al. IDO Expression in brain tumors increases the recruitment of regulatory T cells and negatively impacts survival. Clin Cancer Res 2012; 18: 6110–6121. Liedtke C, Hatzis C, Symmans WF et al. Genomic grade index is associated with response to chemotherapy in patients with breast cancer. J Clin Oncol 2009; 27: 3185–3191.
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
would be able to predict pCR with greater accuracy. The GGI reflects the tumor cell proliferation status, and its association with pCR has been reported previously [25]. We demonstrated in our study that the combination of IRSN-23 and the GGI provides superior prediction of pCR in ER-positive tumors than either signature alone. In particular, this combination identified as many as 61% of the patients with ER-positive tumors who were very unlikely to achieve pCR ( pCR rate = 5%). From a clinical point of view, this information appears to be useful for the decision-making process regarding the indication of NAC for ER-positive tumors because patients with ER-positive tumors classified as Gp-NR with a low GGI, particularly those patients with clinically negative nodes, may be reasonable candidates not for NAC but for neoadjuvant hormonal therapy. In conclusion, we were able to construct IRSN-23, which can predict the response to NAC with greater accuracy than the conventional parameters and independently of the intrinsic subtypes and chemotherapeutic regimens. Our results are consistent with the hypothesis that TILs play an important role in the response to chemotherapy. However, prospective studies are needed to confirm the clinical utility of IRSN-23.
Annals of Oncology
12.
13.
14.
15.
16. Wilke H, Stahl M. Colorectal cancer and metastasectomy: treatment of advanced disease. Ann Oncol 2000; 11(Suppl 3): 45–49. 17. Cersosimo RJ. Lung cancer: a review. Am J Health Syst Pharm 2002; 59(7): 611–642. 18. Gibson L, Lawrence D, Dawson C et al. Aromatase inhibitors for treatment of advanced breast cancer in postmenopausal women. Cochrane Database Syst Rev 2009; 7(4): CD003370. 19. Kopetz S, Chang GJ, Overman MJ et al. Improved survival in metastatic colorectal cancer is associated with adoption of hepatic resection and improved chemotherapy. J Clin Oncol 2009; 27(22): 3677–3683. 20. Dawood S, Broglio K, Gonzalez-Angulo AM et al. Trends in survival over the past two decades among white and black patients with newly diagnosed stage IV breast cancer. J Clin Oncol 2008; 26(30): 4891–4898. 21. Voogd AC, van Oost FJ, Rutgers EJ et al. Long-term prognosis of patients with local recurrence after conservative surgery and radiotherapy for early breast cancer. Eur J Cancer 2005; 41(17): 2637–2644.
Annals of Oncology 25: 100–106, 2014 doi:10.1093/annonc/mdt427
Construction of novel immune-related signature for prediction of pathological complete response to neoadjuvant chemotherapy in human breast cancer Y. Sota1, Y. Naoi1, R. Tsunashima1, N. Kagara1, K. Shimazu1, N. Maruyama1, A. Shimomura1, M. Shimoda1, K. Kishi2, Y. Baba2, S. J. Kim1 & S. Noguchi1* 1
Department of Breast and Endocrine Surgery, Osaka University Graduate School of Medicine, Osaka; 2LS Business Unit, Sysmex Corporation, Kobe, Japan
Received 5 June 2013; revised 3 September 2013; accepted 4 September 2013
Background: The aim of this study was to construct a novel prediction model for the pathological complete response ( pCR) to neoadjuvant chemotherapy (NAC) using immune-related gene expression data.
Patients and methods: DNA microarray data were used to perform a gene expression analysis of tumor samples obtained before NAC from 117 primary breast cancer patients. The samples were randomly divided into the training (n = 58) and the internal validation (n = 59) sets that were used to construct the prediction model for pCR. The model was further validated using an external validation set consisting of 901 patients treated with NAC from six public datasets. Results: The training set was used to construct an immune-related 23-gene signature for NAC (IRSN-23) that is capable of classifying the patients as either genomically predicted responders (Gp-R) or non-responders (Gp-NR). IRSN-23 was first validated using an internal validation set, and the results showed that the pCR rate for Gp-R was significantly higher than that obtained for Gp-NR (38 versus 0%, P = 1.04E−04). The model was then tested using an external validation set, and this analysis showed that the pCR rate for Gp-R was also significantly higher (40 versus 11%, P = 4.98E−23). IRSN23 predicted pCR regardless of the intrinsic subtypes (PAM50) and chemotherapeutic regimens, and a multivariate analysis showed that IRSN-23 was the most important predictor of pCR (odds ratio = 4.6; 95% confidence interval = 2.7–7.7; P = 8.25E−09). Conclusion: The novel prediction model (IRSN-23) constructed with immune-related genes can predict pCR independently of the intrinsic subtypes and chemotherapeutic regimens. Key words: breast cancer, immune signature, intrinsic subtype, neoadjuvant chemotherapy, pathological complete response *Correspondence to: Prof S. Noguchi, Department of Breast and Endocrine Surgery, Osaka University Graduate School of Medicine, 2-2-E10 Yamadaoka, Suita-shi, Osaka 565-0871, Japan. Tel: +81-6-6879-3772; Fax: +81-6-6879-3779; E-mail: noguchi@ onsurg.med.osaka-u.ac.jp
© The Author 2013. Published by Oxford University Press on behalf of the European Society for Medical Oncology. All rights reserved. For permissions, please email: [email protected].
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
analysis of the Munich Cancer Registry. Breast Cancer Res Treat 2011; 128(3): 795–805. Pal SK, Dehaven M, Nelson RA et al. Impact of modern chemotherapy on the survival of women presenting with de novo metastatic breast cancer. BMC Cancer 2012; 12: 435. Dawood S, Haaland B, Albarracin CT et al. Is the proportion of patients with synchronous stage IV breast cancer surviving >2 years increasing over time? J Clin Oncol 2013; 31(Suppl. 12): Abstr 524. Chan S, Friedrichs K, Noel D et al. Prospective randomized trial of docetaxel versus doxorubicin in patients with metastatic breast cancer. J Clin Oncol 1999; 17(8): 2341–2354. Nabholtz JM, Buzdar A, Pollak M. Anastrozole is superior to tamoxifen as first-line therapy for advanced breast cancer in postmenopausal women: results of a North American multicenter randomized trial. J Clin Oncol 2000; 18(22): 3758–3767. Slamon DJ, Leyland-Jones B, Shak S et al. Use of chemotherapy plus a monoclonal antibody against HER2 for metastatic breast cancer that overexpresses HER2. N Engl J Med 2001; 344(11): 783–792.
Annals of Oncology
Annals of Oncology
introduction
patients and methods
previously [7]. The presence of tumor cells in the biopsy samples was estimated through the histological confirmation of their presence in the adjacent tumor biopsy samples, regardless of the percentage of tumor cells in the histological section.
histological evaluation of the response to chemotherapy The surgical specimens were cut into 5-mm slices, and hematoxylin and eosin-stained 3-μm sections were prepared to determine the presence or the absence of tumor cells. The complete disappearance of invasive tumor cells in the breast and negative lymph node were defined as pCR, regardless of the presence or the absence of ductal carcinoma in situ in the breast.
immunohistological examination The ER and Ki67 levels in the tumor biopsy samples were determined immunohistochemically as described previously [7]. The cut-off values for ER and Ki67 were 10 and 20%, respectively. The HER2 amplification was performed through fluorescence in situ hybridization (FISH) using the PathVysion HER-2 DNA Probe Kit (Vysis, Abbott Molecular Inc., Chicago, IL). A tumor was classified as HER2-amplified if the FISH ratio was ≥2.0. FOXP3 and CD8 were determined immunohistochemically as described previously (the 117 patients analyzed in the present study were included in a previous study involving FOXP3 and CD8 [8]).
public datasets of NAC-treated patients The patients who fulfilled the following inclusion criteria were selected from the public datasets (GEO, http://www.ncbi.nlm.nih.gov/geo/): (i) gene expression data were obtained using Affymetrix HG-U133; (ii) NAC consists of anthracycline-based regimens with or without antimicrotubule agents and does not include trastuzumab or hormonal therapy; (iii) information on the pathologically documented response is available and the definition of pCR is the same as described above and (iv) the quality of the gene expression data is sufficiently high. A total of 901 patients who fulfilled the above-mentioned criteria were selected from the six datasets (GSE16446 [9], GSE20194 [10], GSE20271 [11], GSE22093 [12], GSE23988 [4] and GSE41998 [13]) (Table 1) and served as the external validation set.
patients and tumor samples In this study, 117 breast cancer patients (stages II and III) were retrospectively recruited. These patients had been treated with NAC consisting of paclitaxel (Bristol-Myers Squibb Co., New York, USA) 80 mg/m2 weekly for 12 cycles followed by combined 5-fluorouracil (Kyowa Hakko Kirin Co., Tokyo, Japan) 500 mg/m2, epirubicin (Nippon Kayaku Co., Tokyo, Japan) 75 mg/m2 and cyclophosphamide (Shionogi & Co., Osaka, Japan) 500 mg/ m2 every 3 weeks for four cycles (P-FEC) at Osaka University Hospital (OUH) between 2002 and 2010. The patients treated with neoadjuvant trastuzumab or hormonal therapy were excluded from the study. Before NAC, all of the patients underwent a tumor biopsy with a vacuum-assisted corebiopsy instrument (Mammotome 8G; HH Ethicon Endosurgery/Johnson and Johnson Company, Langhorne, PA) under ultrasonographic guidance for histological examination and gene expression analysis. The tumor samples for histological examination were fixed in 10% buffered formaldehyde, and the tumor samples for gene expression analysis were snap-frozen in liquid nitrogen and stored at −80°C until use. Informed consent for the study was obtained from all the patients before the tumor biopsy.
RNA extraction and DNA microarray analysis The RNA extraction and DNA microarray analysis (Human Genome U133 plus 2.0 Array; Affymetrix, Santa Clara, CA) were conducted as described
Volume 25 | No. 1 | January 2014
construction of a prediction model for pCR First, the 934 probes that are classified as ‘immune response’ in the Gene Ontology Biological Process from the HG-U133A annotation file (version na32; http://www.affymetrix.com/) were selected. The 117 patients treated at OUH were randomly divided into the training set (n = 58) and the internal validation set (n = 59) using the outcome-stratified randomization method (Table 1) such that the training set included double the number of pCR cases in the validation set in order to enhance the statistical power for the detection of the differentially expressed genes between the pCR and the non-pCR cases in the training set [14]. All remaining processes were executed by using the training set only. The expression of the 30 of 934 probes was found to be significantly (P < 0.010) different between the pCR and the non-pCR cases through Welch’s t-test. Supervised analysis using DLDA (Diagonal Linear Discriminant Analysis) was employed for the construction of the prediction model using these probes. The sequential forward filtering method, which assesses the leave-one-out cross-validation, was employed to optimize the prediction model, and we found that the DLDA model comprising 23 probes (19 genes; supplementary Table S1, available at Annals of Oncology online) exhibited the highest accuracy in the training set (supplementary Figure S1, available at Annals of Oncology online) and was therefore adopted as the immune-related 23-gene signature for NAC (IRSN-23).
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Recently, an increasing number of breast cancer patients have been treated with neoadjuvant chemotherapy (NAC) to improve not only the operability of locally advanced breast cancer but also the feasibility of breast-conserving surgery for patients with a relatively large tumor. In addition, NAC yields important prognostic information, i.e. a pathological complete response ( pCR) after NAC is strongly correlated with excellent prognosis [1]. Currently, sequential chemotherapy with taxane and anthracycline-based regimens is most commonly used in NAC in clinical practice, but the pCR rate for these regimens is not very high (in the range of 10–30%) [2]. It therefore appears important to develop a predictive factor for pCR that could aid decision-making process concerning the indication for NAC. Various pathological factors, such as negative estrogen receptor (ER), high-Ki67 index, high histological grade (HG) and positive human epidermal growth factor receptor 2 (HER2), have been repeatedly shown to be significantly associated with pCR. In addition, several models for the prediction of pCR have been developed using gene expression analysis with DNA microarrays or other techniques [3]. However, these predictive factors or models are not sufficiently powerful to be incorporated into clinical practice. Recently, Iwamoto et al. [4], who performed a pathway analysis based on gene expression data from DNA microarrays, identified the importance of immune-related pathways in the prediction of pCR. More recently, Denkert et al. [5] reported that tumorinfiltrating lymphocytes (TILs) are associated with a favorable response to NAC, and Schmidt et al. [6] found that an immunerelated signature, which is represented by an IGKC-positive lymphocyte, is important for the prediction of pCR. The aim of our study was therefore to construct a novel prediction model for NAC that focuses on immune-related genes.
original articles
original articles
Annals of Oncology
Table 1. Clinicopathological characteristics of the patients analyzed (training set, internal validation set, and external validation sets) Characteristic Training Internal External validation validation All trials GSE16446 A AT AT (n = 901) (n = 114) (n = 58) (n = 59)
GSE20271 A AT (n = 87) (n = 91)
GSE20194 AT (n = 221)
GSE23988 AT (n = 61)
GSE41998 AT AIx (n = 117) (n = 128)
27 31
30 29
507 394
69 45
48 34
52 35
48 43
105 116
35 26
71 46
79 49
2 46 9 1 0
3 42 9 5 0
59 506 222 113 1
16 79 5 14 0
3 44 21 14 0
5 37 18 26 1
8 39 19 25 0
22 131 34 34 0
2 19 40 0 0
2 77 38 0 0
1 80 47 0 0
17 41 0 0 0
15 44 0 0 0
229 297 80 44 251
52 57 3 2 0
33 32 10 2 5
28 33 22 3 1
31 38 16 6 0
65 105 24 27 0
20 32 5 4 0
0 0 0 0 117
0 0 0 0 128
22 36
22 37
494 407
114 0
41 41
38 49
42 49
83 138
30 31
72 45
74 54
41 17 0
41 18 0
761 114 26
59 29 26
82 0 0
77 10 0
75 16 0
186 35 0
61 0 0
108 9 0
113 15 0
8 37 13 0
9 42 8 0
34 218 349 300
2 20 87 5
3 25 37 17
5 31 36 15
10 30 36 15
13 93 115 0
1 19 38 3
0 0 0 117
0 0 0 128
18 40
9 50
198 703
16 98
24 58
7 80
19 72
45 176
20 41
34 83
33 95
AT, anthracycline plus taxane-based sequential neoadjuvant chemotherapy; A, anthracycline-based neoadjuvant chemotherapy; AIx, anthracycline plus ixabepilone-based sequential neoadjuvant chemotherapy; cT, clinical tumor size; cN, clinical nodal status; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; HG, histological grade; pCR, pathologic complete response.
intrinsic subtype and genomic grade index The 901 samples in the external dataset were classified into luminal A, luminal B, HER2-enriched, basal-like and normal breast-like subtypes by PAM50 [15]. The normal breast-like tumor subtype was excluded from further analysis because this subtype contains only a small amount of tumor cells [16]. The genomic grade index (GGI) was calculated according to the method described by Sotiriou et al. [17].
(logistic regression model). All of the statistical analyses were two-sided, and P-values of <0.050 were considered to be statistically significant. For the hierarchical cluster analysis, Pearson’s correlation coefficients and the averaging method were used, and the results are shown as heat maps.
results
statistics
prediction of pCR by IRSN-23 in the training and internal validation sets
The gene expression data obtained from DNA microarrays were analyzed with the statistical software R (version 2.15.0; http://www.r-project.org/, accessed 2 April 2012) and Bioconductor (http://www.bioconductor.org/, accessed 2 April 2012). The SPSS software package (version 11.0.1; SPSS, Inc., Chicago, IL) was used for the following analyses, i.e. the association among the various parameters (Chi-square test and Fisher’s exact test) and the univariate and multivariate analyses of the association of the various parameters with pCR
Using IRSN-23, the patients were classified as either genomically predicted responders (Gp-R), who are likely to achieve pCR, or genomically predicted non-responders (Gp-NR), who are unlikely to achieve pCR. In the training set, the Gp-R showed a significantly higher pCR rate than the Gp-NR (62 versus 6%, P = 7.12E −06). The accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of IRSN-23
| Sota et al.
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Age ≤50 >50 cT T0–1 T2 T3 T4 Unknown cN N0 N1 N2 N3 Unknown ER Negative Positive HER2 Negative Positive Unknown HG 1 2 3 Unknown pCR Yes No
GSE22093 A (n = 82)
original articles
Annals of Oncology
A Dataset
Gp-R (n) pCR Total
GSE 16446 GSE 20194 GSE 20271 GSE 22093 GSE 23988 GSE 41998
11 35 14 18 12 44
38 77 54 36 22 112
Gp-NR (n) pCR Total 5 10 12 6 8 23
pCR rate (%) Gp-R vs. Gp-NR
76 144 124 46 39 133
29 vs. 45 vs. 26 vs. 50 vs. 55 vs. 39 vs.
7 7 10 13 21 17
Fixed effect model Random effects model
P value
5.8 11.2 3.3 6.7 4.7 3.1
1.8 – 18.2 5.1 – 24.5 1.4 – 7.7 2.3 – 19.6 1.5 – 14.6 1.7 – 5.6
1.19E-03 1.25E-11 4.78E-03 2.62E-04 6.55E-03 1.19E-04
4.8 5.0
3.4 – 6.8 3.2 – 7.9
<0.0001 <0.0001
0.2
80% pCR rate (%)
95%CI
0.5 1 2 Odds ratio
5
20
P = 4.98E-23
60% 40%
11%
40% 20% 0% Gp-R (n = 339)
Gp-NR (n = 562)
Figure 1. Evaluation of IRSN-23 with the external validation set. The predictive value of IRSN-23 for pCR was evaluated using six external validation datasets (n = 901). The odds ratios for pCR are shown in the forest plot according to each dataset (A), and the pCR rates for the Gp-R and Gp-NR in the pooled analysis of the six datasets are shown in (B). Gp-R, genomically predicted responders; Gp-NR, genomically predicted non-responders.
were 79, 89, 75, 62 and 94%, respectively. The IRSN-23 model was then applied to the internal validation set, and the Gp-R showed a significantly higher pCR rate than the Gp-NR (38 versus 0%, P = 1.04E−04). In this set, the accuracy, sensitivity, specificity, PPV and NPV of IRSN-23 were 75, 100, 70, 38 and 100%, respectively. The heat maps were generated through the hierarchical cluster analysis using the genes included in IRSN-23 in the training and internal validation sets (supplementary Figure S2, available at Annals of Oncology online).
relationships between IRSN-23 and clinic pathological parameters The Gp-R were significantly more likely to be more than 50 years of age (P = 4.66E−04), lymph node metastasis-positive (P = 0.034), ER-negative (P = 4.54E−09), HG3 (P = 0.049) and Ki67-positive (P = 0.009) and to exhibit CD8-positive T-cell infiltration (P = 0.024) and FOXP3-positive T-cell infiltration (P = 8.27E−05) compared with the Gp-NR (supplementary Table S2, available at Annals of Oncology online).
external validation of IRSN-23 using public datasets IRSN-23 was further evaluated using six external datasets (II validation set), which included 901 patients treated with NAC. Of these patients, 339 were classified as Gp-R, and 562 were classified as Gp-NR. The pCR rate for the Gp-R was significantly higher than that obtained for the Gp-NR (40 versus 11%; P = 4.98E−23), as shown in Figure 1. The accuracy, sensitivity, specificity, PPV and NPV of IRSN-23 were 70, 68, 70, 40 and
Volume 25 | No. 1 | January 2014
89%, respectively. The multivariate analysis showed that IRSN23 was the most significant [odds ratio (OR) = 4.6; 95% confidence interval (CI) = 2.7–7.7; P = 8.25E−09] predictor of pCR, followed by HER2 (P = 0.012), ER (P = 0.018) and HG (P = 0.049) (Table 2).
prediction of pCR by IRSN-23 according to the intrinsic subtype In the external validation set, the Gp-R showed a significantly higher pCR rate than the Gp-NR in the ER(+)/HER2(−), ER (−)/HER2(+) and ER(−)/HER2(−) subtypes, and a similar tendency was also observed in the ER(+)/HER2(+) subtype (Figure 2A). In each intrinsic subtype determined by PAM50, the Gp-R showed a significantly higher pCR rate than the GpNR (Figure 2B).
prediction of pCR by IRSN-23 according to the NAC regimen In the external validation set, the predictability of pCR by IRSN23 was evaluated with respect to the different NAC regimens, which include epirubicin (n = 114), FAC (5-fluorouracil, doxorubicin and cyclophosphamide) or FEC (5-fluorouracil, epirubicin and cyclophosphamide) (n = 169), sequential anthracycline and paclitaxel (n = 429), sequential anthracycline and docetaxel (n = 61) and sequential anthracycline and ixabepilone (n = 128). As shown in Figure 2C, the Gp-R showed a significantly higher pCR rate than the Gp-NR for each regimen.
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
B 100%
OR
original articles
Annals of Oncology
Table 2. Univariate and multivariate analyses of various predictive factors for pCR using the external validation datasets (n = 901) Characteristics
pCR
Univariate OR
95% CI
P-valuea
Multivariate OR 95% CI
P-valueb
394 507
75 123
1 1.4
0.9–1.9
0.060
335 565 1
71 127 0
1 1.1
0.8–1.5
0.653
403 217 281
80 32 86
1 0.7
0.5–1.1
0.115
407 494 0
42 156 0
1 4.0
2.8–5.8
1.72E−14
2.0
1.1–3.5
0.018
761 114 26
161 35 2
1 1.7
1.1–2.6
0.023
2.1
1.2–3.8
0.012
252 348 301
21 96 81
1 4.2
2.5–6.9
4.23E−09
1.8
1.1–3.2
0.049
562 339
64 134
1 5.1
3.6–7.1
4.98E−23
4.6
2.7–7.7
8.25E−09
pCR, pathological complete response; OR, odds ratio; CI, confidence interval; cT, clinical tumor size; cN, clinical nodal status; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; HG, histological grade; Gp-R, genomically predicted responders; Gp-NR, genomically predicted non-responders. a Chi-square test. Unknown data were not included in the analysis. b Logistic regression analysis. Unknown data were not included in the analysis.
combination of IRSN-23 and the GGI Because the GGI proved to be a important predictor of pCR for ER-positive tumors (P = 6.02E−08) but not for ER-negative tumors (P = 0.071), an analysis combining IRSN-23 and the GGI was performed for the ER-positive tumors only (supplementary Figure S3, available at Annals of Oncology online). The Gp-R with a high GGI showed the highest pCR rate (31%), and the Gp-NR with a low GGI showed the lowest pCR rate (5%). The Gp-R with a low GGI (20%) and the Gp-NR with a high GGI (15%) showed intermediate pCR rates.
discussion Because it is well known that TILs play a important role in the response to chemotherapy for breast cancer [5, 18], it is assumed that the immune-related signature is also associated with the response to chemotherapy [6, 19, 20]. In the current study, we analyzed all of the immune-related genes to select those that differentially expressed in pCR and non-pCR and
| Sota et al.
then used these selected genes to construct the IRSN-23 model using the training set, the diagnostic accuracy of which could be validated with the internal validation set. The IRSN-23 captures the true characteristics of TILs, as was verified through its significant association with the infiltration of CD8-positive T-cells (cytotoxic T-cells) and FOXP3-positive T-cells (regulatory T-cells, T-regs), both of which have been demonstrated to be significantly associated with pCR [8]. It is well known that non-biological experimental variations or batch effects are commonly observed across multiple batches of DNA microarray gene expression experiments, which often renders the analysis of the combined data from the multiple batches difficult [21]. Thus, our prediction model was initially constructed using our data (n = 117), which were obtained under the same experimental conditions to avoid any batch effects, rather than using the public datasets, although these included a much larger number of patients. The external validation of the IRSN-23 model was then performed using six public datasets (n = 901). The IRSN-23 model was able to classify the patients as either Gp-R or Gp-
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Age >50 ≤50 cT T3, T4 T0–1, T2 Unknown cN N1, N2, N3 N0 Unknown ER Positive Negative Unknown HER2 Negative Positive Unknown HG 1, 2 3 Unknown IRSN-23 Gp-NR Gp-R
n
original articles
Annals of Oncology
A ER and HER2
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
ER (+) ER (+) ER (–) ER (–)
12 2 19 99
26 2 12 24
HER2 (–) HER2 (+) HER2 (+) HER2 (–)
B
45 5 35 243
330 27 47 143
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
Luminal A Luminal B HER2-enriched Basal-like
4 10 26 90
19 15 4 13
14 25 52 230
266 124 37 68
27 vs. 40 vs. 54 vs. 41 vs.
8 7 26 17
pCR rate (%) Gp-R vs. Gp-NR 29 vs. 40 vs. 50 vs. 39 vs.
7 12 11 19
OR
95%CI
P value
4.3 8.3 3.5 3.4
2.0 – 9.2 0.8 – 82.9 1.4 – 8.8 2.1 – 5.7
8.93E-05 1.05E-01 7.91E-03 1.07E-06
OR
95%CI
P value
5.2 4.8 8.3 2.7
1.5 – 18.2 1.9 – 12.7 2.6 – 26.6 1.4 – 5.3
4.43E-03 6.59E-04 1.16E-04 2.30E-03
0.2
2 0.5 1 Odds ratio
5
20
0.2
2 0.5 1 Odds ratio
5
20
0.2
2 0.5 1 Odds ratio
5
20
C Chemotherapeutic Regimen
Gp-R (n) pCR Total
Gp-NR (n) pCR Total
Epirubicin FAC or FEC A-Paclitaxel A-Docetaxel A-Ixabepilone
11 22 66 12 23
5 9 32 8 10
38 62 156 22 61
76 107 273 39 67
pCR rate (%) Gp-R vs. Gp-NR
OR
95%CI
P value
29 vs. 7 36 vs. 8 42 vs. 12 55 vs. 21 38 vs. 15
5.8 6.0 5.5 4.7 3.5
1.8 – 18.2 2.5 – 14.1 3.4 – 9.0 1.5 – 14.6 1.5 – 8.1
1.19E-03 1.17E-05 3.90E-13 6.55E-03 3.25E-03
Figure 2. Analysis of IRSN-23 using the external validation set according to the intrinsic subtypes and chemotherapeutic regimens. The odds ratios for pCR are shown in the forest plot according to the ER and HER2 status (A), the intrinsic subtypes determined by PAM50 (B), and chemotherapeutic regimens (C). FAC, 5-fluorouracil, doxorubicin and cyclophosphamide; FEC, 5-fluorouracil, epirubicin and cyclophosphamide; A-Paclitaxel, sequential anthracycline and paclitaxel; A-Docetaxel, sequential anthracycline and docetaxel; A-Ixabepilone, sequential anthracycline and ixabepilone.
NR with pCR rates of 40 and 11%, respectively (P = 4.98E −23). The multivariate analysis showed that IRSN-23 was a highly significant (P = 8.25E−09) predictor of pCR and far superior to HER2 (P = 0.012), ER (P = 0.018) and HG (P = 0.049). Among the immune-related signatures that have been found for the prediction of the response to NAC to date, IRSN-23 appears to have the lowest P-value and the highest OR [6, 19]. Because it is well established that the intrinsic subtypes determined by PAM50 are significantly associated with pCR [22], we also evaluated IRSN-23 for each intrinsic subtype and were able to show that IRSN-23 can predict pCR for each intrinsic subtype. It is noteworthy that IRSN-23 was able to predict pCR even for the luminal A subtype, even though no prediction model has yet been able to do so [19]. In addition, we were able to show that IRSN-23 can predict pCR regardless of the chemotherapeutic regimen. These results appear to indicate that IRSN-23 has a better predictive capability for pCR than the previously reported immune-related signatures and suggest that antitumor immunity is involved
Volume 25 | No. 1 | January 2014
in the response to chemotherapy, regardless of the intrinsic subtypes or chemotherapeutic regimens. IRSN-23 includes CXCL9, which is significantly up-regulated in Gp-R. This observation is consistent with the report by Denkert et al. [5], who found that CXCL9 expression is significantly associated with TILs and the response to NAC. IRSN-23 also includes IDO1, which inhibits the T-cell response by increasing the number of T-regs [23]. Wainwright et al. [24] reported that IDO deficiency reduces T-regs recruitment and enhances T-cell mediated immunity against brain tumors. The association of the intra-tumoral infiltration of T-regs with a good response to chemotherapy for breast cancer has also been reported by us and other investigators [8, 18]. The clinical value of IRSN-23 appears to be limited because its diagnostic accuracy is not sufficiently high. We thus attempted to enhance its accuracy by combining it with another signature. Because the tumor cell proliferation status and the antitumor immunity status are considered important determinants for chemosensitivity, we speculated that the combination of an immune signature with a tumor cell proliferation signature
doi:10.1093/annonc/mdt427 |
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
Intrinsic subtype
pCR rate (%) Gp-R vs. Gp-NR
original articles 7.
8.
9. 10.
11.
12.
13.
14.
funding This work was supported by Grants-in-Aid from the Knowledge Cluster Initiative of the Ministry of Education, Culture, Sports, Science and Technology of Japan and by a cooperative research fund from Sysmex Corporation.
15. 16.
17.
disclosure S.N. and S.J.K. received honoraria from Sysmex Corporation. K.K. and Y.B. are employees of Sysmex Corporation, and their role was to provide technical assistance in the statistical analysis of the DNA microarray gene expression data. All of the remaining authors have no conflicts of interest to declare.
18.
19.
20.
references 1. von Minckwitz G, Untch M, Blohmer JU et al. Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol 2012; 30: 1796–1804. 2. Sachelarie I, Grossbard ML, Chadha M et al. Primary systemic therapy of breast cancer. Oncologist 2006; 11: 574–589. 3. Fuksa L, Micuda S, Grim J et al. Predictive biomarkers in breast cancer: their value in neoadjuvant chemotherapy. Cancer Invest 2012; 30: 663–678. 4. Iwamoto T, Bianchini G, Booser D et al. Gene pathways associated with prognosis and chemotherapy sensitivity in molecular subtypes of breast cancer. J Natl Cancer Inst 2011; 103: 264–272. 5. Denkert C, Loibl S, Noske A et al. Tumor-associated lymphocytes as an independent predictor of response to neoadjuvant chemotherapy in breast cancer. J Clin Oncol 2010; 28: 105–113. 6. Schmidt M, Hellwig B, Hammad S et al. A comprehensive analysis of human gene expression profiles identifies stromal immunoglobulin kappa C as a compatible
| Sota et al.
21. 22.
23.
24.
25.
prognostic marker in human solid tumors. Clin Cancer Res 2012; 18: 2695–2703. Naoi Y, Kishi K, Tanei T et al. Prediction of pathologic complete response to sequential paclitaxel and 5-fluorouracil/epirubicin/cyclophosphamide therapy using a 70-gene classifier for breast cancers. Cancer 2011; 117: 3682–3690. Oda N, Shimazu K, Naoi Y et al. Intratumoral regulatory T cells as an independent predictive factor for pathological complete response to neoadjuvant paclitaxel followed by 5-FU/epirubicin/cyclophosphamide in breast cancer patients. Breast Cancer Res Treat 2012; 136: 107–116. Desmedt C, Di Leo A, de Azambuja E et al. Multifactorial approach to predicting resistance to anthracyclines. J Clin Oncol 2011; 29: 1578–1586. Popovici V, Chen W, Gallas BG et al. Effect of training-sample size and classification difficulty on the accuracy of genomic predictors. Breast Cancer Res 2010; 12: R5. Tabchy A, Valero V, Vidaurre T et al. Evaluation of a 30-gene paclitaxel, fluorouracil, doxorubicin, and cyclophosphamide chemotherapy response predictor in a multicenter randomized trial in breast cancer. Clin Cancer Res 2010; 16: 5351–5361. Hatzis C, Pusztai L, Valero V et al. A genomic predictor of response and survival following taxane-anthracycline chemotherapy for invasive breast cancer. JAMA 2011; 305: 1873–1881. Horak CE, Pusztai L, Xing G et al. Biomarker analysis of neoadjuvant doxorubicin/ cyclophosphamide followed by ixabepilone or paclitaxel in early-stage breast cancer. Clin Cancer Res 2013; 19: 1587–1595. Lerman J. Study design in clinical research: sample size estimation and power analysis. Can J Anaesth 1996; 43: 184–191. Parker JS, Mullins M, Cheang MC et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol 2009; 27: 1160–1167. Elloumi F, Hu Z, Li Y et al. Systematic bias in genomic classification due to contaminating non-neoplastic tissue in breast tumor samples. BMC Med Genomics 2011; 4: 54. Sotiriou C, Wirapati P, Loi S et al. Gene expression profiling in breast cancer: understanding the molecular basis of histologic grade to improve prognosis. J Natl Cancer Inst 2006; 98: 262–272. Ladoire S, Arnould L, Apetoh L et al. Pathologic complete response to neoadjuvant chemotherapy of breast carcinoma is associated with the disappearance of tumorinfiltrating foxp3+ regulatory T cells. Clin Cancer Res 2008; 14: 2413–2420. Ignatiadis M, Singhal SK, Desmedt C et al. Gene modules and response to neoadjuvant chemotherapy in breast cancer subtypes: a pooled analysis. J Clin Oncol 2012; 30: 1996–2004. West NR, Milne K, Truong PT et al. Tumor-infiltrating lymphocytes predict response to anthracycline-based chemotherapy in estrogen receptor-negative breast cancer. Breast Cancer Res 2011; 13: R126. Leek JT, Scharpf RB, Bravo HC et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet 2010; 11: 733–739. Prat A, Parker JS, Fan C et al. PAM50 Assay and the three-gene model for identifying the major and clinically relevant molecular subtypes of breast cancer. Breast Cancer Res Treat 2012; 135: 301–306. Viguier M, Lemaitre F, Verola O et al. Foxp3 expressing CD4+CD25(high) regulatory T cells are overrepresented in human metastatic melanoma lymph nodes and inhibit the function of infiltrating T cells. J Immunol 2004; 173: 1444–1453. Wainwright DA, Balyasnikova IV, Chang AL et al. IDO Expression in brain tumors increases the recruitment of regulatory T cells and negatively impacts survival. Clin Cancer Res 2012; 18: 6110–6121. Liedtke C, Hatzis C, Symmans WF et al. Genomic grade index is associated with response to chemotherapy in patients with breast cancer. J Clin Oncol 2009; 27: 3185–3191.
Volume 25 | No. 1 | January 2014
Downloaded from http://annonc.oxfordjournals.org/ at University of California, San Francisco on December 5, 2014
would be able to predict pCR with greater accuracy. The GGI reflects the tumor cell proliferation status, and its association with pCR has been reported previously [25]. We demonstrated in our study that the combination of IRSN-23 and the GGI provides superior prediction of pCR in ER-positive tumors than either signature alone. In particular, this combination identified as many as 61% of the patients with ER-positive tumors who were very unlikely to achieve pCR ( pCR rate = 5%). From a clinical point of view, this information appears to be useful for the decision-making process regarding the indication of NAC for ER-positive tumors because patients with ER-positive tumors classified as Gp-NR with a low GGI, particularly those patients with clinically negative nodes, may be reasonable candidates not for NAC but for neoadjuvant hormonal therapy. In conclusion, we were able to construct IRSN-23, which can predict the response to NAC with greater accuracy than the conventional parameters and independently of the intrinsic subtypes and chemotherapeutic regimens. Our results are consistent with the hypothesis that TILs play an important role in the response to chemotherapy. However, prospective studies are needed to confirm the clinical utility of IRSN-23.
Annals of Oncology