DNA Methylation Signatures Predicting Bevacizumab Efficacy in Metastatic Breast Cancer

Abstract

Background: Biomarkers predicting response to bevacizumab in breast cancer are still missing. Since epigenetic modifications can contribute to an aberrant regulation of angiogenesis and treatment resistance, we investigated the influence of DNA methylation patterns on bevacizumab efficacy.

Methods: Genome-wide methylation profiling using the Illumina Infinium HumanMethylation450 BeadChip was performed in archival FFPE specimens of 36 patients with HER2-negative metastatic breast cancer treated with chemotherapy in combination with bevacizumab as first-line therapy (learning set). Based on objective response and progression-free survival (PFS) and considering ER expression, patients were divided in responders (R) and non-responders (NR). Significantly differentially methylated gene loci (CpGs) with a strong change in methylation levels (Δβ>0.15 or Δβ<-0.15) between R and NR were identified and further investigated in 80 bevacizumab-treated breast cancer patients (optimization set) and in 15 patients treated with chemotherapy alone (control set) using targeted deep amplicon bisulfite sequencing. Methylated gene loci were considered predictive if there was a significant association with outcome (PFS) in the optimization set but not in the control set using Spearman rank correlation, Cox regression, and logrank test.

Results: Differentially methylated loci in 48 genes were identified, allowing a good separation between R and NR (odds ratio (OR) 101, p<0.0001). Methylation of at least one cytosine in 26 gene-regions was significantly associated with progression-free survival (PFS) in the optimization set, but not in the control set. Using information from the optimization set, the panel was reduced to a 9-gene signature, which could divide patients from the learning set into 2 clusters, thereby predicting response with an OR of 40 (p<0.001) and an AUC of 0.91 (LOOCV). A further restricted 3-gene methylation model showed a significant association of predicted responders with longer PFS in the learning and optimization set even in multivariate analysis with an excellent and good separation of R and NR with AUC=0.94 and AUC=0.86, respectively.

Conclusion: Both a 9-gene and 3-gene methylation signature can discriminate between R and NR to a bevacizumab-based therapy in MBC and could help identify patients deriving greater benefit from bevacizumab.

Introduction

For HER2-negative metastatic breast cancer (MBC), bevacizumab (BEV) in combination with paclitaxel or capecitabine is a treatment option approved by the European Medicines Agency (EMA). Several phase-III studies have shown improvement in progression-free survival (PFS) with bevacizumab when added to chemotherapy, although the extent of improvement in PFS varied ^1-5. The primary endpoint of all of these trials was PFS, thus, conclusions on overall survival (OS) are difficult. However, neither a single trial nor any meta-analysis showed benefit in terms of OS ^6-8. This situation calls for biomarkers allowing clear separation of patients with or without a relevant chance of clinical benefit from this drug. Unfortunately, established markers predicting substantial benefit from this anti-angiogenic monoclonal antibody are still lacking. Promising biomarkers like plasma levels of VEGF-A or VEGFR-2 ^{9, 10}, tissue markers like the VEGFR co-receptor neuropilin-1 (NRP-1) ^11-13, single nucleotide polymorphisms (SNPs) in VEGF-A ¹⁴ or clinical markers like treatment-induced hypertension ^14-16 failed to demonstrate clinical utility or reproducibility. Given the complexity and multifactorial nature of angiogenic processes, probably no single biomarker is able to predict efficacy of anti-VEGF therapy.

DNA methylation is one of the known mechanisms of epigenetic regulation and there is increasing evidence for its involvement in the development of drug resistance ^17-22. By influencing expression levels of genes involved in angiogenesis, a certain methylation pattern could influence the response to an antiangiogenic therapy. Therefore, we aimed to generate a multigene-signature by using a whole-genome approach.

Material and Methods

Patients and study design

Key inclusion criteria for this analysis were histologically confirmed adenocarcinoma of the breast, locally advanced inoperable or metastatic tumor stage, ECOG performance status 0-3, at least one line of chemotherapy (taxane or capecitabine) alone or in combination with bevacizumab for advanced disease, sufficiently documented medical records (allowing calculation of PFS and OS) as well as sufficient tumor material for DNA isolation.

Learning Set

From the 212 breast cancer patients treated with chemotherapy in combination with bevacizumab (first- to fifth-line) at our institution between 2006 and 2012, 116 patients fulfilled the inclusion criteria. From this cohort, a learning set was selected consisting of 36 patients treated with a first-line chemotherapy and bevacizumab combination who clearly did (responder R; n=18) or did not (non-responder NR; n=18) benefit from this treatment based on objective response (ORR), progression-free survival (PFS), overall survival (OS) and estrogen receptor (ER) expression.

Optimization Set

The optimization set consisted of the 80 remaining patients from the bevacizumab-treated cohort that were not included in the learning set. Forty-one of these patients could clearly be classified as R or NR based on ORR, PFS, OS, and therapy-line and considering ER expression.

Control Set

Fifteen patients who never received bevacizumab and who were treated with a taxane or capecitabine as first-line chemotherapy were used as control set and were divided into R (n = 7) and NR (n = 8).

Patient characteristics of all patient sets are provided in Table 1.

Table 1.

Patient characteristics.

Patient characteristics		Learning set (n = 36)		Optimization set (n = 80*)		Control set (n = 15)
Median age (range)		61 (34-81)		60 (29-86)		62 (49-86)
DFS	de novo metastatic	8	22%	13	16%	1	7%
	≤ 24 mo	11	31%	20	25%	5	33%
	> 24 mo	17	47%	47	59%	9	60%
ECOG PS	0-1	35	97%	73	91%	11	73%
	3 or unknown	1	3%	7	9%	4	27%
Grade	1 -2 + unknown	20	56%	52	65%	9	60%
	3	16	44%	28	35%	6	40%
ER/PR status	ER or PR positive	26	72%	54	68%	10	67%
	ER and PR negative	9	25%	26	32%	5	33%
Metastases	visceral	25	69%	58	73%	12	80%
	non-visceral	11	31%	22	27%	3	20%
Adjuvant chemotherapy	yes	17	47%	52	65%	8	53%
	no	19	53%	28	35%	7	47%
Line of therapy	first-line	36	100%	27	34%	15	100%
	second-line	-	-	19	24%	-	-
	> second-line	-	-	34	42%	-	-
Chemotherapy backbone	paclitaxel or docetaxel	22	61%	64	80%	12	80%
	capecitabine	14	39%	15	19%	3	20%
Sample type	primary tumor	29	81%	65	81%	8	53%
	metastases	7	19%	15	19%	7	47%
Outcome months (95%CI)	median PFS	8.8 (5.1-16.6)		7.8 (6.5-9.3)		5.5 (4.3-NA)
	median OS	22.6 (18.4-33.5)		18.9 (16.4-21.7)		11.2 (7.5-NA)
Responder/non-responder	R / NR	18 / 18 (12 / 17)**		16		8 / 7

* One sample was removed because of missing values in the methylation analyses

** Restricted to samples from the primary tumor

DFS: disease-free survival; ECOG PS: Eastern Cooperative Oncology Group Performance Status; ER: estrogen receptor; PR: progesterone receptor; CI: confidence interval; PFS: progression-free survival; OS: overall survival; R: responder; NR: non-responder

Tissue samples and DNA isolation

Formalin-fixed paraffin-embedded (FFPE) tissue blocks containing samples from primary tumors, or if available, from metastatic sites, were selected by an experienced breast pathologist (CH). All tissue samples were collected prior to the start of first-line chemotherapy for metastatic disease. Three to five 10 µm sections were cut from each block without micro- or macro-dissection and placed in sterile Eppendorf tubes. DNA isolation and quantification was performed with the Maxwell 16 FFPE Plus LEV DNA Purification KIT and the PicoGreen dsDNA quantitation assay, respectively. Quality of FFPE-derived DNA was evaluated with Illumina's FFPE QC kit.

Methylation analysis

Learning Set

Genome-wide methylation profiling using the Illumina Infinium HumanMethylation450 BeadChip was performed, which enabled determining the methylation state of 485,577 single CpG sites. DNA was deaminated with the EZ DNA Methylation Kit (Zymo Research) according to Illumina's recommended deamination protocol. Bisulfite conversion was checked by qPCR. One assay targeting a methylated region of DNAJC15 and two assays targeting the GNAS locus (one assay for the unmethylated allele and one assay for the methylated allele) were used for quality control ²³. Deaminated DNA derived from blood was amplified in parallel and served as positive control. A sample passed the quality control when the received ct-value either for the two GNAS loci or the DNAJC15 locus reached the threshold no later than 5 cycles compared to the positive control. All samples passed the quality control. For the detection p-value, the intensity of the 450k-beadchip probe was compared with the intensities of the background signal (intensities of negative control probes). Only probes with detection p-values < 10^-5 were used for the analysis. 544 probes had at least 50% samples with detection p-values > 10^-5. These 544 sites were removed from the normalized data.

Afterwards, Bead Chip analysis was performed. To ensure equal signal distributions for each subarray (samples) and to remove technical bias, the data underwent quantile normalization. Furthermore, a peak correction algorithm was applied to the normalized data ^{24, 25}.

Optimization / Control Set

The methylation status of CpGs (± 200 base pairs) selected in the learning set was determined by targeted bisulfite sequencing (TDBS) ^{26, 27} on an IonTorrent PGM System. The MSP-HTPrimer was used for assay design following a bisulfite sequencing-specific approach, which enriches the region of interest independently of the methylation state ²⁷. Primer design resulted in assays with a length of 118-292 bp (average: 242 bp). The assays designed for bisulfite-assisted genomic sequencing PCR (BSP) were used to enrich the respective target regions in 96 deaminated samples. The EZ-96 DNA Methylation Kit (Zymo Research) was used according to manufacturer's recommendations to deaminate the samples. Target enrichment was done by qPCR in single PCR reactions, followed by pooling of the 48 targets sample by sample. Library preparation and targeted sequencing for the Ion Torrent PGM was done following the manufacturer's protocol. Briefly, after DNA end repair to generate blunt ends, individual barcodes and sequencing adapters (Xpress Barcode Adapters, Thermo Fisher Scientific) were attached to the pooled targets of each sample and purified with AMPure XP magnetic beads (Beckman Coulter). Generated libraries were quantified using the Ion Library Quantitation Kit (Thermo Fisher Scientific). The samples were divided in 3 batches of 32 samples each. For each batch, the barcoded samples were pooled equimolar for the template preparation by emulsion PCR with the Ion PGM HI-Q OT2 KIT on the Ion One Touch 2 Instrument (Thermo Fisher Scientific). Enrichment of Ion Sphere Particles (ISPs) was done on the Ion One Touch ES Instrument (Thermo Fisher Scientific). Quality of ISPs was controlled with the Ion Sphere Quality Control Assay (Thermo Fisher Scientific) on the Qubit 2.0 Fluorometer. The 318 Chip Kit v2 (Thermo Fisher Scientific) was used to sequence the two sample batches with the Ion PGM Hi-Q Sequencing Kit. The sequencing reads were aligned to the hg19 reference genome and fitted to the 48 designed amplicons. A special workflow, which was recently published by Pabinger et al. was used to map the reads from the Ion Torrent PGM to the reference genome ²⁸.

Statistical analysis

All analyses were performed using the statistical software environment R including packages survival, ChAMP, IMA, ROCR as well as BRB array tools for class prediction, and Genesis for hierarchical clustering and heatmap visualization.

Learning Set

After quantile normalization and type I/II peak correction, significantly differentially methylated gene loci (CpGs) between R and NR (P < 0.001) were identified using a moderated t-test, whereby the variance of individual CpGs is calculated using information from all selected CpGs (chip analyses methylation pipeline, ChAMP) ²⁹. Significantly differentially methylated gene loci between R and NR (P < 0.001; |Δβ|>0.10 or |Δβ|>0.15 as stated) were selected and further validated if they showed an AUC ≥ 0.85 by logistic regression analysis, and/or sites in proximity to genes known to be involved in angiogenesis or carcinogenesis, and/or were included in significantly differentially methylated regions as identified by Illumina Methylation Analyzer (IMA) ²⁵ or the Probe Lasso method ³⁰.

Control Set

Cox regression analysis was used to associate methylation with PFS.

Optimization Set

Cox regression was used to associate methylation with PFS and OS. A hazard ratio was calculated for an increase of 10% in methylation level. Receiver operating characteristics (ROC) analyses, class prediction, and (consistent) differential methylation of single CpGs, were used to select CpGs predicting R and NR in the optimization set.

Gene-panel

Logistic regression (LR) models of a 9-gene methylation signature (based on single CpG-ß-values) and all combinations of 1 to 4 of the respective CpGs were learned from the learning set and then, based on leave-one-out cross validation (LOOCV), the model with the highest AUC was selected. As an optimal cutoff the point in the ROC curve closest to the true positive rate of 1 and the false positive rate of 0 was used. Only models with less than 35% missing values in the optimization set were considered. For validation ROC analyses of the logistic regression models and a logrank test to analyze the differences in PFS and OS between predicted R and NR were performed in the learning set and the optimization set.

Multivariate Analysis

Multivariate Cox regression analyses on PFS and OS were performed in the learning set and the optimization set on the predicted R and NR (based on the 3 CpGs/genes logistic regression model) and included the clinicopathological factors length of disease-free survival (DFS) between surgery of the primary tumor and recurrence, adjuvant chemotherapy, ECOG performance score, histologic subtype, tumor grade, receptor status, location of metastases, line of therapy, and origin of the specimen as categorical covariates. A Wald test was used to assess the significant impact of individual covariates (coefficients) and the overall model fit.

Endpoints

Progression-free survival (PFS) was defined as time from treatment initiation until progression or death from any cause, whichever occurred first. Overall survival (OS) was defined as time from treatment initiation until death from any cause. Patients alive (for OS) and who had not experienced progression (for PFS) at the data cutoff date, were censored at the last follow-up date. Overall response rate (ORR) was defined as the proportion of patients who achieved a complete or partial response. Because of the retrospective nature of the study, no categorization according to RECIST was possible. Therefore, partial response was defined as a relevant reduction in tumor burden as per clinical routine assessment and complete response as disappearance of all signs of cancer.

Ethics

The study was approved by the Ethics Committee of the province Salzburg (IRB number: 415-EP/73/67-2011).

Results

The different steps of biomarker discovery are summarized in a flow chart in Figure 1. In the learning set (n = 36) 417 CpGs showed significantly different methylation between R (n = 18) and NR (n = 18) with a Δβ-value > 0.10 or < -0.10 (P < 0.001). In R, 206 sites were hypermethylated and 211 sites were hypomethylated. To avoid a potential confounding by the different sample types (metastases samples are more represented in the R group) only samples from the primary tumor were further investigated (Figure S1). A panel of 48 genes (48 CpGs from Illumina's 450k array) allowed a good separation between R and NR (odds ratio = 101, P < 0.0001; Figure S2). Details on the selection methods and the CpG position for each gene are provided in Table S1. These 48 CpGs ± 200 base pairs (including 361 CpGs in total) were sequenced in the optimization and control set, respectively. At least one methylated cytosine within 26 of 48 analyzed regions was significantly associated with PFS or OS in the optimization set (n = 80) or had an AUC value > 0.7 in a ROC analysis. None of the analyzed CpGs showed a significant association of methylation levels with PFS in the control set (n = 15) in the same direction (Table S2). By using information from the optimization set based on consistency of the methylation and close genomic positions, the methylation panel was further reduced to 9 genes (16 CpGs) (Table S3).

Figure 1.

Data analysis workflow.

Going back to the learning set, unsupervised hierarchical clustering according to the methylation status of these 9 genes could divide patients into 2 clusters predicting R and NR with an odds ratio of 40 (P < 0.001) (Figure 2A). A ROC analysis of a logistic regression model for this 9-gene methylation signature (Table S4) resulted in AUC of 1.0 (AUC LOOCV 0.91) in the learning set (Figure 2B). The hazard ratios (HR) for predicted R and NR in the learning set were 0.18 (95%CI 0.01-NA; P < 0.001) and 0.24 (95%CI 0.07-0.38; P < 0.001) for PFS and OS, respectively (Figure 2C, 1D). In the optimization set from the 17 patients with a complete methylation profile of all 9 CpGs (genes), 9 patients could be predicted as R and 8 patients as NR. Predicted responders were associated with longer PFS (median PFS 9.5 months vs. 5.3 months; HR 0.61; 95%CI 0.21-1.60) and longer OS (median OS 26.7 months vs. 16.8 months; HR 0.56; 95%CI 0.18-NA), however, in both cases not to a significant extent (P = 0.29 and P = 0.21, respectively; Figure S3). To include more patients and hence increase the power, the number of CpGs was reduced by testing all possible logistic regression models including maximal 4 CpGs (genes). The best model with a 3-gene signature (Table S4) was based on the methylation status of single CpGs in MLH1, POLK, and TMBIM6, with an AUC of 0.94 in the learning set (Figure 3A) and an AUC of 0.86 in the optimization set (Figure 3B). The median PFS for predicted R and NR by the 3-gene signature was 19.3 months and 4.9 months in the learning set (HR 0.26, 95%CI 0.07-0.35; P < 0.001) and 9.7 months and 5.5 months in the optimization set (HR 0.29, 95%CI 0.04-0.21; P < 0.001), respectively (Figure 3C, 3D). The median OS was 36.3 months for predicted R and 18.4 months for predicted NR in the learning set (HR 0.30, 95%CI 0.13-0.61; P = 0.001) and 22.8 months for predicted R and 10.8 months for predicted NR in the optimization set (HR 0.36, 95%CI 0.09-0.50; P = 4.8 x 10^-4) (Figure 3E, 3F). The median PFS and OS of patients from the optimization set not classified (n=32 of 80) because of missing values were 7.5 months and 19.9 months, respectively, lying in between the median values for predicted R and predicted NR (Figure 3D, 3F). The association of predicted responders with longer PFS (P = 0.001 and P < 0.001) and longer OS (P = 0.019 and P < 0.001) remained significant in multivariate analysis both in the learning set and the optimization set (Table 2).

Figure 2.

(A) Unsupervised hierarchical clustering of the learning set (450k-chip methylation data) according to the methylation status of the 9 genes deduced from TDBS-derived methylation data of the optimization set allowed a good separation between R (Cluster A) and NR (Cluster B) with an odds ratio of 40 (P < 0.0001). (B) ROC analyses of a logistic regression model for the 9 genes in the learning set resulted in AUC of 1.0 (AUC_LOOCV 0.91). (C) Progression-free survival (PFS) and (D) overall survival (OS) for predicted R and predicted NR from the learning set according to the 9-gene methylation signature.

Figure 3.

(A) ROC analyses of a logistic regression model for the 3 genes in the learning set resulted in AUC of 0.94 and (B) in the optimization set resulted in AUC of 0.86. Progression-free survival (PFS) for predicted R and predicted NR according to the 3-gene methylation signature in the learning set (C) and in the optimization set (D), as well as overall survival (OS) in the learning set (E) and the optimization set (F).

Table 2.

Multivariate Cox proportional hazards regression model for the association of predicted responders with DFS and OS in the learning set (A) and the optimization set (B)

						PFS						OS
		n	n1	n2		coef	HR	95%-CI	p	p*		coef	HR	95%-CI	p	p*
(A) Learning set †
Predicted responders	R vs. NR	36	15	21		-1.755	0.17	0.06-0.51	0.001	0.002		-1.354	0.26	0.08-0.80	0.019	0.003
DFS	M1 vs. <24 mo	36	8	11		-1.000	0.37	0.09-1.43	0.150			-0.044	0.96	0.21-4.45	0.960
	>24 mo vs. <24 mo		17	11		-0.107	0.90	0.35-2.32	0.830			0.560	1.75	0.63-4.87	0.280
Adjuvant chemotherapy	yes vs. no	36	18	18		-0.664	0.51	0.17-1.57	0.240			0.250	1.28	0.45-3.63	0.640
Histology	lobular+other vs. ductal	36	8	28		-0.559	0.57	0.14-2.28	0.430			-0.975	0.38	0.09-1.50	0.170
Tumor grade	3 vs. 1|2|unknown	36	16	20		-0.396	0.67	0.27-1.70	0.400			-0.282	0.75	0.32-1.77	0.520
Receptor status	HR+/HER2- vs. HR-/HER2-	36	26	10		0.030	1.03	0.33-3.26	0.960			-0.749	0.47	0.16-1.42	0.180
Type of metastases	visceral vs. non-visceral	36	25	11		1.039	2.83	0.88-9.11	0.082			1.393	4.03	1.11-14.63	0.034
Origin of specimen	metastasis vs. primary tumor	36	7	29		-0.810	0.45	0.13-1.58	0.210			-0.913	0.40	0.12-1.37	0.140
(B) Optimization set
Predicted responders	R vs. NR	48	34	14		-1.914	0.15	0.05-0.41	<0.001	0.006		-1.573	0.21	0.08-0.51	0.001	0.009
DFS	M1 vs. <24 mo	48	9	12		0.524	1.69	0.39-7.27	0.480			0.069	1.07	0.26-4.49	0.930
	>24 mo vs. <24 mo		27	12		-0.056	0.95	0.33-2.73	0.920			-0.710	0.49	0.17-1.45	0.200
Adjuvant chemotherapy	yes vs. no	48	31	17		0.084	1.09	0.45-2.65	0.850			-0.304	0.74	0.30-1.83	0.510
ECOG PS	≥2 vs. 0|1	48	4	44		1.026	2.79	0.75-10.42	0.130			1.761	5.82	1.50-22.56	0.011
Histology	lobular vs. ductal	48	6	37		1.196	3.31	1.05-10.41	0.041			1.236	3.44	1.07-11.11	0.039
	other vs. ductal		5	37		1.669	5.31	1.73-16.30	0.004			1.160	3.19	1.05-9.70	0.041
Tumor grade	3 vs. 1|2|unknown	48	17	31		1.518	4.56	1.60-13.01	0.005			1.091	2.98	1.14-7.77	0.026
Receptor status	HR+/HER2- vs. HR-/HER2-	48	30	18		0.778	2.18	0.74-6.43	0.160			0.655	1.92	0.68-5.44	0.220
Type of metastases	visceral vs. non-visceral	48	31	17		0.741	2.10	0.86-5.10	0.100			0.508	1.66	0.67-4.11	0.270
Line of therapy	≥2nd vs. 1st	48	33	15		-0.356	0.70	0.31-1.57	0.390			0.135	1.14	0.53-2.47	0.730
Origin of specimen	metastasis vs. primary tumor	48	11	37		0.138	1.15	0.42-3.16	0.790			-0.150	0.86	0.36-2.08	0.740

* overall p-value using Wald test

† “ECOC PS” and “line of therapy” were not included in multivariate analysis in the learning set because all patients had ECOG PS 0 or 1 and all were treated 1^st-line.

DFS: disease-free survival; ECOG PS: Eastern Cooperative Oncology Group Performance Status; ER: estrogen receptor; PR: progesterone receptor; CI: confidence interval; PFS: progression-free survival; OS: overall survival; coef is the coefficient in the Cox regression model for the respective variable; HR: hazard ratio (=exp(coef)); R: responder; NR: non-responder

Discussion

To our knowledge this is the first DNA methylation study investigating efficacy prediction for bevacizumab in cancer, especially in breast cancer patients. Using FFPE-derived DNA from primary tumor sections, we generated both a 9-gene and a 3-gene methylation signature with a high predictive value for bevacizumab efficacy in patients with MBC.

Single molecular alterations like estrogen receptor expression or HER2 overexpression are routinely used to estimate prognosis and treatment efficacy in breast cancer. Other alterations like BRCA1/2 mutation, HER2 mutation or ESR1 mutation are on the verge of getting clinically meaningful. Single gene approaches, however, have several limitations: techniques like immunohistochemistry have a low reproducibility, they allow only a rough classification into different subgroups and they ignore other genes affecting the same biologic pathway. Gene expression studies in breast cancer have demonstrated that mRNA signatures can predict different biological behavior, prognosis and therapeutic success more effectively than singe gene evaluations ^31-34. Such signatures developed by genome-wide approaches do not only reflect discrete differences in specific mRNA expressions but also the complex interplay between different genes, which is part of the so-called interactome. In a similar way, DNA methylation profiles, representing gene regulation on an epigenetic level, can mirror the molecular state of a cell ³⁵. Probably, the complex gene interactions can be even better represented since DNA methylation regulates gene networks rather than single genes ³⁶. In addition, using DNA instead of RNA for profiling has several advantages: compared to RNA or proteins, DNA and the DNA CpG-methylation profile is a much more robust analytical material ^{26, 37}. Furthermore, DNA methylation causes a stable long-term programming of the genome ³⁸, whereas transcription activity is much more influenced by transient signals. Therefore, DNA methylation best fulfills prerequisites for developing robust biomarkers ^{26, 37}.

Several proteins involved in angiogenesis like VEGF-A, VEGF-R2 (KDR) and VEGF-R3 (FLT4) are known to be silenced by promoter hypermethylation. Nevertheless, we used an unsupervised whole genome approach for biomarker generation because resistance or extraordinary sensitivity to anti-VEGF therapy are not automatically driven by genes directly involved in angiogenesis ³⁹. Confining biomarker research to predefined genes based on a biological rationale carries the danger of missing crucial information. VEGF-A plasma levels, for example, showed promising early results ^{40, 41} but failed when prospectively validated ¹⁰.

Furthermore, for the same reason, we did not concentrate on methylation sites within specific genomic regions like promotors or CpG island shores. The impact of the exact location of DNA methylation on gene expression is still not fully understood ⁴² and seems to vary between molecular subtypes of breast cancer ⁴³. Therefore, exclusion of methylation sites apart from promotors or CpG island shores could reduce the success rate of methylation biomarkers. Furthermore, non-protein-coding DNA like DNA transcribed into microRNAs seems to be regulated and dysregulated by methylation as well ⁴⁴, making the regulatory machinery even more complex. Accordingly, most of the genes of our 9-gene signature (PKNOX2, POLK, UNC119, SNRPN, TMBIM6, MLH1 and GNAS) are not linked to angiogenesis yet.

In our analysis, the predicted difference in OS between patients categorized as R and NR according to the methylation pattern is more than 1 year in both cohorts, with a risk reduction for death of more than 60% (HR 0.30 and 0.36 in learning set and optimization set, respectively). When scored according to the European Society for Medical Oncology Magnitude of Clinical Benefit Scale (ESMO-MCBS) ⁴⁵, such a trial result would result in a score of 4 out of 5 points (-1 point because no data on quality of life are available), while the approval study E2100 randomizing between paclitaxel plus bevacizumab and paclitaxel alone ⁴, reached 2 points only ⁴⁵. Knowing that none of the randomized phase III trials in HER2 negative breast cancer comparing bevacizumab plus chemotherapy with chemotherapy alone showed any effect on overall survival ^{1, 3-5, 10, 46, 47}, such a survival difference would, if confirmed, certainly influence clinical practice. The risk reduction for progression or death (PFS) was more than 70% in both cohorts (HR 0.26 and HR 0.29, respectively). Notably, the absolute median PFS and OS values in the two cohorts differed significantly. The main explanation for this observation is that 46% of patients in the optimization set were treated in second-line and beyond. Furthermore, the two cohorts were not matched according to other factors known to influence outcome (e.g., adjuvant chemotherapy, chemotherapy backbone, location and number of metastases, etc.). However, a risk reduction in PFS of this magnitude was never observed in clinical trials investigating bevacizumab efficacy (HR ranging between 0.60 and 0.69 in first-line trials ^{3-5, 10}) and the HR were nearly identical in the two cohorts of patients. If confirmed, this approach could lead to more effective and cost-efficient use of bevacizumab in the clinical setting.

This study has several limitations. First of all, no clear validation of the 48-gene-signature generated in the learning set was possible, because low DNA content of some samples prevented the analysis of all selected CpGs in our validation set (therefore called optimization set). Second, the control set, including patients treated with first-line chemotherapy without bevacizumab, was very small. Third, it is a single-institution, uncontrolled study, which limits any firm conclusion about the utility of the methylation signatures. A further caveat might be that two different techniques were used in the learning and control/optimization sets: bead chip technology vs. targeted bisulfite sequencing. However, previous analyses have shown that generally methylation scores obtained by Illumina 450K analysis replicate well using targeted sequencing approaches ²⁸. In addition, this technical switch allowed a selection of genes that can be robustly analyzed by targeted bisulfite sequencing in archival FFPE samples, which is an important issue for biomarkers when transferred into the clinic.

In conclusion, this is the first study providing predictive DNA methylation signatures for bevacizumab efficacy in MBC: a 9-gene as well as a 3-gene-methylation signature could reproducibly discriminate between responders and non-responders to a bevacizumab-based therapy in two different patient sets. Before such a signature can be implemented in treatment decision-making, these data have to be validated in larger cohorts, ideally in tissue samples from a prospective trial. To demonstrate clinical utility according to the roadmap for development of omics-based tests defined by the United States Institute of Medicine (IOM) ⁴⁸, a prospective retrospective study is planned using DNA samples from a randomized phase III trial comparing chemotherapy plus bevacizumab with chemotherapy alone.

Supplementary Material

Supplementary figures and tables.

Abbreviations

AUC

area under the curve

DNA

deoxyribonucleic acid

FFPE

formalin-fixed paraffin-embedded

HER2

human epidermal growth factor receptor 2

HR

hazard ratio

IMA

Illumina Methylation Analyzer

LOOCV

leave-one-out cross validation

MBC

metastatic breast cancer

MVP

methylation variable position

NR

non-responder

OR

odds ratio

ORR

objective response rate

OS

overall survival

PFS

progression-free survival

qPCR

real-time quantitative PCR

R

responder

RNA

ribonucleic acid

ROC

receiver operating characteristic

SNP

single nucleotide polymorphism

TDBS

targeted bisulfite sequencing

VEGF

vascular endothelial growth factor

VEGFR

vascular endothelial growth factor receptor.