SNP biomarkers of adjuvant anastrozole-induced estrogen suppression in early breast cancer

Abstract

Objective

Based on our previous findings that postmenopausal women with estrone (E1) and estradiol (E2) concentrations at or above 1.3 pg/mL and 0.5 pg/mL, respectively, after six months of adjuvant anastrozole therapy had a 3-fold risk of recurrence, we aimed to identify a SNP-based model that would predict elevated E1 and E2 and then validate it in an independent dataset.

Patients and Methods

The test set consisted of 322 women from the M3 study and the validation set consisted of 152 patients from MA.27. All patients were treated with adjuvant anastrozole, had on-anastrozole E1 and E2 concentrations, and genome-wide genotyping.

Results

SNPs were identified from the M3 genome-wide association study. The best model to predict the E1-E2 phenotype with high balanced accuracy was a support vector machine (SVM) model using clinical factors plus 46 SNPs. We did not have an independent cohort that is similar to the M3 study with clinical, E1-E2 phenotypes, and genotype data to test our model. Hence we chose a nested matched case-control cohort (MA.27 study) for testing. Our E1-E2 model was not validated but we found the MA.27 validation cohort was both clinically and genomically different.

Conclusion

We identified a SNP-based model that had excellent performance characteristics for predicting the phenotype of elevated E1 and E2 in women treated with anastrozole. This model was not validated in an independent dataset but that dataset was clinically and genomically substantially different. The model will need validation in a prospective study.

INTRODUCTION

The third-generation aromatase inhibitors (AIs) anastrozole, exemestane, and letrozole all play a major role in the treatment of postmenopausal women with estrogen receptor (ER) positive breast cancer [1–3] and their value has been established in the prevention setting in high-risk women [4,5]. The assumption is that the mechanism of action of the AIs in inhibiting tumor growth is simply through decreasing estradiol (E2), which is a ligand for the ER. However, no clinical studies had systematically tested whether the degree of estrogen suppression correlates with AI efficacy. To address this question, we performed a matched case-control study [6] in patients from the MA.27 adjuvant trial [7] that randomized patients to anastrozole or exemestane, and from the PreFace adjuvant trial that utilized letrozole, and found that patients whose E1 and E2 were at or above 1.3 pg/mL and 0.5 pg/mL, respectively, after six months of anastrozole therapy had a 3-fold risk of recurrence compared with patients who had either E1 and/or E2 below these thresholds. The E1 and E2 assays utilized in this study had lower limits of quantitation (LLQ) of 1.0 pg/mL and 0.3 pg/mL, respectively. We also performed preclinical laboratory studies that examined mechanisms of action in addition to aromatase inhibition and showed that anastrozole, but not exemestane or letrozole, could directly bind to ERα, activate estrogen response element-dependent transcription, and stimulate the growth of an aromatase-deficient CYP19A^−/−T47D breast cancer cell line [6].

We hypothesized that germline genetic variation is a major contributor to individual variation in anastrozole-related E1 and E2 suppression and performed genome-wide associations studies (GWAS), approximating the phenotype of E1 and E2 suppression identified in the MA.27 case-control study [6], in our Mayo-MD Anderson-Memorial Sloan Kettering (M3) trial of adjuvant anastrozole [8]. In M3, the LLQs for the E1 and E2 assays were 1.56 pg/mL and 0.625 pg/mL, respectively. We developed prediction models in the M3 cohort by applying hold-out and cross-validation methods. Validation was attempted in a second cohort of patients from the MA.27 study [6].

METHODS

Source of patients

Patients included in this study were obtained from the M3 study [8]. These postmenopausal patients had determination of E1 and E2 concentrations after at least 4 weeks of anastrozole therapy, had detectable anastrozole at the time of the blood draw for hormone concentrations, and genome-wide genotyping using the Illumina Human610-Quad platform. To make the M3 sample as similar as possible to the MA.27 case-control study validation set, we also excluded M3 patients who were HER2+ or had prior treatment with tamoxifen. The most promising single nucleotide polymorphisms (SNPs) were identified in the M3 cohort using patients who had both E1-E2 low or both E1-E2 high. Patients with only one phenotype (E1 or E2) high were removed from the analysis. Following the identification of the most promising SNPs, multiple models were examined. This research was performed after approval by local institutional review boards in accordance with assurances filed with, and approved by, the Department of Health and Human Services.

We sought validation of our best model in anastrozole-treated patients from the MA.27 case-control study [6], where genome-wide genotyping utilizing the Illumina Human610 Quad Beadchip [9], and E1 and E2 concentrations after six months of anastrozole therapy had been determined. Details regarding this study are given in Supplementary Material.

Statistical analysis

Logistic regression modeling was applied to the M3 study to determine SNPs associated with the phenotype of E1 ≥1.56 pg/mL and E2 ≥0.625 pg/mL (E1-E2 high) versus both E1 and E2 below these thresholds (E1-E2 low). We modeled the difference between these two extremes and excluded the middle category with those who had only one either E1 or E2 above these thresholds. An additive-genotype model was applied to model the SNPs and the genotypes were coded as 0, 1, 2 to represent the number of minor alleles in a SNP. We evaluated clinical factors [BMI (<25, 25–34.9, ≥35), source (MD Anderson, Mayo and Memorial Sloan Kettering), age (<50, 50–60, ≥60), nodal status (N0, N1-N3), primary tumor status (T1/2, T3/4), prior chemotherapy (Yes, No), smoking history (Yes, No, unknown), current smoker (Yes, No, unknown), associated with the phenotype using a stepwise selection procedure and factors that were statistically associated with the phenotype (p-value < 0.05) were included as adjusting covariates in the GWAS.

We performed a GWAS analysis with the phenotype E1-E2 high vs E1-E2 low in the M3 study. The GWAS was adjusted for clinical variables of BMI, source, and nodal status. The GWAS results showed that there were no markers that met the genome-wide statistical significance (i.e. p-value < 10⁻⁸). However, being unable to reach this p-value does not indicate there is no association between markers and a phenotype [10]. Thus, we utilized novel filtering steps in combination with statistical significance and pruning steps. We selected the SNPs with lowest p-values, a SNP minor allele frequency (MAF) of ≥ 0.1, and an odds ratio (OR) of ≤0.67 or ≥1.5. Linkage disequilibrium (LD)-clumping [11] was then applied to further filter the SNPs sorted by the p-values and MAFs from the GWAS. The SNP with the lowest p-value and highest MAF (index-SNP) was maintained. The thinned SNPs were additionally filtered to those that were also present in the MA.27 cohort as we intended to utilize it for validation.

We applied machine learning methods to the genotype and clinical data and compared the performance of several algorithms (penalized regression including lasso [12], elastic net [13] and ridge regression [14], bagging [15], boosting[16], and kernel-based methods [17] including gradient boosting machine [18,19], random forest [20], adaptive boosting (AdaBoost) [21], and support vector machines (SVM) [22] to determine prediction accuracy in the M3 cohort.

In the M3 cohort, the classifier estrogen suppression was imbalanced; high = 44 (minority class), low =278. The performance of machine learning algorithms is typically evaluated using overall predictive accuracy. However, this is not appropriate when the data are imbalanced because high metrics do not reflect the prediction capacity of the minority class (E1-E2 high). Synthetic Minority Oversampling Technique (SMOTE) [23] creates new instances of the minority class (E1-E2 high) rather than oversampling with replacement. Assessing the performance of a model only by an overall accuracy metric in an imbalanced design setting is often misleading. To address the imbalance issue, we used balanced accuracy [24] rather than overall accuracy. To identify a set of biomarkers that predict the phenotype with maximum balanced accuracy, we examined both hold- out validation and cross-validation in the M3 cohort.

We further performed SMOTE-SVM-RBF (support vector machine [25,26] radial basis function [27]) in different feature sets models and tested for model balanced accuracies: Additional details are provided in Supplementary Material.

Network analysis

Network analysis of the genes, estrone, and estradiol was conducted using Ingenuity Pathway Analysis software (https://www.qiagenbioinformatics.com/products/ingenuitypathway-analysis).

Homogeneity of the M3 and MA.27 cohorts

We evaluated the homogeneity of the MA.27 and M3 cohorts by summarizing the clinical characteristics and genotype distribution.

RESULTS

The participant flow (REMARK) diagram (Supplementary Fig. 1) shows the patients included in, and excluded from, the analysis of the M3 study. The characteristics of the 322 M3 patients are given in Table 1 and show the median age was 68 years, the median BMI was 27.0, the majority (72%) were node-negative, and a minority (36%) had received prior chemotherapy. Of the 322 patients included from the M3 study, 44 were E1-E2 high and 278 were E2-E2 low.

Table 1.

Patient characteristics for M3 and MA.27 trials

	M3	MA.27 Case-Control	p-values for M3 vs MA.27 Case-Control Sample
	n=322	n=152
Age, years			0.32
Median	68.4	64.5
Range	49 – 92	48 – 82
<50	9 (2.8%).	3 (2.0%)
50–60	109 (33.9%)	42 (27.6%)
≥60	204 (63.4%)	107 (70.4%)
BMI			0.03
Median	27.0	28.20
Range	17.0 – 46.7	22.6 – 56.8
<25	114 (35.4%)	38 (25.0%)
25–34.9	162 (50.3%)	96 (63.2%)
35+	46 (14.3%)	18 (11.8%)
T-stage			0.13
1, 2	306 (95.0%)	139 (91.4%)
3, 4	16 (5.0%)	13 (8.6%)
N-stage			<0 .001
N 0	231 (71.7%)	57 (37.5%)
N 1–3	91 (28.3%)	95 (62.5%)
Prior Chemotherapy			< 0.001
No	206 (64.0%)	63 (41.4%)
Yes	116 (36.0%)	89 (58.6%)

GWAS Association Results and Post-GWAS Filtering Steps in M3 study

Genome-wide logistic regression was performed to identify genome-wide markers associated with both E1 and E2 suppression with the GWAS analysis adjusted for BMI, source (M.D. Anderson, Mayo and Memorial Sloan Kettering), and N-stage. The Manhattan plot (Fig. 1) shows that no SNPs achieved genome-wide statistical significance. We took the top 500 SNPs with the lowest p-values from the GWAS after restricting to MAF ≥0.1 and odds ratios (OR) ≤ 0.67 or ≥ 1.5. The arrows on the Manhattan plot (Fig. 1) indicate the chromosomes that contain SNPs that were selected on this basis. The one exception was the SNP (rs9519934) on chromosome 13 that was considered a singleton and not included. After filtering down, 211 SNPs remained. We further filtered using the LD-clumping method by keeping all the SNPs that were within 500 kb of the significant SNP, were in high LD (r²≥0.8) and had a p-value <0.02. LD-clumping reduced our tag SNP set to 51. Out of these 51 tag SNPs from the M3 study, 46 tag SNPs were also present in the MA.27 study genotyping, and these were selected for prediction modeling (Supplementary Table S1). It can be seen that some of the SNPs had substantial OR with 15 having OR ≥3.00 (OR range: 3.03–6.78) and an additional 14 had OR ≤0.33 (OR range: 0.33–0.23. It is of note that none of the 46 SNPs chosen for testing were from chromosome 15 on which CYP19A1 is located.

Figure 1.

Manhattan plot of p-values for conditional logistic regression adjusted for clinical factors. Arrows indicate chromosomes from which SNPs were obtained for the modeling.

Prediction modeling results (hold-out validation and cross-validation in M3)

With the phenotype of E1-E2 high, SMOTE-SVM-RBF performed as the best method. Table 2 shows the prediction accuracy metrics obtained from the four feature-set models that were tested. Focusing on hold-out validation metrics, we evaluated model performance using clinical variables. Using the clinical factors only (BMI, age, N-stage, T-stage, prior chemotherapy status, and progesterone receptor (PR) status), the balanced accuracy was 0.53 with precision, F1, sensitivity, and specificity ranging between 0.43 and 0.80. All three of the models utilizing SNPs showed an improvement in the balanced accuracy and other metrics. Using the top 5 SNPs (model 2), the balanced accuracy increased to 0.66. Using the 46 SNP-set alone (model 3), the balanced accuracy increased to 0.8. We found that model 4, which included clinical factors plus the 46 SNPs, revealed remarkable performance characteristics for predicting the desired phenotype of E1 and E2 at or above the thresholds noted above. That is, the balanced accuracy was 0.93, precision was 0.96, F1 was 0.98, sensitivity was 1.0, and specificity was 0.86.

Table 2:

Model performance for E1-E2 GWAS targeted SNP-set in M3

E1-E2 Suppression Feature-set models			Hold-out Validation in M3 (numbers presented are the mean across 100 iterations)					Cross-Validation in M3 (numbers presented are the mean across 100 iterations)
Model	M3 sample size (high-high/ lowlow)	# of samples (hold out testing)	Balanced Accuracy*	Precision†	F1‡	Sensitivity§	Specificity¶	Balanced Accuracy	Precision	F1	Sensitivity	Specificity
1) Clinical	37/252	7 vs 25	0.53	0.80	0.71	0.64	0.43	0.61	0.60	0.64	0.68	0.55
2) Top 5 SNPs			0.66	0.88	0.71	0.60	0.71	0.79	0.78	0.79	0.81	0.78
3) Only 46 NPs			0.80	0.91	0.90	0.88	0.71	0.92	0.97	0.92	0.88	0.97
4) Clinical + 46 SNPs (Best model)			0.93	0.96	0.98	1.0	0.86	0.88	0.84	0.89	0.93	0.83

Positives are E1-E2 low/low class and Negatives are E1-E2 high/high class.

True positives are E1-E2 low/low samples being correctly predicted as E1-E2 low/low.

True negatives are E1-E2 high/high samples being correctly predicted as E1-E2 high/high samples.

False positives are E1-E2 high/high samples are incorrectly predicted as E1-E2 low/low.

False negatives are E1-E2 low/low incorrectly predicted as E1-E2 high/high

^*Balanced accuracy: (true positives/positives + true negatives/negatives)/2

^†Precision: (positive predictive value): true positives/(true positives + false positives);

^‡F1: 2 X true positives/(2 X true positives + false positives + false negatives) is the harmonic mean of precision and sensitivity and thus takes into account both false positives and false negatives and is a useful metric when there is outcome-class imbalance

^§Sensitivity: the true positive rate.

^¶Specificity: the true negative rate

Cross-validation (Table 2) also showed an improvement in balanced accuracy when SNPs were utilized. Although model 3 utilizing only the 46 SNPs had a slightly higher balanced accuracy than model 4 (clinical plus 46 SNPs), i.e., 0.92 vs. 0.88, in cross-validation, we selected model 4 based on the hold-out validation as it is more robust.

To determine the strength of the association of each SNP with the E1-E2-high phenotype, we performed penalized regressions (ridge regression, lasso regression, and elastic net) using our 46 SNPs and clinical data (T-stage, N-stage, BMI, age, and prior chemotherapy status) (Table 3). The best performing penalized model was ridge regression with a lambda value of 0.19. Lambda in ridge regression is called the tuning parameter that controls the strength of the penalty, and it shrinks the coefficients of correlated predictors towards each other. With ridge regression, the balanced accuracy was 0.86, precision was 0.97, F1 was 0.98, sensitivity was 1.0, and specificity was 0.71. Table 4 shows the coefficients from the ridge regression and the variable importance for each of the 46 SNPs. The coefficients from ridge regression denote the effect of each predictor on the outcome. The variable importance in Table 4 indicates how many times a particular predictor was selected in the 100 cross-validation iterations, indicating the importance in prediction accuracy. As expected, BMI was the most important of the clinical factors.

Table 3:

Performance metrics of ridge regression for both E1-E2 high model

Model	Balanced Accuracy	Precision	F1	Sensitivity	Specificity
Ridge	0.86	0.97	0.98	1.0	0.71
Elastic	0.85	0.96	0.96	0.95	0.75
Lasso	0.80	0.93	0.94	0.96	0.64

Table 4:

Variable Importance of SNPs and coefficients for each SNP from Ridge regression

SNP	Ridge.Coefficient	SVM Variable Importance
(Intercept)	−4.28	NA
agecat<=50	0.13	NA
agecat≥60	0.13	NA
bmicat35+	0.84	NA
bmicat<25	−0.34	NA
rs67798809	0.12	100
rs11704334	−0.26	91.94
rs2601805	0.32	86.21
rs56983565	0.18	83.69
rs7160206	0.18	80.06
rs7289476	0.09	79.33
rs1777688	0.27	76.88
rs8047093	0.2	76.86
rs996241	0.11	76.06
rs10220533	−0.26	75.83
rs1680668	0.11	74.38
rs8057773	−0.22	72.75
rs56035123	0.15	72.23
rs2889231	0.15	71.4
rs11596597	0.18	71.08
rs76795580	0.19	70.25
rs1637565	0.15	69.38
rs61872963	0.32	67
rs1867971	0.17	66.72
rs12379659	0.1	63.1
rs6585603	0.16	62.75
rs9927867	−0.15	62.4
rs9936386	−0.13	61.1
rs11714833	0.1	59.47
rs7259528	0.23	58.19
rs2734363	−0.23	58.07
rs10817663	0.08	57.74
rs12149652	0.28	56.32
rs8111640	−0.23	56.19
rs7140248	0.16	54.29
rs11150474	−0.16	53.79
rs4246902	0.23	53.49
rs321958	0.11	50.11
rs164338	0.18	49.09
rs10426849	0.11	46.91
rs5908985	0.29	44.28
rs12397042	0.25	43.78
rs11716479	0.11	42.68
rs2222933	0.18	42.38
rs1154784	0.09	41.58
rs4683570	0.11	38.9
rs8008126	−0.26	38.75
rs17282759	0.14	37.13
rs78437864	0.13	32.67
bmicat	NA	25.09
rs113471862	0.11	24.64
rs1000772	0.17	20.42
PRstatus	0.15	14.71

To address the question of whether we could identify a smaller SNP set, we performed additional analyses including only SNPs with vatiable importance of ≥30 (45 SNPs), ≥40 (42 SNPs), and ≥50 (34 SNPs) and saw a gradual decrease in performance metrics indicating a loss of predictive power as we remove more low impact SNPs (Supplementary Table 2).

Network Analysis of the 46 SNPs utilized in the model

Network analysis revealed 14 genes are neighboring the 46 tag SNPs utilized in our model. Three of these genes are long non-coding RNAs (GRM7-AS3, LINC00922, LINC00871). We constructed a network with the remaining 11 genes in relation to estrone, estradiol (beta and alpha). The 46 SNPs regulate several genes in estrogen-related pathways. As shown in the network, the SNPs in the 11 genes (purple color) have direct and/or indirect interactions with many important genes such as AKT, CDK4, ELAVL1, PRKAA1, TRIM28, and FSH. Beta-estradiol directly interacts with two (ADCY9 and ATP2B2 [alias: PMCA2]) of these 11 genes.

Comparison of the M3 and MA27 cohorts

We compared the M3 and the MA.27 case-control cohort (REMARK diagram: Supplementary Fig. 2) in terms of clinical characteristics (Table 1) and genomic distribution (Table 5) of the 46 SNPs identified in the M3 analyses. Table 1 shows the M3 cohort had a significantly higher proportion of node-negative patients (p<0.001), a significantly lower proportion of patients with prior chemotherapy exposure (p<0.001), and lower BMI (p=0.03) compared to the MA.27 cohort.

Table 5:

Genomic comparison of MAF of 46 SNPs between M3 and MA.27 case-control cohort with respect to outcome groups of E1-E2 high (cases) and E1-E2 low (controls)

1	2	3	4	5	6	7	8	9	10	11
SNP	M3 Case MAF (N=44)	M3 Control MAF (N=278)	M3 overall MAF	pvalue (case vs control in M3)	MA.27 Case MAF (N=49)	MA.27 Control MAF (N=103)	MA.27 overall MAF	pvalue (case vs control in MA.27)	pval (M3 overall MAF vs MA.27 overall MAF)	pval (M3 case/control MAF vs MA.27 case/control MAF)
rs164338	0.38	0.2	0.22	0.00021	0.23	0.24	0.24	0.88	0.61792	0.00775
rs1000772	0.33	0.18	0.2	0.00088	0.18	0.16	0.16	0.55	0.23932	0.06042
rs2734363	0.19	0.4	0.37	8.00E-05	0.31	0.45	0.41	0.02	0.35162	0.37848
rs1154784	0.41	0.24	0.26	0.00073	0.2	0.29	0.26	0.14	0.91963	0.00069
rs4683570	0.44	0.28	0.3	0.00156	0.25	0.32	0.3	0.22	0.97219	0.00223
rs17282759	0.2	0.08	0.1	0.00047	0.03	0.1	0.08	0.04	0.20604	0.00012
rs11716479	0.26	0.11	0.13	0.00012	0.04	0.14	0.11	0.02	0.24323	2.00E-05
rs11714833	0.46	0.29	0.31	0.00037	0.23	0.36	0.32	0.02	0.84305	3.00E-05
rs1867971	0.4	0.24	0.26	0.00138	0.27	0.23	0.24	0.46	0.58161	0.0841
rs61872963	0.22	0.11	0.12	0.00024	0.17	0.15	0.15	0.59	0.12182	0.06451
rs1637565	0.33	0.19	0.21	3.00E-04	0.2	0.21	0.21	0.77	0.92179	0.00979
rs11596597	0.23	0.11	0.12	0.00049	0.16	0.15	0.15	0.84	0.19238	0.03553
rs6585603	0.62	0.39	0.42	6.00E-05	0.44	0.4	0.41	0.50	0.83798	0.02827
rs56035123	0.44	0.26	0.29	7.00E-04	0.27	0.18	0.21	0.06	0.00809	0.24893
rs10426849	0.31	0.16	0.18	0.00031	0.24	0.21	0.22	0.49	0.12088	0.06009
rs8111640	0.19	0.42	0.39	5.00E-05	0.37	0.39	0.38	0.72	0.88015	0.01127
rs7259528	0.46	0.26	0.29	1.00E-04	0.25	0.21	0.23	0.41	0.03738	0.02922
rs321958	0.28	0.15	0.16	0.00065	0.17	0.16	0.16	0.74	0.95111	0.03654
rs11704334	0.12	0.36	0.32	1.00E-05	0.26	0.3	0.29	0.59	0.25516	0.00829
rs67798809	0.56	0.31	0.35	1.00E-05	0.28	0.31	0.3	0.61	0.18058	0.00052
rs7289476	0.47	0.25	0.28	5.00E-05	0.27	0.25	0.26	0.79	0.3824	0.00802
rs996241	0.39	0.21	0.23	0.00018	0.26	0.22	0.23	0.53	0.93993	0.03294
rs2889231	0.45	0.27	0.3	0.00035	0.3	0.27	0.28	0.57	0.52117	0.03988
rs4246902	0.49	0.28	0.31	6.00E-05	0.32	0.33	0.33	0.74	0.55818	0.00272
rs12379659	0.43	0.23	0.26	0.00011	0.26	0.28	0.27	0.67	0.77185	0.00266
rs10817663	0.49	0.27	0.3	7.00E-05	0.27	0.3	0.29	0.66	0.66076	0.0018
rs2222933	0.61	0.42	0.44	0.00065	0.4	0.43	0.42	0.58	0.51	0.006
rs7140248	0.27	0.13	0.15	0.00012	0.1	0.13	0.12	0.40	0.28342	0.00112
rs1680668	0.23	0.1	0.11	0.00016	0.1	0.11	0.11	0.66	0.67145	0.00317
rs76795580	0.35	0.17	0.2	2.00E-05	0.2	0.22	0.21	0.72	0.61974	0.00206
rs10220533	0.18	0.44	0.4	0	0.41	0.33	0.36	0.22	0.1861	5.00E-05
rs7160206	0.64	0.42	0.45	5.00E-05	0.44	0.44	0.44	0.92	0.65553	0.00569
rs1777688	0.44	0.21	0.24	0	0.2	0.26	0.24	0.27	0.84317	5.00E-05
rs8008126	0.27	0.4	0.39	5.00E-04	0.35	0.46	0.43	0.06	0.11449	0.6855
rs2601805	0.48	0.26	0.29	2.00E-05	0.33	0.25	0.28	0.15	0.7374	0.04789
rs78437864	0.25	0.12	0.14	0.00048	0.06	0.14	0.11	0.03	0.19733	7.00E-05
rs113471862	0.26	0.12	0.13	0.00025	0.05	0.12	0.1	0.05	0.10733	1.00E-04
rs8047093	0.53	0.33	0.35	0.00015	0.29	0.35	0.33	0.29	0.51533	0.00072
rs12149652	0.49	0.34	0.36	0.00027	0.3	0.3	0.3	0.96	0.03228	0.01706
rs9936386	0.25	0.44	0.42	0.00066	0.45	0.47	0.46	0.73	0.18281	0.04324
rs11150474	0.23	0.42	0.4	0.00074	0.41	0.45	0.44	0.59	0.29773	0.05774
rs9927867	0.15	0.36	0.33	8.00E-05	0.42	0.39	0.4	0.62	0.04163	0.00245
rs56983565	0.66	0.43	0.46	3.00E-05	0.41	0.35	0.37	0.29	0.01222	0.03122
rs8057773	0.23	0.47	0.44	1.00E-05	0.54	0.53	0.53	0.86	0.00736	0.00156
rs5908985	0.28	0.13	0.15	1.00E-04	0.14	0.14	0.14	0.97	0.90702	0.00582
rs12397042	0.23	0.11	0.13	0.00194	0.09	0.12	0.11	0.47	0.48086	0.0073

Column 1: dbSNPID for a SNP

Column 2: Minor Allele Frequency (MAF) of a SNP among E1-E2 high group in M3 case-control cohort

Column 3: MAF of a SNP among E1-E2 low group in M3 case-control cohort

Column 4: overall MAF of a SNP in M3 case-control cohort

Column 5:p-value to identify SNP that have significantly different MAF among E1-E2 high vs E1-E2 low group in M3 case-control cohort

Column 6: MAF of a SNP among E1-E2 high group in MA.27 cohort

Column 7: MAF of a SNP among E1-E2 low group in MA.27 cohort

Column 8: overall MAF of a SNP in MA.27

Column 9: p-value to identify SNPs that have significantly different MAF among E1-E2 high vs E1-E2 low group in MA.27

Column 10: p-value to compare the overall MAF of a SNP between M3 and MA.27 case-control cohort.

Column 11:p-value to assess if the case vs control MAF of a SNP is significantly different in MA.27 from M3.

p-values for columns 5 and 9 were obtained from simple linear regression with dosage as an outcome. P-values for column 10 was obtained from one-way ANOVA model with dosage as outcome and cohort as a group. P-values for column 11 was obtained from two- way ANOVA model with two groups, cohort and case/control status along with an interaction term between the 2 groups with dosage as an outcome.

We examined the genomic distribution (MAFs) in the M3 and MA.27 cohorts (Table 5). In the table, we denoted cases as E1-E2 high and controls as E1-E2 low. In the M3 cohort, the MAFs of all 46 SNPs were significantly different (p-value <0.05) (Table 5, column 5) between its cases and control groups, as would be expected. However, in the MA.27 cohort, only 5 (11%) of the 46 SNP MAFs were significantly different among its case-control groups (Table 5, column 9). The overall MAFs were not significantly different between M3 and MA.27 for 40 (87%) of the 46 SNPs (Table 5, column 10). However, table 5, column 11, shows that 38 (82%) out of 46 SNPs were distributed significantly differently between case and control groups within M3 cohorts and within MA.27 nested matched case-control cohorts. This indicates that genomically, in addition to clinically, the MA.27 cohort appeared to be substantially different from M3.

Examination of the best model from M3 study using the MA.27 case-control sample and the phenotype of E1-E2

Despite finding that the M3 cohort and the MA.27 case-control sample were substantially different both clinically and genomically, we examined our best model from the M3 study (model 4, Table 2), in the MA.27 case-control study. Of the 152 patients included from the MA.27 case-control sample, 49 were E1 high-E2 high and 103 were E1 low-E2 low. The SMOTE-SVM-RBF model that performed the best in M3 study with 46 SNPs and clinical variables (Table 2) was used to predict E1-E2 status. Not surprisingly, the balanced accuracy was 0.50.

DISCUSSION

This study was predicated on the hypothesis that a woman’s germline genetic variability as measured by SNPs would be related to the degree of estrogen suppression with AIs. We had previously identified thresholds for estrogen suppression in a case-control study where women with E1 ≥1.3 pg/mL and E2 ≥0.5 pg/mL after six months of anastrozole adjuvant therapy had a three-fold increased risk of an early breast cancer event [6]. The goal of the current study was to identify SNP biomarkers that could predict women destined to have inadequately suppressed E1 and E2 concentrations while receiving adjuvant anastrozole. This study was performed in the M3 study where we had clinical and genome-wide genotype data plus E1 and E2 concentrations while receiving anastrozole. The LLQs of the E1 and E2 assays in the M3 study were slightly higher than in the MA.27 study, i.e., 1.56 pg/mL and 0.625 pg/mL, respectively, in M3 compared with 1.0 pg/mL and 0.3 pg/mL, respectively, in the MA.27 study. The GWAS, with the phenotype of E1 and E2 at or above the identified thresholds, did not identify any genome-wide significant SNPs. However, given that important correlations between genes and phenotypes may be missed using strict adherence to cut-offs for declaring statistical significance [10], we took a novel approach utilizing the most significant p-values and restricting the OR to ≤0.67 or ≥1.5, and MAFs to ≥ 0.1 from the M3 GWAS. Following the identification of 46 SNPs, we explored a variety of models. When clinical characteristics of BMI, age, N-stage, T-stage, prior chemotherapy status, and PR status alone were utilized, e.g., balanced accuracy was only 53%. Given that the phenotype involved estrogens, it was not unexpected that BMI was the most influential clinical factor (Table 4). Utilizing the 46 SNPs alone we identified a balanced accuracy of 0.80, which is substantially higher than that observed for clinical factors alone (0.53) indicating that the genetic factors are more informative than clinical factors. We found that a model including clinical factors plus the 46 SNPs revealed remarkable performance characteristics with a balanced accuracy of 0.93 for predicting the desired phenotype of E1 and E2 at or above the thresholds noted above.

Our network analysis of the 46 SNPs revealed that Beta-estradiol interacted with two genes, ADCY9 and ATP2B2. ADCY9 encodes a membrane-bound adenylate cyclase and has been found to be associated with obesity in prior GWAS and with the regulation of lipid metabolism [28]. The role of ADCY9 in metabolic regulation is associated with the estrogen receptor pathway [29]. Reduction in ATP2B2 has also been shown to interfere with ERα signaling and abrogates the analgesic effects of E2 [30]. In addition, we also observed that the tag SNPs interact with several important genes in three estrogen-related canonical pathways. The three canonical pathways are estrogen receptor signaling (a pathway that refers to all proteins of estrogen function and regulation), estrogen biosynthesis (a pathway that consists of proteins associated with estrogen synthesis), and estrogen-dependent breast cancer signaling (a pathway that includes proteins implicated in estrogen signaling and breast cancer progression). These results indicate that the 46 SNPs are involved in estrogen-related pathways.

We attempted to validate our best model (clinical factors plus 46 SNPs) in a separate sample of patients from an MA.27 nested matched case-control study [6] where we had clinical data, genome-wide genotyping, and E1 and E2 concentrations after six months of anastrozole therapy. This attempt at validation was not successful in that the balanced accuracy was 50% (flip of a coin). A possible explanation, and limitation of our study, is that the MA.27 case-control cohort was substantially different both clinically and genomically. Regarding the latter, in the M3 cohort, the MAFs of all 46 SNPs were significantly different (p-value <0.05, Table 5, column 5) between its cases and control groups, as would be expected. However, in the MA.27 case-control cohort, the MAFs of only 5 (11%) of the 46 SNPs were significantly different among its case-control groups. Thus, one might expect that the MA.27 cohort was not an appropriate validation set because of the selection utilized in the matching process. In addition, the LLQs were higher in M3 compared with MA.27 as noted above. Analysis of the data from our MA.27 case-control study [6], revealed that 9.1% of the anastrozole patients had both E1 and E2 concentrations between the thresholds identified in the MA.27 study and the LLQs of M3. Thus, the phenotype developed with lower LLQs in MA.27 could not exactly be replicated in M3 with its higher LLQs, and we estimate that ~9% of the M3 patients might be misclassified compared to the E1-E2 phenotype definition from MA.27 and this represents a limitation of our study.

For the MA.27 study, a nested matched case-control approach was used where women who developed the breast event within five years of starting the anastrazole therapy are defined as cases and controls were those with disease-free follow-up at least six months longer than the case. In addition, the controls were the women who were within five years of age to a case, had the same disease stage (stage I, II, or III), same BMI category (<25.0, 25.0–34.9, or ≥35), and same adjuvant chemotherapy status (yes or no) as the case. Hence, a possible design limitation could be, that since our best model from M3 included clinical variables such as BMI, age and other features, by virtue of design of the MA.27 nested matched case-control cohort, we might have attenuated the association of BMI and age with E1/E2 phenotype resulting in low accuracies.

An additional limitation of our study was that the SNP-set selection from the genome-wide analysis of the M3 study was from an unbalanced and small cohort (44 high E1-E2 and 278 low E1-E2). This might have led to an overestimation of prediction metrics in the M3 study as we used all 322 samples for the GWAS. Thus an additional independent validation cohort was necessary. Given that there are no other prospective adjuvant anastrozole trials wherein clinical, genome-wide genotyping data, and determination of E1 and E2 concentrations are available, new prospective studies are necessary in order to validate the SNP-based model we identified from the M3 study. We intend to study this model prospectively in a Mayo Clinic Breast Cancer SPORE trial that has been activated (ClinicalTrials.gov Identifier: NCT04294225).

Conclusions

We demonstrated that the addition of 46 SNPs to clinical factors in a model predicted the phenotype of E1 and E2 above the LLQs with markedly improved performance characteristics over clinical factors alone in our M3 test set. This model was not validated in our MA.27 validation set that came from a case-control study, and was substantially different both clinically and genomically.

Figure 2.

Representation of the estrone and estradiol interactions with the genes associated with 46 tag SNPs identified in this study that were associated with estrogen suppression after six months of anastrozole treatment. In the network, symbols corresponding to the 46 tag SNPs are colored in purple; genes colored in white are known interactions but did not display any evidence of selection. Estrone, alpha-estradiol, and beta-estradiol are colored in orange oval boxes. Interactions with three major canonical pathways (CP) were also identified. The links between the CP of Estrogen Biosynthesis and estrone and beta-estradiol are displayed in sky blue color.