Our data showed a high inter-individual variability in tamoxifen and endoxifen concentrations, while age, weight and CYP2D6 genotype were identified as statistically significant predictors of endoxifen concentrations. Consequently, a large proportion of the variability remains unexplained by the model at hand, suggesting the presence of additional unobserved predictors affecting true patient-specific endoxifen concentrations. By means of logistic regression, ROC analyses showed an optimism-adjusted AUC of 90% (95% CI: 0.86–0.95), thereby indicating an excellent predictive accuracy of subtherapeutic endoxifen concentration with CYP2D6 genotype and age as statistically significant predictors. However, the model showed a sensitivity and specificity of 66 and 98 percent, respectively, indicating a high probability of (misclassification) error for the patients with subtherapeutic endoxifen concentrations. Consecutively, those patients will have a high false negative rate and thereby potentially misclassified.
The results of this study confirm that CYP2D6 genotype accounts for a large proportion of the variability and has high predictive properties for identifying subtherapeutic endoxifen concentrations. Comparable results were found in estimates and R2 values by colleagues Teft et al. and Schroth et al. explaining 39 to 58 percent of the inter-individual variability in endoxifen concentrations (12, 36). Additionally, other iso-enzymes of the cytochrome P450 system significantly affects endoxifen formation (13, 16, 23). Puszkiel et al. investigated the effects of CYP3A4*22, CYP2C19*2, and CYP2B6*6 on endoxifen concentrations. CYP3A4*22 homozygous and heterozygous patients were associated with a 16–25% higher endoxifen concentration compared to wild type, irrespective of CYP2D6 genotype. However, no significant differences were found. CYP2C19*2 and CYP2B6*6 showed only marginal effects on endoxifen concentrations, indicating minimal clinical value (37). Thus, CYP2D6 is evidently an important factor in predicting endoxifen concentrations.
Age and weight were identified as statistically significant predictors, but only explained an small (1.5–1.8) percent of the total variability with small effect sizes after adjusting for CYP2D6 genotype, respectively. This is in contrast with recent studies. (37). Conflicting results regarding the impact of age with either no association, increased endoxifen concentrations or decreased endoxifen concentrations in older patients have been reported (36–38).
Weight has been described as a significant predictor for the formation of endoxifen. However, our data implied a negative association in the multiple linear regression analysis between body weight and endoxifen concentration and only explained an additional ≈ 1.5% of the total variability including small effect sizes (-0.005; 95% CI: -0.008 to -0.002). Interestingly, Mueller-Schoell et al. identified participants with high body weight at increased risk of subtherapeutic endoxifen. Up to 13-fold differences in endoxifen concentration were found in heavy young (22 years, 150 kg) and light elderly (95 years, 39 kg). However, analysis was done across extreme values and the risk shrunk after averaging across the population. Additionally, and similarly, after multivariate analysis and identifying CYP2D6 genotype, age and weight as significant predictors: the unexplained inter-individual variability in endoxifen remained high (54%) and therefore predictions made based on their model may deviate in real life (22).
A strength of this study lies in the statistical analysis. MI was used to minimalize bias by missing data (under the assumption of MAR) and results of both primary analysis and CCA were presented. Moreover, prediction model performance was quantified relying on a bootstrap approach to reduce model optimism and to function as internal validation of the prediction model. The results of the primary analysis and the CCA showed marginal differences. These marginal differences are likely caused by the low proportion of missing data (≈ 7%) in the total dataset and even an even lower proportion of missing data (≈ 1%) amongst the predictors included in the model.
Earlier research showed, fairly extensive use of important inhibitors and inducers in the general ER + breast cancer population. Primarily, concurrent treatment with CYP2D6 inducers may significant decrease endoxifen concentration and should be monitored, e.g., selective serotonin reuptake inhibitors (SSRI) are known to decrease endoxifen concentration (12, 13, 16). Therefore, concurrent treatment with, especially, CYP2D6 inhibitors should be monitored and possibly intervened accordingly – to sustain tamoxifen efficacy.
Additionally, 91% of our population showed high adherence to tamoxifen treatment whereas five and six percent of the participants showed medium and low adherence, respectively. Likely, this was an advantage for model building as under these conditions a low proportion of the variability is likely to be attributed by the degree of adherence. However, these results lack external validity as adherence is a major issue in the adjuvant treatment with tamoxifen. Pistilli et al. (15) showed that after one-year the adherence to tamoxifen is 86% based on serum concentration and self-declared adherence (95% CI: 84% − 88%). Nevertheless, real-life adherence estimations range from 15–72% non-adherence over 5 years of tamoxifen therapy (17, 18).
Alternatively, the therapeutic drug monitoring (TDM) strategy might be of benefit for this population (39). Personalizing tamoxifen treatment by dose adjustments based on measured steady state concentration is evident and may benefit the population at risk for subtherapeutic endoxifen. Klopp-Schulze et al. identified CYP2D6, drug-drug interactions and age as significant predictors of endoxifen. However, similarly, the unexplained inter-individual variability in endoxifen concentration remained large (47.2%) and therefore they concluded that therapeutic drug monitoring may be a beneficial strategy. A combination of CYP2D6 predicted phenotype guided dosing and therapeutic drug monitoring at steady state concentration was proposed (38). Thereby indicating that PMs might benefit with a start dosing of 40 mg tamoxifen once daily combined with TDM; and IMs might benefit with the standard dose combined with TDM. For NMs 20 mg once daily tamoxifen might be sufficient without TDM for most of the tamoxifen users.
In conclusion, the inter-individual variability of endoxifen concentration could largely be explained by CYP2D6 genotype and for a small proportion by age and weight. However, the remaining unexplained inter-individual variability is high and therefore model informed tamoxifen dosing combined with therapeutic drug monitoring approach – by directly measuring endoxifen concentration – should be a practical tool for personalized tamoxifen treatment.
Table 1
– Patient characteristics at baseline, n = 303.
Characteristic | N (%) or Median (IQR) |
Age, years | 56 | (18) |
Weight, kg | 73.5 | (18.8) |
BMI, kg·m− 2 | 26 | (6.9) |
Tamoxifen, nmol/L | 308 | (137.5) |
Endoxifen, nmol/L | 26.2 | (18.4) |
Natural Log (Endoxifen), nmol/L | 3.3 | (0.7) |
CYP2D6, phenotype PM IM NM Missing data | 25 99 177 2 | (8.2) (32.7) (58.4) (0.7) |
CYP2D6 inhibitor Weak inhibitor Strong inhibitor No inhibitor | 4 1 298 | (1.3) (0.3) (98.4) |
CYP3A4*22 genotype CC CT/TT Missing data | 269 28 6 | (88.8) (9.2) (2) |
CYP3A4/5 inhibitors Weak inhibitor No inhibitor | 12 291 | (4) (96) |
BMI = Body Mass Index, CYP2D6 strong inhibitor: quinidine; CYP2D6 weak inhibitors: escitalopram, sertraline, and citalopram; CYP3A weak inhibitors: pantoprazole, prednisone and omeprazole. PM = poor metabolizer; IM = intermediate metabolizer; NM = normal metabolizer, respectively. |
Table 2
– Primary analysis, multiple linear regression estimates and average-based optimism adjusted R-squared.
Coefficient | Estimate | Std. error | Pr (>|t|) | 95% CI lower bound | 95% CI upper bound | Bootstrap CI lower bound | Bootstrap CI upper bound |
Intercept | 2.178 | 0.172 | < 0.001*** | 1.843 | 2.500 | 1.870 | 2.497 |
CYP2D6 IM/PM IM/IM NM/PM NM/IM NM/NM | 0.275 0.734 1.060 1.116 1.375 | 0.178 0.152 0.149 0.152 0.170 | < 0.001*** < 0.001*** < 0.001*** < 0.001*** < 0.001*** | 0.034 0.424 0.867 0.937 1.208 | 0.492 1.029 1.221 1.283 1.536 | 0.034 0.404 0.910 0.965 1.220 | 0.487 1.033 1.202 1.263 1.506 |
Age | 0.006 | 0.002 | < 0.001*** | 0.003 | 0.010 | 0.003 | 0.012 |
Weight | -0.005 | 0.002 | < 0.002*** | -0.007 | -0.002 | -0.008 | -0.002 |
Adjusted R² | 0.571 | | | | | 0.491 | 0.635 |
Table 3
– Sensitivity analysis and complete case analysis.
Coefficient | Estimate | Std. error | Pr (>|t|) | 95% CI lower bound | 95% CI upper bound | Bootstrap CI Lower bound | Bootstrap CI Upper bound |
Intercept | 2.178 | 0.172 | < 0.001*** | 1.839 | 2.517 | 1.884 | 2.465 |
CYP2D6 IM/PM IM/IM NM/PM NM/IM NM/NM | 0.275 0.734 1.060 1.116 1.374 | 0.124 0.162 0.090 0.090 0.084 | < 0.027* < 0.001*** < 0.001*** < 0.001*** < 0.001*** | 0.031 0.414 0.890 0.942 1.207 | 0.519 1.053 1.234 1.291 1.542 | 0.045 0.435 0.888 0.960 1.232 | 0.459 0.982 1.183 1.251 1.505 |
Age | 0.006 | 0.001 | 0.001** | 0.002 | 0.010 | 0.003 | 0.010 |
Weight | -0.005 | 0.001 | 0.002** | -0.008 | -0.001 | -0.008 | -0.002 |
Adjusted R² | 0.556 | | | | | 0.469 | 0.641 |
Table 4
– Logistic regression for subtherapeutic endoxifen levels.
Coefficient | Estimate | Standard error | Pr (>|t|) | 95% CI Lower bound | 95% CI Upper bound | Bootstrap CI lower bound | Bootstrap CI upper bound |
Intercept | -2.088 | 1.050 | 0.04* | -4.267 | -0.110 | -17.029 | -0.146 |
CYP2D6 NM/IM NM/PM IM/IM IM/PM PM/PM | 1.717 1.793 3.950 5.780 22.223 | 0.690 0.680 0.946 0.965 1293.752 | 0.01* 0.008** < 0.001*** < 0.001*** 0.986 | 0.460 0.559 2.182 4.108 -24.296 | 3.256 3.321 5.978 7.990 496.5 | 0.143 0.153 -0.099 3.825 20.370 | 3.396 3.500 20.908 22.310 24.420 |
Age | -0.028 | 0.017 | 0.10 | -0.062, | -0.005 | -0.056 | 0.001 |