Differential effects of predictors of warfarin dose according to race/color categories in the admixed Brazilian population

    loading  Checking for direct PDF access through Ovid


Pharmacogenomics dosing algorithms are consistently more accurate in predicting warfarin dose requirement for patients of European ancestry compared with those of predominant or exclusive African descent 1–4. In the extensively admixed Brazilian population, however, a pharmacogenetic algorithm showed comparable accuracy in predicting stable warfarin doses, irrespective of self-reported ‘race/color’ 5. Prompted by the observation that predictors of warfarin dose differ between European Americans and African Americans 6, we applied multivariate regression modeling to carry out ‘race/color-stratified analyses for White (N=196) and Black (N=76) Brazilians 5. Initially, we assessed the association between individual ln-transformed warfarin dose in each race/color group and the demographic, clinical, and pharmacogenetic variables identified previously as predictors of warfarin dose in the combined cohort (Supplementary Table 1, Supplemental digital content 1, http://links.lww.com/FPC/B192) 5. The final regression models for White and Black patients (Fig. 1) included age, weight, prosthetic valves, amiodarone use, CYP2C9*3, and VKORC1 3673 G>A genotypes as covariates, whereas possession of CYP2C9*2 and simvastatin use were retained in the final model for White, but not Black patients. The predictive power of algorithms derived from the final regression models for the two cohorts differed little. In White patients, the correlation coefficient (R2) and the mean absolute error (MAE) between predicted and observed doses were 0.58 and 6.1 mg/week, respectively and for Black patients those were 0.61 and 7.0 mg/week, respectively.
Our results for the race/color stratified analyses in Brazilians are in agreement with those reported for European Americans versus African Americans 6, in that CYP2C9*2 was a predictor of warfarin dose in White patients, but not in Black Brazilians or African Americans. The differential effect of CYP2C9*2 in Brazilians may be tentatively ascribed to the significantly lower frequency of this allele in Black (0.066), compared with White (0.155) patients (Supplementary Table 1, Supplemental digital content 1, http://links.lww.com/FPC/B192). A similar explanation might apply to the American patients studied by Limdi et al. 6. By contrast, the differential impact of simvastatin use and possession of heart valve prosthesis observed in our patients were not reported for the American cohorts 6. It is noteworthy that statins were examined as predictors of warfarin dose in the latter study, but prosthetic heart valves apparently were not.
Next, we carried out multivariate regression modeling of the individual warfarin dose in the combined cohort of White and Black patients using as candidate variables those listed in Supplementary Table 1 (Supplemental digital content 1, http://links.lww.com/FPC/B192) plus interaction terms of each of the variables with self-reported race/color (e.g. race/color×age, etc.). The final regression model retained all the independent covariates identified in our original analyses of this cohort 5, namely, age, weight, prosthetic valves, simvastatin and amiodarone use, CYP2C9*2, CYP2C9*3, and VKORC1 3673 G>A genotypes, plus one interaction term, namely race/color×prosthetic valve. Possession of prosthetic heart valves was associated with significantly larger warfarin dose requirement in Black compared with White patients. There is no apparent explanation for this differential effect, however, it is noteworthy that prosthetic heart valves were carried by 65 and 53% of Black and White patients, respectively. An algorithm based on the final regression model, comprising the race/color×prosthetic valve interaction term (Supplementary Fig. 1, Supplemental digital content 2, http://links.lww.com/FPC/B193), performed equally well in White (R2=0.60; MAE=6.4 mg/week) and Black (R2=0.60; MAE=7.4 mg/week) patients, but showed little gain in accuracy relative to the original nonstratified algorithm derived for this cohort 5 or the race-stratified algorithms described above, all of which did not comprise race/color interaction terms.
    loading  Loading Related Articles