Validation and comparison of pharmacogenetics-based warfarin dosing algorithms for application of pharmacogenetic testing

Abstract

Warfarin is a widely prescribed drug that is difficult to use because of its narrow therapeutic window. Genetic polymorphisms associated with warfarin metabolism have been identified, but the clinical utility of genetic testing in warfarin dosing has not been established. External validation of published algorithms is critical to determine the best prediction for warfarin dosing in prospective trials. We used two independent datasets totaling 1095 patients to evaluate four published algorithms and a simple prediction algorithm developed in this study based on the CYP2C9*2, CYP2C9*3, and VKORC1 -1639 polymorphisms in 150 patients taking warfarin. Predicted warfarin doses were calculated and compared for accuracy with actual maintenance doses. All evaluated pharmacogenetics-based dosing algorithms performed similarly for both datasets. The proportion of variation explained (R(2)) was high (60% to 65%) in the small white-only Connecticut dataset but low (36% to 46%) in the large dataset on a diverse ethnic population from the International Warfarin Pharmacogenetics Consortium (IWPC). When comparing the percentage of patients whose predicted dosage are within 20% of actual, the IWPC algorithm performed the best overall (45.9%) for the two datasets combined while other algorithms performed nearly as well. Because no algorithm could be considered the best for all dosing ranges, it may be important to consider the nature of a local service population in choosing the most appropriate pharmacogenetics-based dosing algorithm.

Figure 1

Characteristics of the study population. A: Distribution histogram of the therapeutic maintenance dose of warfarin among study subjects. The x axis represents the warfarin dose in mg/week. The y axis represents the number of subjects for the respective dosing range. Four subjects (weekly dose of 75 to 95 mg) are the main cause for deviation from a normal distribution. All four subjects are wild-type. B: A box-and-whisker plot of the therapeutic maintenance dose for each genotype category for the study subjects. CYP2C9 genotypes are according to standard nomenclature. The VKORC1 genotypes are based on haplotype nomenclature: BB represents wild-type; AB, heterozygote; AA, homozygote.

Figure 2

Development of a new regression algorithm for predicting warfarin dosing (UCHC model). A: Average stable maintenance doses in mg/week based on the number of genotype variants. Number of patients in each group: wild-type (0 variant), 35 (28%); 1 variant, 41 (33%); 2 variants, 36 (29%); 3 variants, 11 (8.8%); and 4 variants, 2 (1.6%). Dose differences across groups are highly significant (P < 0.001). B: Correlation analysis of the actual versus predicted doses using the regression algorithm shown in Table 2. Each patient is represented by a dot. The solid line is the linear regression, and the dotted line is the line of perfect prediction.

Figure 3

Comparison of four published algorithms with the UCHC model for determining warfarin dosing. Shown are the scatterplots of the actual versus predicted doses by each of the algorithms for 974 patients in the IWPC. Solid lines are the least squares regression, and dotted lines represent the line of perfect prediction.

Figure 4

Comparison of all genotype-based dosing models based on the percentage of patients with predicted doses within 20% of actual. The x axis shows three actual dose groups: low (less than 21 mg/week), intermediate (21 to 49 mg/week), and high (greater than 49 mg/week). The y axis depicts the percentage of patients within each dose-group whose predicted dose is ideal (within 20% of actual). A: Comparison in the UCHC dataset. B: Comparison in the IWPC validation dataset.

Figure 5

Sensitivity of dosing algorithms based on a sliding window of ±10 mg of actual doses. A: Comparison in the UCHC dataset. B: Comparison in the IWPC validation dataset.