We read, with interest, the paper by Damon et al on predictive models based on selected gene variants that explain variability in tacrolimus dose-corrected trough levels (C 0 ) in renal recipients (1). We want to share some of our critiques regarding methodology and interpretation of data.First, we suggest not using the term "predictive modeling" in a setting of multiple gene variant sets that variably associate with concurrent tacrolimus C 0 at different time points after transplantation. Validation of predictive models would require accurate estimation of tacrolimus dose requirements in independent populations either at similar, separate time points (PLS1 models) or for all time points (PLS2 models; e.g. in the "Necker cohort," although this was not performed). The authors presume that the identified sets of gene variants (because of their distribution with oxidoreductase/monooxygenase activity) are causally linked to tacrolimus metabolism and thus exposure. It is important to note, however, that in the first 3 mo after transplantation, a number of clinical factors known to affect tacrolimus bioavailability closely covary with tacrolimus dosing and influence dose-corrected exposure, for example, graft function or dysfunction, gastrointestinal motility, hematocrit, albumin, corticosteroid tapering, and concomitant drugs (drug-drug interactions) (2). The fact that these factors were not corrected for increases the risk of overfitting the PLS1 models by modeling variability that is likely to be determined by clinical factors. Why would the majority of gene variants that were relevant on day 60 no longer be so on day 90? With the exception of previously identified genetic variants of enzymes involved in tacrolimus metabolism (figure 2 in Damon et al; singlenucleotide polymorphisms [SNPs] are shown in dark red/orange for the PSL1 model, and underlined CYP3A5 and CYP3A4 SNPs contributed to the PLS2 model), the sets of SNPs that are linked at only one or two time points with tacrolimus exposure are purely associative (78 of 85 SNPs) and are not explanatory, let alone predictive, unless validated. The collapse of complexity in the PSL2 models and the concurrent drop in performance (R² = 0.28) illustrate this point quite clearly.Second, the rationale for separate analyses in the fixed-dose subgroup and the CYP3A5 nonexpressers in the adapted dose group is not clear. The majority of white patients (80% CYP3A5*3/*3) are initially overexposed with a standard tacrolimus loading dose (0.2 mg/ kg) (3-5). Because all tacrolimus concentrations (whether in the fixed-or adapted-dose group) are considered to be in steady state on day 10 (no dose changes allowed until then), the differences in PSL1 model performance in these subgroups (supplementary table 3 in Damon et al) cannot be explained by different dosing strategies, as the authors suggest (1). For later time points (after day 10), this could theoretically be the case because more dose adaptations were performed in the fixed-dose subgroup, resulting in more "non-steady-sta...