There is growing interest in the utility of pharmacokinetic gene polymorphism screening in psychopharmacological treatment, particularly with antipsychotic medications and older antidepressant agents 
. Others have argued that the efficacy and toxicity of most psychotropics could be influenced by DNA variants in pharmacokinetic genes, and that drug selection and dosage should ideally be based on genotypic information 
. There is growing consensus that there is little data that suggests that assessment of cytochrome P450 polymorphisms may be clinically useful for guiding SSRI therapy 
The flat dose-response curve and wide therapeutic index of SSRIs argue against a strong relationship between plasma levels and clinical response 
and there is little evidence regarding how plasma levels of citalopram influence clinical efficacy 
. This appears to be the case for citalopram, which has few drug-drug interactions based on in vitro
and in vivo
. Nevertheless, the pharmacokinetics of many SSRIs, including citalopram, are affected by CYP2D6 and CYP2C19 genotype status, as polymorphisms in these enzymes do alter citalopram disposition 
. For example, CYP2C19 poor metabolizers showed a 42% decrease in citalopram clearance when compared to homozygous extensive metabolizers, yet there was no difference in the subject's side effect profile 
. In another study of seven non-responders to citalopram, six of seven were extensive metabolizers for CYP2D6 and all seven were CYP2C19 extensive metabolizers 
. When given an inhibitor of these two enzymes, citalopram serum levels rose in all seven subjects, with six of them showing substantial clinical improvement. These data suggest that enzymes involved in citalopram metabolism may contribute to response, at least in some extensive metabolizers. There are no similar data regarding side effects, although a large (n
749) Swedish study found no difference in citalopram or desmethylcitalopram levels between those experiencing a number of common side effects compared and those who did not, suggesting that side effects are influenced primarily by pharmacodynamic rather than pharmacokinetic factors 
. A study by Murphy et al. also found pharmacodynamic gene variation to be important in antidepressant intolerance 
. Recently, a study was reported involving genotyping of ABCB1 variants in persons taking antidepressants in which an association between several of these variants and response in ~114 persons taking ABCB1 substrates, but not in ~85 persons taking drugs that are not substrates for the protein encoded by ABCB1 
. The three ABCB1 variants that we genotyped for this report were genotyped by Uhr et al., and just as in our study, no association with response phenotypes were noted. For the eleven SNPs found to be associated by Uhr et al., ten are in very strong linkage disequilibrium. Three of these markers were genotyped as part of our unpublished genome-wide association study of antidepressant response in our sample, and these three adequately tag the ten correlated markers based on CEU HapMap Phase II data (r2
1.0 for seven or five of seven remaining markers, respectively). These three markers, rs10280101, rs2235040, and rs12720067, showed p-values of 0.58, 0.30, and 0.56, respectively, for the remission phenotype. Urh et al. reported one additional marker as being associated with antidepressant response. This marker, rs2235015, was not genotyped by us, nor did we genotype a SNP tagging this marker. Thus it is an open question if this last marker shows association to treatment response phenotypes in our sample.
The size of the STAR*D study provides a clinical sample with statistical power to detect moderately sized genetic influences. In this study, we detected no significant association between any of the polymorphisms and our treatment phenotypes. Our two-stage analysis allowed us to control Type I error by requiring validation of our results in a second sample. However, by splitting our sample as such, we sacrificed statistical power, and thus increased the risk of Type II error. For our response phenotype in the discovery set, we had 80% power to detect a minimum detectable odds ratio of 1.9 assuming an allele frequency of 0.05 and 5% significance level using our Caucasian sample. The minimum detectable odds ratio increased to 2.74 for the tolerance phenotype. There are multitudinous potential analyses that can be carried out given the richness of the phenotypic data. In this study, we did not formally correct our results for multiple comparisons, although our two-stage design serves to control Type I error, lending further support to the overall negative results. The availability of our genotype data at the NIMH Center for Collaborative Genetic Studies on Mental Disorders facilitates additional exploratory hypothesis testing.
Our study has several limitations. Given the many differences in SNP allele frequencies and phenotype classification among self-reported ethnic groups, population stratification may be a potential explanation for our negative findings, with true associations being obscured by unobserved population sub-structure. This is particularly relevant given the wide differences in allele frequencies between populations for many of the genes studied here 
. However, population studies have found that self-reported ethnicity is a close surrogate for underlying genetic ancestry information 
, thus we sub-grouped our analyses based on self-reported ethnicity in order to limit potential confounding. By analyzing the ethnicity groups separately and using a two-stage association approach, we had reduced power to detect associations in the African American subgroup, and thus cannot entirely dismiss these loci in this subgroup. We limited our genotyping of pharmacokinetic candidate genes to known, deleterious alleles that are common in Caucasian populations. In order to comprehensively screen these genes, rare and functionally unknown variants would need to be genotyped. The STAR*D clinical study, while large and broad in scope, was not explicitly designed for pharmacogenetic studies of this type. For instance, citalopram was chosen partly due to its lower potential for influence by pharmacokinetic polymorphism. Citalopram dosage was also not fixed, though the majority of subjects (78%) were receiving 40–60 mg per day at the end of the study. The final citalopram dosage prescribed was not influenced by the subject's genotype status. This is consistent with work carried out with many of the same functional DNA variants in the Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) study, in which there was no association to dosing, efficacy, or tolerability to five antipsychotics (David Goldstein, personal communication). This observation is particularly interesting in that others have noted a strong correlation between maximum prescribed dose of phenytoin or carbemazepine in epilepsy and genetic variants in CYP2C9 or SCN1A , suggesting clinical adjustment of dose in response to genotype 
. Reflecting the “real-world” treatment focus of the STAR*D study, patients were not drug naïve and certain concomitant medications for general medical conditions were allowed during treatment. Unfortunately, systematic data on concomitant medications was not collected during the trial, and thus we were unable to control for this theoretical drug-drug interaction effect. It is noteworthy that the analysis of the CATIE study indicates that using concomitant medications known to alter metabolic status did not alter the results (David Goldstein, personal communication). While clinical outcome, not alteration of pharmacokintetic profiles, was our study endpoint, circulating concentrations of citalopram or citalopram metabolites would have been a useful proxy measure of compliance. Unfortunately, plasma citalopram levels were not obtained from any STAR*D subjects, thus unmeasured compliance is a limitation of this study and consequently reduces our statistical power. The design of genetics component of STAR*D was not entirely prospective, with some subjects consenting for DNA collection any time after initiating treatment, raising the possibility that those consenting for genetic analysis do not represent all subjects. We adjusted our analyses for the time period between starting the trial and donating blood for DNA, and found no effect on the results (data not shown). Finally, our findings regarding citalopram may not be generalizable to other SSRI's, each of which has a unique metabolic disposition. Any broadly administered pharmacogenetic test will have to tolerate similar limitations in order to be useful in “real-world” clinical settings. Thus, at least for citalopram, it may be premature to advocate pharmacokinetic gene analysis for dose adjustment or clinical decision making.