Investigating the clinical factors and comedications associated with circulating levels of atorvastatin and its major metabolites in secondary prevention

Aims The lipid‐lowering drug, atorvastatin (ATV), is 1 of the most commonly prescribed medications worldwide. The aim of this study was to comprehensively investigate and characterise the clinical factors and comedications associated with circulating levels of ATV and its metabolites in secondary prevention clinical practice. Methods The plasma concentrations of ATV, 2‐hydroxy (2‐OH) ATV, ATV lactone (ATV L) and 2‐OH ATV L were determined in patients 1 month after hospitalisation for a non‐ST elevation acute coronary syndrome. Factors were identified using all subsets multivariable regression and model averaging with the Bayesian information criterion. Exploratory genotype‐stratified analyses were conducted using ABCG2 rs2231142 (Q141K) and CYP2C19 metaboliser status to further investigate novel associations. Results A total of 571 patients were included; 534 and 37 were taking ATV 80 mg and 40 mg daily, respectively. Clinical factors associated with ATV and/or its metabolite levels included age, sex, body mass index and CYP3A inhibiting comedications. Smoking was newly associated with increased ATV lactonisation and reduced hydroxylation. Proton pump inhibitors (PPIs) and loop diuretics were newly associated with modestly increased levels of ATV (14% and 38%, respectively) and its metabolites. An interaction between PPIs and CYP2C19 metaboliser status on exposure to specific ATV analytes (e.g. interaction P = .0071 for 2‐OH ATV L) was observed. Overall model R2 values were 0.14–0.24.ConclusionMultiple factors were associated with circulating ATV and metabolite levels, including novel associations with smoking and drug–drug(–gene) interactions involving PPIs and loop diuretics. Further investigations are needed to identify additional factors that influence ATV exposure.


| INTRODUCTION
Statins are amongst the most highly prescribed medications worldwide. Atorvastatin (ATV) is currently the guideline-recommended first-line hypolipidaemic drug for primary and secondary prevention of cardiovascular (CV) disease (CVD). Statins reduce major CV events bỹ 20-30% per 1.0 mmol/L reduction in low-density lipoproteincholesterol (LDL-C), across a range of baseline CV event risks. 1 Whilst efficacious and generally well tolerated, statins can cause adverse effects in a small proportion of patients including incident diabetes mellitus 2 and statin-associated myotoxicity (SAM). 3 ATV is administered as a calcium salt of its active carboxylic acid form 4 and its metabolism is shown in Figure 1. Briefly, ATV can undergo hydroxylation and/or lactonisation, mediated principally by cytochrome P450 3A4 (CYP3A4) 5 and uridine 5 0 -diphosphoglucuronosyltransferases (UGTs), 6 respectively. ATV and its acid metabolites, 2-hydroxy (2-OH) ATV and 4-OH ATV, all inhibit 3-hydroxy-3-methyl-glutaryl-coenzyme A reductase (HMGCR) to lower LDL-C. All 3 lactone (L) metabolites (ATV L, 2-OH ATV L, 4-OH ATV L) are inactive against HMGCR but can be hydrolysed both nonenzymatically and via plasma esterases and paraoxonases to their corresponding acids. 7,8 ATV and its metabolites are eliminated predominantly via the biliary system. 4 All statins, including ATV, can result in SAM, which ranges from common myalgias (in~5% patients 9 ) where causal attribution may be difficult, through to myopathies of increasing severity with elevated plasma creatine kinase levels (~0.1% patients) to rare rhabdomyolysis (0.1-8.4/100 000 patient-years). 3 Although the mechanisms remain incompletely understood, factors that increase systemic statin exposure appear to increase the risk of SAM. These factors include higher statin dose, advanced age, low body mass index (BMI), and comedications that inhibit CYP3A4 in patients on a statin that is a CYP3A4 substrate (ATV, simvastatin lovastatin). 3,10,11 Statin lactones are considered more myotoxic than statin acids. 12 The aim of this work was to determine the plasma levels of ATV, and importantly its main acid and lactone metabolites, in a large cohort of patients on high dose ATV following a non-ST elevation acute coronary syndrome (NSTE-ACS), and to identify the clinical factors associated with these levels. To the best of knowledge, a real-world assessment of ATV levels on this scale has not been previously reported.

| PhACS study
This investigation harnessed the Pharmacogenetics of Acute Coronary Syndrome (PhACS) study, described previously. 13  The protocol was approved by the Liverpool (adult) research Ethics Committee, UK; site-specific approval was granted at all sites involved and local informed consent was obtained from all study subjects in accordance with the Declaration of Helsinki.

| Determination of ATV and metabolite plasma concentrations
The concentrations of ATV and its 3 main metabolites (2-OH ATV, ATV L and 2-OH ATV L) 14 were determined in EDTA plasma samples collected at V2 by high-performance liquid chromatography-tandem mass spectrometry, using an assay validated to European Medicines Agency criteria, as previously reported. 15 The plasma aliquots used here had not been used before, minimising freeze-thaws. Briefly, 50-μL plasma samples were first mixed with 20 μL of a composite internal standard solution containing a deuterated internal standard for each analyte, and then diluted further by addition of 180 μL 100% acetonitrile containing 0.3% acetic acid. Following centrifugation, the supernatant was diluted 1:2 in water and a 5-μL aliquot injected for analysis. Analytes were separated using a 2.

| Cohort selection
The inclusion criteria for this study were: all PhACS participants on 80 mg ATV daily at both baseline and V2, or on ATV 40 mg at both baseline and V2, and both drug adherence data and a stored plasma EDTA sample available at V2.
Four exclusion criteria were applied: (i) participants that failed standard genetic QC procedures 16,17 ; (ii) participants that had 2 or more ATV analyte levels <0.5 ng/mL (i.e. less than the uniform lower limit of quantification); (iii) participants with ATV concentrations ≥31 ng/mL (80 mg ATV) or ≥ 15.5 ng/mL (40 mg ATV), and; (iv) participants missing any data in the clinical variables considered (see below).
Exclusion criterion (i) was applied because, although a genomewide association study was not conducted here, adherence to its QC steps should improve the quality of the clinical data used (e.g. excluding participants with discordant clinical vs X-chromosome-derived sex) and would facilitate reliable use of specific single nucleotide polymorphisms selected for the genotype-stratified analyses (see below). The PhACS genetic QC and imputation steps have been reported before. 17 Participants with ≥2 analytes <0.5 ng/mL (exclusion criterion (ii)) were excluded because they were significantly more likely to be statin nonadherent (self-reported to have missed at least 1 statin pill in the preceding week, P = .001) compared to participants with no low ATV analyte concentrations. However, those with only 1 analyte <0.5 ng/mL were not statin nonadherent (P = .88), might be exhibiting unorthodox ATV disposition/metabolism, and so were retained. As adherence data had been used to help select the cohort, it was not used in subsequent modelling. Exclusion criterion (iii) was applied because, whilst sample collection time was known, time of last ATV administration was not; time since last ATV was estimated by assuming all participants took their last ATV dose at 2200 the previous evening. A ceiling of 31 ng/mL was set, based on the mean ATV concentration plus 3 standard deviations (SDs) for ATV 80 mg from a previous ATV pharmacokinetics patient study 18 to exclude participants with morning administration and near C max levels; this ceiling concentration was halved for ATV 40 mg patients. Exclusion criterion (iv) was applied to ensure a complete dataset so all models produced for a given endpoint could be fairly compared.

| Clinical variables
To characterise the cohort, baseline demographics and comorbidities, V2 drug status and blood sampling conditions were described as follows: ATV dose, sex, age, BMI, smoking (current vs previous/never), hypertension, hyperlipidaemia, diabetes mellitus, chronic kidney disease (CKD, defined as a serum creatinine >150 μmol/L), previous CVD  Table S1 lists the CYP3A-modifying drugs considered), the estimated time since last ATV dose (see above), and sample storage duration. All variables were considered in model selection (see below) except hypertension, hyperlipidaemia, and prior CVD.

| Statistical analysis
Continuous variables are described by their mean (SD) or median (interquartile range) depending on the normality of their distribution, and categorical variables as their frequency (percentage).
As expected, raw ATV analyte concentrations showed a rightskew distribution, and so the correlation between ATV analyte (untransformed) concentrations was assessed by Spearman's rank correlation. Prior to modelling, all endpoints were log 10 transformed to normalise the distribution of (observed) residuals and optimise homoscedasticity. Clinical covariate multicollinearity was assessed using the variance inflation factor, which was <1.5 for all covariates, indicating negligible multicollinearity.
Multivariable linear regression was used for covariate selection, as has been done previously with sparse ATV sampling. 18 For a given endpoint, first the maximum multivariable linear regression model (containing all 20 covariates) was constructed. Second, all subsets linear regression was performed from this maximum multivariable model, limited by the condition that all derived models include the time since last ATV dose covariate. Third, the derived models were ranked according to the Bayesian information criterion (BIC), which is more conservative than the corrected Akaike information criterion because BIC carries a larger penalty for increasing model complexity. 19 Fourth, all models with a BIC within 2 of the model with the lowest BIC (BIC Δ < 2) underwent model averaging to produce model averaged (unstandardised) coefficients and standard errors for all the covariates included in this set of models, to enable multimodel inferences. The model selection constraint of Δ < 2 was chosen because the empirical evidence supporting each of these models is considered essentially equivalent. 19 Conditional averaging was used, and so for a given covariate, its coefficient and standard error were averaged from only those models with BIC Δ < 2 that contained the covariate. 20 This approach was taken because of the interest in assessing the relationship between individual covariates and the endpoints, with overall model predictive capability of secondary importance. 21 The relative importance of each of the selected covariates to the endpoint was assessed by summing the BIC-derived model weight for all models that underwent model averaging and contained the covariate. The Pvalue for each covariate in the final model is provided for reference only. The utility of the final produced model for describing the observed variation in the endpoint was assessed using the unadjusted R 2 value. This process was undertaken for all endpoints. The above described covariate selection and model averaging was carried out using the R statistical package, MuMIn. 20

| Secondary analysis
To compare model selection approaches, a second analysis was undertaken using multivariable linear regression with stepwise covariate selection based on the probability of F. A covariate was included in a step if its probability of F was both the lowest and <0.05, and previously selected covariates were removed if their probability became >0.10.

| Sensitivity analyses
To investigate the smoking covariate and identified drug-drug interactions further, patients with endpoint values outside the mean ± 2SDs were excluded and the multivariable model re-tested. Secondly, novel plausible associated drug classes were substituted for 1 or more individual drugs from the class, and the multivariable models re-run.

| Genotype-stratified analyses
Exploratory genotype-stratified analyses were carried out to further assess key identified drug-drug interactions. Specifically, the adjusted association of PPIs with endpoints was determined following cohort stratification by CYP2C19 predicted metaboliser status, because PPIs are metabolised by CYP2C19. 22 RMs, for analysis. Only 1 individual was suggested to be a CYP2C19*3 (rs4986893) carrier, and so this variant was not considered further.
Secondly, the impact of ABCG2 rs2231142 (Q141K; imputed) on the adjusted associations of loop diuretics with the endpoints was considered. This is because loop diuretics (furosemide, bumetanide) inhibit breast cancer resistance protein in vitro (BCRP, encoded by ABCG2), 24,25 and the rs2231142 minor allele is a reduction-offunction variant 14 ; thus a BCRP-mediated loop diuretic-ATV interaction may be attenuated in 141 K carrier patients.
The sensitivity and genotype-stratified analyses were only conducted on those endpoints where the perpetrator comedication of interest had been included in the endpoint's main analysis model, and P < .05 was taken to indicate significance. All statistical testing was carried out in the R computing environment, 26

| Nomenclature of targets and ligands
Key protein targets and ligands in this article are hyperlinked to corresponding entries in http://www.guidetopharmacology.org, the common portal for data from the IUPHAR/BPS Guide to PHARMACOLOGY.

| RESULTS
The study cohort included 571 patients; 534 and 37 were prescribed ATV 80 and 40 mg daily, respectively. Table 1 (Tables S2 and   S3). In general, the relative importance of ATV dose, demographics, comorbidities and blood sample characteristics in the models was high, whereas the importance of comedications was more variable.
ATV dose was positively associated with the concentration of each ATV analyte and their total, but not with analyte ratios. Time since last ATV was strongly inversely correlated with analyte and total concentrations, but positively associated with 2-OH ATV/ATV and ATV L/ATV ratios, indicating progressive ATV metabolism over time.
Median sample storage time was 5.9 years; therefore, it was not unexpected that storage duration was newly associated with analyte concentrations and ratios. Concomitant use of a strong/moderate CYP3A drug inhibitor, or amiodarone itself, was associated with higher ATV concentrations and lower 2-OH ATV/ATV and 2-OH ATV L/ATV L ratios, reflecting inhibition of CYP3A-mediated hydroxylation.
Smoking was newly associated with reduced hydroxylation and increased lactonisation ratios. Overall, the proportion of variability explained was modest (R 2 values of 0.14-0.24). Reassuringly, the  Table S4).
The multivariable models suggested novel drug-drug interactions.
PPIs and loop diuretics were both associated with higher concentrations of ATV analytes and their total (Figure 2), with relative importance values ranging from 0.55-1.00 and 0.06-1.00, respectively. P2Y 12 inhibitor use was associated with increased hydroxylation and potentially reduced ATV lactonisation (importance 0.11-0.87).
Although, aspirin and ACEI/ARBs were also included in the hydroxylation ratios, their relative importance was low.

| Sensitivity analyses
When outliers were removed, all identified associations were significant for smoking and PPIs, and the majority for loop diuretics. The single association between aspirin and increased ATV L    .000060

P-value B (SE)
.031 Analysis was conducted limited to those endpoints for which loop diuretic or PPI had been included in the main analysis multivariable model. CYP2C19 poor and intermediate metabolisers were grouped together because there were only 10 patients with the poor metaboliser (*2/*2) genotype, and CYP2C19 rapid and ultra-rapid metabolisers were combined as there were only 24 ultra-rapid metabolisers (*17/ *17). The genotype groups for CYP2C19 considered in interaction testing were rapid/ultra-rapid metaboliser (RM/UM) vs non-RM/UM. hydroxylation remained, and P2Y 12 inhibitor use (99% on clopidogrel, 1% on prasugrel) was still associated with reduced ATV lactonisation. However, the links between P2Y 12 inhibitors or ACEI/ARBs and the 2 hydroxylation ratios became insignificant (supplementary Tables S5 and S6).
Within the cohort, of those patients on a loop diuretic at V2, 90% were taking furosemide and 10% bumetanide. For patients taking a PPI at V2, 58% were taking lansoprazole, 38%  Table S7).

| Genotype-stratified analyses
Overall, the evidence favouring potential novel associations between ATV analytes/ratios with PPIs and loop diuretics was considered strongest, and so were explored further in genotypestratified analyses (Table 5). Interestingly, the association between PPIs and ATV analytes appeared restricted to CYP2C19 RM/UM patients, with significant interaction P-values (PPI*CYP2C19 RM/UM vs non-RM/UM) for the lactone analytes and sum analyte total (Table 5). Stratified analysis suggested a possible influence of ABCG2 rs2231142 on the effect of loop diuretics on ATV analyte levels, but formal interaction testing was insignificant (Table 5).

| DISCUSSION
The main findings of this study were: confirmation of the impact of

| Clinical factors
Women had lower ATV and ATV L concentrations but increased 2-OH ATV/ATV and 2-OH ATV L/ATV L ratios, in keeping with the known~2-fold increased CYP3A4 protein levels 27 and increased CYP3A activity 28 in female compared to male livers.
Increasing age was associated with higher concentrations of ATV analytes, in keeping with previous findings. 29 Some, 28,30 but not all 31 studies have found that CYP3A activity does not change with normal human ageing. Thus, the increased hydroxylation ratios seen here with increasing age are more suggestive of reduced clearance of the hydroxylated metabolites. Higher BMI was associated with lower concentrations of all analytes and total, except for parent ATV, in keeping with increased analyte distribution. Interestingly, a previous sparse pharmacokinetics study limited to parent ATV similarly found no association between ATV and BMI. 18 CKD was not selected in any of the multivariable models except for 2-OH ATV, where it was associated with higher 2-OH ATV levels.
However, CKD was only present in 1 of the 13 2-OH ATV models that underwent model averaging and so its relative importance value of 0.05 was negligible. This is in keeping with the observation that <2% of an ATV dose is excreted through the kidneys. 4 Chronic alcoholic liver disease markedly increases ATV and ATV acid metabolite levels. 4 Nevertheless, hepatic disease was not prominent here, probably because only 5 participants reported liver disease.
Smoking was most strongly associated with an increase in the ATV L/ATV ratio.

| Co-medications
CYP3A drug inhibitors were associated with a higher ATV concentration and lower 2-OH ATV/ATV and 2-OH ATV L/ATV L ratios, which was expected as CYP3A4 is the major enzyme responsible for ATV and ATV L hydroxylation. 5,36 In this study, the low number of patients concurrently on a CYP3A inducer likely accounts for that lack of signal. Amiodarone is a CYP3A inhibitor of unspecified strength and inhibits P-glycoprotein 37 ; cases of muscle toxicity including rhabdomyolysis have been reported in patients on both amiodarone and a statin. [38][39][40] Importantly, this present study not only confirmed the impact of amiodarone on ATV hydroxylation, but extended the findings to demonstrate that amiodarone also inhibits ATV L hydroxylation, resulting in elevated ATV L exposure.
Several potential highly novel drug-drug interactions were identified in this study. However, based on the sensitivity analyses and total number of ATV analytes affected, loop diuretics and especially PPIs were considered the most robust novel associations. Both PPIs and loop diuretics were associated here with increased plasma concentrations of all ATV analytes and their sum total, including the more myotoxic lactone metabolites. 12 Nevertheless, the increases in ATV analyte levels were modest; for example, median ATV concentration increased~14% (4.9 to 5.6 ng/mL) and~38% (4.9 to 6.8 ng/mL) with PPIs and loop diuretics, respectively.
Interestingly, ATV severe myotoxicity cases involving a possible interaction with a PPI (predominantly omeprazole) have been reported [41][42][43][44] and myotoxicity is a possible PPI adverse class effect. 43 Furthermore, it has been suggested that PPIs may modestly boost statin-mediated LDL-C reduction, in keeping with elevated intrahepatic statin levels. 45 Although lansoprazole and omeprazole are substrates and inhibitors of CYP2C19 and CYP3A4, 22,23,46,47 CYP3A4-mediated inhibition would not account for the uniform increase in all ATV analytes observed here; possible explanations include an increase in ATV bioavailability secondary to a PPI-induced rise in gastric pH, or inhibition of an elimination transporter. Given the exploratory finding that the impact of PPIs on ATV analyte levels appeared restricted to CYP2C19 RM/UM patients, it is also plausible that a PPI CYP2C19 product(s) could inhibit a transporter involved in ATV (biliary) efflux, such as BCRP or P-glycoprotein. 48 A recent in vitro evaluation of furosemide identified it as a substrate for OAT1, OAT3, BCRP, OATP1B1 and OATP1B3. 24 Moreover, furosemide and bumetanide inhibit BRCP in vitro, with furosemide being the more potent inhibitor. 24,25 Although formal interaction testing between ABCG2 rs2231142 (Q141K) and loop diuretics was not significantly associated with analyte levels, these are exploratory analyses with a low power to detect an interaction; for instance, the number on a loop diuretic and 141 K was just 22 (Table 5). Thus, follow up functional studies remain of interest. Alternatively, loop diuretics may increase ATV analyte levels through depleting intravascular volume.
Neither CYP2C19 status nor rs2231142 were associated themselves with ATV levels (not shown).
Overall, the magnitude of the identified PPI and loop diuretic associations were modest and of doubtful clinical importance at the population level. However, in specific individuals already likely to have elevated ATV levels due to other risk factors, such as elderly frail patients with comorbidities and polypharmacy, the additional increase in ATV exposure due to these novel drug interactions could plausibly contribute to incident SAM.

| Study limitations
To the best of our knowledge, this study constitutes the largest study of ATV systemic exposure to date, and was conducted in a real-world patient cohort. Furthermore, the main acid and lactone metabolites of ATV were analysed alongside parent ATV. However, there are limitations to this study. First, time since last ATV administration was estimated. Nevertheless, time since last dose is not accurately recorded in many observational pharmacokinetic studies. 49 Moreover, the mean sampling time (14 h) in this study is on the low curvature aspect of the analyte concentration-time profiles 8 and so the effect of sampling time measurement error is expected to be low. 49 Overall, time was very strongly associated with the majority of endpoints, and several other covariates (e.g. ATV dose, sex, CYP3A inhibitors) behaved similarly as expected in this study, providing confidence in the results. Second, other potentially relevant clinical covariates, such as grapefruit juice and alcohol consumption, were not recorded, and so could not be tested. Third, variation in certain genes (e.g. SLCO1B1) is recognised to directly influence ATV levels. However, clinical factors were prioritised here because of their ready availability in clinical practice.

| Conclusions
This study has demonstrated multiple factors including demographics,