Evaluation of a weighted genetic risk score for the prediction of biomarkers of CYP2A6 activity

Abstract

The nicotine metabolite ratio (NMR; 3-hydroxycotinine/cotinine) is an index of CYP2A6 activity. CYP2A6 is responsible for nicotine’s metabolic inactivation and variation in the NMR/CYP2A6 is associated with several smoking behaviors. Our aim was to integrate established alleles and novel genome-wide association studies (GWAS) signals to create a weighted genetic risk score (wGRS) for the CYP2A6 gene for European-ancestry populations. The wGRS was compared with a previous CYP2A6 gene scoring approach designed for an alternative phenotype (C2/N2; cotinine-d2/(nicotine-d2 + cotinine-d2)). CYP2A6 genotypes and the NMR were assessed in European-ancestry participants. The wGRS training set included N = 933 smokers recruited to the Pharmacogenetics of Nicotine Addiction and Treatment clinical trial [NCT01314001]. The replication cohort included N = 196 smokers recruited to the Quit 2 Live clinical trial [NCT01836276]. Comparisons between the two CYP2A6 phenotypes and with fractional clearance were made in a laboratory-based pharmacokinetic study (N = 92 participants). In both the training and replication sets, the wGRS, which included seven CYP2A6 variants, explained 33.8% (P < 0.001) of the variance in NMR, providing improved predictive power to the NMR phenotype when compared with other CYP2A6 gene scoring approaches. NMR and C2/N2 were strongly correlated to nicotine clearance (ρ = 0.70 and ρ = 0.79, respectively; P < 0.001), and to one another (ρ = 0.82; P < 0.001); however reduced function genotypes occurred in slow NMR but throughout C2/N2. The wGRS was able to predict smoking quantity and nicotine intake, to discriminate between NMR slow and normal metabolizers (AUC = 0.79; P < 0.001), and to replicate previous NMR-stratified cessation outcomes showing unique treatment outcomes between metabolizer groups.

Clinical Trial Registrations

NCT01314001 and NCT01836276.

1 |. INTRODUCTION

Nicotine is the primary psychoactive compound in cigarettes, responsible for tobacco’s addictive properties.¹ Nicotine undergoes CYP2A6-mediated metabolism to cotinine (COT)² and COT is further metabolized to 3-hydroxycotinine (3HC) exclusively by CYP2A6.^3,4 The 3HC/COT ratio, known as the nicotine metabolite ratio (NMR), is a well-established index of CYP2A6 activity.³ The NMR is strongly associated with CYP2A6 genotype, highly correlated with total nicotine clearance,³ and is associated with smoking behaviors including acquisition,⁵ cigarettes/day,^6–8 smoking topography,⁹ nicotine dependence,^7,8,10 and cessation outcomes.^11–15

Heritability estimates for NMR derived from European-ancestry twins range from 60% to 80%.^16,17 Among European-ancestry individuals, common loss- (*2, *4) and decrease- (*9, *12) of-function * alleles [pharmvar.org] explain approximately 20% of total NMR variation.¹⁷ Since these * alleles are primarily associated with haplotypes that lead to reduced activity, they are less effective in stratifying faster metabolizers, resulting in a wide range of variability among normal metabolizers.

Several genome-wide association studies (GWAS) of the NMR have shown most (more than 98%) of the genome-wide significant variants are concentrated in or around CYP2A6.^16,18–20 Among European-ancestry cohorts, the top replicated hit is rs56113850, a CYP2A6 intronic SNP, explaining approximately 14% to 22% of NMR variation.¹⁶ Putative independent signals (via conditional analysis) included rs56113850, rs113288603, and esv2663194 (ie, CYP2A6*12) among Finnish smokers,¹⁶ and rs56113850, rs2316204, and rs1801272 (ie, CYP2A6*2) among European-ancestry smokers recruited to the Pharmacogenetics of Nicotine Addiction and Treatment (PNAT2) trial (unpublished observations). In addition, novel CYP2A6 diplotypes constructed from rs28399453, rs150298687, rs7260629, and rs57837628 have been identified through next generation sequencing.²¹

Measuring ad libitum NMR requires that individuals are smoking at regular intervals so that cotinine remains in steady state. A CYP2A6 genetic risk score would improve assessment in instances where ad libitum NMR is unavailable, including among nonsmokers, intermittent smokers, and former smokers, as well as in studies in which DNA is available, but other biological matrices are not. Genetic risk scores have been described for multiple complex diseases and behaviors, by aggregating multiple genetic variants into a single predictive measure.^22,23 CYP2A6 is involved in the pathogenesis of multiple diseases and metabolism of several clinical substrates. For example, CYP2A6 activates tobacco-specific nitrosamines; greater CYP2A6 activity is associated with an increased risk for lung cancer^8,24 and several tobacco-related illnesses (eg, COPD and type-2 diabetes).^25,26 Among CYP2A6 substrates also include tegafur, letrozole, metronidazole, and efavirenz.^27,28 A recent set of polygenic risk scores (PRSs) developed by Chen and colleagues demonstrated an ability to capture 9.2% to 16% of the variation in nicotine metabolism markers, but these PRSs were unable to predict either smoking quantity or cessation.²⁹ Improved estimation of the impact of different CYP2A6 genotypes on enzyme activity will enhance the clinical utility of available genotype data. Specifically, improving interpretation of the effect of CYP2A6 gene variants on CYP2A6 enzyme activity will facilitate the ability to use genetics-based approaches to study the influence of CYP2A6 on disease risk and/or drug metabolism in the absence of a measured phenotype.

A CYP2A6 multiplicative model, also referred to as the CYP2A6 metric,²⁹ was constructed by Bloom and colleagues on an alternative CYP2A6 phenotype (COT-d2/(NIC-d2 + COT-d2); abbreviated C2/N2, also referred to as the metabolism proportion²⁹), where COT-d2 represents dideutero-cotinine and NIC-d2 represents dideutero-nicotine.³⁰ The model used seven CYP2A6 polymorphisms (*1A(51A), *1D-Y351H, *2, *4, *9, *12, and *14) and explained approximately 70% of the variance in the C2/N2 measure. Because of nicotine’s short half-life (1–2 h) and cotinine’s long half-life (16–19 h),¹ C2/N2 is usually quantified in a laboratory setting where deuterated compounds are consumed orally. In contrast, because of the long half-life of cotinine and formation dependence of 3-hydroxycotinine on cotinine,^1,3 NMR can be reliably derived from nicotine derived from ad-libitum smoking^31–33 and is minimally affected by other enzyme pathways.^34–36 We are unaware of studies directly associating the laboratory phenotype C2/N2 with smoking behaviors, including response to smoking cessation therapies. It is also unclear how C2/N2 correlates with nicotine clearance or NMR, or whether the multiplicative metric, which was primarily designed to fit the C2/N2 measure, would be effective in predicting NMR.

This study seeks to (1) assess how well a previous CYP2A6 multiplicative gene scoring approach based on the C2/N2 phenotype predicts NMR; (2) develop an improved genotype model, a weighted genetic risk score (wGRS), specifically designed to predict NMR among those of European-ancestry through the integration of NMR GWAS signals with established * alleles and evaluate this wGRS in an independent cohort to validate model generalizability; (3) compare C2/N2 and NMR directly with each other and with nicotine clearance; (4) evaluate the wGRS’s relationship to smoking quantity and nicotine intake; (5) examine the ability of the wGRS to discriminate between slow and normal metabolizers based on NMR cut-points implicated in clinical outcomes; and lastly, (6) compare the wGRS with the NMR in predicting smoking cessation outcomes.

2 |. MATERIALS AND METHODS

2.1 |. Study populations

Each study was approved by institutional review boards at all participating sites and at the University of Toronto. Participants providing written informed consent for DNA sample collection and release of de-identified information to investigators underwent genotyping.

2.1.1 |. Training set

A total of N = 933 treatment-seeking European-ancestry smokers were recruited to the PNAT2 clinical trial [NCT01314001], where ancestry was determined from GWAS data using principal components analysis¹⁹; 96.8% of European-ancestry smokers and 98.5% of African-ancestry smokers in the PNAT2 trial had genetic ancestries concordant with self-reported ancestry.¹⁹ Study details are described elsewhere.¹² The NMR was measured from whole blood³² collected at intake. Smoking quantity in PNAT2 was assessed at intake using self-reported cigarettes per day (CPD), assessed both as a continuous variable, and by a 4-level ordered grouping strategy (CPD ≤ 10; 11 ≤ CPD ≤ 20; 21 ≤ CPD ≤ 30; and CPD ≥ 31) as previously described.²⁹ Nicotine intake at this time point was assessed by the sum of cotinine and 3-hydroxycotinine (COT+3HC), a superior biomarker to cotinine alone^6,37 as cotinine can overestimate smoking quantity in CYP2A6 slow metabolizers because of reduced COT metabolism to 3HC.³⁸

2.1.2 |. Replication cohort

A total of N = 196 treatment-seeking European-ancestry smokers were recruited to the Quit-2-Live (Q2L) clinical trial [NCT01836276], where ancestry was self-reported. Study details are described elsewhere.³⁹ The NMR was measured from whole blood collected at intake.

2.1.3 |. Laboratory-based pharmacokinetic study

A total of N = 92 European-ancestry participants (N = 44 smokers and N = 48 nonsmokers) were recruited to the Pharmacogenetic Study of Nicotine Metabolism, where ancestry was self-reported. Study details are described elsewhere.³ Fractional clearance to cotinine, C2/N2, and the NMR were derived from single-dose oral administration of deuterium-labeled nicotine (nicotine-d2). We refer to this version of the NMR as NMR-d2, which has also been referred to as experimentally ingested NMR (eNMR) by others.²⁹ Metabolite concentrations for C2/N2 and the NMR were determined from plasma samples collected at 30 minutes and 6 hours following oral administration of nicotine-d2, respectively, as previously described.^3,30

2.2 |. Genotyping

Copy number variants were determined through TaqMan copy number assays (ThermoFisher Scientific, Waltham, Massachusetts, USA) according to the manufacturer’s protocol and previous studies.⁴⁰ A gene conversion variant in the CYP2A6 3′ flanking region (CYP2A6*1B) was determined by a two-step PCR assay as described previously.⁴¹ For the training set, single nucleotide variants were genotyped using an Illumina HumanOmniExpressExome-8 version 1.2 array with a custom add-on containing more than 2500 additional variants; details on genotyping, quality control procedures, and imputation are found elsewhere.¹⁹ For the replication cohort and the laboratory-based pharmacokinetic study, single nucleotide variants were genotyped by quantitative polymerase chain reaction (qPCR) approaches including TaqMan (rs56113850, rs113288603, rs1801272 (CYP2A6*2)) and two-step PCR (rs2316204, rs1137115 (CYP2A6*1A), rs28399433 (CYP2A6*9), rs28399435 (CYP2A6*14)) assays (ThermoFisher Scientific) as described.⁴¹ Genotype frequencies for each of the SNPs were in Hardy-Weinberg equilibrium (P > 0.05) in each of the cohorts. Genotyping methods were cross-validated in a subset of training set samples (N = 141) to confirm validity of comparisons between SNPs that were imputed in the training set (ie, rs113288603, rs56113850, and rs2316204) and qPCR results. rs113288603 and rs2316204 yielded 100% concordance between genotyping methods; rs56113850 yielded a 94% concordance rate, with mismatches found predominantly in samples with a CYP2A6 deletion (CYP2A6*4).

2.3 |. Gene scoring models

2.3.1 |. Multiplicative model

The multiplicative model was employed as described previously using described variants and parameters.⁴² Briefly, model parameters were fit to the C2/N2 phenotype and included genotypes for six CYP2A6 * alleles (*1A(51A) (rs1137115), *2 (rs1801272), *4 (gene deletion), *9 (rs28399433), *12 (CYP2A6/2A7 hybrid), and *14 (rs28399435)), with final scores ranging from 0.44 to 0.90. CYP2A6*1D-Y351H is a rare allele that was not identified in our training set.

2.3.2 |. Weighted genetic risk score model

An additive wGRS was constructed to the variation in the NMR in the training set, where the selection of variants came from four variant sets. Set 1 were independent signals identified by conditional analysis in the training set. Details on GWAS methodology and conditional analysis for the training set (variant set 1) were identical to methods used previously for the African-ancestry participants recruited to the PNAT2 clinical trial.¹⁹ Set 2 were independent signals identified from another large-scale European-ancestry GWAS of the NMR.¹⁶ Set 3 were CYP2A6 * alleles not identified in sets 1 and 2, but common in European-ancestry populations, and set 4 were SNPs identified in a recent CYP2A6 sequencing study.

Using the combination of independent signals (variant sets 1 & 2) as the initial baseline set of variants for the wGRS model, additional established CYP2A6 gene variants that are frequently studied in candidate gene analyses were tested. These were from set 3 CYP2A6 * alleles (*4 (gene deletion), *9 (rs28399433), *1A(51A) (rs1137115), *1B (58 base-pair gene conversion in the 3′ UTR of CYP2A6), and *14 (rs28399435)), as well as set 4 SNPs identified in a recent CYP2A6 sequencing study (rs28399453, rs150298687, rs7260629, and rs57837628; Tanner et al., 2017). These additional variant sets (3 & 4) were investigated by evaluating the variance (R²) captured (in combination with the independent signals from variant sets 1 & 2). Only the variants which contributed an additional source of variance to the NMR phenotype (ie, increased the R²) were added into the wGRS model. Scores were created by summing the number of risk alleles weighted by their unstandardized effect sizes. The use of unstandardized betas is to retain the unit of measurement from the GWAS analysis, where standardized variables invite bias because of sampling error. Betas were estimated from frequentist additive linear regression models (using SNPTEST, version 2.5.2)⁴³ of the NMR phenotype in the training set, adjusted for principal components 1 and 2, and unstandardized through multiplying betas by the standard deviation (SD) of the NMR in the training set (SD = 0.205). To evaluate an individual’s wGRS, the number of risk alleles were summed using their assigned weights (Table 1); this summed value was then adjusted by the addition of 2.0 to create a final score with a positive range of values, assigning individuals with the CYP2A6 diplotype *1/*1, ie, those without the included wGRS decrease or increase of function variants, a score of 2.0 to resemble other CYP gene activity scores, eg, CYP2D6⁴⁴ or CYP2C19⁴⁵ and as outlined by the Clinical Pharmacogenetics Implementation Consortium (CPIC) to assign phenotypes based on CYP genotypes.⁴⁶ This will enhance the clinical implementation of CYP2A6 genetic data by facilitating the conversion of the wGRS to an analogous CYP activity score. For example, an individual with the risk allelesT/C for rs56113850, T/T for rs2316204, A/T for CYP2A6*2, and reference alleles for the remaining genotypes would be as assigned an activity score of 2.046 (ie, [{0.135 + 0.16 − 0.25} + 2.0]).

TABLE 1.

The seven variants included in the weighted genetic risk score (wGRS)

Variant	Reference Allelec	Risk Allele	Location with Respect to CYP2A6 Gene	Beta per Risk Alleled	Weight per Risk Allelee
rs56113850a,b	T	C	Intron 4	+0.657	+0.135
rs2316204a	C	T	5 kb 3′	+0.388	+0.080
rs113288603b	C	T	6 kb 5′	−0.122	−0.025
*2 (rs1801272)a	A	T	Exon 3 (L160H)	−1.221	−0.250
*9 (rs28399433)	A	C	Promoter (TATA box)	−0.778	−0.160
*4 (CYP2A6 deletion)	-	Deletion	Deletion of exons 1–9	−1.707	−0.350
*12 (CYP2A6/2A7 hybrid)b	-	Hybrid	Translocation of exons 1–2	−1.329	−0.272
*1 (no wGRS variants)	-	-	-	-	0.000

^aIndependent signals identified from conditional analyses in the nicotine metabolite ratio (NMR) genome-wide association studies (GWAS) in Pharmacogenetics of Nicotine Addiction and Treatment (PNAT2) European-ancestry smokers (training set).

^bIndependent signals identified from conditional analyses in the meta-GWAS of the NMR in Finnish European-ancestry smokers.¹⁶

^cReference alleles are in relation to the positive strand of the genome orientation.

^dBetas were estimated from frequentist additive linear regression models (using SNPTEST, version 2.5.2)⁴³ of the NMR phenotype in the training set, principal components 1 and 2 were included as covariates.

^eThe change in NMR, and thus the “weight per risk allele” was estimated by multiplying the standard deviation (SD) of NMR (SD = 0.205 in training set sample) by the effect size (ie, beta) of the risk allele.

2.4 |. Statistical analysis

All statistical analyses were completed using SPSS version 20 (IBM Corporation) and MedCalc version 17.4 (MedCalc Software). The Shapiro-Wilk test was used to test for normality in dependent variables. Variables which were not normally distributed were log-transformed.³³ Linear regression was used to assess NMR variation accounted for by the gene scoring models, and the overall contribution to NMR variation was assessed after controlling for known NMR covariates (sex, age, and body mass index [BMI])⁴⁷ reflecting factors that influence the NMR in addition to CYP2A6 genetic variants. Correlations between variables in the laboratory-based pharmacokinetic study were assessed using Spearman rank correlation coefficients. Linear regression and Spearman rank correlations were used to evaluate the relationship between the wGRS and measures of smoking quantity and nicotine intake. Receiver operating characteristic (ROC) curve analyses were conducted with two NMR definitions of normal metabolizers (NMR ≥ 0.26 and ≥ 0.31) coding slow and normal dichotomously as the outcome variable, with the various gene scoring models included as continuous predictors. The Youden’s J index was used to determine the criterion for the optimal cut-point in the wGRS. Logistic regression was used to evaluate end-of-treatment quit rates (nicotine patch vs. varenicline) within slow and normal metabolizers defined by NMR or the wGRS. An interaction between treatment and metabolizer group was evaluated as the ratio of odds ratios (ORRs).¹²

3 |. RESULTS

The multiplicative model, originally developed to predict C2/N2 following oral nicotine administration,³⁰ explained 9.2% of the variance (R²) in log-transformed NMR (log-NMR) in the training set (Figure 1). Including non-CYP2A6 covariates (sex, age, and BMI) known to influence NMR⁴⁷ in the model modestly increased the amount of log-NMR variability captured (R² = 12.4%).

FIGURE 1.

Linear regression analysis of the relationship between the multiplicative model score as described previously⁴² and log-transformed nicotine metabolite ratio (NMR) in the training set (N = 933). The multiplicative model explained 9.2% of the variance in log-NMR. When non-CYP2A6 covariates (sex, age, and BMI)⁴⁷ were additionally included, the model explained 12.4% of the variance in log-NMR

The final wGRS model included seven variants (Table 1), derived from the four variant sets comprising five putative independent signals identified using conditional analyses in two exclusively European-ancestry NMR GWASs, as described in variant sets 1 and 2, and two additional functionally relevant CYP2A6 * alleles common (minor allele frequency > 1%) in European-ancestry populations [pharmvar.org], as described in variant set 3. Versions of the model including other common CYP2A6 * alleles (*1A(51A), *1B, and *14; variant set 3), and novel diplotypes (constructed from rs28399453, rs150298687, rs7260629, and rs57837628; variant set 4)²¹ yielded poorer fit to the NMR phenotype and were consequently excluded, suggesting the impact of these excluded variants is captured by the independent signals. The wGRS model was based on an additive genetic model (see individual variant weights, Table 1), which reflects the observed influence of CYP2A6 variant alleles on the NMR, as illustrated for rs56113850 with incrementing allele dosages (Figure S1).

The wGRS was assessed in an independent cohort of treatment-seeking smokers (replication cohort) to validate the model’s generalizability. The developed wGRS model explained 33.8% of the log-NMR variance in the replication cohort. When sex, age, and BMI were included as covariates, the overall model explained 41.1% of the log-NMR variance (Figure 2A). In the training set, the variance in log-NMR explained exclusively by the wGRS model was similar (R² = 33.8%). After including known NMR covariates (sex, age, and BMI), the overall model explained 37.6% of the log-NMR variance in the training set (Figure 2B).

FIGURE 2.

Linear regression analysis of the relationship between the 7-variant weighted genetic risk score (wGRS) and log-transformed NMR in the A, replication cohort (N = 196) and B, training set (N = 933). The wGRS explained 33.8% of the variance in log-NMR in both cohorts. When non-CYP2A6 covariates (sex, age, and BMI)⁴⁷ were additionally included, the model explained 41.1% and 37.6% of the variance in the replication cohort and training set, respectively

Variations of the optimal 7-variant wGRS were then tested in the replication cohort and training set using a subset of the seven variants (Table 2). A 4-variant wGRS including only the common and functionally-relevant CYP2A6 * alleles (*2, *4, *9, and *12) was tested, representing variants commonly used to genotype CYP2A6 and excluding for the three additional variants in the 7-variant wGRS to come from the recent GWASs (rs56113850, rs2316204, rs13288603). The 4-variant wGRS explained 23.1% of the log-NMR variance (31.3% including covariates of sex, age, and BMI) in the replication cohort and 29.6% (33.2% including covariates) in the training set. A 6-variant version was also tested; the 6-variant model may be applicable in instances where copy-number genotyping for *4 (CYP2A6 gene deletion) or *12 (CYP2A6/CYP2A7 hybrid allele) is not feasible. This 6-variant model excluded *4, and used rs28399442 as a proposed surrogate marker for *12.⁴⁸ This version of the model was marginally inferior to the full 7-variant wGRS, explaining 29.3% of the log-NMR variance (33.0% including covariates of sex, age, and BMI) in the training set; rs28399442 was not genotyped in the replication cohort. In evaluating the performance of rs28399442 at capturing the *12 hybrid allele, the rs28399442 surrogate marker correctly identified 34/41 (82.9%) of the *12 alleles previously identified using qPCR copy number variation analysis in the training set, with no false positives. Furthermore, there were negligible differences in the log-NMR variance explained between a 6-variant model that included the *12 surrogate SNP (ie, rs28399442) and a 6-variant model that included the *12 allele determined through qPCR copy number variation analysis: R² = 29.3% versus 28.9%, respectively (excluding covariates), and R² = 33.0% versus 32.7%, respectively (including covariates).

TABLE 2.

Linear regression analysis of the relationship between alternative weighted genetic risk score (wGRS) models and log-transformed nicotine metabolite ratio (NMR) in the replication cohort (N = 196) and training set (N = 933)

	R2; P value		R2; P value
	(without covariates)		(with covariates)e
Model	Replication	Training	Replication	Training
7-variant wGRSa	33.8%; P < 0.001	33.8%; P < 0.001	41.1%; P < 0.001	37.6%; P < 0.001
4-variant wGRSb	23.1%; P < 0.001	29.6%; P < 0.001	31.3%; P < 0.001	33.2%; P < 0.001
6-variant wGRSc	N/A	29.3%; P < 0.001	N/A	33.0%; P < 0.001
6-variant wGRSd	33.3; P < 0.001	28.9%; P < 0.001	40.4%; P < 0.001	32.7%; P < 0.001

Abbreviation: N/A (Not Applicable)—rs28399442 (*12 Surrogate) not genotyped in replication cohort

^a*2, *4, *9, *12, rs56113850, rs2316204, rs113288603.

^b*2, *4, *9, *12.

^c*2, *9, rs28399442 (*12 Surrogate), rs56113850, rs2316204, rs113288603.

^d*2, *9, *12, rs56113850, rs2316204, rs113288603.

^eCovariates included: sex, age, and BMI.⁴⁷

In contrast to previously used methods to classify individuals by CYP2A6 * alleles into slow, intermediate, or normal metabolizers (Figure 3A), the semi-continuous range of values from the wGRS can be used either to replicate similar broad categorizations as demonstrated by splitting the wGRS scale by tertiles (Figure 3B), or more refined divisions using, for example, quintiles (Figure 3C).

FIGURE 3.

Tukey box-and-whisker plots of nicotine metabolite ratio (NMR) distributions. Data is from the training set (N = 933) split as a function of A, * allele groupings; SM, slow metabolizers; IM, intermediate metabolizers; NM, normal metabolizers. V/V, two variant * alleles B, wGRS scale (1.300–2.430) split into tertiles (T, tertile) and C, weighted genetic risk score (wGRS) scale (1.300–2.430) split into quintiles (Q, quintile)

The C2/N2 and NMR-d2 ratios were compared following oral nicotine administration in the laboratory-based pharmacokinetic study. Following oral nicotine-d2, the C2/N2 and NMR-d2 ratios were significantly and strongly correlated with each other (ρ = 0.82) (Figure 4A). Furthermore, both the C2/N2 and NMR-d2 ratios were significantly and strongly correlated with nicotine fractional clearance to cotinine (ρ = 0.79 and ρ = 0.70, respectively) (Figures 4B, 4C). Decrease/loss of function CYP2A6 * alleles (*2, *4, and *9) were distributed throughout the range of C2/N2 values (Figure 5A), in contrast to the NMR where they were concentrated exclusively in the slower half of the NMR range (Figure 5B). The lack of association of these alleles with lower C2/N2 suggests C2/N2 does not represent genetic variation in CYP2A6, which may contribute to the relatively poor ability of the multiplicative model, based on C2/N2, to predict NMR (Figure 1).

FIGURE 4.

Correlations between nicotine kinetic parameters in the laboratory-based pharmacokinetic study (N = 92). Correlations between the CYP2A6 phenotype ratios C2/N2 and NMR-d2 measured at 30 and 360 minutes, respectively are shown in A. Correlations between nicotine fractional clearance (L/min) and C2/N2, and between nicotine fractional clearance (L/min) and NMR-d2 are shown in B, and C, respectively

FIGURE 5.

The distribution of A, C2/N2 and B, NMR-d2 in the laboratory-based pharmacokinetic study (N = 92) as a frequency histogram color-coded by genotype class. Decrease/loss of function CYP2A6 * alleles (*2, *4, and *9) were distributed throughout the range of C2/N2 values A, these CYP2A6 * alleles (*2, *4, and *9) were concentrated exclusively in the slower half of the NMR-d2 range B

The wGRS was significantly associated with smoking quantity, as defined by a 4-level grouping order of CPD (R² = 0.8%, P = 0.006), as well as by Spearman correlation when entering CPD as a continuous variable (ρ = 0.085, P = 0.009). Likewise, the wGRS was significantly associated with nicotine intake, as defined by a log-transformed sum of COT + 3HC (R² = 2.5%, P < 0.001), as well by Spearman correlation to COT + 3HC as an untransformed variable (ρ = 0.132, P < 0.001).

ROC curve analyses were performed to assess the ability of the wGRS to discriminate between slow and normal metabolizers in the replication cohort, using the NMR cut-point of 0.31 (Figure 6) used in the original PNAT2 smoking cessation clinical trial to prospectively stratify slow and normal metabolizers for treatment randomization.¹² The wGRS model, excluding covariates, showed fair to good diagnostic ability to discriminate between slow and normal metabolizers at the 0.31 cut-point, yielding a significant area under the curve (AUC) of 0.78 (95% confidence interval (CI), 0.71–0.85) in the replication cohort. The Youden index J statistic indicated an optimal cut-point wGRS-2.14 to best identify normal metabolizers based on the NMR ≥ 0.31 definition. The wGRS yielded a similar AUC and Youden index J statistic in the training set when dichotomizing NMR metabolism groups based on the 0.31 cut-point: 0.79 (95% CI, 0.76–0.82) and 2.19, respectively (Figure S2). The diagnostic ability of the wGRS appeared superior to the multiplicative model in the training set, which in comparison yielded an AUC of 0.60 (95% CI, 0.55–0.64) for the NMR 0.31 cut-point (Figure S2). Moreover, the wGRS showed consistent diagnostic validity at another cited NMR cut-point used to evaluate smoking cessation treatment outcomes (NMR: 0.26¹³), yielding an AUC of 0.78 (95% CI, 0.71–0.86) and 0.81 (95% CI, 0.78–0.85) in the replication cohort and training set, respectively, compared with the multiplicative model which yielded an AUC of 0.65 (95% CI, 0.60–0.71) in the training set.

FIGURE 6.

Receiver-operating-characteristic (ROC) curve of the weighted genetic risk score (wGRS) (dashed line) in discriminating CYP2A6 slow and fast metabolizer groups using an NMR cut-point of 0.31 in the replication cohort. The wGRS for an nicotine metabolite ratio (NMR) cut-point of 0.31 (slow: NMR < 0.31, normal: NMR ≥ 0.31)¹² yielded an area under the curve (AUC) estimate of 0.78; P < 0.001 (95% confidence interval (CI), 0.71–0.85). The wGRS for an NMR cut-point of 0.26 (slow: NMR < 0.26, normal: NMR ≥ 0.26)¹³ yielded an AUC estimate of 0.78; P < 0.001 (95% CI, 0.71–0.86) (not shown)

In the placebo-controlled PNAT2 clinical trial, focusing on the active treatment arms as the main hypothesis,¹² N = 838 multiracial smokers were randomized to varenicline or nicotine patch based on pretreatment NMR. Normal metabolizers (NMR ≥ 0.31) experienced significantly higher end-of-treatment quit rates on varenicline compared with the nicotine patch, while slow metabolizers (NMR < 0.31) had similar quit rates on varenicline and the nicotine patch resulting in a significant NMR-by-treatment interaction (ratio of odds ratio, ORR = 1.89; 95% CI, 1.02–3.45; Figure 7A). Compared with the complete multiracial cohort receiving varenicline or nicotine patch (N = 838), in the genetically-determined European-ancestry subset (N = 404 training set smokers of the 838 that were randomized to the varenicline or nicotine patch treatment arms), a similar NMR metabolism group-by-treatment interaction (ORR = 2.08; 95% CI, 0.88–4.89) on quitting was observed (Figure 7B). Moreover, substituting NMR with our wGRS (Figure 7C) (normal metabolizers defined as wGRS ≥ 2.14, as described above) produced a similar ORR of 2.30 (95% CI, 0.94–5.58).

FIGURE 7.

End-of-treatment quit rates by treatment group and metabolizer group. Odds ratios (OR) with 95% confidence intervals (CI) comparing the efficacy of varenicline versus the nicotine patch. Metabolizer-by-treatment interaction effects on end-of-treatment quit rates evaluated by the ratio of odds ratios (ORR) with 95% CI. A, NMR stratification (slow: NMR < 0.31, normal: NMR ≥ 0.31) in the complete varenicline and nicotine patch treatment arms from intent-to-treat dataset (N = 838).¹² B, NMR stratification (slow: NMR < 0.31, normal: NMR ≥ 0.31) in the genetically determined European-ancestry subset of the varenicline and nicotine patch treatment arms (N = 404). (C) wGRS stratification (slow: wGRS < 2.14, normal: wGRS ≥ 2.14) in the genetically determined European-ancestry subset of the varenicline and nicotine patch treatment arms (N = 404)

Likewise, the relative treatment effects within metabolizer group in the N = 404 subset were comparable to the observations from the N = 838 dataset (Figure 7A) when stratified by the NMR (Figure 7B), and the wGRS (Figure 7C), where normal metabolizers showed significantly higher quit rates on varenicline versus nicotine patch (OR = 2.54, P = 0.002 by the NMR or OR = 2.55, P = 0.001 by the wGRS compared with OR = 2.17, P = 0.001 by the NMR in the N = 838 dataset), while slow metabolizers demonstrated similar quit rates between treatments (OR = 1.16, P = 0.65 by the NMR or OR = 1.11, P = 0.77 by the wGRS compared with OR = 1.13, P = 0.56 by the NMR in the N = 838 dataset) (Figures 7A–C). Using the Youden J Statistic identified in the training set (wGRS≥2.19) versus the replication cohort (wGRS≥2.14), a similar treatment-by-group effect was observed (Figure S3).

4 |. DISCUSSION

We present a simple 7-variant approach to translate CYP2A6 genotypes into a semi-continuous CYP2A6 genetic measure for use in European-ancestry populations. The model improves on previous approaches, such as traditional broad categorization of composite CYP2A6 genotypes based on CYP2A6 * alleles into slow, intermediate, and normal metabolizers, by translating CYP2A6 genetic information to a semi-continuous metric predictive of the NMR. Moreover, our findings replicated in an independent cohort of treatment-seeking smokers, as demonstrated by a similar proportion of log-NMR variance explained by the wGRS. This successful replication suggests that parameter estimates for the wGRS are precise and may be extended to further studies of European-ancestry populations seeking to classify CYP2A6 genotypes.

Our wGRS model is superior to the multiplicative metric, originally modeled on C2/N2, for predicting NMR variation. This finding is intriguing since, following oral nicotine, both C2/N2 and NMR-d2 were significantly correlated with each other (Figure 4A) and to nicotine fractional clearance to cotinine (Figures 4B and 4C). The discrepancy in NMR prediction may be due, at least in part, to the observation that well-characterized decrease/loss of function CYP2A6 * alleles (*2, *4, and *9) are found across the C2/N2 distribution, in contrast to the NMR where these variant alleles are concentrated exclusively in the slower half of the NMR range (Figures 5A and 5B). This finding suggests that C2/N2, unlike the NMR, does not exclusively reflect CYP2A6 activity, and may be less suitable for phenotyping CYP2A6 and for fitting CYP2A6 genotypes for genetic scoring. Furthermore, the wGRS model, using the same variants and weights as described to predict the NMR, revealed comparable predictive power (12%) to the multiplicative model (11%) in predicting C2/N2 (Figure S4), substantially lower than previously reported for the multiplicative model (70%).³⁰

After including known NMR covariates, the wGRS model explained approximately 35% of NMR variation suggesting, based on NMR heritability estimates of 60% to 80%, that additional genetic variation is yet to be characterized in CYP2A6, and perhaps in additional regulatory or pharmacokinetic genes. Despite the noted linkage disequilibrium between the reduce-of-function CYP2A6 * alleles and the independent signals identified through conditional analysis in GWASs,¹⁶ the inclusion of these independent GWAS signals explained more of the variation in the NMR than the CYP2A6 * alleles alone (Table 2), suggesting that some of the contribution of these GWAS hits to NMR variability is independent of their linkage to CYP2A6 * alleles. Conditional analysis, while informative for identifying independent signals in GWASs, is a relatively conservative approach; many additional CYP2A6 genetic variants not included here likely contribute to the variation in the NMR. In all, 719 genome-wide significant variants were identified in a meta-GWAS of the NMR in Finnish smokers¹⁶; assessment of these variants using approaches beyond traditional univariate analyses may explain a larger portion of the variation in the NMR. Emerging computational approaches, such as those involving more sophisticated predictive modeling (eg, regularized regression)⁴⁹ may help account for the complex linkage disequilibrium and haplotype structures between associated variants, aiding in the prediction of the NMR. One study limitation was the lack of principal components-based ancestry determination in the replication cohort. However, the high concordance between PC-based and self-reported ancestry in the training set (96.8%), as observed in other studies (eg, 96.9% in Sucheston et al⁵⁰), and the high similarity in wGRS fit between the training set and replication cohort (Figure 2) suggests this is a minimal limitation. However, this may contribute to a poorer fit of the wGRS in other studies, especially when the degree of genetic admixture among study participants is high.

The wGRS significantly associated to measures of smoking quantity and nicotine intake as assessed by the relationships to CPD and COT+3HC, respectively. Through ROC analyses, we demonstrated that the wGRS was favorable to the multiplicative approach in distinguishing slow from normal NMR metabolizers based on NMR cut-points that have previously reflected distinct smoking cessation outcomes (Figure 6 and Figure S2).^12,13 The wGRS was able to replicate the within-metabolizer cessation outcomes observed by baseline pretreatment NMR, where normal (NMR ≥ 0.31 or wGRS ≥ 2.14) but not slow metabolizers show significant differences in quit outcomes between varenicline and nicotine patch treatments (Figure 7). Similar interaction effect sizes (ORRs) were noted between NMR and the wGRS approaches, but we were underpowered in the European-ancestry subset (N = 404 of 838) to observe statistical significance; similar and significant effect sizes (ORs) were noted between treatments within normal metabolizer groups, where no differences between treatments were observed among slow metabolizers (Figure 7A–C).

Here we have focused on developing a CYP2A6 gene-scoring approach in European-ancestry populations. Because of the unique differences in SNP rankings based on association strength (P values) and linkage disequilibrium patterns between different ancestral populations,¹⁸ developing unique gene-scoring approaches according to ancestry is likely necessary. Of note, in a meta-GWAS for the NMR in a cohort of exclusively African-ancestry smokers,¹⁹ a distinct list of independent signals was yielded compared with those identified in European-ancestry smokers, and only approximately 40% of the overall significant hits overlapped those identified in the GWAS of Finnish smokers.¹⁶ Bayesian fine-mapping approaches will likely aid in the identification of the causal SNPs that these independent signals are tagging and may give rise to a unified approach in translating CYP2A6 gene-scoring approaches across multiple ancestries, and individuals of mixed ancestry. However, we demonstrated that our current approach of combining independent GWAS signals with functional CYP2A6 * alleles provides a good and immediately available method of translating CYP2A6 genetic variants and warrants extension to other ancestries.

In summary, we have developed an original genomics approach to translate a small subset of seven CYP2A6 genetic variants into a single semi-continuous genetic score. This model replicated in an external cohort, indicating generalizability, associated with measures of smoking quantity and nicotine intake, and showed the ability to replicate NMR-based clinical outcomes for slow and normal metabolizers. Our wGRS approach represents a practical approach for diverse studies seeking to understand the contribution of CYP2A6 genetic variation to tobacco dependence, as well as tobacco-related diseases, and potentially to the metabolism of other clinical substrates.