Introduction

High-dose busulfan is commonly used as the cornerstone of chemotherapy regimen prior to hematopoietic stem cell transplantation (HSCT) [1]. Targeted busulfan is necessary because of the definite dosage–response relationship and narrow therapeutic index: subtherapeutic busulfan systemic exposure, expressed as area under the curve (AUC) < 900 μmol × min/L, results in relapse or graft failure, while supratherapeutic AUC > 1500 μmol × min/L, is associated with fatal toxicities, such as hepatic veno-occlusive disease (VOD) [2, 3]. Since very few busulfan AUCs achieve the therapeutic range after their initial standard doses [4, 5], an individualized busulfan initial dose is warranted for the targeted AUCs to improve clinical efficiency.

Noteworthily, studies [6,7,8] showed that glutathione S-transferase (GST) genotypes played a positive role in the individualization of busulfan dose for optimal clinical outcomes. GSTs, which are the enzymatic superfamily of detoxification, catalyze the conjugation of busulfan with glutathione (GSH) in liver [9]. Of all, GSTA1 is the predominant GST isoenzyme of busulfan metabolism, with minor contributions by GSTM1 and GSTP1 in vitro studies [10]. GST polymorphisms, gene deletion or single nucleotide polymorphism (SNP), cause a decreased metabolic activity of busulfan in patients with genetic variants [11]. In pediatric patients, Nava et al. [6] showed that GSTA1 diplotypes linked to poor busulfan metabolism and associated with AUCs within the toxic range. Consistently, GSTA1 diplotypes influenced the prediction errors of the weight- and age-based methods which are used to calculate the initial doses of busulfan. However, busulfan initial dosing guided by GST genotypes is not recommended in clinical practice due to inconsistent effect of the genetic polymorphism on the pharmacokinetic, efficacy, or toxicity of busulfan [12]. Studies about the influence of GST polymorphisms on busulfan pharmacokinetic (PK) are scant in Asian patients. The PK study of Yin et al. [8] demonstrated that GSTA1 and GSTP1 polymorphisms are significantly associated with busulfan exposure in Chinese adult patients. In addition, both Yin et al. [8] and our previous study [13] showed low busulfan exposure after the first dose in the Chinese population. We found that about half of the first AUCs were below the therapeutic range. These are a necessity to personalize the initial dose of busulfan to improve clinical outcome in Chinese patients.

This study aimed to quantify the impact of GSTA1, GSTP1, and GSTM1 genotypes on busulfan PK variability by population pharmacokinetic (PPK) analysis, and to clarify the role of genetic factors on the individualization of busulfan initial dose in Chinese adult patients undergoing HSCT.

Materials and methods

Patients

Forty-three adult patients who underwent HSCT for malignant diseases were enrolled in the modeling group at the department of Blood Marrow transplantation of Ruijin Hospital, Shanghai, China, from May 2011 to October 2014. Nineteen adult patients were enrolled in the validation group from March 2017 to August 2017. All procedures performed in this study were approved by the Ruijin Hospital Research Ethics Committee and in accordance with the 1964 Helsinki Declaration. Before the study, informed consent was obtained from the enrolled patients.

All patients received intravenous busulfan (Busulfex; Kyowa Hakko Kirin Co., Ltd, Tokyo, Japan) at 1.6 mg/kg twice daily. Busulfan was given as part of the BUCY, BUFLU, and CBV regimens. For BUCY regimen, busulfan was given over 2 h for 4 days on days − 8 to − 5 followed by intravenous cyclophosphamide (CY) 60 mg/kg once daily on days − 3 and − 2. For BUFLU regimen, intravenous fludarabine (FLU) 30 mg/m2 was given once daily with busulfan on day − 6 and − 3. The CBV regimen in validation group included 3 days of busulfan from day − 8 to − 6, VP16 at 400 mg/m2 daily on day − 5 and − 4 and CY 50 mg/kg daily on day − 3 and − 2. Then, gemcitabine was given at 75 mg/m2 on day − 9 and at 10 mg/m2 on day − 5, respectively.

Blood sampling and analysis

Blood samples were collected before busulfan infusion and at 0.5 h, 1 h, 1.5 h, 2 h, 2.5 h, 3 h, 4 h, 5 h, 6 h, 8 h, and 10 h after starting the first dosing. All the blood samples were drawn from a peripheral vein in the arm opposite to the central line where busulfan was infused. 3 mL of whole blood sample was collected in EDTA anticoagulant glass tubes. 1 ml of whole blood for DNA extraction was taken from the samples obtained before busulfan infusion. The residual whole blood was centrifuged at 3000 rpm for 10 min at 4 °C for plasma separation. Busulfan plasma concentrations were analyzed using a validated analytical method [13] by liquid chromatography–electrospray–tandem mass spectrometry. High-performance liquid chromatography separations were performed on an Agilent Eclipse XDB-C18 analytical column (100 mm × 2.1 mm, 3.5 μm) with a column temperature of 40 °C. The plasma was eluted using a mobile phase composed of 2 mmol/L ammonium formate solution (0.1% formic acid)-acetonitrile (60:40, V/V) at a flow rate of 0.2 mL/min. Electrospray ionization mass spectrometry was performed in multiple reaction monitoring mode with the target ions m/z 264.2 → 151.1 (busulfan) and m/z 278.3 → 69.2. The linearity range of busulfan standard curve was 25–2500 ng/mL (R = 0.999). The intraday and inter-day accuracies of busulfan quality control samples were from 96.4 to 109.2% and from 99.2 to 111.2%, respectively. The extraction recovery of busulfan was from 102.0 to 113.7%.

DNA extraction and glutathione S-transferases genotyping

The whole blood samples for DNA extraction were obtained before HSCT. DNA was extracted from peripheral blood lymphocytes by TIANamp Blood DNA Kit (Tiangen Biotech Co., Ltd, Beijing, China). The following genetic variants were determined: GSTA1 -69 C/T (rs3957357), GSTP1 313A/G (rs1695), GSTM1 (null allele). The GSTA1 haplotype (GSTA1*A and GSTA1*B) was determined by the -69 C/T variation in the prompter region of GSTA1. This SNP was previously shown in complete linkage disequilibrium with -631T/G, -567T/G, and -52G/A [14]. The SNPs in the GSTA1 and GSTP1 genes were genotyped by SNaPshot assay according to the manufacturer’s protocols (ABI SNaPshot Multiplex kit, CA, USA). The promoter region of the GSTA1 gene was amplified with the forward primer GSTA1-F (5′-GCTCGACAACTGAATTCCAGGTC-3′) and the reverse primer GSTA1-R (5′-CCCTAGTCTTTGCACCCAACTCAT-3′). The promoter region of the GSTP1 gene was amplified with the forward primer GSTP1-F (5′-CTCATCCTTCCACGCACATCCT-3′) and the reverse primer GSTP1-R (5′-TTTCTTTGTTCAGCCCCCAGTG-3′). Analysis was carried out using Genemapper software (version 4.1; Life Technologies). The GSTM1 gene was PCR amplified (215 bp) together with albumin gene (268 bp) as an internal control [15]. The primer pairs were as follows: GSTM1-F, 5′-GAA CTC CCT GAA AAG CTA AAG C-3′ and GSTM1-R, 5′-GTT GGG CTC AAA TAT ACG GTG G-3′, which yield a 215 bp fragment and β-actin-F, 5′- CAA CTTCAT CCA CGT TCA CC-3′ and β-actin-R, 5′- GAA GAG CCAAGG ACA GGT AC-3′, which yield a 268 bp fragment. PCR products were analyzed on a 2% agarose gel.

Development of a population pharmacokinetic model

A PPK model was developed by nonlinear mixed-effect modeling using Phoenix NLME software, Version 1.4 (Pharsight, A Certara Company, USA), according to the best practices and guidance [16]. The first-order conditional estimation method with extended least squares method (FOCE ELS) was adopted in the whole process to estimate PK parameters and their variability.

Structure model

Plasma concentration data were natural logarithm transformed before PPK modeling. Initially, a one-compartment model with first-order elimination was used as the base model, according to our previous study [13]. The model was parameterized in terms of total body clearance (CL) and volume of distribution (Vd). The inter-individual variability of the PK parameters was estimated using an exponential error model, where Pi represents the individual parameter estimate, P is the typical parameter estimate, and ηi is the unexplained inter-individual variability, which is assumed to follow a normal distribution with a mean of 0 and variance ω2:

$$ P_{\text{i}} = \, P \times e^{\eta i} . $$

A residual variability (additive, proportional, exponential and mixed random intra-individual residual error) model was explored to account for the intra-individual variability according to the log-likelihood (LL) difference between models [the improvement of objective function value (OFV)].

Covariate analysis

The effects of age, sex, actual body weight, body surface area (BSA), diagnoses, hepatic function markers [alanine transaminase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), total bilirubin (TBI)], GST genotypes, and conditioning regimen on pharmacokinetic parameters were investigated as potential covariates . The effects of continuous covariates were modeled in a linear or nonlinear relationship using the following equations:

$$ \begin{aligned} P_{\text{i}} = \, tv\left( {\text{P}} \right) \, + \, \theta \times \, \left( {{\text{covariate}}/{\text{typical value}}} \right) \hfill \\ P_{\text{i}} = \, tv\left( {\text{P}} \right) \, + \, \theta \times \, \left( {{\text{covariate}} - {\text{typical value}}} \right) \hfill \\ P_{\text{i}} = \, tv\left( {\text{P}} \right) \, \times \, \left( {{\text{covariate}}/{\text{typical value}}} \right)^{\theta } \hfill \\ \end{aligned} $$

where tv (P) is the typical value of P, Pi is the value of P for individual, and θ is the estimated effect of the covariate on P.

For categorical covariates, the effect on P was modeled using the following equation:

$$ \begin{aligned} P_{\text{i}} = tv\left( P \right) {\text{Indicator }} = \, 0 \hfill \\ P_{\text{i}} = tv\left( P \right) \times \theta {\text{Indicator }} = { 1} \hfill \\ \end{aligned} $$

where indicator represents the binary covariate being assigned to 0 or 1.

The stepwise covariate modeling and likelihood ratio test were used to test the effect of each variable. The selection of covariates was initially selected by data visualization and biological plausibility. Further screening of potential covariates was performed by forward addition (P < 0.05) and backward elimination (P < 0.01) based on the changes in OFV: any decreases in OFV by more 3.84 and 6.63 (1 degree of freedom) denote an improved fit at P < 0.05 and P < 0.01, respectively. An additional criterion for covariate retention was reduction in the inter-individual variability and improvement in the parameter estimate precision. Clinical significance was also considered for retention of a potential covariate.

Model validation

Accuracy and stability of prediction of the final covariate model were validated both internally and externally.

Graphical inspection of the final model adopted goodness-of-fit plots, including observed concentrations (OBS) versus population predicted concentrations (PRED), OBS versus individual predicted concentrations (IPRED), conditional weighted residuals (CWRES) versus PRED, and CWRES versus time after the first dose. The CWRES results were summarized graphically using the quantile–quantile (Q–Q) plot and the histogram. Bootstrap analysis was performed for internal validation of model. 1000 resamples from the original data were performed. Mean values and 95% confidence interval (CI) of bootstrap parameters were compared with estimates of the final model. The final model was also evaluated graphically and statistically by visual predictive checks (VPC). The final model parameters were used to simulate a data set for 200 replicates and the 5th and 95th CIs of simulated values were processed for VPC.

To validate the developed model externally, the model was further evaluated in an independent group of adult patients undergoing HSCT. The individual PK parameters were predicted by Bayesian estimation (MAXEVAL = 0 in the estimation step, where MAXEVAL is the maximum number of model evaluations that can be used) with Phoenix NLME using the parameters of the final model. The predictive performance was evaluated by calculating the prediction error (PE) and absolute prediction error (APE) using the following equations:

$$ \begin{aligned} {\text{PE}}\% = \, \left( {{\text{Bayesian simulated}} - {\text{ Observed}}} \right)/{\text{ Observed}} \hfill \\ {\text{APE}}\% = \left| {{\text{Bayesian simulated}} - {\text{ Observed}}} \right|/{\text{ Observed}} \hfill \\ \end{aligned} $$

Statistical analysis

The continuous variables were represented as mean with standard deviation or 95% CI. Statistical significance of the difference between groups was calculated by Chi square test. P values < 0.05 were considered significant.

Results

Patients’ characteristics and GST genetic frequencies

The mean age of patients was 31.5 years, and 74.4% of them were male. The mean weight was 64.3 kg in the modeling group (n = 43). There were 19 enrolled patients in the validation group. The age, sex, and weight were not significantly different (P > 0.05) between the validation group and modeling group. Patients’ characteristics are presented in Table 1. Most of the enrolled patients received allogenous HSCT from a human leukocyte antigen-matched donor using peripheral blood stem cells for their hematological malignancies. Five patients in the validation group received autologous HSCT.

Table 1 Characteristics of patients in the modeling group and validation groups
Full size table

The GST genetic frequencies are shown in Table 2. 10 of 43 (23.2%) patients were heterozygous genotype of GSTA1 (GSTA1*A*B), and only 1 patient (2.4%) was homozygous GSTA1 (GSTA1*B*B). Genetic frequency of GST A1*B haplotype was 14.0%. Both GSTA1 and GSTP1 genetic frequencies of patients in modeling group were in Hardy–Weinberg equilibrium (HWE). 2 of the 19 patients in validation group were heterozygous and homozygous GSTA1*B, respectively. The GSTA1 genetic frequency was not in HWE.

Table 2 GST genetic frequencies of modeling group and validation group
Full size table

Structure model

A total of 488 busulfan concentrations from 43 patients were obtained for model development. A one-compartment model with first-order elimination best described the PK of intravenous busulfan. The intra-individual variability of the plasma concentration was estimated using the proportional error model finally. The estimated CL and Vd of basic model were 14.2L/h and 64.1L, respectively.

Covariate analysis

Age, actual body weight, body surface area, and hepatic function markers (ALT, AST, ALP, TBI) were tested as continuous covariates. Sex, diagnoses, GST genotypes, and conditioning regimen were tested as categorical covariates, while non-Hodgkin’s lymphoma, myelodysplastic syndromes, and hybrid acute leukemia patients were analyzed together with acute lymphoblastic leukemia patients relative to the patients with acute myelocytic leukemia and chronic myelocytic leukemia due to their small number. The patients with GSTA1*B*B or GSTP1*G*G were excluded to avoid statistical deviation.

GSTA1 genotype was a covariate on busulfan CL in a statistically significant manner. OFV decreased by 15.64 from 4942.75 (basic model) to 4927.11 (final model), adding the covariate of GSTA1 genotype. The individualized CL was estimated by the following equation:

$$ {\text{CL}} = tv{\text{ CL }} \times \, \left( { 1+ {\text{CL}}_{\text{part}} \times G_{GSTA1} } \right) \times e^{{\eta {\text{ij}}}} , $$

where GGSTA1= 1 for GSTA1*A*B, 0 for GSTA1*A*A. Population-estimated CL of the final model was 15.0 L/h for a typical patient with GSTA1*A*A wild type. The inter-individual variability of CL declined from 17.8 to 16.8%, after inclusion of GSTA1 genotype as a covariate. The estimated PPK parameters and bootstrap results are listed in Table 3.

Table 3 Population pharmacokinetic parameters of intravenous busulfan and bootstrap results
Full size table

Model evaluation

Goodness-of-fit plots of the final model in Fig. 1 showed that the PRED and IPRED were in reasonable agreement with OBS. Distribution of CWRES versus predicted concentration and time after dose were shown to be close to zero and uniformly distributed within the range (− 2 to 2) in the final model. Additionally, the CWRES distribution and histogram indicated that the assumption of normal distribution of the differences between PRED and OBS was acceptable, as shown in Figure S1. The robustness of the final model was internally validated by the bootstrap results. The population-estimated parameters were similar to the simulation values and within 95% CI from bootstrap analysis, as shown in Table 3. The VPC in Fig. 2 showed that the observed values were almost all positioned within the 5th and 95th CIs of simulated values. The results demonstrated the accuracy of the final model.

Fig. 1
figure 1

Goodness-of-fit plots of the final model. a Observed concentrations versus population predicted concentrations; b observed concentrations versus individual predicted concentrations; c conditional weighted residuals versus population predicted concentrations; d conditional weighted residuals versus time after the first dose. OBS means observed concentration; PRED means population predicted concentrations; IPRED means individual predicted (IPRED) concentrations; CWRES means conditional weighted residuals

Full size image
Fig. 2
figure 2

The visual predictive check (VPC, n = 1000) of final model. The dots represent observed concentrations. The lines represent observed concentrations and the shadow represents the 5th and 95th percentiles of the simulated values

Full size image

Busulfan concentrations from 19 patients in the external validation group were used to validate the final GSTA1 genotype-based model. The predictive performance of the final PPK model was evaluated by the comparison between the observed values (CLobs and AUCobs) and the simulated values (CLsim and AUCsim),as shown in Fig. 3. The Bayesian simulated PK parameters highly correlated with the observed data (r2 = 0.98 for the CLobs with CLsim and r2 = 0.84 for the AUCobs with AUCsim). The mean PE (MPE,  %) and mean APE (MAPE,  %) of the CLsim were 4.04% (95%CI: 4.02% to 4.06%) and 4.39% (95% CI: 4.57% to 4.51%). The MPE and MAPE of the AUCsim were − 2.55% (95%CI: − 2.58% to − 2.52%) and 4.93% (95%CI: 4.92% to 4.94%).

Fig. 3
figure 3

GSTA1 genotype-based simulations of CL and AUC in validation group (n = 19). a The box and whisker plot of observed CL and simulated CL. b The box and whisker plot of observed AUC and simulated AUC

Full size image

Discussion

In this study, we developed the PPK model of intravenous busulfan, incorporating GST A1 genotype, and externally validated the accuracy and stability of the model in Chinese adult patients undergoing HSCT. The estimated CL of intravenous busulfan was 14.2 L/h (3.64 mL/min per kg) and consistent with the CL of 4.02 mL/min per kg in another study [8] of Chinese adult patients, higher than 1.9 mL/min per kg and 3.34 mL/min per kg in Caucasian patients [17, 18]. Inclusion of GSTA1 genotype as a covariate accounted for 1.1% of the inter-individual variability of busulfan CL in the final model.

The fast metabolism of busulfan might relate with the ethnic difference of GSTA1 genotype between Chinese patients and Caucasian patients. The patients with GST A1*B haplotype were the poorer metabolizers with lower CL of busulfan compared to those with wild type. The studies of Kusama et al. [19] and Kim et al. [20] showed significant impact of GST A1 genotype on busulfan CL. Genetic frequency of GST A1*B haplotype was 14.0% with 1 patient of homozygous excluded to avoid statistic deviation in our study. The distribution was consistent with the rarity of haplotype *B in the Asian population (42.9% versus 13.9% in the Caucasians and Asians) [21]. GSTA1-69C/T (rs3957357), with -631T/G, -567 T/G, and -52 G/A, is located in the promoter region of GSTA1 gene. GSTA1-69C/T can adequately identify GSTA1*B haplotype which showed significant decline of GSTA1 protein expression in vitro [22]. Ansari et al. [23] further distinguished *B 1b, *B 1a and *B 2 haplotypes by − 513 (rs11964968), − 1142 (rs58912740). Then, Nava et al. [6] classified *A*B and *B*B diplotypes into normal and poor metabolizers of busulfan. Based on this accurate GSTA1-genotype classification, the PPK model excellently predicted the initial busulfan doses and achieved the target AUC in 85.2% of the pediatric patients (95%CI 78.7–91.7%) [7]. However, this metabolic classification of busulfan was not appropriate for Chinese patients due to the low genetic frequencies of above GSTA1*B variants in the Asian population [21].

PPK analysis has shown an advantage of quantifying the effects of covariates on PK variability to assess the clinical relevance of those effects [24]. Our results showed that GSTA1 genotype explained a tiny part of inter-individual variability from 17.8% of the basic model to 16.7% of the final model. The low genetic frequency of GST A1*B may lead to minimal influence of GSTA1 genotype and limited variability of busulfan CL between Chinese adult patients. Additionally, that also might attribute to different GSTA1 enzymatic activities in patients at different age. Recent studies [6, 25, 27] reported that GSTA1 genotype played a pivotal role in prediction of initial busulfan doses in pediatric patients. Meanwhile, Abbasi et al. [28] showed that GSTA1 did not influence busulfan CL obviously in both oral and intravenous administration in adult patients. A higher impact of GSTA1 genotype on busulfan CL in pediatric patients could be explained by age-specific GSTA1 enzymatic activity [29, 30]. For GSTA1, one of drug-processing genes in liver, the highest fold increase in the mRNA expression was induced by the prototypical ligands of xenobiotic-sensing transcription factors during adolescent age. Li et al. [29] highlighted that the drug-processing genes regulation was critical to predict drug PK and to decrease drug toxicity in pediatric patients. This in vitro study demonstrated the importance of GSTA1 genotype to recommend the initial busulfan dose and to avoid toxicity in children.

We did not find that the other factors, such as GSTP1, GSTM1, conditioning regimen, body surface area (BSA), and weight, showed a significant impact on busulfan PK. Pharmacogenomics data about busulfan PK in different populations are summarized in Table 4 [6,7,8, 19, 20, 25,26,27,28, 31,32,33,34,35,36,37,38,39,40,41,42,43,44,45]. Of all, GSTM1 null genotype and GSTP1*G diplotype were commonly reported to have a significant association with busulfan CL [20, 34]. However, Ten Brink et al. [39] did not find an association between the two GST genotypes and busulfan CL. This may be related to their minor contribution to busulfan metabolism. For conditioning regimen, Yeh et al. [46] reported a greater interdose variability of busulfan CL in the targeted busulfan–FLU regimen compared to the targeted busulfan–CY regimen (P = 0.0016). The impact of FLU on busulfan CL was not found in the study of Perkins et al. [47]. PPK analysis of Wu et al. [48] found that BSA significantly influenced the CL and Vd of busulfan in Chinese patients (P < 0.001, n = 53). The PPK or PK studies of busulfan in Asian patients were scanty and limited by small sample size and single center. A large-scale and multicenter clinical trial was needed to develop the busulfan PPK model to validate the influencing factors on busulfan PK and to be applicable for more Asian patients.

Table 4 Pharmacogenomics data about busulfan pharmacokinetics in the patients undergoing hematopoietic stem cell transplantation
Full size table

The GSTA1 genotype-based PPK model of intravenous busulfan was successfully developed and externally validated in Chinese adult patients. However, the influence of GSTA1 genotype on busulfan CL is tiny and unlikely to be clinically relevant. We did not suggest the guidance of GSTA1 genotype on initial busulfan dose in Chinese adult patients undergoing HSCT.