Proposed Therapeutic Range of Treosulfan in Reduced Toxicity Pediatric Allogeneic Hematopoietic Stem Cell Transplant Conditioning: Results From a Prospective Trial

Treosulfan is given off‐label in pediatric allogeneic hematopoietic stem cell transplant. This study investigated treosulfan's pharmacokinetics (PKs), efficacy, and safety in a prospective trial. Pediatric patients (n = 87) receiving treosulfan‐fludarabine conditioning were followed for at least 1 year posttransplant. PKs were described with a two‐compartment model. During follow‐up, 11 of 87 patients died and 12 of 87 patients had low engraftment (≤ 20% myeloid chimerism). For each increase in treosulfan area under the curve from zero to infinity (AUC(0‐∞)) of 1,000 mg hour/L the hazard ratio (95% confidence interval) for mortality increase was 1.46 (1.23–1.74), and the hazard ratio for low engraftment was 0.61 (0.36–1.04). A cumulative AUC(0‐∞) of 4,800 mg hour/L maximized the probability of success (> 20% engraftment and no mortality) at 82%. Probability of success with AUC(0‐∞) between 80% and 125% of this target were 78% and 79%. Measuring PK at the first dose and individualizing the third dose may be required in nonmalignant disease.

disease being treated. Malignant conditions tend to be treated with high intensity myeloablation, whereas nonmalignant conditions may be treated with lower intensity (lower dosing and/or fewer agents). Nevertheless, even with reduced toxicity conditioning, transplant-related morbidity and mortality remain significant. [1][2][3][4][5][6][7][8][9] Busulfan is commonly used in allo-HSCT conditioning and studies have demonstrated therapeutic drug monitoring (TDM) and dose adjustment are associated with reduced transplant-related mortality. [10][11][12] The target area under the curve (AUC) and therapeutic range of busulfan was recently revised in a pharmacokinetic/ pharmacodynamic (PK/PD) meta-analysis and methods for personalizing exposure by measuring busulfan PK after the first dose and adjusting later dose(s) is now well established. 10 Treosulfan is a busulfan analogue but, although busulfan causes direct DNA alkylation, treosulfan is a prodrug with alkylating activity mediated by its main epoxybutane derivatives. 13 Since the first report of treosulfan-based conditioning in pediatric allo-HSCT in 2002, it has been increasingly used off-label in children, largely due to a perceived wider therapeutic index and a lower propensity to cause veno-occlusive disease/sinusoidal obstruction syndrome than busulfan. 14 Data on treosulfan PK and toxicity in childhood are limited to mainly observational or retrospective studies, [15][16][17][18][19] meaning the therapeutic range in this population is poorly defined.
This study aimed to characterize the PK/PD profile of treosulfan in children undergoing allo-HSCT in an investigator-initiated, multicenter phase II clinical trial. The primary end point was to measure treosulfan PK and the secondary end point was to assess its association with short-term toxicity, graft failure, and mortality.

Ethics and patient recruitment
This was a prospective, open-label, phase II study (Clini calTr ials. gov, NCT02048800; EudraCT number 2013-003257-20) conducted in accordance with the principles of Good Clinical Practice and the Declaration of Helsinki. Patients aged 28 days to 18 years old were eligible if they were scheduled to receive treosulfan-fludarabine conditioning prior to allo-HSCT. Patients and/or their legal guardians were asked to provide written informed consent and assent where appropriate at two centers in the United Kingdom: the Bone Marrow Transplant Department in Great Ormond Street Hospital, London, and the Bone Marrow Transplant Department in Great North Children's Hospital, Newcastle upon Tyne.
The study was split into an initial pilot phase and the main trial. In the pilot phase, PK sampling was undertaken following the first dose, whereas in the main study PK samples were taken after the first and third dose and detailed study of short-term toxicity was performed. All patients were followed up for at least 1 year for survival and engraftment. Around 50 patients are required to capture important covariate effects in PK studies 20 so we aimed to recruit at least 50 to the main study.

Study conditioning regimen
The chemotherapy protocol consisted of treosulfan and fludarabine for all patients. Treosulfan was administered by 2-hour i.v. infusion on days -7, -6, and -5 prior to allo-HSCT at a total dose of 42 g/m 2 (14 g/m 2 / dose) in children aged > 12 months, 36 g/m 2 (12 g/m 2 /dose) in children aged 3-12 months, and 30 g/m 2 (10 g/m 2 /dose) in children ≤ 3 months. Fludarabine was given from day -7 to day -3 prior to allo-HSCT, at a total dose of 150 mg/m 2 . In vivo T-cell depletion with alemtuzumab or antithymocyte globulin (ATG) was administered according to donor type and stem cell source. Further details of transplant procedures are given in the Supplementary Materials.

Toxicity monitoring
In addition to mortality, in the main study, acute transplant-related toxicity was assessed up to 1 month post-allo-HSCT, graded according to the National Cancer Institute Common Terminology Criteria for Advance Events criteria. 21 The diagnosis of acute graft-vs.-host disease (GVHD) was made clinically and confirmed pathologically with skin, mucosal, or liver biopsy whenever possible. Grading of acute GVHD was performed according to the Seattle criteria. 22 Chronic GVHD was assessed and scored according to the National Institute of Health (NIH) criteria. 23 Blood sample collection and treosulfan determination Patients had indwelling multilumen central venous catheters in situ. Treosulfan was administered over 2 hours down one lumen and the line flushed. Following the end of the flush, blood was taken from a different lumen of the central venous catheter. A minimum of 3 mL of dead space blood was drawn and discarded prior to sampling. Initially, samples were drawn at the following times after completion of the flush postinfusion: 5, 15, and 30 minutes, and 1, 2, 3, 4, and 8 hours after the end of the infusion. To limit invasiveness and after confirming PK parameters (in particular AUC (0-∞) ) could still be estimated, an interim analysis showed sampling could be reduced to 4 postdose samples at: end of infusion, and 1, 2, and 4 hours after the end of the infusion.
Treosulfan concentrations in plasma were determined using a validated reverse-phase high-performance liquid chromatography method with refractometric detection in the Chemical Pathology Laboratory at Great Ormond Street Hospital. 15 Further details are given in the Supplementary Materials.

PK model building
Parameters for both one-compartment and two-compartment models assuming linear or nonlinear (Michaelis-Menten) elimination were estimated using nonlinear mixed effects modeling with NONMEM version 7.4, using the first order conditional estimation algorithm. 24 Interindividual variability was tested for all parameters assuming a log-normal distribution, and interoccasion variability was tested for clearance and central volume. The residual error included additive and proportional terms.
Allometric size scaling of clearance and volume terms were added a priori, and addition of a sigmoidal postmenstrual age maturation function tested. 25 Biomarkers relating to hepatic function (bilirubin and ALT), renal function (serum creatinine), and blood pH were tested on clearance. These covariates entered the model in the following form: where p i is the individual parameter of interest, c i is the individual value of the covariate and c is the typical value of the covariate in the population. In the fixed allometric weight scaling, c i was the individual body weight, c was set to 70 kg, and c was 0.75 for clearance and intercompartmental clearance, and 1 for central and peripheral volume. For bilirubin, ALT and pH, c was set to the median observed value, whereas for serum creatinine (because it is known to change with age) c was set to the median expected for age, as reported by Ceriotti et al. 26 In adolescents aged 15-18, a sex-specific linear extrapolation was used to link the end of the Ceriotti et al. 26 function and the adult expected values, as previously reported by Johansson et al. 27 The sigmoidal age function scaling clearance used postmenstrual age (assumed a gestational age of 40 weeks when this was unavailable) and contained two estimated parameters, so the model took the following form: where a i is an individual's postmenstrual age in weeks, a is the age at which clearance is 50% mature, and is a shape parameter.
The following categorical covariates were also tested: use of T-cell depletion in the conditioning regime, whether patients were in the pilot or the main study, and study site. These categorical covariates entered the model as follows: with I the indicator taking values of 1 when the covariate is present, and zero otherwise, and c now being the fractional parameter change in the presence of the covariate, and allowed to take values of ≥−1.
For nested models, significance of the additional parameters was evaluated with the likelihood ratio test, the difference in -2 log-likelihood (objective function value (OFV) in NONMEM) of the models being asymptotically 2 1 distributed. Covariates were added if the likelihood ratio test indicated a significant improvement in fit at the level of P < 0.01. Further model evaluation consisted of plotting predictions vs. observations, and standardized residuals vs. time and predictions, a visual predictive check (1,000 samples) and a nonparametric bootstrap (1,000 samples). A cumulative AUC (0-∞) calculated from the sum of all three doses administered divided by the individual clearance estimate was generated for each patient.

Statistical analysis of PDs
Two Cox proportional hazard survival analyses were performed in R, one to assess time to graft failure (chimerism in myeloid engraftment ≤ 20%) and one to assess time to mortality. Covariates considered were: cumulative treosulfan AUC (0-∞) for the three doses, age, use of T-cell depletion (alemtuzumab or ATG) in conditioning, donor source and matching, diagnosis, and CD34 dose. Univariable analysis was performed, and if two or more were significant (P < 0.05) these were taken forward to a multivariable analysis.
Upon finding low AUC (0-∞) to be associated with engraftment ≤ 20%, and high AUC (0-∞) to be associated with mortality, a therapeutic target was derived by fitting a quadratic model of cumulative AUC (0-∞) vs. probability of success (defined as being alive at last follow-up, with a myeloid engraftment > 20%). The linear predictor was defined as follows: where logAUC was the natural logarithm of cumulative AUC (0-∞) . This model was fitted to the binomial probability of success (defined as engraftment > 20% and being alive) with a generalized linear model in R. Logit, probit, and complimentary log-log canonical link functions were tested, the model with the lowest Akaike Information Criteria being chosen. Target concentration was defined as the logAUC, which maximized the probability of success, which by differentiating the expression above, yields: where AUC max is the natural logarithm of cumulative AUC (0-∞) , which maximizes probability of success. A nonparametric bootstrap with 10,000 samples was used to derive a 95% confidence interval (CI) on AUC max .
In the main study, an analysis investigating the relationship between cumulative AUC (0-∞) and National Cancer Institute (NCI) common toxicity criteria grade (0-5) for all major toxicity types was undertaken. The relationship with AUC (0-∞) and NCI grade was analyzed using the Kruskal-Wallis test by rank.

Dosing simulations
During the course of our study, the company manufacturing treosulfan (Medac) suggested the following body surface area (BSA)-based dosing scheme: 42 g/m 2 , reduced to 36 g/m 2 for those with a BSA < 1 m 2 , and 30 g/m 2 for those with a BSA < 0.5 m 2 . Ten thousand hypothetical patient characteristics (age 1 month-16 years) were generated using a published weight for age model 28 and each was randomly assigned a serum creatinine value based on their age and sex by sampling from the model by Ceriotti et al. 26 Using the PK model, the cumulative AUC (0-∞) was simulated for our dosing scheme, the Medac scheme and doses derived from our final model accounting for age, weight, and serum creatinine, or only age and weight. These were compared with the target concentration derived from the quadratic model.

Patients, donors, and transplant characteristics
A total of 87 children (30 in the pilot phase and 57 in the main study) receiving treosulfan as the sole alkylating agent in conditioning for allo-HSCT between January 2013 and December 2016 were enrolled and followed up for at least 1 year posttransplant. The median follow-up was 16 months (range 1247 months for surviving patients), baseline characteristics are detailed in Table 1. A total of 633 PK samples were obtained following the first and third doses, and no sample was below the assay lower limit of quantification. A total of 10 patients underwent the full PK sampling schedule (8 postdose samples), the remaining patients contributing 4 PK samples per occasion following the prespecified interim PK analysis.
A two-compartment model provided a superior fit to the one-compartment (P < 0.01). The MichaelisMenten elimination did not result in successful minimization (Kaplan-Meier value became very large) or lower OFV, indicating linear clearance in the dose range studied. The addition of a sigmoidal maturation function decreased the OFV by 29 points. Serum creatinine was the only other covariate that significantly (P < 0.01) improved model fit. A scatter plot of correlations in the continuous covariates is given in Figure S1. Table 2 gives PK model parameters, a visual predictive check is given in Figure 1, and further goodness-of-fit and covariate plots are shown in the Supplementary Figure S2-S4. Parameter estimates are provided in Table 2.

PD modeling and dosing simulations
Survival and ≤ 20% engraftment were modeled in a stepwise manner using Cox proportional hazards. First, univariable analysis was performed and significant covariates (P < 0.05) taken forward to a multivariable analysis ( Table 3). For mortality, two covariates were significant (AUC (0-∞) ) and being in receipt of a mismatched donor, whereas for engraftment there was a trend for low AUC (0-∞) to be associated with poor engraftment (hazard ratio (95% CI) = 0.61 (0.361.04); P = 0.072).
Upon finding low treosulfan AUC (0-∞) associated with poor engraftment and high treosulfan AUC (0-∞) with mortality, we   Figure 3a and parameter estimates are presented in Table 2. The probability of success was maximized at 82% for a treosulfan AUC (0-∞) of 4,829 mg hour/L (cumulative of the 3 doses). A nonparametric bootstrap revealed    ARTICLE this estimate to be unbiased but imprecise with bootstrap median (95% CI) of 4,876 (1,623-10,839) mg hour/L. The target was, therefore, rounded to two significant figures to 4,800, which also gives an 82% probability of success, whereas the interval between 3,863 and 6,037 mg hour/L represents the treosulfan AUC (0-∞) interval suggested for narrow therapeutic index drugs, 29 which gives corresponding probabilities of success of 78% and 79%, respectively. In our study, only 57% of children achieved this AUC (0-∞) range, whereas 16% had a treosulfan AUC (0-∞) below the lower cutoff (3,840 mg hour/L) and 26% patients had a treosulfan AUC (0-∞) above the upper cutoff (6,000 mg hour/L).  Transplant-related mortality in patients with treosulfan AUC (0-∞) > 6,000 mg hour/L was 39% compared with 3% in those with a treosulfan AUC (0-∞) < 6,000 mg hour/L. The corresponding survival was significantly lower (P < 0.0001) in patients with AUC (0-∞) values above and below this cutoff (Figure 3b).
Simulated treosulfan AUC (0-∞) for our dosing scheme (30,36, and 42 g/m 2 for patients < 3 months, 3-12 months, and > 12 months, respectively) and the Medac dosing scheme (30, 36, and 42 g/m 2 for BSA ≤ 0.5, 0.5-1 and > 1 m 2 , respectively) are shown in Figure 4. In addition, calculating the dose from our  This study ARTICLE model with and without the creatinine covariate is presented. This was achieved by taking the typical clearance for a patient based on their covariates and defining the dose to target a cumulative AUC (0-∞) of 4,800 mg hour/L as follows: where with wt i being the individual's weight in kg, a i the postmenstrual age in weeks, secr i the individual's serum creatinine in µmol/L, and mscr is the median creatinine for age predicted from the Ceriotti model. 26

DISCUSSION
In a prospective clinical trial of treosulfan PK in pediatric allo-HSCT, treosulfan AUC (0-∞) was associated with poor donor engraftment and mortality. Because all but two patients in our study had nonmalignant disease, our findings should be inferred only to apply to this group. This has facilitated the proposal of a therapeutic target of cumulative AUC (0-∞) 4,800 mg hour/L. Being within 80-125% of this target cannot be met in ~ 50% of patients through dosing by covariates alone (Figure 4), hence, the major finding is that a TDM-guided treosulfan dose adjustment should be explored. Children with a cumulative treosulfan AUC (0-∞) > 6,000 mg hour/L had transplant-related mortality of 39%, whereas patients with AUC (0-∞) below this had transplant-related mortality of 3%. The only other significant relationship with mortality in the univariable analysis was receiving a mismatched donor, but upon multivariable analysis the strength of this association was reduced ( Table 3). Our choice of using all transplant-related mortality could be questioned given that three patients died beyond 100 days of viral-associated complications. It could be argued that because treosulfan is mainly myeloablative, high AUC (0-∞) may not be related to these cases, and fludarabine, lymphocyte depletion, or the patient's underlying immune deficiency are more important. However, treosulfan does have broader immunosuppressive effects than simple myeloabaltion 30 and its role in the establishment and longer term effects from viral complications cannot be completely ruled out. Truncating the survival analysis at 100 days and 6 months shows the effect is less strong early on but by 6 months is similar to the overall effect (100day hazard ratio (95% CI): 1.18 (0.93, 1.51); 6-month hazard ratio: 1. 3 (1.06, 1.6)).
Although an observational study of 77 children by van der Stoep et al. 19 did not find a correlation between treosulfan AUC (0-∞) and mortality, this may be due to the heterogeneity of conditioning as 67.5% patients also received thiotepa. In contrast, children enrolled in our prospective clinical trial received homogeneous conditioning of treosulfan and fludarabine only. Recently, an observational study in children undergoing treosulfan conditioning for allo-HSCT in thalassemia major found a trend of 82% survival in patients with an AUC (0-∞) of < 5,484 mg hour/L compared with only 68% in patients above this threshold. 31 Taken together with our result possibly indicates the need to individualize doses in patients with nonmalignant disease. To draw firm conclusions on causation a prospective study is required, because patients who enter the conditioning period with lower treosulfan clearance may have comorbidities predisposing them to mortality, which would not be prevented by lowering treosulfan AUC (0-∞) .
The association between conditioning drug exposure and clinical outcome in pediatric allo-HSCT has been explored in a number of studies. Busulfan studies have shown an association among exposure and toxicity and engraftment 32,33 with TDM and personalization utilized for a number of years. The US Food and Drug Administration (FDA) proposes a target of 900-1,350 µM minutes 34 whereas the European Medicines Agency (EMA) proposes 900-1,500 µM minutes. 35 Surprisingly, these targets are based on small observational studies. 11,12 A larger but retrospective study on 674 patients with malignant and nonmalignant conditions recently derived a higher target of 1,225-1,575 µM minutes. 10 Likewise, Admiraal et al. 36 recently demonstrated that an optimal exposure to ATG is associated with higher event-free survival and lower risk of acute GVHD in adults undergoing allo-HSCT.
Our target was found to be rather imprecise (95% CI 1,623-10,839 mg hour/L) upon nonparametric bootstrap but the median (4,876 mg hour/L) was close to our estimate, suggesting it is unbiased. The imprecision is likely due to the small number of events but for now our data remain one of the largest to date. It has been proposed that an acceptable range is being within 80% and 125% of a target value, and if the log-normal distribution is assumed this translates to 90% of patients achieving that range if unexplained variability is 13.6% coefficient of variance. 29 Comparing our interindividual and interoccasion variability values on clearance shows dosing by covariates alone will not achieve this target (Figure 4), because the first dose AUC could be measured and the interoccasion variability is 14%, it is likely the cumulative AUC (0-∞) for the three doses could readily be targeted. Our future work will include a detailed optimal design and simulation-estimation study to evaluate the potential of treosulfan TDM.
The clearance estimate in our model was scaled by both weight and age. Weight scaling used a fixed allometric model, which approximately follows BSA and for older children and has recently been shown to apply for most drug classes. 37 Hence, BSA-based dosing should give similar AUC (0-∞) for children older than around 2 years. It is also well known that in the first year of life BSA-based dosing leads to higher AUC (0-∞) due to immaturity in clearance. 37 There are a number of ways to model declining clearance with younger age and recently it was shown that most give equivalent results, 25 and, hence, we used the standard method proposed by Holford et al. 38 The major benefit of using this standard method is that it is then very straightforward to compare clearance values between different studies. Our ARTICLE estimate of clearance was 17.31 L/hour/70 kg, which is similar to that in a recent treosulfan observational study (17.9 L/hour/70 kg), 19 and the surface-area scaled value in the recent thalassemia study (20.07 L/ hour/1.73 m 2 ), 31 all of which are somewhat higher than the value recently estimated by Danielak et al. 18 (14.7 L/hour/70 kg). The likely reason is the latter study only included 15 patients with a wide age range, the youngest of whom was < 6 months old, yet no age-related maturation term was used. Recently, a model-based reanalysis of the data by van der Stoep et al. in 2017 19 found a similar maturation half-time to ours (38 weeks) with a lower shape parameter (1.2). 39 It is likely our shape parameter is more reliable because our patients were, on average, 19 months old whereas in that study the median age was 52 months, 39 and, furthermore, the busulfan maturation half time and the shape parameter were 40 weeks and 2.2 in a large meta-analysis. 40 A possible reason for decreased clearance in younger patients is immaturity in glomerular filtration rate because around 40% is excreted renally. During the covariate analysis we found serum creatinine to be inversely correlated with clearance. This was modeled by multiplying clearance by the ratio of serum creatinine to the age-expected serum creatinine raised to an estimated power. 27 Although this relationship was statistically significant and so retained in the model, the covariate power estimate of −0.3 means even in a child with a twofold higher than age-expected creatinine, this would only decrease clearance by around 19%.
Target attainment through dosing by covariates was similar from each of the tested dosing regimens (text probabilities in Figure 4). Because the simulated population had a uniform distribution of ages, it seems that surface area or allometric dosing gives very similar target attainment in patients aged > 2 years. The differences come in the younger age groups where dosing by age and weight seems optimal, the addition of creatinine adding little. The Medac scheme showed a trend for reduced overexposure compared with dosing in our study. However, because all covariate-based dosing gives target attainment of around 50%, TDM will still be required.
In conclusion, the PK of treosulfan in children have been characterized and an association with high AUC (0-∞) and mortality and low AUC (0-∞) and poor engraftment was found. A prospective study on TDM-guided personalization is warranted.

SUPPORTING INFORMATION
Supplementary information accompanies this paper on the Clinical Pharmacology & Therapeutics website (www.cpt-journal.com).