PKPD-Based Translational Modeling of Calaspargase Pegol Preclinical Activity in Hepatocellular Carcinoma Authors: Eef Hoeben, Vincent Madelain, Chayan Acharya, Pierre Barbier Saint Hilaire, Ethika Tyagi, Malek Okour, Sylvain Fouliard Published: 2026 ABSTRACT AbstractBackground and ObjectiveCalaspargase pegol (CalPEG) is a PEGylated conjugate of L-Asparaginase (L-Asp) and an asparagine specific enzyme approved by the Food and Drug Administration (FDA) as a component of a multi-agent chemotherapeutic regimen for the treatment of acute lymphoblastic leukemia (ALL) in children and young adults. Although CalPEG is an established and effective treatment for childhood ALL, its potential in solid tumors is still underexplored. Recent reports illustrated that L-Asp may also have potential for the treatment of certain aggressive solid tumors, including hepatocellular carcinoma (HCC)[1–5]. In order to investigate this potential, clinical and preclinical data were integrated using model-informed drug development (MIDD) approaches, which are among the most adequate tools to facilitate the translation of preclinical into clinical data [6–9]. This work explores the potential of pharmacokinetic/pharmacodynamic tumor growth inhibition (PKPD TGI) modeling to maximize the use of in vivo preclinical data and PK data in patients with ALL for translation of preclinical observed antitumor activity to clinical efficacy.MethodsA PKPD TGI model was developed that described tumor volumes (TVs) versus time in a preclinical hypermethylated xenograft mouse model relevant for HCC. Translation of the PKPD TGI model to human was performed by linking the structural TGI model developed in mice with human PK model of CalPEG. This model was developed on PK data, i.e., plasma Asp activity (PAA), obtained in pediatric and adult patients with ALL and was used to predict TGI profiles of CalPEG in adult patients with HCC.ResultsThe developed PKPD TGI model adequately described the TV versus time profiles in tumor bearing (TB) mice and quantified the PKPD relationship between PAA and TGI in TB mice. The structural mouse TGI model was linked to predicted human PK parameters and this human PKPD TGI model was used to perform simulations and to predict antitumor activity of CalPEG at different doses/schedule in patients with HCC.ConclusionsThe results indicate that PKPD TGI modeling can be used to build a preclinical to clinical translational framework to predict the exposure–response relationship in patients with HCC and to inform dosing strategies or trial designs for solid tumors.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s13318-026-01010-4. Key Points CalPEG is a well-established ALL treatment with unexplored potential in solid tumors.A translational PK/PD modeling framework is described, aiming at leveraging preclinical information to predict the potential activity of CalPEG in patients with solid tumors.The findings can help inform dosing strategies or trial designs for solid tumors. Introduction L-Asparaginase (L-ASNase) is an enzyme that hydrolyses the nonessential amino acid asparagine (Asn) into aspartic acid and ammonia and is currently a cornerstone drug in the treatment of childhood and adult acute lymphoblastic leukemia (ALL). ALL lymphoblasts, in contrast to healthy cells, largely lack the enzyme asparagine synthetase (ASNS) expression, and thus depend on extracellular sources of Asn for their survival. Depletion of circulating Asn by L-ASNase results in starvation and selective apoptosis of the leukemic blasts. More recently, accumulating evidence suggests that L-ASNase could also have clinical potential for the treatment of other subtypes of leukemia or certain aggressive solid tumors, including hepatocellular carcinoma (HCC) [10]. However, one of the main reasons preventing the expansion of today’s FDA-approved L-ASNases to solid cancers is their high toxicity, especially in adults, which includes severe immunological side effects (hypersensitivity) as well as nonimmune related toxicities, such as pancreatitis, liver toxicities, coagulopathy, and neurotoxicity [11, 12]. The design of optimized and novel L-ASNases molecules/formulations provided opportunities to address both the short half-life and the unwanted toxicities of L-ASNases. Calaspargase pegol (S095015, SC-PEG Escherichia. coli L-asparaginase, CalPEG) is a stabilized formulation of L-ASNase and results from the covalent conjugation of monomethoxypolyethylene glycol (PEG) to lysine residues on the enzyme via a succinimidyl carbonate linker [13, 14]. The addition of the PEG-tag delays elimination from the body and increases the half-life of PEG-L-ASNase in humans towards 16.1 days. CalPEG received USA marketing authorization in 2018 as a component of a multi-agent chemotherapeutic regimen for the treatment of ALL in children and young adults up to 21 years of age. Although CalPEG is an established and effective treatment for childhood ALL, its potential in solid tumors is still underexplored. Therefore, preclinical in vitro and in vivo experiments were performed to evaluate whether CalPEG could also have clinical potential for the treatment of solid tumors, including HCC [15, 16]. It has been shown that ASNS promoter-hypermethylated HCC cell lines with the lowest ASNS mRNA expression are the most responsive to CalPEG treatment in vitro [17]. An in vivo efficacy study in female BALB/c nude mice bearing human ASNS promoter-hypermethylated cell line derived xenograft (CDx) model relevant for HCC (JHH5) showed that CalPEG treatment inhibited tumor growth (TG) in a dose-dependent manner and resulted in tumor rejections at the highest doses. Cancer remains an unmet medical need, not only to develop new drugs, but also to improve the speed and efficiency of the development of oncology drugs. PKPD modelling has become a key tool in early understanding and quantification of the exposure–response relationship and associated variability in order to inform dosing strategies and designs for clinical trials. Therefore, an integrated quantitative translational PKPD modeling approach was used to facilitate the translation of preclinically observed antitumor activity into clinical efficacy. Active CalPEG exposure in patients with HCC was predicted on the basis of exposure–response data generated in this JHH5 xenograft mouse model and available PK information in pediatric and adult patients with ALL integrated into a population PK (popPK) model [18]. popPK models have been developed for PEG-conjugated asparaginase in children with ALL [19, 20] and are used to predict PAA levels in childhood patients with ALL in case those data are lacking due to clinical practice (missing values, monitoring gaps) or noncompliance and to adapt treatment schedules. In this PKPD analysis, owing to the lack of published popPK data in solid tumors, PAA in HCC patients was predicted using a popPK model, developed on PAA derived from pediatric and adult patients with ALL, under the assumption of similar PAA in patients with ALL and HCC. Although preclinical xenograft models represent a simplification of a human malignancy and their predictive value have been challenged, this translational approach was found to be predictive for clinical efficacy for many small molecule oncology compounds [21–23] and was therefore considered to be useful to understand the degree of antitumor activity associated with CalPEG and explore its potential clinical activity in the HCC indication. The PKPD TGI model flow, as presented in the paper, can be applied to optimize early clinical trial designs in HCC, including dosage optimization, for a more efficient selection of doses for later phases and for potential combination strategies. Methods Data Collection The in vivo PK and therapeutic efficacy of CalPEG was evaluated in the treatment of a subcutaneous (s.c.) JHH-5 HCC xenograft model in 40 female BALB/c mice. Each mouse was inoculated s.c. at the right flank region with JHH-5 tumor cells (1 × 107) in 0.2 mL of phosphate-buffered saline (PBS) mixed with Matrigel (1:1) for tumor development. On the day of treatment, the mean tumor size was 175 mm3. Vehicle and CalPEG were administered i.v. every 4 days (Q4D) for 28 days at 0 IU/kg (vehicle) 20, 40, and 100 IU/kg (CalPEG). Plasma samples for PK, i.e., plasma asparaginase activity (PAA), were collected 5 min before the 3rd administration on D8 (N = 10 mice/group) and 5 min before the 5th administration on D16 (N = 10 mice/group). At termination, plasma PK samples were collected before the 8th administration on D28 for the first N = 5 mice/treated group and 2 h after the 8th administration on D28 for the remaining N = 5 mice/treated group (Supplementary Table S1). The mice from the vehicle group were prematurely killed (before day 28) as their tumor volume exceeded the ethical limit of 2000 mm3. Asparaginase activity in mouse plasma was measured indirectly using a coupled enzymatic assay in which asparaginase activity was the rate-limiting step [17]. TVs were measured twice per week before randomization (D0) and were measured every 2 days after randomization (D0) up to the last dosing day on D28. TVs were measured in two dimensions using a caliper and the volume was expressed in mm3 using the formula V = (L × W × W)/2, where V is tumor volume, L is tumor length (the longest tumor dimension) and W is tumor width (the longest tumor dimension perpendicular to L). Nonmeasurable TVs were reported as < 1mm3. A total of 40 mice (N = 10/dose group, including vehicle group) were finally included in the population PKPD TGI analysis for the JHH5 model. The final dataset included 527 TV observations and 87 PAA measurements for PK. PK/PD Model Development A stepwise PKPD TGI modeling approach was considered to translate preclinically observed antitumor activity into clinical efficacy. Step 1 consisted of the description of the plasma asparaginase activity (PAA, expressed in IU/L) versus time profiles in female nongrafted and grafted BALB/c nude mice using a PK model. Step 2 consisted of the development of a PKPD TGI model to describe the TV versus time profiles in mice xenografted with JHH5 cells. Step 3 consisted of the prediction of human PK of PAA using an existing population PK model of CalPEG developed in pediatric and adult patients with ALL. Step 4 consisted of the development of a human PKPD TGI model, linking the structural mouse TGI and human population PK model, and use the integrated model for prediction of clinical antitumor activity in patients with HCC. PK Model and Prediction of PAA in Nongrafted and Grafted Female BALB/c Nude Mice A PK model describing the kinetics of PAA after i.v. administration of CalPEG in female nongrafted and grafted BALB/c nude mice had been developed earlier on the basis of a rich dataset (344 PK samples in 90 mice administered with doses ranging from 6 to 300 IU/kg) and was considered for the description of the PAA–time profiles in these experiments (Table 1, Supplementary Table S2, Supplementary Figs. S1 and S2). Briefly, the final PK model was a one-compartment model with nonlinear Michaelis–Menten elimination. To account for a time-dependency after multiple administration, an indirect response model was used. In this model, it was assumed that the elimination of asparaginase is through proteolysis and that the elimination rate of CalPEG was dependent on a (relative) amount of enzyme (Ae) in the system (Ae = 1: elimination at baseline, Ae = 0: no elimination). CalPEG was assumed to decrease Rin of the enzyme in a concentration-dependent manner, i.e., in the presence of CalPEG, Ae will decrease and as a consequence, the elimination rate of CalPEG will decrease as well. With this parameterization, Ae cannot go above 1 and only a decrease of CalPEG elimination rate was observed. Random effects were identified in the model parameters for V, Vmax, and Km, describing the inter-individual variability (IIV) of these parameters. Residual variability was modeled using a proportional error model. This PK model was initially assessed for its ability to describe PAA in the current experiment through an external visual predictive check (VPC). The design of the PK part of the study was simulated using this model, then median and 90% prediction interval (PI) were superimposed with measured data. As the VPCs showed that the existing PK model was able to correctly predict the new PAA observations in the xenograft efficacy JHH5 model, no further model development was deemed needed; the estimated population PK parameters, including IIV, was fixed in the PKPD TGI model for a sequential PKPD approach [24]. Table 1PK parameter estimates using final PK and PKPD TGI model in mice and population PK model in patients with ALL (age range 1–73 years)ModelParameterUnitFixed effect estimateRSE (%)IIV (SD)RSE (%)Shrinkage (%)Mice PKVL/kg0.0636–0.0765––ModelVmIU/kg/h5.07–0.164––KmIU/L4000–0.0625––kout1/h0.000813––––EC50IU/L107––––Human PKVcL1.78–0.150––ModelCLL/h0.00245–0.441––VmaxTRU/h21.4––––KmTRU/L58.8––––k211/h0.0395–0.519––VmaxU/h0.583––––KmU/L0.00100––––Corr. CL:k21––0.582––BSA on Vc1.23––––Female on Vc– 0.0928––––BSA on CL1.19––––BSA on VmaxTR0.726––––BSA on Vmax1.24––––AGE on Vmax–0.0166––––Mice PKPDkg1/h0.00708110.2245437.4TGI Modelkglmm3.h−170.019–––slpL/IU/h × 10−53.22180.3597028.7tauh41.13.6–––W0mm31524.10.11410248.9Ψ–20 FIX––––Error modelRUV1SD additive component0.8273.2RUV2SD proportional component0.4378.5ALL acute lymphoblastic leukemia, BSA body surface area, CL absolute clearance, EC50 asparaginase activity needed to reduce the amount of enzyme to 50%; IIV inter-individual variability,k21 inter-compartmental rate constant,kg growth rate corresponding to exponential growth of the tumor, kgl growth rate corresponding to linear growth of the tumor, Km concentration at half maximum elimination capacity, KmTR concentration at half maximum peripheral distribution capacity, Kout first order constant for loss of enzyme, PD pharmacodynamics,PK pharmacokinetics,RSE relative standard error,RUV residual unexplained variability,SD standard deviation,tau transit time,TGI tumor growth inhibition,V volume of the central compartment, Vm maximum capacity,Vmax maximum elimination capacity,Vc central volume of distribution, VmaxTR maximum peripheral distribution capacity,W0 tumor volume at time zero PK parameter estimates using final PK and PKPD TGI model in mice and population PK model in patients with ALL (age range 1–73 years) RUV1 SD additive component RUV2 SD proportional component ALL acute lymphoblastic leukemia, BSA body surface area, CL absolute clearance, EC50 asparaginase activity needed to reduce the amount of enzyme to 50%; IIV inter-individual variability,k21 inter-compartmental rate constant,kg growth rate corresponding to exponential growth of the tumor, kgl growth rate corresponding to linear growth of the tumor, Km concentration at half maximum elimination capacity, KmTR concentration at half maximum peripheral distribution capacity, Kout first order constant for loss of enzyme, PD pharmacodynamics,PK pharmacokinetics,RSE relative standard error,RUV residual unexplained variability,SD standard deviation,tau transit time,TGI tumor growth inhibition,V volume of the central compartment, Vm maximum capacity,Vmax maximum elimination capacity,Vc central volume of distribution, VmaxTR maximum peripheral distribution capacity,W0 tumor volume at time zero PK/PD TGI Modeling in Mice The PKPD TGI model building was performed in three steps: first, selection of the structural model describing the PAA concentration effect relationship; second, selection of the IIV model, and third, selection of the residual error model. A semi-mechanistic TGI model [25] was used to describe simultaneously the tumor growth in control (Eq. 1) and treatment groups (Eq. 2). Where TS is the amount (volume) of replicating tumor cells, kg and kgl are the growth parameters corresponding to exponential and linear growth, respectively, and φ is a shape parameter of the model usually fixed to a high value (e.g. 20). PAA was assumed to increase the rate of tumoral cell death, with a delay between kinetic of PAA and cell death introduced as three transit compartments, where cells were assumed as dying and no longer replicating. where TS2, TS3, and TS4 are nonreplicating dying cells, Ed is the drug effect as a function of PAA (Aspa), τ is the transit time driving the time needed to observe cell death, and slp is the slope. Different concentration effect relationships were evaluated to capture concentration effect Ed over time, including at least for Model selection was based on different criteria. For nested models, the objective function value (OFV) was used in a likelihood ratio test (LRT) and for non-nested models, the model with the lowest Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) was chosen. Model selection was also based on other criteria including goodness-of-fit (GOF) plots, individual fits, and accuracy of the parameter estimates. PK Model in Pediatric and Adult Patients with ALL A population PK model of CalPEG (legacy model) was initially developed in pediatric and young patients with ALL (age range 1–26 years) after administration of calaspargase pegol [18]. A description of the studies included for population PK model development is provided in Supplementary Tables S3A, S3B, and S3C. The structural model is a two-compartmental disposition model with saturable transport to the peripheral compartment and combined linear and nonlinear elimination from the central compartment. A model representation is given in Sect Human Efficacious Dose Prediction. The model demonstrated a close to directly proportional relationship between body surface area (BSA) and the linear clearance (CL), the maximum capacity of the nonlinear elimination (Vmax) and the volume of the central compartment (Vc). In addition, a statistically significant difference in Vc between male and female subjects was identified. IIV terms were supported on Vc, CL, and k21 (rate constant from the peripheral to the central compartment). The residual variability for CalPEG was described by a combined error model on log-transformed data. The legacy model was later updated with additional data from adult patients with ALL (age range 24–73 years). The final model (age range 1–73 years) is a two-compartment model characterized by a saturable distribution and a combined saturable and linear elimination, similar to the legacy model. Previously identified covariate–parameter relationships were estimated with similar effect sizes. An additional covariate–parameter relationship of Vmax decreasing with increasing age was identified. The residual variability in the final PK model was simplified to a proportional error model (additive on the log scale). A summary of the PK model parameters is given in Table 1. The VPCs for external qualification of the PK model are given in Supplementary Figs. S3A–S3C. The analysis dataset used to build the final model included data after administration of CalPEG from doses of 1000 IU/m2 up to 2500 IU/m2 and contained, in total, 144 subjects with observations, i.e., 1680 observations and 11.7 observations per subject. Human Efficacious Dose Prediction A “human” PKPD TGI model was developed by linking the structural mouse TGI model with the human population PK model. The assumption was made that the PKPD relationship remained similar in mice and patients, and that the PK (PAA) in adult patients with HCC would be similar to the PK in adult patients with ALL. This model allowed us to perform simulations of TV versus time profiles and to project active doses in adult patients with HCC. The population PK model was used to perform simulations (N = 1000) in adult patients with HCC after a 2-h infusion administered every 3 weeks (Q3W) at doses of 500, 750, 1000, 1250, 1500, 2000, and 2500 IU/m2 and to predict exposure (Ctrough, Cinf, and Cavg) after the 1st, 3rd, and 5th administration. The typical population values of the model parameters with their corresponding IIV (CL, Vc and k21) were used for the simulations. Covariate values were set on the basis of typical values in patients with HCC. BSA was set to the values reported by the National Institute of Health for the corresponding age group(s) (i.e., 2.05 m2 (males) and 1.81 m2 (females) for age range of 50–69 years); simulations were performed separately for males and females and as the majority of HCC patients are > 55 years of age, simulations were performed for a 55-year-old patient. A surrogate of clinical response in these simulations was defined as a 90% reduction in TV versus baseline, i.e., complete tumor regression [26]. Data Analysis and Software Nonlinear mixed-effects modeling (NONMEM®) [27, 28] was performed using NONMEM version 7.4.4 software (ICON Development Solutions, Ellicott City, Maryland, USA) [29]. The first-order conditional estimation with interaction option (FOCE-I) algorithm was used to estimate population model parameters in NONMEM and the $COVARIANCE routine was used to provide precision of parameter estimates. Dataset exploration and visualization, as well as graphical analysis were performed using R statistical software for Windows version 4.3.1 (R Development Core Team, Vienna, Austria). The imputation method was used for TV measurements that were reported as < 1 mm3 and were equated to the value of the limit of quantification (LoQ)/measurability set to 1 mm3. Inter-individual variability (IIV) was tested on model parameters by a backward approach. In this backward step, all IIV included in the model were removed separately and kept in the model if this removal was significant (p = 0.01, chi-squared distribution, one degree of freedom (df)). For residual unexplained variability, different error models were tested, including a proportional error model and a combined additive and proportional error model. The ability of the model to describe and/or predict data were assessed by standard GOF plots, normalized prediction distribution error (NPDE) [30], and visual predictive check (VPC). Epsilon (ε)-shrinkage was reported in order to assess the possibility to diagnose the model using individual fits, and was computed as follows: where IWRES is an individual weighted residual [31]. The uncertainty of the parameter estimates, relative standard error (RSE), and 95% confidence interval (CI) of the population parameters (fixed effects, IIV and residual error parameters) were computed as follows: where SE(θ) is the standard error of the considered parameter θ, as provided by the $COVARIANCE routine in NONMEM, and θ the parameter estimate. The 95% CI were computed using the standard error of each parameter. Results PK Analysis in Nongrafted and Grafted Female BALB/c Nude Mice The PK model built on extensive PK data of PAA in female BALB/c nude mice (Supplementary Table S2) was used to predict the PAA observations in the JHH5 mouse model. The population PK parameters, including IIV, estimated by this PK model were fixed in the PKPD TGI model (Table 1). An external VPC showed that this PK model was able to correctly predict the new PAA observations in the JHH5 model as the observed PAA observations were adequately predicted and were equally distributed around the predicted median without over or underestimation of the variability (Fig. 1). Therefore, no further development of the PK model was performed. Visual predictive check on linear-scale (left) and on logarithmic-scale (right) performed using existing PK model and the model building dataset simulated 1000 times for PAA (IU/L) in the JHH5 model. Red symbols: observations; solid black line: median; grey shaded area: 90% prediction interval of simulations. PK pharmacokinetics, PAA plasma asparaginase activity, IU international unit PKPD TGI Modeling with In Vivo Efficacy Data A semi-mechanistic TGI model was used for PKPD TGI modeling [25]. Individual TV were fitted simultaneously for vehicle and treated groups. In the absence of treatment, the model assumed that the TG occurred in two phases, i.e., initially, an exponential growth phase characterized by kg (exponential rate constant of TG) followed by a linear growth phase kgl (zero-order rate constant of TG). A linear concentration effect relationship was selected for the final model and a slope (slp) estimated, as this model resulted in the lowest BIC value and provided the best fit to the data. A delay model with three transit compartments, i.e., tumor cells pass through progressive stages of damage because of the drug’s action, allowed to model the process of cell death, which occurred at later times and was characterized by the model parameter tau. Tau was estimated to be 41.1 h and resulted in an average time for a cell to die of about 5–6 days. Exponential random effects were added to TV at baseline (w0), slp, and kg. Exclusion of IIV on slp and w0 resulted in an increase in the OFV of > 6.63 (p < 0.01, chi-squared distribution, df = 1) and were kept in the final model. Removal of IIV on kg resulted in model instability and was therefore kept in the model. A combined additive and proportional error model best described the residual error. Final CalPEG population PKPD TGI parameters are given in Table 1. The uncertainty on estimates for fixed effects was small, with RSE < 20%, and the uncertainty on IIV was moderate. GOF plots showed a good agreement between observations and individual predictions of TV in untreated groups and groups treated with CalPEG. Scatter plots did not show any substantial and systematic deviations from the identity line, except for the low population predicted TV values as indicated by the light grey circles in Fig. 2a, corresponding to time points where 4 mice out of 10 had data below LoQ, whereas 6 mice had measurable TV up to the end of the study (day 28). As only measurable TV values (and not data below LoQ) are represented on the GOF plots, this deviation is not unexpected. Overall, the GOF plots presented in Fig. 2b show a good model fit, although a slight misspecification can be observed. GOF plots were scattered evenly around the zero line when represented for CWRES and NDPE versus PRED, respectively (Fig. 2b). The IWRES, CWRES, or NDPE plots were scattered evenly around the horizontal zero line on the y-axis, when represented versus time with most of the points within −2 to 2 (Fig. 2c). No outlier data, i.e., observations with CWRES greater than 5, were observed on CWRES plots. The value of epsilon-shrinkage for the CalPEG population PK/PD model was low (4.2%), indicating no risk of perfect-fit. a Goodness-of-fit plots of final JHH5 model. Observations versus population (left) and individual (right) predicted TV on logarithmic-scale. Dashed black line: identity line; solid black line: trend of the data; black symbols: observed TV; light grey circles: TV values from 4 mice out of 10 dosed with CalPEG at 40 IU/kg and with TV measurements above LoQ. The misspecification in the GOF plot (observed versus population predictions) can be explained by the fact that values below LoQ are not represented on the plot. GOF Goodness-of-fit, IU international unit, LoQ lower limit of quantification, TV tumor volume. b Goodness-of-fit plots of final JHH5 model. IWRES versus individual predicted TV (top), CWRES (middle) and NPDE (bottom) versus population predicted TV. Solid black line: regression line; dotted black horizontal lines: values of IWRES, CWRES and NPDE should be approximately N(0,1) and hence concentrated between –2 and 2. CWRES conditional weighted residuals, IWRES individual weighted residuals, NPDE normalized prediction distribution error, TV tumor volume. c Goodness-of-fit plots of final JHH5 model. IWRES (top), CWRES (middle), and NPDE (bottom) versus time. Solid black line: regression line; dotted black horizontal lines: values of IWRES, CWRES, and NPDE should be approximately N(0,1) and hence concentrated between –2 and 2. CWRES conditional weighted residuals, IWRES individual weighted residuals, NPDE normalized prediction distribution error The VPCs stratified by dose of CalPEG were overall satisfactory as the observed data were equally distributed around the predicted median without over or underestimation of the variability (Fig. 3). The individual fits are presented in Supplementary Fig. S4. A linear concentration effect relationship was selected for the final model as this model resulted in the lowest BIC value and provided an adequate fit to the data. Visual predictive checks on linear-scale (left) and on logarithmic-scale (right) stratified by dose, performed using the final PKPD TGI model. Solid colored lines: observations; solid grey line: median; grey shaded area: 90% prediction interval of simulations; colored dots are observations below the lower limit of quantification; IU international unit, TV tumor volume, VPC visual predictive check Evaluations presented above showed that the final PKPD TGI models adequately described the TV data after CalPEG administration in tumor bearing BALB/c mice and that the models were suitable for translation to humans. PKPD Translation to Humans A population PK model of CalPEG was developed in pediatric and adult ALL and used to predict PAA in adult patients with HCC under the assumption that there was no difference in PAA between patients with ALL and patients with HCC after imputation of relevant covariate values in this population. The parameter estimates of PAA by the final population PK model are provided in Table 1 and predicted PAA versus time profiles are presented in Fig. 4 (males) and Supplementary Fig. S5 (females). The predicted median PAA exposure metrics after the first, third, and fifth administration are presented in Table 2 (males) and Supplementary Table S4 (females). Predicted PAA versus time profiles in patients using final population PK model (SEX = 1 for Males, BSA 2.05 m2, AGE 55 years old). Solid red line: Ctrough of 100 IU/L sufficient to ensure therapeutically adequate Asn depletion in patients with ALL; solid black line: median PAA; grey shaded area: 90% prediction intervals; vertical dotted black lines: tri-weekly drug administrations. ALL acute lymphoblastic leukemia, Asn asparagine, BSA body surface area, IU international unit, PAA plasma asparaginase activity, PI prediction interval, PK pharmacokinetics, Q3W every 3 weeks Table 2Predicted PAA (median) after 1st, 3rd, and 5th administration of CalPEG in males1st Administration3rd Administration5th AdministrationDose, IU/m2Ctrough, IU/LCinf, IU/LCavg, IU/LCtrough, IU/LCinf, IU/LCavg, IU/LCtrough, IU/LCinf, IU/LCavg, IU/L50019.318846.357.223210383.926714075043.628390.1133383220177450287100076.3385146219554348276651428125010948120131272247837884957815001495732603978795994721032706200023578039159012518716911455102925003269595197431562110485117961269Cavg average concentration, Cinf concentration at the end of infusion, Ctrough trough concentration, IU international unit, PAA plasma asparaginase activity Predicted PAA (median) after 1st, 3rd, and 5th administration of CalPEG in males Cavg average concentration, Cinf concentration at the end of infusion, Ctrough trough concentration, IU international unit, PAA plasma asparaginase activity By linking the structural mouse TGI model with the human population PK model, a human PKPD TGI model was developed under the assumption that the growth characteristics, potency, and PKPD relationship are relevant for translation from mice to patients. A schematic of the human PKPD TGI model is presented in Fig. 5. The model was used to simulate human TGI versus time profiles for different doses/schedules. Approximately 1000 simulations were performed using the PD parameters of the JHH5 model at doses of 500 up to 2500 IU/m2 Q3W in males (Fig. 6) and females (Supplementary Fig. S6). The simulations (median) showed that complete tumor regression (i.e., 90% reduction in TV versus baseline) was observed at 750, 1000, and 1250 IU/m2 Q3W after the 5th, after the 3rd, and between the 2nd and 3rd administration, respectively. Schematic of the human PKPD TGI model. Aspa plasma asparaginase activity in U, CL absolute clearance, IV intravenous, k21 inter-compartmental rate constant, kg exponential growth rate of tumor, kgl linear growth rate of the tumor, Km concentration at half maximum elimination capacity, KmTR concentration at half maximum peripheral distribution capacity, PD pharmacodynamics, PK pharmacokinetics, Slp is slope of linear relationship, TGI tumor growth inhibition, Vc central volume of distribution, Vmax maximum elimination capacity, VmaxTR maximum peripheral distribution capacity, τ transit time, X1 volume of replicating tumor cells, X2, X3, and X4 nonreplicating dying cells Predicted TGI versus time profiles in patients using PD parameters from the JHH5 model (SEX = 1 for Males, BSA 2.05 m2, AGE, 55 years old). Solid red line: 90% tumor volume regression; solid black line: median TV; grey shaded area: 90% prediction intervals; vertical black dotted lines: days of drug administration. BSA body surface area, IU international unit, PD pharmacodynamics, PI prediction interval, TGI tumor growth inhibition, TV tumor volume In summary, based on the preclinical model, the following doses are predicted to be efficacious in patients with HCC: 750 IU/m2 (low), 1000 IU/m2 (medium), and 1250 IU/m2 (high). Discussion It has been shown that a relatively high fraction of HCC is hypermethylated at the asparagine (Asn) synthetase (ASNS) promoter [17]. Testing CalPEG across a panel of HCC cell lines demonstrated that ASNS promoter-hypermethylated cell lines with the lowest ASNS mRNA expression are the most responsive to CalPEG treatment in vitro. An in vivo efficacy study in mice bearing the human ASNS promoter-hypermethylated xenograft model of HCC (JHH5 model) showed that CalPEG treatment inhibited TG in a dose-dependent manner and resulted in tumor rejections at the highest doses. However, it has to be taken into account that this hypermethylated JHH5 model is not necessarily representative for the whole HCC population, but might represent a subpopulation that is more prone to responding to CalPEG. The inhibition of TG in xenografted female BALB/c nude mice receiving CalPEG has been studied through PKPD TGI modeling using a population approach and TV data from this hypermethylated CD× JHH5 model. In this model, several dose levels of CalPEG were evaluated allowing for a better description and quantification of the dose–PAA–efficacy relationship. A mathematical framework was built to translate this relationship from animals to humans and to predict clinical efficacy of CalPEG in a (sub)population of patients with HCC. An existing population PK model of PAA in female BALB/c mice was used to predict the PAA in the JHH5 xenograft efficacy study. An external VPC showed that all treatment groups were adequately described (Fig. 1), and no difference in PK between experiments was identified, thus the estimated PK model parameters were used to simulate PAA of CalPEG for the regimen in the xenograft efficacy study during the PKPD TGI modeling process. A semi-mechanistic TGI model [25] was used to fit the TV versus time profiles from the xenograft efficacy study. From two primary model parameters i.e., kg (related to the growth characteristics of the tumor) and the slope (slp) of the linear PKPD relationship (measure of the potency of the compound), a secondary parameter, CT (i.e., minimal threshold PAA for tumor regression; CT = kg/slp), was derived [32, 33]. The estimated minimal PAA of CalPEG needed to reach tumor regression in mice was 220 IU/L. On day 12, the predicted median predose PAA values were 71.3, 167, and 596 IU/L at 20, 40, and 100 IU/kg, respectively. PAA values were around or above CT for the entire treatment duration and resulted in complete tumor regression in 6 out of 10 mice dosed at 40 IU/kg (day 14) and in all mice at 100 IU/kg (day 12). The reason for the tumor regrowth observed after CalPEG treatment at 40 IU/kg in the remaining four mice could be owing to development of resistance under CalPEG treatment, which was not accounted for by the model. Integration of a time-varying effect in the model due to resistance or adding a resistance term to the model structure may further have improved the model performance [34]. Genotyping of the xenograft tumors in mice, to assess alterations that could lead to drug resistance, was not performed;therefore, the reason for resistance development in those four mice is unknown. To account for a delay in the death of the tumor cells upon treatment, a model with three transit compartments was used. The number of transit compartments between cell proliferation and cell death is empirical and was selected on the basis of a model that provided the best fit to the data. The final TGI model fitted the observed TV data well in every experimental condition and at all tested doses, although some model misspecifications were observed in the GOF plot (Fig. 2a) for the low population predicted TV values, which could be explained by the values below LoQ, and represented 36% of the total number of observations. The imputation method was used to handle TV measurements below LoQ (set to 1 mm3) in order to decrease the time it took for the model to converge and might (partly) explain the model misspecification. Therefore, the final model was rerun using the Laplacian conditional estimation with an interaction method for final parameter estimation and the M3 method for handling of the LoQ values [35]. TV measurements that were reported as 0 were flagged as censored values (with a limit of quantification (LoQ)/measurability set to 1 mm3). A summary of the modeling results of this rerun is given in Supplementary Table S5 and Supplementary Figs. S7A–S7C. Additional bias could also have been introduced by the technique used to measure the TV [36]. It is standard practice to measure subcutaneous TV by manual measurement of the length and width of the tumor using calipers. These measurements introduce biases that are primarily owing to observer subjectivity, skin thickness, subcutaneous fat, and variations in tumor shape that are often irregular instead of an assumed spheroid. The concentration effect relationship of CalPEG was found to be linear. Additional relationships (power model, Emax, sigmoid Emax) were evaluated as well, but did not improve the fit through the data and resulted in an increase of BIC value. For translation to patients, the estimated PD parameters from the JHH5 model in tumor bearing mice were linked to the predicted PAA in patients with HCC. As currently no PK data are available in patients with HCC, an existing PK model [18] that was developed on PAA data obtained in young and adult patients with ALL was used to predict PAA in patients with HCC under the assumption that there is no difference in PK between patients with ALL and HCC. The results of the simulations must be interpreted in light of the model assumptions made and take into account the limitations of the preclinical xenograft model. The human model assumed that the growth characteristics, potency, and PKPD relationship (linear) of CalPEG were similar between mice and patients with HCC. However, tumor volume doubling times (TVDT) in patients with HCC (~120–150 days) are higher than the TVDT in tumor-bearing mice (~2.17 days) [37]. In addition, it has been shown that HCC tumors are very heterogeneous, with a wide variation in tumor growth patterns depending on viral liver diseases and tumor size. The TV in patients with HCC is approximately 40.2 cc which is much larger than the TV at inoculation (~0.165 cc) in mice. Both factors could have been used for the prediction of TV in patients with HCC and would have improved prediction of active doses in patients with HCC. It should also be noted that xenograft mouse models are a simplification of a human cancer and does not take into account tumor metastasis, host immunity, tumor heterogeneity, and development of tumor resistance. Nevertheless, the relationship between drug exposure and response from xenograft models are still useful to understand the degree of TGI of the drug and for translation of antitumor activity from mice to humans. In addition, it should be recognized that the knowledge of current mathematical modeling in oncology is mainly built on experience with small molecules [21, 22] and that for immune-oncology agents, differences between mouse and human immune systems would need to be taken into account. The predicted TV versus time curves (Fig. 6; Males, BSA 2.05 m2 and age, 55 years old) illustrated that complete tumor regression (i.e., 90% TV reduction) was observed at doses > 750 IU/m2. Visual inspection of the data in Fig. 6 illustrates that the steady-state of CalPEG is reached after approximately 9–12 weeks, which is consistent with the half-life of CalPEG (t1/2 of 16.1 days), indicating that after 12 weeks of treatment, no further increase in PK is expected. Based on all preclinical data, the predicted clinical doses associated with complete tumor regression were 750 IU/m2 Q3W (after dose 5), 1000 IU/m2 (after dose 3), and 1250 IU/m2 (dose 2–3). This means that it will take approximately 15 weeks, 9 weeks, or 6 weeks before complete tumor regression will be observed with a Q3W regimen of CalPEG, provided via a 2-h infusion, at 750, 1000, or 1250 IU/m2, respectively. The time to reach complete tumor regression could be shortened by further increasing the dose (> 1250 IU/m2); however, this might not be feasible due to dose-limited toxicities. Alternatively, before reaching the target dose level and to reach steady-state earlier, the dose frequency could be increased and a dose escalation scheme could be proposed. In females (Supplementary Fig. S6; BSA 1.81 m2 and age, 55 years old), similar clinical doses associated with complete tumor regression were identified, but since exposure in females was slightly higher than in males, complete tumor regression was achieved earlier. These model-based dose predictions were in line with human equivalent doses derived via the PK matching approach, i.e., exposures at 20 IU/kg (mice) ~750 IU/m2 Q3W (patients), 40 IU/kg ~1000 IU/m2 Q3W and 100 IU/kg ~2000 IU/m2 Q3W. The above findings show that population PKPD TGI models can help inform on dosing strategies and in designing successful clinical trials. Conclusions A PKPD TGI model that was developed adequately described the TV versus time profiles in a preclinical hypermethylated xenograft model relevant for HCC. This TGI model was used to quantify the PKPD relationship between PAA and TGI in mice. Then, a human PKPD TGI model was developed by linking the structural mouse TGI model with the human population PK model. This human PKPD TGI model was used to perform simulations and to predict antitumor activity of CalPEG at different doses/schedule in patients with HCC. This work adequately illustrates how legacy clinical data together with newly generated preclinical data can be integrated into a PKPD TGI model to build a preclinical to clinical translational framework in order to be able to predict the exposure–response relationship in newly considered indications. In addition, the translational modeling work as presented in the paper can also be used to predict clinical response in cases of combination therapies and for quantitative prediction of the exposure–safety relationship. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 1086 KB) Funding This study was funded by Servier. Declarations Conflict of Interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests. Eef Hoeben reports a relationship with Servier Pharmaceuticals LLC that includes: consultant. Vincent Madelain reports a relationship with Servier Pharmaceuticals LLC that includes: employment. Chayan Acharya reports a relationship with Servier Pharmaceuticals LLC that includes: consultant. Pierre Barbier Saint Hilaire reports a relationship with Servier Pharmaceuticals LLC that includes: employment. Ethika Tyagi reports a relationship with Servier Pharmaceuticals LLC that includes: employment. Malek Okour reports a relationship with Servier Pharmaceuticals LLC that includes: employment. Sylvain Fouliard reports a relationship with Servier Pharmaceuticals LLC that includes: employment. Ethical Approval All studies were conducted in accordance with ethical principles of the Declaration of Helsinki, and were consistent with International Conference on Harmonisation, Good Clinical Practices, and applicable regulatory requirements. Availability of Data The data that support the findings of this study are available from the corresponding author upon reasonable request. Qualified researchers may request access to study documents that support the methods and findings reported in this manuscript. Consent to Participate All participants provided consent prior to enrolling in these studies. Consent for Publication Not applicable Code Availability Not applicable. Author Contributions Eef Hoeben and Sylvain Fouliard planned and executed the modelling activity; Vincent Madelain was involved in study design; Ethika Tyagi and Pierre Barbier Saint Hilaire were involved in data generation; Chayan Acharya was involved in clinical modelling of PAA; Malek Okour and Sylvain Fouliard were involved in study oversight. All authors contributed to the interpretation of results, development of the manuscript and critical revision of its content and approved the final version for submission. References Associated Data Supplementary Materials Supplementary file1 (DOCX 1086 KB)