The iMATRIX-atezolizumab study (NCT02541604) was a phase I/II, multicenter, open-label study to assess the safety and PK of atezolizumab in pediatric patients and young adults. The study enrolled patients with solid tumors with known or expected PD-L1 pathway involvement for which prior treatment was proven to be ineffective or intolerable, and for whom there was no curative standard-of-care treatment. Patients with Hodgkin lymphoma (HL), non-Hodgkin lymphoma (NHL), or other rare tumors with/without documented expression of PD-L1 on tumor cells or immune infiltrating cells were eligible. Patients with history of any autoimmune disease were excluded. However, patients with a history of autoimmune-related hypothyroidism on a stable dose of thyroid-replacement hormone, or those with controlled Type 1 diabetes mellitus on a stable insulin regimen were eligible. Patients received atezolizumab q3w using a weight-adjusted dose at 15 mg/kg for patients aged < 18 years (maximum dose 1200 mg), and a flat dose of 1200 mg for patients aged ≥ 18 years. Atezolizumab was administered by intravenous infusion on day 1 of each cycle, with an infusion duration of 60 min in cycle 1 and 30 min in subsequent cycles. The study was conducted in accordance with the Declaration of Helsinki and Good Clinical Practice Guidelines, following approval by the ethics board in each institution. Informed consent was obtained from each patient or each patient’s authorized representative.

The PK/anti-drug antibody (ADA) sampling schedule following atezolizumab administration was designed to describe its distribution, elimination, and immune response. The sparse PK/ADA sampling schedule was used to assess PK and ADA after single and repeated dosing. Pharmacokinetic and ADA sampling of atezolizumab was performed at the end of infusion on day 1 of cycles 1 and 4 (PK only), and C_min and ADA were collected prior to infusion on day 1 of cycles 2, 3, 4, 8, 12, 16, and every 8 cycles thereafter. Atezolizumab was quantified by enzyme-linked immunosorbent assay (ELISA). The lower limit of quantification (LOQ) for the atezolizumab assay in human serum was 60 ng/mL. Samples for ADA analysis were evaluated using a bridging ELISA assay with positive samples in screening further confirmed by titer. Further details for PK and immunogenicity assays have been reported previously [27].

Exploration and visualization of the data, as well as descriptive statistics, were performed using R v3.3.1 with additional CRAN packages. The dataset included 520 samples; 81 samples that occurred prior to the first dose were below the LOQ and were excluded. Data manipulation was limited to flagging data records not used for analysis, imputing missing variables to median values, and excluding patients with no dose information (n = 1). Missing covariate values were imputed to median values for continuous covariates or to the most frequent category for categorical covariates.

The popPK analysis was performed using a non-linear mixed-effects modeling approach with NONMEM v7.3 (ICON Development Solutions, Ellicott City, Maryland, USA) in conjunction with Perl-Speak-NONMEM (PsN) (v3.7.6, Uppsala University, Uppsala, Sweden). A prior two-compartment intravenous infusion input adult popPK model of atezolizumab was used as a basis to model pediatric data. The typical clearance (CL; L/day) of atezolizumab for an adult patient i was: CLi=0.200∙ALBUi40−1.12∙BWTi770.808∙TUMi630.125∙1.159ifADAis positive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ {\mathrm{CL}}_i=\left(0.200\bullet {\left(\frac{ALBU_i}{40}\right)}^{-1.12}\bullet {\left(\frac{BWT_i}{77}\right)}^{0.808}\bullet {\left(\frac{TUM_i}{63}\right)}^{0.125}\right)\bullet \left(1.159\ if\ ADA\ is\ positive\right) $$\end{document}

BWT: Body weight (kg); ALBU: Albumin (g/L); TUM: Tumor burden (mm); ADA: Post-baseline status of anti-drug antibodies.

The typical volume of the central compartment (V1; L) and volume of the peripheral compartment (V2; L) of atezolizumab for an adult patient i were: V1i=3.28∙BWTi770.559∙ALBUi40−0.350∙0.871if female\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ V{1}_i=\left(3.28\bullet {\left(\frac{BWT_i}{77}\right)}^{0.559}\bullet {\left(\frac{ALBU_i}{40}\right)}^{-0.350}\right)\bullet \left(0.871\ if\ female\right) $$\end{document}

V2i=3.63∙0.728if female\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ V{2}_i=3.63\bullet \left(0.728\ if\ female\right) $$\end{document}

An extensive list of covariates reflecting cancer status/type, organ dysfunction, and race/region tested in the prior adult PK model were not re-tested in the pediatric and young adult model. To provide consistency between the adult and pediatric popPK analyses, the adult popPK model was fitted to the pediatric PK data, utilizing the same structure, but re-estimating each parameter. Proportional changes (i.e., for ADA and sex) in the pediatric and young adult model were parameterized θ.cov, where θ indicates the proportional change and cov was either ADA or sex (both coded 0 or 1); this differed from the prior adult model.

The performance of the model was evaluated using standard diagnostic plots to evaluate the observed dependent variable (atezolizumab concentration) versus population predictions, dependent variable versus individual predictions, conditional weighted residuals (CWRES) versus population predictions, CWRES versus time, quantile-quantile plot of CWRES, random effect distributions, and correlations of random effects between parameters. The predictive performance of the popPK model was also evaluated with a prediction-corrected visual predictive check with 500 replicates [31, 32].

Individual empirical Bayesian estimates of PK parameters were used to compute atezolizumab exposure variables based on the nominal dose regimen including area under the curve (AUC), maximum concentration (C_max), and minimum concentration C_min, in cycle 1 and at steady-state. The cycle 1 and steady-state PK profile for each individual based on the starting dose was simulated using individual empirical Bayesian estimates of PK parameters based on the final model. The following time points were used for simulations: 0, every 0.01 day for the first 3 days, every 0.5 days until 21 days post dose, and 20.99 days post dose at cycle 1, and a similar schedule at steady-state (cycle 10). Atezolizumab exposure metrics including C_max, C_min, and AUC (cycle 1) were derived from the simulated individual PK profiles, and AUC at steady-state was derived as dose/CL. The resulting metrics were compared and stratified by age group using box-plots.

The exposure-response analysis of safety was conducted using data from all atezolizumab-treated patients for whom exposure data were available. p(AE) is the observed probability of an adverse event (AE) versus atezolizumab AUC in cycle 1. Exposure levels of atezolizumab were binned based on the quantiles of the log-transformed AUC. A mean curve obtained from averaging each exposure record in the data set and binning boundaries by quartiles of exposure was established. Bootstrapped replicates (n = 100) were used to plot the 90% confidence band for the mean fit curve. The overall analysis represented findings across 69 pediatric patients.

The primary efficacy outcome measures were objective response rate (ORR) and progression-free survival (PFS). ORR was defined as the proportion of patients with measurable disease at baseline who achieved a complete or partial response, with response on two consecutive occasions ≥ 4 weeks apart, as determined by the investigator using Response Evaluation Criteria in Solid Tumors (RECIST) v1.1. PFS was defined as the time from initiation of study drug to the first documented occurrence of disease progression, as determined by the investigator using RECIST v1.1 criteria.

The presence of ADAs to atezolizumab during the study, relative to baseline, and in relationship to the serum concentration of atezolizumab at specified time points was determined. Characterization of immunogenicity was performed for all patients with at least one ADA assessment. Patients were considered ADA positive if they were ADA negative or had missing baseline data but developed an ADA response following study drug exposure, or if they were ADA positive at baseline and the titer of one or more post-baseline samples was ≥ 0.60 titer-units greater than that of the baseline sample. Patients were considered ADA negative if they were ADA negative or had missing baseline data and all post-baseline samples were negative, or if they were ADA positive at baseline but did not have any post-baseline samples with a titer that was ≥ 0.60 titer-units greater than that of the baseline sample.

A total of 431 atezolizumab serum concentrations from 87 relapsed-refractory pediatric and young adult patients enrolled in the iMATRIX-atezolizumab study were used for the popPK analysis. The dataset comprised predominantly patients aged < 18 years, including two infants aged < 2 years, with a wide body weight (8.7–154 kg) and age range (7 months–29 years). Median age and weight was 12 years and 38.9 kg, respectively, across the 69 pediatric patients, and 22 years and 61.0 kg, respectively, across the 18 young adults (Table 1 and Additional file 1: Figure S1).Tab1

Patient baseline demographic and clinical characteristics

Descriptive statistics of patient characteristics and covariates by age group are summarized in Table 1. Baseline demographics were balanced in terms of gender. Although the sample size was limited, no apparent difference in albumin or ADA response to atezolizumab by age was seen (P > 0.05). Multiple tumor types were present including Ewing sarcoma, neuroblastoma, non-rhabdomyosarcoma soft tissue sarcoma, osteosarcoma, rhabdomyosarcoma, Wilms’ tumor, HL, NHL, malignant rhabdoid tumor, atypical teratoid/rhabdoid tumor, and other rare tumors. The number of tumor types was diverse across the age groups, with median tumor burden increasing with age. The majority of patients had a Lansky/Karnofsky performance score ≥ 80%.

A pediatric and young adult model was established with the current study data utilizing the same structure as the adult popPK model to allow for consistency, while re-estimating each parameter. The prior adult model was a two-compartment model with an intravenous infusion input.

Parameter estimates from the modeling are displayed in Table 2. Parameters were estimated with good precision. Parameter estimates for CL and V of 0.217 L/day, and 3.01 L, respectively, including covariate effects, were generally in line with the prior adult popPK model. Two exceptions were the estimates for V2 and inter-compartmental clearance (Q), which were not weight-normalized, and decreased in pediatric patients. As a sensitivity analysis, the inclusion of weight and age on V2 and Q resulted in estimates closer to, but still less than, those achieved in adults. Sex effects had minimal impact on the objective function. Between-subject and residual variability were acceptable given the relatively small numbers of patients and sparse PK sampling.Tab2

Parameter estimates in pediatric and young adult patients

Graphical evaluations of the final popPK model are displayed in Fig. 1. The plots suggest that the model is adequate with respect to structure and covariate parameterizations. In particular, relationships of the random effects for CL and V (eta.CL and eta.V1) did not show any bias with age (smooth curve showing a horizontal linear relationship around zero) (Fig. 1d) suggesting that the body weight effects in these parameters captured the difference between adults and pediatric patients. The prediction-corrected visual predictive check (Fig. 1a) suggested that the model captured the central tendency and the variability in PK. Given the interest in body surface area (BSA)-based dosing for pediatric patients, a plot of the random effects of CL and V1 by BSA was also explored (Additional file 2: Figure S2). No bias was revealed, suggesting that covariates including body weight in the model also account for changes in BSA, highlighting the appropriateness of weight-based dosing.Fig1Fig. 1

(a) Prediction-corrected visual predictive check, (b) goodness of fit diagnostic plots, (c) Eta distributions, and (d) random effect correlations to covariates. Prediction-corrected visual predictive check (a): the gray solid and dashed lines represent the observed median and the 10th and 90th percentiles, respectively, while the two shades of blue represent overlap between the empirical 95% prediction intervals. Goodness of fit diagnostic plots (b): the gray solid line indicates fitted values from a nonparametric smoother. Dashed lines indicate the line of unity (top plots), or zero lines and boundary lines for conditional weighted residuals (bottom). Eta distributions (c): the blue solid line represents a density curve. Random effect correlations to covariates (d): for continuous covariates, the blue solid line represents fitted values from a nonparametric smoother. The dashed line indicates the zero line, the box-plot indicates the median and interquartile range (25th to 75th percentile), the whiskers indicate 1.5 times the interquartile range. Abbreviations: ADA anti-drug antibody, CL clearance, V1 volume of the central compartment, V2 volume of the peripheral compartment

Summaries of the individual exposure metrics are displayed in Fig. 2, based on individual model predictions across the 87 patients at cycle 1 and steady-state. Overall, AUC and C_max increased from children to adolescents to young adults, whereas C_min was comparable across age groups, especially at steady-state. The expected inter-quartile range (IQR) of exposure in 1000 simulated adult patients (median age: 62 years) using the adult popPK model are also shown. Following the 15 mg/kg simulated regimen in adults, a median cycle 1 C_min of 53.0 μg/mL with an interquartile interval (Q1 and Q3) of 44.6 and 64.7 μg/mL was predicted.Fig2Fig. 2

Cycle 1 and steady-state (cycle 10) exposure metrics by age group: (a) C_max, (b) C_min, and (c) AUC. Expected interquartile range (IQR) from simulated distributions (n = 1000) based on reported geometric means and %CVs. The box-plots indicate the median and IQR (25th to 75th percentile). The whiskers indicate 1.5 times the IQR. Abbreviations: AUC area under the curve, C _min minimum concentration, C _max maximum concentration

The exposure-safety analysis was performed in all pediatric patients aged < 18 years with exposure data (n = 69). The incidence of grade ≥ 3 AEs and AEs of special interest (AESI) versus atezolizumab cycle 1 AUC is shown in Fig. 4. AESI categories included: rash, hepatitis, aspartate transferase/alanine aminotransaminase elevations, infusion-related reactions, hypothyroidism, blood thyroid-stimulating hormone increase, pancreatitis, diabetes mellitus, colitis, hyperthyroidism, and meningoencephalitis. Grade ≥ 3 AEs and all-grade AESI occurred at an incidence of 33% (events in 69 patients) and 46% (events in 69 patients), respectively. Exposure metrics within the first treatment cycle were used rather than steady-state to isolate potentially confounding factors on exposure such as time-varying clearance [33]. No exposure-response relationship with atezolizumab AUC in cycle 1 was detected.Fig4Fig. 4

Incidence of grade ≥ 3 AEs (a) and any-grade AESI (b). AEs and AESI are displayed by open blue circles. Solid black circles with standard error bars (y-value: binned probability of having an event from observations; x-value: median exposure value within the bin). Red line: mean model fitted curve (obtained from averaging the fitted curve for each exposure record in the data set). Dashed green lines: binning boundaries. Exposure levels are binned based on the quantiles of the log transformed exposure variable levels. Blue shaded area: based on 100 bootstrap replicates, depicting the 90% confidence band for the mean model fitted curve. Plot is based on 69 patients. Abbreviations: AE adverse event, AESI adverse event of special interest, AUC area under the curve

Among 87 patients, there were 4 responders (4.6%), all of whom had a partial response; 1 of these patients had malignant rhabdoid tumor, 2 had HL, and 1 had NHL. In total, 63 patients (72.4%) had disease progression, 10 (11.5%) had stable disease, 2 (2.3%) were not evaluable, and 8 (9.2%) had missing post-baseline assessments. Median PFS was 1.3 months (95% confidence interval [CI], 1.2–1.4). Overall, 63 patients were evaluable for PD-L1 expression, of whom 18% had high PD-L1 expression (IC2/3), including all of the 4 responding patients. Interpretation of atezolizumab exposure and biomarker expression with outcomes was not performed due to the low number of responders.

Ten patients had missing ADA records, which were imputed as the median (ADA negative) for the popPK analysis. The number of imputed records by age group was: 1/2 (< 2 years), 5/29 (2 to < 12 years), 3/38 (12 to < 18 years), and 1/18 (≥ 18 years). Imputed records were not expected to impact the outcome given that < 20% of total ADA records in any given age group (apart from infants which were not interpretable) were imputed. Observed exposure and safety by ADA were interpreted using non-imputed records.

Overall, 11/87 (13%) patients were treatment-emergent ADA-positive to atezolizumab, which included 0/2 (0%) patients aged 0 to < 2 years, 5/29 (17%) patients aged 2 to < 12 years, 4/38 (11%) patients aged 12 to < 18 years, and 2/18 (11%) patients aged ≥ 18 years. The observed geometric mean peak and trough exposure of atezolizumab by ADA status in PK-evaluable patients across multiple cycles is provided in Additional file 4: Table S1. The geometric mean cycle 1 C_min of atezolizumab was comparable between ADA-positive (57.0 μg/mL) and ADA-negative (62.5 μg/mL) patients.

The incidence of serious AEs was broadly similar between ADA-positive (36.4%) and ADA-negative (34.8%) patients, as was the incidence of grade 3/4 AEs (63.6 and 56.1%, respectively). Overall, 7/11 (63.6%) ADA-positive and 29/66 (43.9%) ADA-negative patients experienced ≥ 1 immune-related AESI. Interpretation of any effect of ADA on AE/AESI incidence or severity in pediatrics was limited by the low number of ADA-positive patients.

The PK and safety profile was generally comparable between ADA-positive and ADA-negative patients. The relationship between pediatrics and young adults in terms of demographics, disease, immune status, and genetics that could influence ADA production remains unknown given the small size of the ADA-positive population.