Cancer Chemotherapy and Pharmacology

, Volume 70, Issue 4, pp 591–601

A population pharmacokinetic/pharmacodynamic model of thrombocytopenia characterizing the effect of trastuzumab emtansine (T-DM1) on platelet counts in patients with HER2-positive metastatic breast cancer


  • Brendan C. Bender
    • Genentech, Inc.
    • Department of Pharmaceutical BiosciencesUppsala University
  • Franziska Schaedeli-Stark
    • F. Hoffman-La Roche Ltd.
  • Reinhold Koch
    • F. Hoffman-La Roche Ltd.
  • Amita Joshi
    • Genentech, Inc.
  • Yu-Waye Chu
    • Genentech, Inc.
  • Hope Rugo
    • UCSF Helen Diller Family Comprehensive Cancer Center
  • Ian E. Krop
    • Dana–Farber Cancer Institute
    • Genentech, Inc.
  • Lena E. Friberg
    • Department of Pharmaceutical BiosciencesUppsala University
  • Manish Gupta
    • Genentech, Inc.
    • Bristol–Myers Squibb
Original Article

DOI: 10.1007/s00280-012-1934-7

Cite this article as:
Bender, B.C., Schaedeli-Stark, F., Koch, R. et al. Cancer Chemother Pharmacol (2012) 70: 591. doi:10.1007/s00280-012-1934-7



Trastuzumab emtansine (T-DM1) is an antibody-drug conjugate in the development for the treatment of human epidermal growth factor receptor 2-positive cancers. Thrombocytopenia (TCP) is the dose-limiting toxicity of T-DM1. A semimechanistic population pharmacokinetic/pharmacodynamic (PK/PD) model was developed to characterize the effect of T-DM1 on patient platelet counts.


A PK/PD model with transit compartments that mimic platelet development and circulation was fit to concentration-platelet–time course data from two T-DM1 single-agent studies (TDM3569g; N = 52 and TDM4258g; N = 112). NONMEM® 7 software was used for model development. Data from a separate phase II study (TDM4374g; N = 110) were used for model evaluation. Patient baseline characteristics were evaluated as covariates of model PD parameters.


The model described the platelet data well and predicted the incidence of grade ≥3 TCP. The model predicted that with T-DM1 3.6 mg/kg given every 3 weeks (q3w), the lowest platelet nadir would occur after the first dose. Also predicted was a patient subgroup (46 %) having variable degrees of downward drifting platelet–time profiles, which were predicted to stabilize by the eighth treatment cycle to platelet counts above grade 3 TCP. Baseline characteristics were not significant covariates of PD parameters in the model.


This semimechanistic PK/PD model accurately captures the cycle 1 platelet nadir, the downward drift noted in some patient platelet–time profiles, and the ~8 % incidence of grade ≥3 TCP with T-DM1 3.6 mg/kg q3w. This model supports T-DM1 3.6 mg/kg q3w as a well-tolerated dose with minimal dose delays or reductions for TCP.


Trastuzumab emtansineT-DM1ThrombocytopeniaPopulation pharmacokinetic/pharmacodynamic modelSemimechanisticCumulative TCP


Trastuzumab emtansine (T-DM1) is an antibody-drug conjugate (ADC) in development for the treatment of human epidermal growth factor receptor 2 (HER2)-positive cancers [14]. It is composed of the potent antimicrotubule maytansinoid derivative DM1 conjugated to the HER2-targeted monoclonal antibody trastuzumab via a stable thioether linker, MCC ([N-maleimidomethyl] cyclohexane-1-carboxylate). T-DM1 binds to HER2-overexpressing cells and undergoes receptor-mediated internalization, resulting in the intracellular release of Lys–MCC–DM1 and subsequent tumor cell death [1]. In two phase II studies [3, 4], T-DM1 3.6 mg/kg administered intravenously every 3 weeks (q3w) demonstrated activity against HER2-positive metastatic breast cancer (MBC) that was previously treated with HER2-directed therapy, with objective response rates of 25.9 and 34.5 % by independent radiologic review.

Thrombocytopenia (TCP) was among the more frequently observed toxicities in clinical studies of T-DM1 [24]. In the phase I dose-escalation study (TDM3569g) [2], the maximum tolerated T-DM1 dose was 3.6 mg/kg q3w, with grade 4 TCP as the dose-limiting toxicity (DLT) at 4.8 mg/kg q3w. Platelet counts of most patients treated at the maximum tolerated dose followed a characteristic pattern: after the administration of T-DM1, platelet counts decreased to a nadir by day 8, with recovery to baseline levels by day 1 of the next cycle. A slow downward drift in the platelet–time profile was noted in some patients over multiple cycles. In three early clinical studies of single-agent T-DM1, platelet declines were mostly grade 1 or 2 and no clinically significant bleeding events were reported [24]. The mechanisms of T-DM1-induced TCP are not known. In contrast to platelets, other hematologic lineages were relatively spared from T-DM1; leukopenia, neutropenia, and anemia were all observed at much lower incidences compared with TCP.

This report describes the development of a semimechanistic population pharmacokinetic/pharmacodynamic (PK/PD) model, which was designed to (1) describe the time course of platelet response to T-DM1, (2) test model structures to support hypotheses regarding the mechanism(s) of the effect of T-DM1 on platelet counts, (3) evaluate patient baseline characteristics as covariates of model PD parameters, and (4) predict platelet response and incidence rates of grade ≥3 TCP in future clinical studies of T-DM1.

Materials and methods

Patient population

The PK/PD model was based on 4,340 assessments of platelet concentration from 164 patients enrolled in two studies: the phase I study TDM3569g (N = 52) and the phase II study TDM4258g (N = 112). Patients in the phase I study received T-DM1 intravenously qw (weekly) or q3w. Doses in the q3w regimen were 0.3 mg/kg (n = 3), 0.6 mg/kg (n = 1), 1.2 mg/kg (n = 1), 2.4 mg/kg (n = 1), 3.6 mg/kg (n = 15), and 4.8 mg/kg (n = 3). Doses in the qw regimen were 1.2 mg/kg (n = 3), 1.6 mg/kg (n = 3), 2.0 mg/kg (n = 3), 2.4 mg/kg (n = 16), and 2.9 mg/kg (n = 3). Patients in the phase II study received T-DM1 3.6 mg/kg q3w. Data from 110 patients enrolled in the phase II study TDM4374g, evaluating T-DM1 3.6 mg/kg q3w (1,841 assessments of platelet concentration), were used for external evaluation of the model.

Baseline demographic and disease characteristics from studies TDM3569g, TDM4258g, and TDM4374g have been reported elsewhere and are similar across studies [25]. Demographic values (mean ± SD) are summarized briefly here: age (54 ± 10 years), weight (72.1 ± 17.7 kg), observed baseline platelet count (267 ± 105 × 1,000/μL), creatinine clearance (96 ± 34 mL/min), serum creatinine (75 ± 56 μmol/L), albumin (39 ± 5 g/L), total protein (70.4 ± 7.3 g/L), alanine transaminase (ALT) (31 ± 21 IU/L), aspartate transaminase (AST) (39 ± 28 IU/L), total bilirubin (8.1 ± 5.3 μmol/L), tumor burden (based on the sum of the longest diameters of target lesions reported at baseline) (9.3 ± 7.2 cm), and HER2 expression (84.4 ± 126 ng/mL). Patient races were of the following percentages: White, 79 %; Black, 8 %; Hispanic or Latino, 7 %; American Indian or Alaska Native, 4 %; Asian, 2 %. Patients received a median duration of seven cycles of T-DM1 (range 1–34 cycles). In all three studies, T-DM1 treatment continued until progressive disease or unacceptable T-DM1 toxicity. Dose delays or dose reductions to 3.0 mg/kg due to TCP or other reasons occurred in 24 patients (~10 %) in the 3.6 mg/kg q3w group. TCP of grades 1–4 was reported in 31 % (73/237) of patients treated at 3.6 mg/kg q3w, with approximately 8 % experiencing grade ≥3 TCP.

Platelet measurements

Hematologic sampling was conducted at multiple time points throughout the studies. Patients with baseline platelet counts <100 × 1,000/μL were excluded from the studies. For patients receiving q3w regimens in study TDM3569g, platelet counts were assessed on days 1, 2, 4, 7–8, 11, and 18 in cycle 1. Typical collection times in subsequent cycles were days 1, 4, and 7–8. For patients receiving qw regimens, platelet counts were assessed on days 1, 2, 4, 7–8, 11, 14–15, and 18 in cycle 1, with predose weekly sampling beginning on day 21. For TDM4258g, platelet counts were assessed every 7 days in all cycles. For TDM4374g, platelet counts were assessed weekly during cycle 1 and on days 1 and 8 of all subsequent cycles. When hematologic toxicity occurred, more frequent sampling was conducted in accordance with the clinical protocol.

T-DM1 population PK modeling

A T-DM1 population PK model was previously developed from the phase I and II studies [5]. Briefly, plasma T-DM1 concentrations were measured using an enzyme-linked immunosorbent assay (ELISA), which captured any trastuzumab molecule conjugated to DM1 (1–8 DM1 molecules per antibody). In cycles 1 and 4, frequent PK sampling post-T-DM1 infusion was conducted, while PK sampling was limited to pre- and post-T-DM1 infusion in other cycles. A linear two-compartment PK model with first-order elimination from the central compartment best described the T-DM1 plasma concentration–time data. PK parameter estimates and interindividual variability (IIV) were: clearance (CL) = 0.7 L/day (21), central compartment volume (V1) = 3.33 L (13 %), peripheral compartment volume (V2) = 0.89 L (50 %), and inter-compartmental clearance (CLd) = 0.78 L/day (IIV not estimated). Shrinkage values for CL, V1, and V2 were 8, 18, and 35 %, respectively. For PK/PD modeling, the post hoc Bayesian estimates of individual PK parameters were used as input to model the respective patient platelet response.

T-DM1 population PK/PD modeling of platelet response

A schematic of the semimechanistic population PK/PD model that was developed to describe the T-DM1-driven platelet–time course is shown in Fig. 1. This model structure was based on the PK/PD model of myelosuppression proposed by Friberg et al. [6, 7], which has been widely used to characterize the time course of leukocyte, neutrophil, and platelet counts during cytotoxic drug treatment [812]. The final model consisted of a T-DM1 drug effect inhibiting the proliferation rate of a proliferative platelet pool (PP) compartment, three transit compartments (T1, T2, and T3), and a circulating platelet (PLT) compartment. As shown in Fig. 1, individual patient PK parameters from the population PK model provided T-DM1 central compartment concentrations (C), which acted as linear inhibitory drug effects (C × Slope) on the proliferation rate (KPROL) of the PP compartment. Other system-related parameters that were estimated included the mean transit time (MTT; h), equal to the number of intercompartmental transits divided by the transit rate constant (4/Ktr; h−1); BASE (×1,000/μL), the total baseline platelet count at time = 0; GAM, a dimensionless feedback term that increases the proliferation rate when platelet counts (PLT) drop below BASE.
Fig. 1

Schematic of the semiphysiologic PK/PD model describing the time course of platelet response after T-DM1 administration (modified from Friberg et al. [6]). PP the proliferative platelet pool compartment; T1, T2, and T3 transit compartments; PLT circulating platelet compartment; BASE baseline platelet count at time = 0, modeled as BASE1 + BASE2 in the proliferating platelet compartment; BASE/PLT baseline platelet count/platelet count at time (t); BASE1 amount of baseline proliferating PP that is nondepletable; BASE2 amount of baseline proliferating PP that is depletable by the Kdeplete × Cavg rate; Cavg average T-DM1 concentration over dosing intervals; CL clearance; CLd intercompartmental clearance; C × Slope T-DM1 drug effect; GAM feedback parameter; Kdeplete rate of depletion of BASE2 PP; Kel rate of physiologic elimination of circulating platelets; KPROL rate of PP proliferation; Ktr transit rate constant between transit compartments; MTT mean transit time through transit compartment chain; Slope drug potency parameter; V1 T-DM1 central volume of distribution; V2 T-DM1 peripheral volume of distribution

Modifications to the myelosuppression model [6] (i.e., the addition of structural components and parameters) were made taking into account notable observations from the clinical studies. First, in most patients, platelet reduction to nadir and return to baseline appeared stable throughout the course of treatment; however, in some patients, platelet–time profiles drifted slowly down over multiple cycles of T-DM1. The rate and extent of this decline was modeled by the introduction of an additional slow T-DM1-related drug effect on a depletable fraction of the proliferating PP. Specifically, the PP compartment was modeled as a nondepletable (BASE1) proliferating pool and as a depletable (BASE2) proliferating pool of platelets. The drug effect (Cavg × Kdeplete) was incorporated to slowly deplete the BASE2 pool over time, where Kdeplete (L/mg × week−1) is the depletion rate constant and Cavg (mg/L) is the average T-DM1 concentration over dosing intervals. BASE2 is a function of time (t), where BASE2(t) = BASE2 × EXP(−Kdeplete × Cavg × t). Given that BASE = BASE1 + BASE2, the total “baseline” platelet count at time (t) is calculated as BASE(t) = BASE1 + (BASE − BASE1) × EXP(−Kdeplete × Cavg × t). Thus, BASE1 and Kdeplete were additional system- and drug-related parameters to be estimated, respectively. During model building, Kdeplete parameter estimates resulted in a skewed, apparently bimodal distribution with a tenfold difference between the mode values. A mixture model implementation in NONMEM® 7 software (ICON, Dublin, Ireland) was used to estimate the probability of a lower (POP1) or higher (POP2) value of Kdeplete, resulting in two patient subgroups with an apparent stable platelet–time profile (POP1) or with an apparent decline (POP2).

The second modification was applied in consideration of the clinical observation that platelet count nadirs were generally lowest after the first T-DM1 dose, seen primarily with the q3w regimen. This phenomenon was modeled using two separate Slope parameters (i.e., Slope1 for the first dose only and Slope2 for all subsequent doses) for both q1w and q3w data.

The final PK/PD model differential equations describing the platelet–time course following T-DM1 dosing are shown below. A(1) and A(2) represent T-DM1 mass in the PK central and peripheral compartments, respectively. The T-DM1 concentration–time course is described as A(1)/V1. Platelet count (PLT) was the dependent variable (DV) in the model.
$$ {\text{d}}A\left( 1 \right)/{\text{d}}t = {\text{ CLd}}/V_{2} \times A\left( 2 \right) \, - {\text{ CLd}}/V_{1} \times A\left( 1 \right) \, - {\text{ CL}}/V_{1} \times A\left( 1 \right) $$
$$ {\text{d}}A\left( 2\right)/{\text{d}}t = {\text{ CLd}}/V_{ 1} \times A\left( 1\right) \, - {\text{ CLd}}/V_{ 2} \times A\left( 2\right) $$
$$ {\text{Concentration }}\left( t \right) \, = A\left( 1\right)/V_{ 1} $$
$$ {\text{Drug}}\,{\text{Effect}}_{ 1} = {\text{ Slope}} \, \times \,{\text{Concentration }}\left( t \right) $$
$$ {\text{Drug}}\,{\text{Effect}}_{ 2} = K_{\text{deplete}} \times C_{\text{avg}} $$
$$ K_{\text{tr}} = { 4}/{\text{MTT}} $$
$$ {\text{BASE}}\left( t \right) \, = \, ({\text{BASE}} - {\text{BASE}}_{ 1} ) \times {\text{EXP}}( - {\text{Drug}}\,{\text{Effect}}_{ 2} \times t) + {\text{ BASE}}_{ 1} $$
$$ {\text{dPP}}/{\text{d}}t = - K_{\text{tr}} \times {\text{PP }} + K_{\text{tr}} \times {\text{PP}} \times ( 1- {\text{Drug}}\,{\text{Effect}}_{ 1} ) \times \left( {{\text{BASE}}\left( t \right)/{\text{PLT}}} \right)^{\text{GAM}} $$
$$ {\text{d}}T_{ 1} /{\text{d}}t = - K_{\text{tr}} \times T_{ 1} + K_{\text{tr}} \times {\text{PP}} $$
$$ {\text{d}}T_{ 2} /{\text{d}}t = - K_{\text{tr}} \times T_{ 2} + K_{\text{tr}} \times T_{ 1} $$
$$ {\text{d}}T_{ 3} /{\text{d}}t = - K_{\text{tr}} \times T_{ 3} + K_{\text{tr}} \times T_{ 2} $$
$$ {\text{dPLT}}/{\text{d}}t = - K_{\text{tr}} \times {\text{PLT }} + K_{\text{tr}} \times T_{ 3} $$

Data analysis

PK/PD model building was performed using the first-order conditional estimation (FOCE) method with INTERACTION using NONMEM 7 [13]. Platelet count data were log-transformed. Log-normal parameter distributions were used for IIV, where the parameter for an ith patient was represented by Parameteri = Typical Value × exp(ηi), where ηi represents the IIV. The residual error was modeled as a proportional error, which is additive in the log domain. The objective function value (OFV) was used for the comparison of hierarchical models, using the log-likelihood ratio test. A difference in OFV of >3.84, corresponding to a significance level of P < 0.05, was used for discrimination between two nested models that differed in one parameter.

Model building

During the model-building process, a linear (Slope × C) versus a nonlinear ([Emax × C]/[EC50 + C]) drug effect, a drug effect on PP, T1, T2, T3, or PLT compartments, as well as variable numbers of transit compartments (n = 2–5), were considered. With regard to the downward drift in platelet profiles observed in some patients, numerous approaches were taken: (1) modeling was done with and without mixture model implementation, (2) the platelet proliferation pool was not divided and thus only a single BASE parameter was estimated, (3) Kdeplete,POP1 was fixed at 0, (4) the downward drift was modeled as driven by cumulative T-DM1 exposure, (5) the downward drift was modeled as driven by a time effect on the feedback (GAM) parameter, and (6) with regard to the low cycle 1 platelet count nadirs, a single Slope parameter was tried, as well as using an intraoccasion variability (IOV) on Slope; a similar IOV approach was performed with the GAM parameter. All models described above were compared with respect to their OFV, diagnostic plots, successful parameter estimation, and mechanistic plausibility.

Covariate analysis

A covariate analysis strategy was developed a priori to identify patient baseline characteristics that might explain sources of IIV on the model drug-related parameters of Kdeplete and Slope, as well as system-related parameters of BASE, BASE1, and MTT. Tested baseline covariates included patient age, weight, race, observed baseline platelet count, creatinine clearance, serum creatinine, albumin, total protein, ALT, AST, total bilirubin, tumor burden (based on the sum of the longest diameters of target lesions reported at baseline), HER2 expression, and prior myelosuppressive chemotherapies (i.e., paclitaxel, docetaxel, or carboplatin).

The covariate analysis was performed using the linearized FOCE method with stepwise covariate model building [14] to screen all available covariates for significance on any parameter. If significant covariate parameter correlations were found, they were assessed by stepwise addition, followed by backwards elimination using the nonlinear FOCE method.

Model evaluation

Model evaluation was performed using standard diagnostic plots, visual predictive checks (VPCs), external evaluation, and predictability of TCP. For VPCs, prediction intervals with 90 % confidence intervals were obtained by simulating 100 data sets from the model using the original data set. Simulations were constrained to 210 days and for the two largest-dose groups in the model-building dataset, 3.6 mg/kg q3w (n = 127) and 2.4 mg/kg qw (n = 16). The predicted versus observed platelet response were examined to determine whether observed 5th, 50th, and 95th percentiles of data fell within the confidence intervals around these percentiles (Fig. 2).
Fig. 2

Visual predictive check of the final model simulations at a T-DM1 dose of 3.6 mg/kg q3w with model dataset observations. The solid red line represents the median of the observed data (open circles). The red shaded region represents the 95 % confidence interval of the model simulated 50th percentile. The outer blue shaded regions represent the 95 % confidence interval around the model simulated 5th and 95th percentiles. The stippled red lines represent the 5th and 95th percentiles of the observed platelet observations

To assess model predictability of patient platelet response in subsequent studies, platelet observations from the phase II study TDM4374g (n = 110; 3.6 mg/kg q3w) were used as an external evaluation dataset. For simulations, results from the population PK analysis [5] were incorporated with results from this PK/PD analysis. Specifically, the model-simulated T-DM1 concentrations were based on PK parameter estimates, PK variability, the effect of body weight on CL and V1, and the correlation between CL and V1 as reported [5]. Platelet response simulations were, in turn, driven by these T-DM1 concentrations using the final PD parameters and IIVs shown in Table 1. Patient body weights were also simulated by the model using a median weight of 69.6 kg and an IIV of 25 %, as calculated from the model-building dataset, which contained 164 patients. This approach was taken in order to predict patient platelet response a priori and assumes a similar patient weight distribution in future studies. This was confirmed for patients in the evaluation dataset, TDM4374g, where patients had a median weight of 69.7 kg and an IIV of 25 %.
Table 1

Population parameter estimates for the final model


Parameter descriptions



RSE (%)

IIV (%)



T-DM1 drug effect for first dose







T-DM1 drug effect for subsequent doses







Mean transit time







Feedback term




Baseline platelet count at time = 0







Baseline platelet count not depleted







Baseline platelet count depleted by (Cavg × Kdeplete) rate




Depletion rate of BASE2 platelet pool for population 1 patients

L/mg × week−1






Depletion rate of BASE2 platelet pool for population 2 patients

L/mg × week−1






Probability of patient in POP1



Probability of patient in POP2



Res err

Residual error

18.4 %


Cavg average T-DM1 concentration (mg/L) over dosing intervals, IIV interindividual variability, RSE relative standard error, T-DM1 trastuzumab emtansine

aThe OMEGA SAME option was used for IIV on Kdeplete

One hundred simulation replicates were generated at a dose of 3.6 mg/kg q3w, with 110 simulated patients per replicate to match the patient number from TDM4374g. Nominal time points and dosing were used without patient dropout or dose reduction. Observed platelet count data from TDM4374g were then overlaid with model simulations to assess whether observed 5th, 50th, and 95th percentiles of data fell within the confidence intervals around these percentiles. The median line through the observed data was plotted and compared with the 50th percentile window. The median line from the model-building dataset observations was also plotted for comparison.

For a secondary evaluation, the PK/PD model was applied to this dataset with MAXEVAL = 0 in order to assign patients from TDM4374g to either the Kdeplete,POP1 or Kdeplete,POP2 subgroups; the MAXEVAL = 0 approach was used as a rapid method to stratify patients only for visualization purposes, and parameter values for the evaluation dataset were not re-estimated. Simulations of platelet response were also stratified by the Kdeplete,POP1 and Kdeplete,POP2 subgroups, and the 90 % prediction interval and 50th percentile were plotted with respective 95 % confidence intervals. The observed platelet count data from the Kdeplete,POP1 or Kdeplete,POP2 subgroups were then overlaid with model simulations to assess whether observed 5th, 50th, and 95th percentiles of data fell within the confidence intervals around these percentiles. The median line through the observed data was plotted and compared with the 50th percentile window. The median line from the model-building dataset observations was also plotted for comparison.

Finally, the ability of the model to predict the incidence of grade ≥3 TCP by day 63 was evaluated. Day 63 was chosen as the end point since most patients were still on study, and grade ≥3 TCP (when observed) was usually reached within three treatment cycles. The observed incidence of grade ≥3 TCP from the model dataset and evaluation dataset was overlaid with the simulated predictions. These probabilities were further separated by quartiles of observed baseline platelet count, T-DM1 Cmax (calculated by Dose/V1), and T-DM1 area under the curve (AUC) (calculated by Dose/CL). Simulations included patient dosing history from the model dataset.


Final population PK/PD model

The final population PK/PD model parameter estimates based on the model-building dataset (TDM3569g and TDM4258g) are shown in Table 1. There were no apparent trends in PK/PD model parameter estimates across doses (0.3–4.8 mg/kg) or between schedules (qw and q3w). The relative standard errors (RSE) were below 25 % for all parameters, indicating that parameters could be estimated with good precision. Values for system-related parameters GAM, MTT, and BASE were 0.135, 37.4 h, and 255 × 1,000/μL, respectively. BASE was composed of a 46 % BASE1 nondepletable proliferating PP (118 × 1,000/μL) and a 54 % BASE2 depletable proliferating PP (137 × 1,000/μL).

The implementation of two separate Slope parameters (for the initial dose and for subsequent doses) caused a drop of OFV by 325 points compared with a model with a single Slope parameter. The T-DM1 drug effect was greater after the initial T-DM1 dose (Slope1 = 0.00297 L/mg) versus subsequent doses (Slope2+ = 0.00182 L/mg), corresponding to a typical platelet count nadir of 115 × 1,000/μL in the first cycle and then a higher platelet nadir of 145 × 1,000/μL in the second cycle. Eighty-five percent of patients in the model-building dataset demonstrated this pattern of lowest nadir in the first cycle, with variable degrees of increased nadir counts in the second and subsequent cycles. This observation is illustrated in Fig. 3b for patient ID = 143 and in Fig. 4b.
Fig. 3

Representative model fits for data from patients treated with T-DM1 either weekly (a) or once every 3 weeks (b). Vertical hashes indicate dosing times. DV observed platelet count, IPRE model-predicted individual platelet count, PRED model-predicted population platelet count, qw once weekly, q3w once every 3 weeks
Fig. 4

Model-predicted 90 % window and observed platelet counts versus time for the overall patient platelet response and stratified by two patterns of platelet response; nominal time points and dosing were used for the simulations: a overall platelet response, b platelet count nadir followed by a return to baseline between treatment cycles, c downward drift in nadir and postnadir counts over time. Gray line indicates cutoff for grade 3 thrombocytopenia (50 × 1,000/μL)

The rates of the downward drift in platelet–time profiles were quantified by the Kdeplete parameter multiplied by the Cavg, and patients were assigned either to POP1 with a low Kdeplete or to POP2 with a higher Kdeplete by the mixture model. The probability of POP1 was estimated at 55 %, and the typical values for Kdeplete were 0.000625 L/mg × week−1 and 0.00842 L/mg × week−1 for the two subgroups. The actual percentages of patients in the model-building dataset assigned to POP1 and POP2 were 61 % and 39 %, respectively. For the typical patient in POP2 treated with 3.6 mg/kg T-DM1 q3w (Cavg = 17.1 mg/L), the rate for BASE2 depletion (Kdeplete × Cavg) is 0.144 week−1, corresponding to a time to reach steady state of 24 weeks. Therefore, the platelet–time profile is expected to stabilize after 24 weeks (eight treatment cycles). For the typical patient in POP1 treated with 3.6 mg/kg q3w, the rate for BASE2 depletion is 0.0107 week−1, predicting extremely slow declines of platelet–time profiles that stabilize after 324 weeks. This is longer than current T-DM1 treatment periods, and these platelet–time profiles are considered stable. Approximately 22 % of all patients had Kdeplete × Cavg values >0.144 week−1 (i.e., above the typical value for POP2); their platelet–time profiles drifted down faster and stabilized earlier than eight cycles. Representative patients, ID = 143 (POP1) and ID = 223 (POP2), from each subgroup are shown in Fig. 3b. After stabilizing, the model predicts that patients have a nondepletable PP (BASE1) and that platelet–time profiles will have less amplitude of platelet drop and return to baseline. This is also illustrated in Fig. 3b (patient ID = 223).

Covariate analysis

None of the available patient baseline covariates were found to significantly correlate with any PD parameter. Notably, because of eligibility criteria, patients in study TDM4258g had received numerous prior treatments, including known myelosuppressive agents. However, patients with prior paclitaxel treatment (55 %), docetaxel treatment (54 %), or carboplatin treatment (44 %) did not have lower predose platelet counts (↓ BASE), higher drug sensitivity (↑ Slope), or greater platelet–time profile declines (↑ Kdeplete) compared with those who had not received paclitaxel, docetaxel, or carboplatin.

Model diagnostics

The PK/PD model described all platelet profiles, which included multiple dose levels and regimens, and no indications of model misspecifications were evident from standard model diagnostic plots (data not shown). The VPC on the largest cohort of the modeling dataset (3.6 mg/kg q3w) indicated that the model described the data well (Fig. 2). A VPC was attempted for the 2.4 mg/kg qw (n = 15), but the low number of patients caused inflation of the variability and made the VPC noninformative. However, the median line through the observations was within the 50th percentile interval.

Representative model fits for individual patients are shown for qw regimens (Fig. 3a) and q3w regimens (Fig. 3b). Results of model simulations with the external evaluation dataset are shown in Fig. 4. The simulations, which used nominal dosing and time points, described the platelet response for all patients and patient subpopulations well. Figure 4a shows model simulations and patient data for all patients. Figure 4b shows model simulations and patient data of the Kdeplete,POP1 subgroup, that is, those with stable platelet–time profiles. Figure 4c shows model simulations and patient data of the Kdeplete,POP2 subgroup, that is, those with downward drifting platelet–time profiles. Patient platelet responses were similar between the model-building dataset (TDM3569g and TDM4258g) and the evaluation dataset (TDM4374g). Given this similarity, there is no need to update the model-building dataset with TDM4374g data.

Predicted incidence of grade ≥3 TCP

Figure 5a shows box plots of the model-predicted probability of grade ≥3 TCP overlaid with observed probabilities from the model-building and evaluation datasets. Box plots are shown for data from all patients (All), and by quartiles of patient baseline platelet count (lower 25, 25–75 %, and upper 75 %). The model predicted the ~8 % incidence rate of grade ≥3 TCP for this patient population, as well as the incidence rates, by quartiles, of baseline platelet count. A higher incidence of grade ≥3 TCP occurs when baseline platelet counts are ≤200 × 1,000/μL (i.e., in the lowest quartile of baseline platelet counts, compared with the entire patient dataset). This is illustrated by the correlation between observed baseline platelet counts and observed nadirs shown in Fig. 5b. The upper quartile of T-DM1 exposure metrics (i.e., Cmax and AUC) did not show such a profoundly increased incidence of grade ≥3 TCP compared with the entire patient dataset (data not shown).
Fig. 5

aBox plots are stratified by “All” baseline platelet counts and by quartiles of observed baseline platelet counts (lower 25 %, 25–75 %, and upper 75 %). b Correlation is shown between the observed baseline platelet count and platelet nadir at 3.6 mg/kg q3w. SD standard deviation, TCP thrombocytopenia


In most patients, T-DM1 q3w administration has been associated with a predictable cyclic pattern of platelet decline to nadir and return to baseline between doses. Platelet counts reach grade ≥3 TCP levels in only a minority of patients (7–8 %) [3, 4, 15] and have not been associated with serious hemorrhage. This cyclic pattern is similar to that of some conventional chemotherapeutic agents, which are often associated with myelosuppression [6]. In addition to this acute drop in platelet count to nadir by day 8, which appeared lowest during cycle 1, a slow downward drift in the platelet–time profile was noted in some patients receiving repeated doses of T-DM1. The mechanisms of patient platelet response to T-DM1 are currently unknown.

The semimechanistic PK/PD model proposed by Friberg et al. [6, 7] has been used extensively in drug development to quantify and describe drug-induced myelosuppression [812]. The term “semimechanistic” derives from compartments representative of bone marrow, blood cell maturation, and circulating blood cells; typically, the model incorporates a drug effect on the bone marrow progenitor compartment, mimicking chemotherapeutic mechanisms of toxicity. Using this model as a template, structural modifications were incorporated to test hypotheses of the potential mechanisms of T-DM1-mediated effects on platelets and to test patient baseline characteristics as covariates of PK/PD model parameters.

As an ADC, T-DM1 can be viewed as being composed of multiple components with regard to PK drivers of toxicity or efficacy: (1) the ADC, T-DM1; (2) the parent antibody, trastuzumab; and (3) the cytotoxic agent, DM1. For PK/PD modeling, T-DM1 was used as the PK driver of platelet response for several reasons. First, trastuzumab, as a single agent, has not been associated with myelosuppressive effects [16]. Second, although there is little clinical experience with DM1 administered as a free drug, maytansine, the parent drug of DM1 and of similar structure, showed no substantial myelosuppressive effects in phase I [17] and phase II [18] studies. Third, TCP has not been reported as a significant toxicity with other DM1-containing ADCs currently in clinical development [1921]. Preclinical investigations are ongoing to characterize the effect of T-DM1 treatment on platelets and to confirm the T-DM1 moiety responsible.

The data on platelet concentrations, obtained from multiple dose levels and schedules of T-DM1, provided variable T-DM1 concentration–time and platelet profiles. These data, together with the proposed model, support a linear effect of T-DM1 on the (nonobserved) platelet proliferating pool in line with the observed delayed nadir on day 8 relative to drug administration. Other hypotheses of the relationship between T-DM1 concentrations and platelet response, including a direct drug effect on the (observed) PLT compartment, were also tested but were not well supported by the modeling results. Incorporating two separate Slope parameters best captured the observation that cycle 1 platelet nadirs are lower than in subsequent cycles. This is an empirical modification of the model that was justified by a highly significant improvement in the model fit, which was confirmed with the evaluation data set; the underlying mechanism is currently unknown.

The downward drift in platelet–time profiles over time, readily observable in some patients, was modeled using an additional T-DM1-related effect. Although the true mechanism or moiety involved is unclear, the incorporation of a T-DM1-related effect on the platelet baseline (BASE2) is supported by the data. This phenomenon is not without precedent, since patterns of the decline of platelet and neutrophil counts over time have previously been described as cumulative myelosuppression for cytotoxic drugs. In patients with advanced breast cancer who received FLAC (5-fluorouracil, leucovorin, doxorubicin, and cyclophosphamide) [22], cumulative TCP was the DLT. In work by Maze et al. [23], the cumulative TCP with nitrosoureas noted clinically was reproduced in a murine model and a reduced number of primitive hematopoietic proliferation cells was reported.

To model this downward drift in platelet–time profiles, an additional mechanistic assumption was hypothesized and incorporated into the structure of the model proposed by Friberg et al. [6, 7]. The PP compartment was modeled as two fractions, hypothesized as sensitive and nonsensitive fractions of the platelet proliferation pool. In this model, the sensitive lineage is affected by drug exposure and is depleted with a time-dependent rate of decline. This long-term drug effect on the baseline platelet count is modeled on top of the acute drug effect that is hypothesized to cause the oscillation in platelet count. Due to the slower nature of the platelet profile decline, the average T-DM1 concentration over dose intervals was used as a measure of drug exposure rather than the T-DM1 concentration–time profile, to drive this effect. However, the true PK driver is unknown.

The structural model modifications captured not only the rate of platelet profile decline and the new apparent steady-state platelet baseline but also the reduction in the amplitude of platelet decline and rebound, further supporting this model structure. The mixture model implementation for the parameter Kdeplete captured the apparent bimodal distribution of the observed platelet–time profile decline, with an extremely slow decline in about 55 % of patients (POP1) and a notable, highly variable decline in about 45 % of patients (POP2), which typically stabilizes by week 24 (eight treatment cycle). Overall, 22 % of patients are predicted to have a more rapid platelet profile decline, which stabilizes sooner than eight treatment cycles and which is readily observed without modeling so long as there is adequate platelet sampling. Although it was not possible to correlate the two patient subgroups with any of the investigated covariates, the mixture model implementation allowed us to adequately quantify the rate and extent of decline as well as the proportion of patients likely to experience such a decline.

Given these modeling results, and evidence in the literature for this phenomenon [22, 23], it is plausible that T-DM1 affects a platelet proliferation lineage or cofactor; when this lineage is depleted, the new platelet baseline is derived from a lineage that is less sensitive to T-DM1. The model structure predicts that patients’ PPs are not entirely depleted, and, as shown in Fig. 4c, the nondepletable PP is enough to maintain platelet count nadirs >80 × 1,000/μL, above grade 3 TCP (50 × 1,000/μL) [24]. It should be noted that this model is currently based on data from patients receiving continuous T-DM1 treatment. It is unknown how discontinuation of T-DM1 administration affects platelet response(s).

To assess whether a patient’s platelet response to T-DM1 (i.e., magnitude of platelet count drop to nadir or downward drift in platelet–time profile) could be identified a priori, a covariate analysis was performed using the final model. Ultimately, no baseline demographic or pathophysiologic characteristic was identified as a covariate relating to platelet response, though some hypotheses were ruled out. These included (1) baseline hepatic transaminase (ALT/AST) levels on Kdeplete, with which patients with compromised liver function may have had reduced thrombopoietin or other cofactor(s) that would have affected platelet regulation; (2) BASE on Slope, in which a larger pool of baseline platelets may have correlated with decreased drug effect; and (3) prior chemotherapies on Slope or Kdeplete, in which myelosuppressive agents (e.g., paclitaxel, docetaxel, and carboplatin) taken before T-DM1 treatment may have exacerbated the platelet response to T-DM1.

Although assessed covariates did not contribute to the understanding of the mechanisms underlying platelet responses, model predictions did provide useful information regarding patient propensity for grade ≥3 TCP. Prior to this modeling analysis, there was concern that patients with greater T-DM1 exposure at studied doses might have an increased likelihood of developing grade ≥3 TCP. However, platelet observations and model simulations at 3.6 mg/kg q3w showed that patients with the greatest (upper quartile) T-DM1 exposure (T-DM1 Cmax or T-DM1 AUC) did not have a markedly increased incidence of grade ≥3 TCP compared with the entire population. Based on the current dataset, only baseline platelet counts ≤200 × 1,000/μL were associated with an increased risk of grade ≥3 TCP, suggesting that these patients should be carefully monitored.

The model also suggests that more frequent T-DM1 dosing can modulate toxicity. As shown in Fig. 3b (ID = 143) and Fig. 3a (ID = 152), the population predicted platelet counts (PRED) returned to baseline prior to the next T-DM1 dose with the 3.6 mg/kg q3w regimen, while the dose frequency of the 2.4 mg/kg qw regimen allowed only a partial return to baseline, though within the normal range. For each of these cohorts, platelet nadirs were approximately 150 × 1,000/μL, yet the 2.4 mg/kg qw regimen resulted in roughly twice as much overall T-DM1 exposure. Given the potential for exposure-driven T-DM1 antitumor activity, a weekly T-DM1 treatment regimen may be similarly beneficial to patients, without any clinically significant exacerbations in TCP.

The population PK/PD model reported here describes the clinical platelet response to T-DM1 well and can potentially be used to predict the incidence of TCP, thereby optimizing the safety of T-DM1. Utilizing structural model modifications to the standard PK/PD myelosuppression model [6, 7], our model suggests that (1) TCP is less pronounced after the first T-DM1 dose, (2) platelet–time profiles that drift slowly downward over time will eventually stabilize above grade 3 TCP, and (3) an individual patient’s platelet response to T-DM1 cannot be predicted a priori from any baseline characteristics tested. We conclude that T-DM1 3.6 mg/kg q3w is safe in patients with HER2-positive breast cancer and necessitates minimal dose delays and reductions due to clinically significant TCP. Planned and ongoing clinical studies are evaluating T-DM1 in combination with various chemotherapeutic agents, some of which demonstrate thrombocytopenic effects; the application of this T-DM1 PK/PD model in the combination therapy setting is currently under investigation.


The study was funded by Genentech, Inc. Support for third-party writing assistance was provided by Genentech, Inc.

Copyright information

© Springer-Verlag 2012