Translational pharmacokinetic–pharmacodynamic modelling; application to cardiovascular safety data for PF-00821385, a novel HIV agent

Abstract

AIM

To assess the translation of pharmacokinetic–pharmacodynamic (PK–PD) relationships for heart rate effects of PF-00821385 in dog and man.

METHODS

Cardiovascular telemetric parameters and concentration data were available for animals receiving active doses (0.5–120 mg kg⁻¹, n= 4) or vehicle. PF-00821385 was administered to 24 volunteers and pharmacokinetic and vital signs data were collected. PK–PD models were fitted using nonlinear mixed effects.

RESULTS

Compartmental models with linear absorption and clearance were used to describe pharmacokinetic disposition in animal and man. Diurnal variation in heart and pulse rate was best described with a single cosine function in both dog and man. Canine and human heart rate change were described by a linear model with free drug slope 1.76 bpm µM⁻¹[95% confidence interval (CI) 1.17, 2.35] in the dog and 0.76 bpm µM⁻¹ (95% CI 0.54, 1.14) in man.

CONCLUSIONS

The preclinical translational of concentration–response has been described and the potential for further interspecies extrapolation and optimization of clinical trial design is addressed.

WHAT IS ALREADY KNOWN ABOUT THIS SUBJECT

• While animal toxicology is routinely used for safety screening of compounds prior to their being tested in man, there is very little in the literature on the quantitative translation of cardiovascular drug effects from animal to man.

WHAT THIS STUDY ADDS

• This study compares pharmacokinetic–pharmacodynamic analysis of dog and human effects of PF-00821385 on heart rate.

• This paper considers how quantitative translational knowledge of drug effects in dogs may be used to optimize future human studies.

Introduction

Human immunodeficiency virus (HIV)-1 particles are decorated with a network of densely arranged envelope spikes formed of a trimer of heterodimers of the gp120 surface and the gp41 transmembrane glycoproteins. These molecules mediate HIV-1 entry into target cells, initiating the HIV-1 replication cycle and serve as targets for potential anti-HIV-1 therapies [1]. 5-{(1S)-2-[(2R)-4-Benzoyl-2-methyl-piperazin-1-yl]-1-methyl-2-oxo-ethoxy}-4-methoxy-pyridine-2-carboxylic acid methylamide (PF-00821385, molecular weight = 440.49 Da) was developed as a specific inhibitor of HIV-1 gp120-mediated cell–cell fusion for the treatment of HIV-1 infection.

In preclinical in vitro screens PF-00821385 had little or no interaction with physiologically important receptors, binding sites, enzymes, or ion channels. Weak functional activity was observed at the human κ-opioid receptor where the IC₅₀ for binding at the receptor was 29 µM (K_i= 9.6 µM) and the functional agonist EC₅₀ in a human recombinant κ-opioid assay was 58 ± 5.4 µM. PF-00821385 had no effect on the binding of [³H]dofetilide to the human ether-a-go-go related gene (hERG) potassium channel up to 10 µM. In isolated Purkinje fibres from dogs, PF-00821385 (100 µM) reduced action potential amplitude, but the magnitude of these effects was small.

The first oral single dose-escalating study in dogs showed increased heart rate (HR; ∼110 bpm change from baseline) and reduced systolic blood pressure (SBP) at a dose of 20 mg kg⁻¹ (unbound C_max= 56 µM) with similar changes observed at 40, 60, 80, 100 and 120 mg kg⁻¹. Ventricular tachycardia was observed at doses ≥60 mg kg⁻¹ and these events were associated with myocardial necrosis and coronary vascular inflammation. Increased HR, ventricular tachycardia and dose-related decreases in SBP were also observed in a subsequent 7-day multiple dose study with peak unbound concentrations on day 7 of 55 (30 mg kg⁻¹), 99 (50 mg kg⁻¹) and 201 (90 mg kg⁻¹) µM. Post mortem observations indicated subacute coronary arteriopathy from 30 mg kg⁻¹, and multifocal subacute myocardial necrosis/fibrosis from 50 mg kg⁻¹. The 14-day toxicology study examined active doses of 5, 15 and 50 mg kg⁻¹. Considerable increases in HR occurred at all dose levels with a concurrent reduction in SBP. Histopathological findings were similar to the 7-day study at 50 mg kg⁻¹ and the no adverse effect level was 15 mg kg⁻¹ (mean unbound day 14 C_max= 28.5 µM). Three more studies were conducted in anaesthetized, restrained and ambulatory dogs to assess further the safety pharmacology of PF-00821385 and elucidate the potential mechanism of the coronary arterial lesions observed in the toxicology studies. Both pharmacokinetic (PK) and pharmacodynamic (PD) data were collected and the analyses of these data are discussed further in this study.

Based on the results from these preclinical studies, a cautious first-in-human (FIH) study design with appropriate safety margins was adopted (see Methods). Following the conclusion of the study a retrospective analysis was performed to assess how PK–PD models of preclinical cardiovascular safety may have predicted the cardiovascular changes in man and have been used to prospectively optimize early clinical study designs, as outlined by Danhof [2].

This paper describes the PK–PD models developed for the preclinical and clinical datasets and the lessons learned regarding the potential for translation of concentration–response between dog and man.

Methods

Ambulatory canine pharmacodynamic experimental methods and data collection

PF-00821385 was synthesised in Pfizer Laboratories. It was dissolved in 0.5% (w/v) methylcellulose +0.1% (v/v) Tween 80 in purified water prior to administration. Vehicle and PF-00821385 at 1.5, 5 and 15 mg kg⁻¹ were administered as a single dose by gavage, on separate occasions, to four conscious, freely moving Beagle dogs (Pfizer laboratory colony) of similar weight (range 16.3–16.4 kg). Each animal was implanted with pre-calibrated miniature pressure transducers (Konigsberg Instruments Inc., Pasadena, CA, USA) in the thoracic aorta and left ventricle. The electrocardiogram was recorded between the transducer tip in the aorta and the electrode, which was sutured subcutaneously. Cardiovascular parameters were measured using remote telemetry. For each treatment in each animal the 1 min mean data produced by the data acquisition system was transferred to Excel (Microsoft, Reading, UK). In order to keep the datasets to a manageable size, data were pre-processed by averaging into 15-min time bins, yielding 96 data points per experiment per animal per dose. Periods of feeding were excluded from the analysis as this resulted in large spikes in HR.

Canine pharmacokinetic experimental methods and data collection

Subsequent to the PD experiments described above, each animal (n= 4) was dosed with 5 mg kg⁻¹ of compound and venous blood samples were taken at 0.5, 1, 1.5, 2, 4, 6 and 24 h post dose. Additional satellite PK experiments were carried out in eight Beagle dogs (Pfizer laboratory colony). PF-00821385 was administered orally by gavage at 0.5 (n= 2), 20 (n= 1), 40 (n= 1), 100 (n= 2) and 120 mg kg⁻¹ (n= 2). In addition, two animals in this group were also subsequently given 0.5 mg kg⁻¹ intravenously. Venous blood samples were taken at 0.5, 1, 1.5, 2, 4, 6 and 24 h post dose. Compound concentration in plasma was estimated as follows; extraction was achieved via Water-tert-butyl methyl ether liquid–liquid extraction of 250 µl of dog plasma. Extracted analytes were then separated from endogenous material by high-performance liquid chromatography using a 50 × 4.6 mm i.d. Chromolith SpeedROD RP-18e column. (Merck KGaA, Darmstadt, Germany) and detected using a Sciex API4000 mass spectrometer (Perkin-Elmer Sciex, Foster City, CA, USA). The limit of quantification was 1 ng ml⁻¹.

Canine pharmacokinetic–pharmacodynamic analysis

Data were analysed using a nonlinear mixed-effects modelling approach with the NONMEM software system, version VI (Globomax, Hanover, MD, USA) [3]. Diagnostic plots were produced using S-Plus (Insightful, Seattle, WA. USA) and the Xpose library [4].

Following administration, absorption was rapid and serum concentrations declined in an exponential fashion. Nontransformed concentrations were modelled as a one-compartment disposition model with first-order absorption.

The HR in animals receiving placebo varied over time, with a sustained peak in the evening, approximately 10 h after morning dosing (09.00 h) (Figure 1). This variation in heart rate was modelled using a single cosine function with a period of 24 h (Equation 1).

Figure 1.

A visual predictive check (VPC) of the canine pharmacokinetic–pharmacodynamic model's ability to predict the heart rate data. Open circles are the observed data points; the solid grey line represents the 50th quantile of the simulated data, while the short broken black lines represent the 95% prediction interval obtained from the simulations

(1)

where HR_{placebo, ij} is the heart rate (in beats min⁻¹) in the i^th individual at a given time t_j in the absence of drug, BASE_i is the baseline HR, AMP is the amplitude of the circadian phase and PEAK_i is the timing of the peak relative to the vehicle administration.

The relationship between the drug concentrations in the plasma compartment and HR was modelled using a linear (Equation 2) or an E_max model with or without a delay;

(2)

where HR_ij denotes the predicted heart rate for the i^th individual at time t_j, SLOPE denotes the individual rate of change in pulse rate and F_ij denotes the corresponding predicted concentration based on the PK model. PK and PD parameters were estimated sequentially with individual Bayesian post hoc estimates from the PK model serving as input for the PD model with the assumption that the individual PK and PD parameters specified in the models were constant over time.

The interindividual variability (IIV) of PK and PD parameters was modelled using multiplicative exponential random effects. Residual variability was described either using a combined proportional and additive error model for PK data or an additive error model for the PD data. Interoccasion variability was not included in the PD model.

First-in-human study design and data collection

This was a randomized, double-blind, placebo-controlled study in two cohorts of 12 healthy volunteers. Cohort 1 received three of four single escalating oral doses of PF-00821385 (3, 10, 30 or 100 mg) and one administration of placebo, while fasted. Cohort 2 received three of four single escalating doses of PF-00821385 (250, 500, 1000 or 1300 mg) and one administration of placebo, while fasted. The wash-out period between doses was 7 days. All volunteers in Cohort 2 then received a repeat of a previous PF-00821385 dose (500 mg) in the fed state. Volunteers were healthy men aged between 21 and 55 years with a body mass index of approximately 18–30 kg m^–2 and a body weight >50 kg. PF-00821385 was administered as an oral suspension and PK samples were collected predose, and at 0.5, 1, 1.5, 2, 3, 4, 6, 8, 12, 16, 24 and 36 h post dose. Samples were analysed using the same validated analytical method as described above for the analysis of plasma in the dog. Supine pulse rate was recorded using an automated device predose, and at 1, 2, 4, 6, 8, 12, 24 and 48 h post dose. Based on findings from the regulatory repeat-dose toxicology studies, stopping criteria were employed such that dose escalation would be stopped or limited if: (i) mean peak total plasma concentrations exceeded 28400 ng ml⁻¹, (ii) more than two subjects experienced a change in standing SBP of >30 mmHg from baseline, or (iii) more than two subjects experienced a change in supine pulse rate of >25 bpm from baseline.

Written informed consent was obtained from each volunteer. The study was approved by the Ethics Committee of Singapore General Hospital and by the Health and Sciences Authority of Singapore, and was conducted in accordance with the protocol, International Conference on Harmonization Good Clinical Practice guidelines, and applicable local regulatory requirements and laws.

Human pharmacokinetic–pharmacodynamic analysis

A two-compartment model with first-order absorption was fit to the PF-00821385 plasma concentrations and FOCE INTER was used to estimate all parameters. IIV in the pharmacokinetic parameters was modelled using multiplicative exponential random effects and residual variability was modelled as an additive error on the log-transformed concentrations. A nonparametric bootstrap of 500 iterations was used to provide estimates of the standard errors and the 95% confidence intervals (CI) of the estimated parameters [5, 6]. A visual predictive check was employed to characterize the model's simulation properties.

The supine pulse rate in volunteers receiving placebo varied over time with a sustained peak in the evening, approximately 12 h after morning dosing (∼08.00 h) (Figure 3). As with the dog data, this variation in HR was modelled using a single cosine function with a period of 24 h (Equation 1). The rate of change in pulse rate was described as a linear function of concentration. Individual Bayesian post hoc estimates served as a sequential input for the PD model.

Figure 3.

A visual predictive check (VPC) of the pharmacokinetic–pharmacodynamic model's ability to predict the pulse rate data. Open circles are the observed data points; the solid grey line represents the 50th quantile of the simulated data, while the short broken black lines represent the 95% prediction interval obtained from the simulations. The solid black line represents the 50th quantile of the simulated human data using the slope of drug effect from the canine model

IIV in the pharmacodynamic parameters was modelled using multiplicative exponential random effects and residual variability was modelled as an additive error. Inter-occasion variability was not included in the model.

Results

Canine pharmacokinetic–pharmacodynamic analysis

A total of 88 plasma concentrations were available from 16 dogs following oral administration at varying dose levels and 22 plasma concentrations from two dogs following intravenous administration. All data were used to construct a population PK model, which allowed for a more robust estimation of PK parameters as well as exploration of possible exposure-driven nonlinearities. A one-compartment disposition model described the data well with model predictions following the shape of the observed data. IIV was included on the following structural parameters: absorption rate constant (k_a), volume of distribution of the central compartment (V/F), and clearance from the central compartment (CL/F). Diagnostic plots showed no particular bias or trends and no difference in the concentration–time profiles of the animals from the PD experiment and those from the satellite PK experiment.

Baseline HR was measured prior to dosing of drug or vehicle in all cases and the canine population estimate for HR was 68 bpm with a variability of approximately 7.5%. Inspection of the raw PD data indicated that in all animals and following all treatments, a time-dependent change in the HR was observed. The HR started low, tended to increase toward a maximum at about 10 h and then decrease back to the initial level over the following 12 h or so. The placebo data were modelled using a single cosine function with an estimated peak time of 10.4 h post morning dose. A PK–PD model incorporating a time-dependent circadian effect and a direct drug effect without delay was found to describe the data adequately. IIV was included on the following structural parameters: baseline HR and amplitude of the circadian effect. Parameter estimates are presented in Table 1 and the final canine PK–PD model predicted that the typical drug effect was 1.76 bpm µM⁻¹ (95% CI 1.17, 2.35) free drug.

Table 1.

Final parameter estimates, standard errors and 95% confidence intervals for the canine pharmacokinetic–pharmacodynamic (PK–PD) model estimated by NONMEM

	NONMEM final estimate
Population parameter	Mean	SE	95% confidence interval
F	1.01	0.003	1.00, 1.02
CL/F (l h−1)	1.02	0.135	0.755, 1.28
V/F (l)	7.05	0.640	5.80, 8.30
ka (h−1)	2.42	0.520	1.40, 3.44
PK additive residual error (ng ml−1)	9.43	3.53	2.51, 16.3
PK proportional residual error	0.181	0.022	0.138, 0.224
Variance for CL/F	0.311	0.102	0.111, 0.511
Variance for V/F	0.136	0.068	0.003, 0.269
Variance for ka	0.404	0.141	0.128, 0.680
BASE (bpm)	68.2	1.01	66.2, 70.2
AMP (bpm)	12.4	0.706	11.01, 13.8
PEAK (h)	10.4	0.246	9.91, 10.88
SLOPE (bpm µM−1)	1.76	0.29	1.17, 2.35
PD additive residual error (bpm)	14.1	4.17	5.93, 22.3
Variance for BASE	0.0057	0.0017	0.0024, 0.0090
Variance for PEAK	0.0134	0.0046	0.0044, 0.0224

Human pharmacokinetic–pharmacodynamic analysis

A total of 969 plasma concentrations were available from 24 healthy male volunteers following oral administration. Unchanged PF-00821385 was the major circulating component detected in the human plasma extracts. A two-compartment disposition model described the data well and IIV was included on the following structural parameters: relative bioavailability (F), V/F, CL/F. Diagnostic plots showed no particular bias or trends and, together with the results of the visual predictive check (Figure 2), qualified the use of the individual post hoc parameter estimates from the model as a suitable input for PK–PD modelling.

Figure 2.

A visual predictive check (VPC) of the human pharmacokinetic model's ability to predict the concentration data. Open circles are the observed data points; the solid grey line represents the 50th quantile of the simulated data, while the short broken black lines represent the 95% prediction interval obtained from the simulations

Haemodynamic parameters are known to show diurnal variation and this was captured using a single cosine function with an estimated peak time of 14.3 h for pulse rate in man (similar to the dog). This model adequately described the rate of change in pulse rate in the absence of drug (Figure 3). Baseline pulse rate, defined as the pulse rate measured in the morning prior to drug administration, in this population was estimated to be 61 bpm with IIV approximately 10%, which is consistent with previous healthy volunteer studies (data on file). Drawing on the results of the preclinical toxicology studies, a cautious study design was employed with stopping criteria based on exposure and relative changes in supine pulse and standing SBP. As a result, the dataset was truncated, but did allow the fitting of a PK–PD model incorporating a time-dependent circadian effect, and a direct linear drug effect. IIV was included on the following structural parameters: baseline HR and peak time of the circadian effect. Final parameters are presented in Table 2 and the predictive performance of the model is presented in Figure 3. This figure shows that the change in supine pulse rate is slightly under-predicted at the highest dose. This could, in part, be a consequence of data truncation with the majority of data lying in the area of low plasma concentrations and minimal changes in pulse. This in turn limited our ability to fit an E_max type model and the rate of change in pulse rate appears to be steeper than is captured by the simple linear model. The final human PK–PD model predicts that the drug effect is 0.76 bpm µM⁻¹ (95% CI 0.54, 1.14) free drug.

Table 2.

Final parameter estimates, standard errors and 95% confidence intervals for the human pharmacokinetic–pharmacodynamic (PK–PD) model estimated by NONMEM and via bootstrap resampling

	NONMEM final estimates		Bootstrap resampling
Population parameters	Mean	SE	Median	95% confidence interval
F	1 (FIXED)	–	–	–
ka (h−1)	0.599	0.0101	0.599	0.582, 0.616
CL/F (l h−1)	36.7	2.55	36.8	32.8, 40.9
Vc/F (l)	18.4	3.08	18.4	14.2, 24.3
Vp/F (l)	6.88	1.03	6.92	5.51, 8.89
Q (l h−1)	0.704	0.117	0.705	0.546, 0.945
PK residual error (ng ml−1)	0.658	0.0615	0.650	0.546, 0.759
Variance for CL	0.039	0.028	0.034	8.73 × 10–9, 9.35 × 10–4
Variance for Vc	0.332	0.116	0.314	0.164, 0.505
Variance for F	0.075	0.030	0.073	0.017, 0.123
BASE (bpm)	61.5	1.36	61.6	59.3, 63.6
AMP (bpm)	4.01	0.416	4.00	3.33, 4.82
PEAK (h)	14.3	0.55	14.3	13.3, 15.2
SLOPE (bpm µM−1)	0.76	0.16	0.76	0.54, 1.14
PD additive residual error (bpm)	4.77	0.309	4.75	4.28, 5.23
Variance for BASE	0.0094	0.0026	0.0087	0.0050, 0.0133
Variance for PEAK	0.0206	0.0055	0.0194	0.0099, 0.0299

Discussion

On completion of the single and multiple dose (7 and 14 day) toxicology studies in dogs it was clear that PF-00821385 carried a potential cardiovascular risk with increasing exposures. While not common, it is not unusual to progress investigational compounds with this profile. Prior to progression, however, additional telemetry studies were conducted to elucidate and understand the underlying mechanism in order to project potential clinical risks for PF-00821385. However, these projections lacked a quantitative framework and predictive capacity, which a retrospective analysis of the data aimed to address. The purpose of the work described was therefore to use PK–PD modelling approaches to facilitate the comparison of concentration–response between the dog and human volunteer studies and to evaluate these results in the context of the available translational pharmacology literature. Ultimately, such analyses can be used to develop better methods for more efficient and safe clinical trials of cardiovascular safety and improved consensus in the processes used [7].

From the PK–PD modelling we concluded that a direct linear concentration–response change in HR/pulse was observed in both the dog and in man for this compound over the dose ranges studied (Figure 4). This facilitated a comparison and led to the conclusion that the response was very similar between the species, relative to the variability. To our knowledge, this is the first such report describing the comparison of HR PK–PD for the same drug in man and dog. PK–PD approaches have been applied to understand the extent of concordance of the impact of drugs on cardiovascular parameters within subpopulations of the same species [8–10]. However, reports of interspecies correlation of cardiovascular PK–PD effects are currently limited and the relationship of exposure–response across species for any given cardiovascular measurement is not well understood [11]. Indeed, the area of cross-species scaling of pharmacology in general is still in its infancy, although some recent examples have suggested that allometric scaling may be applicable to predict not only PK but also PD responses in humans from data obtained in preclinical models [12, 13]. An interesting open question is therefore the generality of our observations. At present, we do not know the mechanism underlying the cardiovascular effects of PF-00821385. The compound was found to have in vitroκ-opioid agonist effect with a functional EC₅₀ of 58 µM. With the caveat that censoring prevented determination of saturation of effect in animals and man, the observed PK–PD relationships showed linear changes in this concentration range. In addition, κ-opioid agonists have been reported to elicit increases in HR in the dog [14] and hence this pharmacology may explain the effects we have observed. However, similar changes have not been reported in man with other κ-opioid agonists, where the predominant effect reported was diuresis [15, 16]. Other gp120 antagonists such as FP-21399 [17] and BMS 378806 [18] have been evaluated in man and rat, respectively, but there are no reports of HR effects and therefore the HR increases noted would not appear to be a class effect.

Figure 4.

Observed heart/pulse rate in dogs (open inverted triangle) and humans (closed square) with increasing free concentration. The lines compare the predicted mean rate of change for dogs (short broken line) and humans (solid line)

A further relevant consideration related to the generality of cross-species PK–PD concordance is that of the optimal species, and indeed baseline cardiovascular parameters, for predicting clinical outcome. According to the principles of allometry, resting HR for mammals scales as Y = a.M^b, where Y = heart rate (s⁻¹), M = body mass (kg), a = intercept when M = 1 (= 4.02 s⁻¹) and b is the exponent =–0.25 [19]. Based on this relationship, the resting HR for a typical 15-kg dog would be predicted to be 20–30 bpm higher than for a typical 70-kg man. In this study we found that the resting HRs were rather similar (68 for the dog and 61.5 for man). Although interpretation of data such as resting HR measures is complicated by biological variability, experimental design and measurement technique, a survey of the available literature suggested the following average resting dog HRs: 110 bpm [20], 98 bpm [21], and 85 bpm [22]. Thus, it would appear the unique inbred Pfizer animal colony have atypically low resting HRs. Such variability in resting HR as a function of strain has been reported in other species such as the mouse [23]. Potentially, this could influence the extent of correlation between the PK–PD relationship for PF-00821385 in dog and man, and concordance in resting HR between species may increase the probability of observing concordance in concentration–response to a drug. The allometric relationship for maximum HR is as for resting, but with Y = 6.5 s⁻¹ and b =–0.15 [19]. Hence, the difference between maximum and resting HR can calculated (Figure 5). It can be seen that up to ∼200 bpm resting HR, higher resting HR would lead to an increased difference between resting and maximum HR and therefore the potential for an apparently more sensitive response to drug in terms of absolute bpm. However, the magnitude of the difference between human and dog is relatively small (c. 38 bpm) and the influence of this difference on sensitivity to drugs across man, dog and rat is therefore also likely to be small. Hence, similarity of pharmacological response between the species would appear to be more critical to the translational pharmacology than the initial resting HR.

Figure 5.

The difference in maximum and resting heart rate as a function of resting heart rate. H, D and R represent resting heart rates for human, dog and rat, respectively, as predicted by the allometric equations

An understanding of the concentration–response relationship also helps hypothesis building as to the origin of the effect. This is essential information if it is necessary to improve the therapeutic index of a candidate drug over any cardiovascular effect. Moreover, the understanding of concentration–response in both species should be a strong platform from which to explore the wider question of which parameter, and indeed species, is optimal for projection of effect to man, e.g. as has been reported for the prediction of the relationship between in vitro binding to hERG and QT prolongation in man [24]. To achieve this, further examples of the comparison of concentration–response across species will be required.

In the context of the design of clinical trials we see a number of advantages to this ‘model-based approach’ (for a more detailed discussion of this see [25]) and how the information available at the clinical transition point may aid in the design of the first-in-man study. This may include the determination of the initial dose, a dose escalation plan, optimization of PD sampling points and appropriate stopping rules. While a number of factors play a role in determining the starting dose, using simulation to predict a pharmacological effect on pulse rate relies on one key assumption: identical PK–PD behaviour for man and dog. However, uncertainty around this assumption may limit the utility of simulation in initial dose selection. Nevertheless, this work presents a general approach that could be pursued. For example, as a study progressed and the amount of PK data available in man increased, the PK model could have been updated and the PK–PD model used to simulate outcome, under the assumption that man and dog are equally sensitive to HR changes (i.e. slope of PD effect is the same). This is demonstrated in Figure 3 (solid black line), where the PK–PD model in man has been used to simulate changes in pulse rate using the slope from the dog. Based on this, a less conservative dose escalation could, in retrospect, have been employed; 3, 10, 50, 250, 750 and 1300 mg, for example, saving two dosing periods. Furthermore, individual prediction of the number of subjects predicted to exceed the stopping criteria for HR could further improve decision making with regard to the magnitude of dose escalation. More generally, we envisage that the preclinical PK–PD could be used in this adaptive way. Provided the PD response is measured, the preclinical PK–PD model can be updated as human data are generated and the clinical design adjusted accordingly. This should realise benefits in terms of efficiency and subject safety, as exemplified by the work described here.

PF-00821385 was safe and well tolerated in the FIH study. The most commonly reported treatment-related adverse events were headache (four subjects) and diarrhoea (three subjects). There were single incidences of treatment-related vomiting, dizziness and hypotension. All treatment-related adverse events were mild except for hypotension, which was moderate. All treatment-related adverse events were resolved. A PK–PD disease model previously developed to describe the viral load–time profile in HIV patients [26] predicted that the dose required to result in a 1.5 log₁₀ decrease in viral load following 10 days of dosing was close to, or above the maximum tolerated dose (1300 mg) [27]. Based on the outcome of the safety–exposure and efficacy–exposure analyses, further development of PF-00821385 will not be pursued.

Competing interests

All authors are currently Pfizer employees and declare no further conflict of interest.

We would like to thank David H. Williams for compound synthesis. The statistical input from Gary Layton into the design of the clinical study is also recognized. We would also like to thank Mick Sutton for his expert advice.