Tutorials › Biostatistics › PK/PD Modeling for Dose Selection

Home  /  Tutorials  /  Pharmacometric Modeling

PK/PD & Pharmacometric Modeling

PK/PD Modeling for Dose Selection

A practical guide to connecting dose, pharmacokinetic exposure, and pharmacodynamic response to select doses for clinical development, including PK models, exposure–response relationships, Emax models, indirect response models, simulation, target attainment, and dose selection.

Advanced 20 min read

What You'll Learn

  • How PK models connect dose to drug exposure
  • How PD models connect exposure to pharmacologic response
  • How exposure–response models support dose selection
  • How Emax and other nonlinear PD models describe dose–response relationships
  • How simulation can compare candidate doses and quantify uncertainty
  • How PK/PD modeling can translate exposure targets into a rational dose recommendation

Introduction

Selecting a clinical dose is one of the most important decisions in drug development.

The challenge is that the administered dose is usually not the quantity that directly produces the pharmacologic effect.

Instead, the sequence is typically:

1
Dose: the amount of drug administered to the patient.
2
PK: the dose produces a time-varying drug concentration or exposure.
3
PD: exposure produces a pharmacodynamic effect.
4
Clinical response: pharmacologic effects contribute to the observed biomarker or clinical endpoint.

PK/PD modeling provides a quantitative framework for connecting these steps.

Key idea: PK/PD modeling allows investigators to move from “What dose was given?” to “What exposure did that dose produce, what response did that exposure produce, and what dose is most likely to achieve the desired response?”

The Dose–Exposure–Response Relationship

The fundamental structure of a PK/PD model can be represented as:

$$ \text{Dose} \longrightarrow \text{Exposure} \longrightarrow \text{Pharmacodynamic Effect} $$

More formally, dose determines the PK trajectory, and the PK trajectory drives the PD model.

$$ D \longrightarrow C(t) \longrightarrow E(t) $$

where:

  • \(D\) is dose.
  • \(C(t)\) is drug concentration over time.
  • \(E(t)\) is pharmacodynamic effect over time.

The ultimate objective is not necessarily to identify the dose that produces the largest concentration.

Instead, the objective is usually to identify the dose that provides an appropriate balance of:

  • Pharmacologic efficacy
  • Probability of achieving the desired exposure
  • Safety and tolerability
  • Interpatient variability
  • Dosing convenience
  • Uncertainty in the model

Why Dose Alone Is Often a Poor PD Predictor

Two patients receiving the same dose can have substantially different concentrations.

For example, differences in:

  • Body size
  • Clearance
  • Bioavailability
  • Renal function
  • Hepatic function
  • Drug interactions
  • Genetic factors
  • Concomitant medications

can produce different exposures.

Consequently, the relationship:

$$ \text{Dose}\rightarrow\text{Response} $$

may be substantially weaker than:

$$ \text{Exposure}\rightarrow\text{Response} $$
Important: When exposure is a more direct driver of pharmacologic effect than dose, an exposure–response model can provide a substantially more informative basis for dose selection.

Pharmacokinetic Models

The first component of a PK/PD model is usually the pharmacokinetic model.

A simple one-compartment IV bolus model can be written as:

$$ C(t) = \frac{D}{V} e^{-kt} $$

where:

  • \(D\) = dose
  • \(V\) = volume of distribution
  • \(k\) = elimination rate constant

The elimination rate constant is related to clearance by:

$$ k=\frac{CL}{V} $$

For oral administration, absorption must also be represented.

$$ C(t) = \frac{F D k_a}{V(k_a-k)} \left( e^{-kt}-e^{-k_at} \right) $$

where:

  • \(F\) = bioavailability
  • \(k_a\) = absorption rate constant
  • \(CL\) = clearance
  • \(V\) = volume of distribution

Key PK Exposure Metrics

Several exposure measures may be considered when developing an exposure– response relationship.

Exposure Metric Typical Interpretation
\(C_{\max}\) Peak concentration
\(C_{\min}\) Trough concentration
AUC Total systemic exposure over a specified interval
\(C_{\mathrm{avg}}\) Average concentration over an interval
Time above threshold Duration for which concentration exceeds a target

The appropriate exposure metric depends on the pharmacology.

Do not automatically assume AUC is the correct exposure metric. For some mechanisms, peak concentration, trough concentration, time above a threshold, or another exposure summary may be more closely associated with pharmacodynamic response.

Pharmacodynamic Models

The PD component describes how drug exposure produces a pharmacologic effect.

A basic linear model is:

$$ E=E_0+S C $$

where:

  • \(E_0\) = baseline effect
  • \(S\) = slope
  • \(C\) = concentration

A linear model is useful when the observed exposure range is approximately linear.

However, pharmacologic effects frequently approach a maximum.

The Emax Model

A commonly used nonlinear exposure–response model is the Emax model:

$$ E(C) = E_0+ \frac{E_{\max}C} {EC_{50}+C} $$

where:

  • \(E_0\) = baseline response
  • \(E_{\max}\) = maximum drug effect
  • \(EC_{50}\) = concentration producing 50% of the maximum drug effect
  • \(C\) = drug concentration

As concentration approaches zero:

$$ E(C)\rightarrow E_0 $$

As concentration becomes very large:

$$ E(C)\rightarrow E_0+E_{\max} $$
Interpretation: The Emax model captures diminishing returns. Once concentrations become sufficiently high, additional exposure produces progressively smaller increases in pharmacodynamic effect.

The Hill Emax Model

The relationship can be generalized using a Hill coefficient:

$$ E(C) = E_0+ \frac{E_{\max}C^\gamma} {EC_{50}^\gamma+C^\gamma} $$

where \(\gamma\) controls the steepness of the exposure–response curve.

When:

$$ \gamma=1 $$

the model reduces to the standard Emax model.

Why the Emax Model Is Important for Dose Selection

Suppose a treatment produces increasing response as exposure increases.

At low exposure, increasing dose may produce substantial improvements.

At higher exposure, the incremental benefit may become progressively smaller.

This means that the dose producing the maximum possible exposure may not be the optimal clinical dose.

Instead, investigators may seek a dose that lies near the region where most of the achievable pharmacodynamic benefit has already been obtained.

From Exposure–Response to Dose–Response

The PK and PD components can be combined.

Suppose:

$$ C_{\mathrm{avg}} = \frac{F D}{CL\tau} $$

for a linear PK model at steady state.

Substituting this exposure into an Emax model gives:

$$ E(D) = E_0+ \frac{ E_{\max} \left( \frac{FD}{CL\tau} \right) }{ EC_{50} + \frac{FD}{CL\tau} } $$

The resulting function directly connects dose with expected pharmacodynamic effect.

This is the central PK/PD idea: The dose–response relationship can be derived from the dose → PK → exposure → PD chain rather than estimated as a purely empirical relationship between dose and response.

A Simple Worked Example

Suppose a drug is administered once daily and has approximately linear PK.

Assume:

Parameter Value
Bioavailability \(F\) 0.80
Clearance \(CL\) 5 L/h
Dosing interval \(\tau\) 24 h
Baseline effect \(E_0\) 20 units
Maximum drug effect \(E_{\max}\) 80 units
\(EC_{50}\) 4 mg/L

For a dose of 100 mg:

$$ C_{\mathrm{avg}} = \frac{FD}{CL\tau} $$

Substituting the values:

$$ C_{\mathrm{avg}} = \frac{(0.80)(100)} {(5)(24)} $$
$$ C_{\mathrm{avg}} \approx0.667\text{ mg/L} $$

The expected PD effect is then:

$$ E = 20+ \frac{80(0.667)} {4+0.667} $$
$$ E\approx31.43 $$

The model therefore predicts an effect of approximately 31.4 units at this dose under the assumed parameters.

Evaluate Multiple Candidate Doses

Dose selection should rarely be based on a single candidate dose.

Suppose the candidate doses are:

Dose Predicted Average Exposure Predicted PD Effect
50 mg Calculate Calculate
100 mg Calculate Calculate
200 mg Calculate Calculate
400 mg Calculate Calculate

The resulting table illustrates a key feature of nonlinear pharmacology: doubling the dose does not necessarily double the pharmacodynamic effect.

What Dose Should Be Selected?

The dose-selection problem can be framed as:

$$ \text{Select }D \text{ such that } P(\text{desired efficacy}\mid D) \text{ is sufficiently high} $$

while simultaneously maintaining an acceptable probability of toxicity:

$$ P(\text{toxicity}\mid D) \text{ is sufficiently low} $$

Thus dose selection is often a balance between efficacy and safety.

Exposure Targets

Suppose pharmacology suggests that an exposure of at least \(C^*\) is required for adequate activity.

The dose-selection problem becomes:

$$ P(C\ge C^*\mid D)\ge q $$

where:

  • \(C^*\) = target exposure
  • \(D\) = candidate dose
  • \(q\) = desired probability of target attainment

For example, an investigator might seek a dose for which at least 90% of patients are predicted to achieve the target exposure.

Population variability matters. A dose that produces the target exposure in the typical patient may fail to produce the target exposure in a substantial fraction of patients if interindividual PK variability is large.

Interindividual Variability

Population PK models typically represent parameters as varying between individuals.

For example:

$$ CL_i = CL_{\mathrm{pop}} e^{\eta_{CL,i}} $$

where:

  • \(CL_i\) = clearance for individual \(i\)
  • \(CL_{\mathrm{pop}}\) = typical population clearance
  • \(\eta_{CL,i}\) = individual deviation from the population value

Similarly:

$$ V_i = V_{\mathrm{pop}} e^{\eta_{V,i}} $$

This variability can have a major impact on dose selection.

Covariates Can Explain Variability

Patient characteristics may explain part of the variability in PK parameters.

For example:

$$ CL_i = CL_{\mathrm{pop}} \left( \frac{WT_i}{70} \right)^{\theta_{WT}} e^{\eta_{CL,i}} $$

Other possible covariates include:

  • Body weight
  • Age
  • Renal function
  • Hepatic function
  • Sex
  • Concomitant medications
  • Disease status

A better explanation of PK variability can improve predictions of exposure at candidate doses.

Indirect Response Models

Not every pharmacodynamic response follows concentration instantaneously.

For biomarkers with turnover, an indirect response model may be more appropriate.

A basic turnover model is:

$$ \frac{dR}{dt} = k_{\mathrm{in}} - k_{\mathrm{out}}R $$

where:

  • \(R\) = response or biomarker level
  • \(k_{\mathrm{in}}\) = production rate
  • \(k_{\mathrm{out}}\) = loss rate

At baseline equilibrium:

$$ R_0 = \frac{k_{\mathrm{in}}}{k_{\mathrm{out}}} $$

Drug exposure can then modify either production or loss.

Why Time Matters in PK/PD Models

A static exposure–response model may be insufficient when the timing of drug exposure and response differ.

For example:

$$ C(t) \longrightarrow E(t) $$

may involve a delay.

An effect-compartment model can represent this delay:

$$ \frac{dC_e}{dt} = k_{e0} \left(C-C_e\right) $$

and the PD model can then use \(C_e\) rather than the plasma concentration directly.

Modeling principle: When plasma concentration and pharmacodynamic effect are not synchronized, a delay mechanism may be necessary rather than forcing the PD model to respond instantaneously to plasma concentration.

Biomarker-Based Dose Selection

PK/PD models are particularly useful when a biomarker provides an early measure of pharmacologic activity.

The development chain may be:

1
Dose produces exposure.
2
Exposure changes the biomarker.
3
Biomarker changes are associated with clinical benefit.
4
The model predicts which doses are likely to produce the desired biomarker response.

Exposure–Response Models in Clinical Trials

Clinical exposure–response analyses can use individual exposure metrics such as:

  • Individual AUC
  • Individual \(C_{\max}\)
  • Individual \(C_{\min}\)
  • Average concentration
  • Time above concentration threshold

The response model might be:

$$ P(Y_i=1) = \operatorname{logit}^{-1} \left( \alpha+\beta C_i \right) $$

for a binary clinical outcome.

Alternatively, a continuous endpoint could be modeled as:

$$ Y_i = E(C_i)+\epsilon_i $$

Binary Clinical Response

For a binary response endpoint, a logistic exposure–response model may be appropriate:

$$ \operatorname{logit}(P_i) = \alpha+ \beta C_i $$

where:

$$ \operatorname{logit}(P_i) = \log \left( \frac{P_i}{1-P_i} \right) $$

A nonlinear exposure effect can also be incorporated.

Exposure–Response Versus Dose–Response

Feature Dose–Response Exposure–Response
Predictor Administered dose Drug exposure
Accounts for PK variability Limited Yes
Mechanistic interpretation Lower Higher
Useful for individualized prediction Limited Often stronger
Can inform target exposure Indirectly Directly

Simulation for Dose Selection

Once a PK/PD model has been developed, simulation can be used to evaluate candidate doses.

A typical simulation workflow is:

1
Generate virtual patients from the population PK model.
2
Assign candidate doses.
3
Simulate individual concentration–time profiles.
4
Calculate exposure metrics.
5
Apply the PD model.
6
Evaluate efficacy and safety targets.
7
Compare candidate doses.

Probability of Target Attainment

A particularly useful simulation quantity is the probability that a patient achieves a predefined exposure target.

$$ PTA(D) = P(C_{\mathrm{metric}}\ge C_{\mathrm{target}}\mid D) $$

where PTA denotes probability of target attainment.

A dose might therefore be selected if:

$$ PTA(D)\ge0.90 $$

subject to an acceptable safety profile.

Balancing Efficacy and Safety

Dose selection should generally consider both efficacy and safety.

For example, define:

$$ P_E(D) = P(\text{efficacy target}\mid D) $$

and:

$$ P_S(D) = P(\text{safety threshold exceeded}\mid D) $$

A candidate dose might need to satisfy:

$$ P_E(D)\ge0.90 $$ and:

$$ P_S(D)\le0.10 $$

The exact thresholds depend on the development program and therapeutic context.

A Candidate Dose Selection Table

Dose Median Exposure PTA Efficacy Probability Safety Probability
Dose A Low 55% 60% 2%
Dose B Moderate 82% 84% 5%
Dose C High 94% 92% 9%
Dose D Very high 98% 95% 21%

Dose C may represent a reasonable compromise because it provides a high probability of achieving the desired exposure while maintaining an acceptable predicted safety profile.

Important: The dose with the highest predicted efficacy is not necessarily the dose that should be selected. Dose selection is a decision problem involving efficacy, safety, variability, feasibility, and uncertainty.

Model Uncertainty

PK/PD predictions are conditional on the model.

Important sources of uncertainty include:

  • Parameter uncertainty
  • Interindividual variability
  • Residual variability
  • Model misspecification
  • Covariate uncertainty
  • Uncertainty in the exposure–response relationship
  • Uncertainty in the translation from biomarker to clinical outcome

Therefore, dose-selection simulations should ideally propagate relevant uncertainty rather than treating estimated parameters as known constants.

Parameter Uncertainty

Suppose the estimated EC50 is:

$$ EC_{50}=4\text{ mg/L} $$

A point estimate alone does not indicate how uncertain this estimate is.

For simulation, one may instead draw:

$$ EC_{50}^{(s)} \sim p(EC_{50}\mid\text{data}) $$

and propagate this uncertainty through the dose-selection simulation.

Bayesian PK/PD Modeling

A Bayesian framework can be particularly useful when prior knowledge is available.

The posterior distribution is:

$$ p(\theta\mid y) \propto p(y\mid\theta)p(\theta) $$

where:

  • \(\theta\) represents model parameters.
  • \(y\) represents observed data.
  • \(p(\theta)\) is the prior distribution.
  • \(p(\theta\mid y)\) is the posterior distribution.

Posterior samples can then be propagated through the PK/PD model to produce probabilistic predictions for candidate doses.

PK/PD Modeling and Dose Selection Are Different From Dose Escalation

Dose escalation asks a different question from dose selection.

Problem Primary Question
Dose escalation How should doses be increased during early clinical development?
Dose selection Which dose or doses should be taken forward?
PK modeling What exposure does a dose produce?
PD modeling What effect does exposure produce?
PK/PD modeling How do dose, exposure, and effect connect?

Population PK/PD Models

Population modeling allows data from many individuals to be analyzed simultaneously.

A typical hierarchical structure is:

$$ \theta_i = \theta_{\mathrm{pop}} e^{\eta_i} $$

followed by an observation model such as:

$$ Y_{ij} = f(\theta_i,t_{ij}) + \epsilon_{ij} $$

This structure allows the model to distinguish:

  • Typical population behavior
  • Between-subject variability
  • Within-subject residual variability

Model Validation

A PK/PD model should not be accepted simply because the parameter estimates look reasonable.

Model evaluation may include:

  • Goodness-of-fit diagnostics
  • Visual predictive checks
  • Posterior predictive checks
  • Bootstrap analysis
  • Prediction-corrected visual predictive checks
  • External validation
  • Predictive performance in later studies
Prediction matters more than fit alone. A model that fits historical observations well but performs poorly when predicting new data may be a poor basis for dose selection.

R Implementation of an Emax Model

E0 <- 20
Emax <- 80
EC50 <- 4

emax <- function(C) {

  E0 +
    (Emax * C) /
    (EC50 + C)
}

emax(0.667)

This evaluates the predicted PD response for a given concentration.

Calculate Exposure From Dose

F <- 0.80
CL <- 5
tau <- 24

avg_concentration <- function(Dose) {

  F * Dose /
    (CL * tau)
}

avg_concentration(100)

Combine PK and PD

dose_response <- function(Dose) {

  Cavg <- F * Dose /
    (CL * tau)

  E0 +
    (Emax * Cavg) /
    (EC50 + Cavg)
}

dose_response(50)
dose_response(100)
dose_response(200)
dose_response(400)

The resulting function provides a simple model-based prediction of PD response across candidate doses.

Evaluate Candidate Doses

doses <- c(
  50,
  100,
  200,
  400
)

results <- data.frame(
  Dose = doses,
  Exposure = sapply(
    doses,
    avg_concentration
  ),
  Effect = sapply(
    doses,
    dose_response
  )
)

results

Simulating Interindividual Variability

set.seed(123)

n <- 10000

CL_i <- rlnorm(
  n,
  meanlog = log(5),
  sdlog = 0.30
)

Dose <- 200

Cavg_i <- F * Dose /
  (CL_i * tau)

Effect_i <-
  E0 +
  (Emax * Cavg_i) /
  (EC50 + Cavg_i)

mean(Effect_i)

quantile(
  Effect_i,
  c(0.05, 0.50, 0.95)
)

This illustrates how variability in clearance translates into variability in exposure and ultimately pharmacodynamic response.

Probability of Achieving a PD Target

Suppose the desired pharmacodynamic effect is:

$$ E\ge70 $$

The simulated probability of target attainment is: 

mean(
  Effect_i >= 70
)

The resulting quantity estimates:

$$ P(E\ge70\mid D=200) $$

This is often more useful for dose selection than reporting only the mean predicted response.

Comparing Candidate Doses by Simulation

evaluate_dose <- function(Dose) {

  Cavg <- F * Dose /
    (CL_i * tau)

  Effect <-
    E0 +
    (Emax * Cavg) /
    (EC50 + Cavg)

  data.frame(
    Dose = Dose,
    MeanExposure = mean(Cavg),
    MedianExposure = median(Cavg),
    PTA = mean(Cavg >= 2),
    MeanEffect = mean(Effect),
    PDTarget = mean(Effect >= 70)
  )
}

rbind(
  evaluate_dose(50),
  evaluate_dose(100),
  evaluate_dose(200),
  evaluate_dose(400)
)

From Model Output to a Dose Recommendation

The final step is not simply to identify the dose with the highest model prediction.

Instead, investigators should define the decision criteria before reviewing the final simulations.

For example:

  • At least 90% probability of achieving the exposure target
  • At least 80% probability of achieving the desired PD response
  • Acceptable predicted probability of toxicity
  • Acceptable interindividual variability
  • Practical dosing frequency
  • Compatibility with observed clinical safety data

A Practical Dose-Selection Workflow

1
Define the pharmacologic mechanism and clinically relevant endpoint.
2
Develop an appropriate PK model.
3
Characterize interindividual variability and relevant covariates.
4
Identify the exposure metric most closely related to pharmacodynamic activity.
5
Develop the PD or exposure–response model.
6
Evaluate model adequacy and predictive performance.
7
Define efficacy and safety targets.
8
Simulate candidate doses across virtual patient populations.
9
Calculate exposure and pharmacodynamic target attainment.
10
Quantify uncertainty around model predictions.
11
Compare candidate doses using prespecified decision criteria.
12
Select the dose or dose range that provides the most appropriate benefit–risk profile for further development.

Common Dose-Selection Mistakes

  1. Choosing the highest tolerated dose automatically. Higher exposure does not necessarily produce proportional increases in efficacy.
  2. Using dose instead of exposure when PK variability is substantial. The same dose can produce very different exposures across patients.
  3. Ignoring nonlinear pharmacology. Emax-type relationships can produce diminishing returns at higher exposure.
  4. Ignoring interindividual variability. The typical patient may not represent the tails of the exposure distribution.
  5. Using a poorly justified exposure metric. AUC is not automatically the best predictor of every pharmacodynamic effect.
  6. Focusing only on the mean prediction. Probability of target attainment may be more relevant for decision making.
  7. Ignoring model uncertainty. Dose recommendations should reflect uncertainty in model parameters and structural assumptions.
  8. Confusing model fit with predictive ability. A model can describe historical data well while making poor prospective predictions.
  9. Ignoring safety exposure–response relationships. An efficacy model alone is insufficient for benefit–risk dose selection.
  10. Treating simulation as proof. Simulations quantify what the model predicts; they do not eliminate uncertainty or replace clinical evidence.

PK/PD Modeling and the Development Program

PK/PD modeling can contribute at multiple stages of drug development.

Development Stage Potential PK/PD Contribution
Preclinical Translate exposure and pharmacology across species
First-in-human Predict exposure and support dose escalation
Phase I Characterize PK, tolerability, and pharmacologic response
Phase II Characterize exposure–response and select doses
Phase III Support dose justification and exposure–response interpretation
Regulatory Provide quantitative evidence supporting dose selection

Why PK/PD Modeling Is Powerful

Traditional dose selection may rely heavily on comparing observed outcomes across a small number of tested doses.

PK/PD modeling allows the investigator to ask a richer set of questions.

  • What exposure does each dose produce?
  • How much variability exists between patients?
  • What exposure is associated with meaningful pharmacodynamic activity?
  • What fraction of patients achieve that exposure?
  • Where does the exposure–response curve begin to plateau?
  • What happens if the dose is increased?
  • What happens if the dose is decreased?
  • How does uncertainty affect the dose recommendation?
  • What dose provides the best predicted benefit–risk balance?

The Most Important Concept

The central concept is the distinction between dose, exposure, and response.

A dose is what is administered.

Exposure is what the body experiences.

The pharmacodynamic response is what the drug does.

$$ \boxed{ \text{Dose} \rightarrow \text{PK} \rightarrow \text{Exposure} \rightarrow \text{PD} \rightarrow \text{Clinical Response} } $$

PK/PD modeling provides the quantitative bridge connecting these components.

Bottom line: PK/PD modeling provides a quantitative framework for dose selection by linking dose to exposure and exposure to pharmacodynamic response. A useful dose-selection analysis combines an appropriate PK model, an exposure–response or mechanistic PD model, interindividual variability, uncertainty, and simulation. The goal is not simply to select the highest dose that works, but to identify the dose or dose range that provides a sufficiently high probability of achieving the desired pharmacologic and clinical benefit while maintaining an acceptable safety profile.

References

Holford, N.H.G. & Sheiner, L.B. (1981). Understanding the dose-effect relationship: clinical application of pharmacokinetic-pharmacodynamic models. Clinical Pharmacokinetics, 6, 429–453.
Sheiner, L.B. & Ludden, T.M. (1992). Population pharmacokinetics/dynamics. Annual Review of Pharmacology and Toxicology, 32, 185–209.
Gabrielsson, J. & Weiner, D. Pharmacokinetic and Pharmacodynamic Data Analysis: Concepts and Applications.
Mould, D.R. & Upton, R.N. (2012). Basic concepts in population modeling, simulation, and model-based drug development. CPT: Pharmacometrics & Systems Pharmacology.
Ette, E.I. & Williams, P.J. Pharmacometrics: The Science of Quantitative Pharmacology. Wiley.
EMA. Guideline on the reporting of physiologically based pharmacokinetic (PBPK) modelling and simulation.
FDA. Model-Informed Drug Development Paired Meeting Charter.

Related Tutorials

Related Tools

  • → All Statistical Tools
  • → PK Parameter Calculator (coming soon)
  • → Emax Model Calculator (coming soon)
  • → Dose Simulation Tool (coming soon)
← Back to All Tutorials