Tutorials › Pharmacometrics › Model-Based Dose Selection for Phase 2
Pharmacokinetics · MIDD · Phase 2 Development

Model-Based Dose Selection for Phase 2

Learn how pharmacokinetic, exposure-response, and pharmacodynamic models can integrate early clinical evidence to support Phase 2 dose selection—and how simulation can evaluate the consequences of alternative doses before a larger confirmatory trial begins.

Intermediate PK/PD Model-Based Development Phase 2
01 · The development question

1. Why Is Dose Selection So Important in Phase 2?

Phase 2 development often represents the transition from demonstrating that a drug can produce a pharmacologic effect to determining how it should be dosed in a larger patient population. The central question is not simply whether a dose produces an effect, but how dose, exposure, response, tolerability, and patient variability should be integrated to choose doses for subsequent development.

A dose-ranging study may evaluate several dose levels, but the observed responses at those doses are only part of the available evidence. Pharmacokinetic and exposure-response models can combine information across dose levels and time points to estimate relationships that are difficult to see from simple comparisons of treatment means.

Dose Dose levels PK / PD model dose → exposure exposure → response variability + uncertainty Dose selection simulation-informed Observed clinical data → model → simulated development scenarios

Model-based dose selection uses quantitative relationships between dose, exposure, response, and variability to evaluate candidate doses.

Core idea: the objective is not to find the dose with the highest observed response in a small study. It is to characterize the exposure-response and exposure-tolerability relationships well enough to make a scientifically justified dose-selection decision under uncertainty.
02 · From dose to response

2. Why Dose Alone Is Often Not Enough

A dose is an administered quantity. Pharmacologic effects, however, are generally related more directly to drug exposure at the site of action than to the nominal administered dose.

Two patients receiving the same dose can have different concentrations because of differences in clearance, absorption, body size, organ function, concomitant medications, or other sources of pharmacokinetic variability. Consequently, a dose-response relationship may appear weaker or more variable than the underlying exposure-response relationship.

Quantity Question Role in dose selection
Dose How much drug was administered? Defines the intervention but may not fully characterize pharmacologic exposure.
Concentration What systemic concentration was achieved? Provides a closer link to pharmacologic activity and toxicity.
AUC What was the overall exposure? Can characterize relationships between cumulative exposure and response.
Cmax What peak exposure was achieved? May be relevant to efficacy, acute pharmacology, or tolerability.
Ctrough What concentration remains before the next dose? Can be useful when sustained exposure is related to efficacy or safety.
Exposure How much drug reaches the relevant systemic exposure range? Provides a quantitative bridge between dosing and observed clinical effects.

The resulting framework is often written as:

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

This distinction is central to model-based dose selection. A dose may be selected because it reliably produces an exposure associated with a desirable balance of efficacy and tolerability—not simply because it is the highest dose tested.

03 · Phase 2 evidence

3. What Evidence Is Available in Phase 2?

A Phase 2 dose-selection analysis can integrate several sources of information. The exact data available depend on the development program, but commonly include pharmacokinetic observations, pharmacodynamic biomarkers, clinical endpoints, adverse events, and patient characteristics.

Evidence Information provided Potential modeling role
PK concentrations Drug exposure over time Estimate PK parameters and individual exposure.
Biomarkers Pharmacologic activity Characterize concentration-effect relationships.
Clinical efficacy endpoint Patient-level therapeutic response Estimate exposure-response relationships.
Adverse events Tolerability and safety Characterize exposure-safety relationships.
Patient covariates Sources of PK or response variability Explain differences among individuals.
Prior knowledge Previous clinical and nonclinical evidence Inform assumptions, priors, or model structure where appropriate.

The resulting analysis is therefore broader than a traditional dose-response comparison. The model attempts to explain how the administered dose produces exposure and how exposure is associated with both benefit and risk.

04 · PK foundation

4. The PK Model Comes First

A model-based dose-selection framework usually begins with a pharmacokinetic model. The PK component describes how dose produces concentration over time and how individual characteristics may alter that relationship.

A simple one-compartment model with first-order elimination can be written as:

$$C(t)=\frac{F D}{V}e^{-CLt/V}$$

where \(F\) is bioavailability, \(D\) is dose, \(V\) is apparent volume of distribution, and \(CL\) is clearance.

More realistic Phase 2 analyses may use multi-compartment models, absorption models, nonlinear elimination, time-varying clearance, or population PK models with between-subject variability.

Why this matters: the PK model converts an administered dose into an estimated exposure for each individual. That exposure can then be linked to efficacy and safety endpoints.
05 · Patient variability

5. Why Population PK Is Often Important

Phase 2 trials generally enroll heterogeneous patients. A population PK model can describe a typical patient's PK parameters while also quantifying between-subject variability and the influence of patient characteristics.

A simplified population PK model for clearance might be represented as:

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

Here \(CL_i\) is the clearance for individual \(i\), \(CL_{\text{pop}}\) is the typical population clearance, \(WT_i\) is body weight, and \(\eta_{CL,i}\) represents unexplained between-subject variability.

Other covariates may include renal function, age, sex, disease status, concomitant medications, or other scientifically justified characteristics.

Covariate relationships are particularly useful when the goal is to predict the exposure distribution that future Phase 2 or Phase 3 patients may experience.

06 · Exposure

6. Turning PK Into Individual Exposure

Once a population PK model has been developed, individual exposure metrics can often be derived from model-based predictions. Depending on the scientific question, these may include AUC, Cmax, Cmin, trough concentration, average concentration, or time above a target concentration.

For example, under linear PK:

$$AUC_{0-\infty}=\frac{F D}{CL}$$

This equation illustrates why variability in clearance translates directly into variability in exposure. At a fixed dose, an individual with lower clearance will generally have greater exposure under the assumptions of the linear model.

Key transition: once individual exposure is estimated, the analysis can ask whether efficacy and safety are more closely associated with exposure than with the administered dose.
07 · Exposure-response

7. Modeling the Exposure-Efficacy Relationship

The next component is an exposure-response model. The appropriate model depends on the endpoint. Continuous endpoints, binary response, count outcomes, time-to-event outcomes, and longitudinal measurements may require different statistical models.

A simple Emax model for a continuous response can be written as:

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

In this model, \(E_0\) represents baseline response, \(E_{\max}\) represents the maximum model-predicted drug effect, and \(EC_{50}\) is the exposure associated with half of the maximum drug effect.

The model can be extended when the observed relationship is not adequately described by a simple Emax function. For example, a Hill model introduces a shape parameter:

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

The purpose is not to choose the most sophisticated function automatically. The model should be appropriate for the endpoint, supported by the available data, and sufficiently identifiable for the decision being made.

08 · Exposure-safety

8. Modeling Exposure and Tolerability

Dose selection should generally consider safety as well as efficacy. A dose that increases response but also produces substantially greater exposure-related toxicity may not provide the desired benefit-risk balance.

Safety endpoints can be modeled in several ways. For example, a binary adverse event probability might be described using a logistic exposure-response model:

$$\operatorname{logit}\{P(AE=1)\} =\alpha+\beta C$$

where \(C\) represents a relevant exposure metric.

More complex analyses may account for repeated adverse events, time to first event, exposure duration, dose interruptions, or multiple safety endpoints.

Benefit-risk principle: dose selection is generally a joint efficacy-and-safety problem. Modeling only efficacy can omit an important part of the information required to select a dose.
09 · Integration

9. Bringing Efficacy and Safety Together

A model-based framework can produce predicted efficacy and safety outcomes over a range of candidate exposures. This allows investigators to examine where the expected benefit begins to plateau and where additional exposure may produce diminishing benefit or increasing toxicity.

Efficacy Safety burden Exposure Model-predicted outcome candidate exposure

Conceptually, model-based dose selection seeks an exposure range that provides adequate efficacy while maintaining acceptable tolerability. The actual relationships must be estimated from program-specific data.

In practice, the decision may involve several candidate doses rather than a single optimal point. The analysis can therefore focus on identifying an exposure range that is sufficiently effective and tolerable while accounting for uncertainty and variability.

10 · Dose mapping

10. From Target Exposure Back to Dose

One of the most useful features of model-based development is that the reasoning can work in both directions.

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

Once an exposure associated with an acceptable efficacy-safety profile has been identified, the PK model can be used to determine which dose is expected to produce that exposure in the target population.

Under a simple linear IV model:

$$AUC=\frac{D}{CL}$$

Therefore, if a target AUC is \(AUC_{\text{target}}\), an approximate dose can be obtained from:

$$D_{\text{target}}=AUC_{\text{target}}\times CL$$

Real development programs are usually more complicated. Interindividual variability in clearance means that one dose produces a distribution of exposures, not one deterministic exposure. This is why population simulation is often more informative than calculating a single nominal dose.

11 · Dose range

11. Selecting More Than One Dose

Model-based dose selection does not necessarily result in one dose. A development program may choose multiple doses when the uncertainty in the exposure-response relationship remains substantial or when different doses provide useful information about the shape of the relationship.

For example, a Phase 2 study might investigate:

  • A lower dose expected to provide meaningful exposure with a favorable tolerability profile.
  • A middle dose expected to be near the region of greatest incremental efficacy.
  • A higher dose intended to characterize whether additional exposure provides additional benefit.
  • A placebo group to establish the clinical response in the absence of active treatment.

The spacing between doses should be informed by the expected exposure relationship, PK variability, prior evidence, and the scientific objectives of the study.

Important distinction: a dose-ranging study can serve two purposes: estimating treatment effects at specific doses and learning the shape of the dose-exposure-response relationship. Model-based design attempts to use both.
12 · Worked example

12. Worked Example: Translating Target Exposure Into a Phase 2 Dose

Consider a hypothetical drug with approximately linear PK over the exposure range of interest. Suppose the Phase 1 and early Phase 2 data suggest that an average daily AUC of approximately 100 mg·h/L is associated with a clinically relevant exposure range.

Suppose the population PK model predicts a typical clearance of 5 L/h.

Step 1: Estimate the dose for the target exposure

$$D=AUC_{\text{target}}\times CL$$
$$D=100\text{ mg·h/L}\times5\text{ L/h}=500\text{ mg}$$

Step 2: Consider a lower candidate dose

If a 300 mg dose is considered:

$$AUC=\frac{300}{5}=60\text{ mg·h/L}$$

Step 3: Consider a higher candidate dose

If a 700 mg dose is considered:

$$AUC=\frac{700}{5}=140\text{ mg·h/L}$$
Dose Predicted typical AUC Interpretation
300 mg 60 mg·h/L Below the illustrative target exposure.
500 mg 100 mg·h/L Near the illustrative target exposure.
700 mg 140 mg·h/L Above the illustrative target exposure.

This calculation alone does not establish that 500 mg is the correct clinical dose. It simply maps a target exposure to a nominal dose under a simplified PK assumption.

A complete model-based analysis would then ask whether the predicted exposure distribution at each dose produces adequate efficacy and acceptable safety across the intended patient population.

13 · Variability

13. Why the Typical Patient Is Not Enough

The previous example uses a typical clearance of 5 L/h. Real patients vary. Suppose clearance has substantial between-subject variability. The same 500 mg dose could then produce substantially different exposures in different patients.

Under a linear model:

$$AUC_i=\frac{D}{CL_i}$$

Therefore, variability in \(CL_i\) translates into variability in exposure. A dose-selection analysis should consider this distribution rather than relying solely on the typical value.

This is particularly important when the exposure-response relationship is steep or when safety events become more frequent at higher exposures.

Population perspective: the relevant question is often not "What exposure does this dose produce?" but "What distribution of exposures will this dose produce in the population we intend to study?"
14 · Simulation

14. Using Clinical Trial Simulation for Dose Selection

Once PK, efficacy, and safety models have been developed, they can be used to simulate hypothetical future trials. This allows investigators to compare candidate dosing strategies before conducting the study.

A simulation may repeatedly generate:

  1. Patient characteristics from a representative population.
  2. Individual PK parameters from the population model.
  3. Concentration-time profiles under each candidate dose.
  4. Exposure metrics derived from those profiles.
  5. Efficacy outcomes using the exposure-response model.
  6. Safety outcomes using the exposure-safety model.
  7. The planned statistical analysis of the simulated trial.

Repeating this process many times provides a distribution of possible trial outcomes rather than a single predicted result.

$$\text{Model}+\text{Population}+\text{Design} \rightarrow\text{Simulated trials} \rightarrow\text{Decision characteristics}$$

Simulation can therefore evaluate questions such as whether a proposed dose range is sufficiently informative, how often a dose would meet a predefined decision criterion, or how uncertainty in model parameters affects the decision.

15 · Decision criteria

15. What Does "Good Dose" Mean?

A model does not define the clinical decision by itself. The development team must first define the decision criteria that the model will help evaluate.

Depending on the program, these criteria may involve:

  • A target probability of achieving a clinically meaningful efficacy response.
  • A target exposure range associated with pharmacologic activity.
  • An acceptable probability of an important adverse event.
  • A desired separation between efficacy and safety exposure distributions.
  • A target probability of achieving a predefined endpoint in a future trial.
  • Operational considerations such as dosing frequency, formulation, or adherence.

These criteria should be specified in a manner consistent with the scientific and clinical objectives of the development program.

Modeling does not replace clinical judgment. It provides a quantitative framework for expressing assumptions, integrating evidence, and evaluating the consequences of alternative decisions.
16 · Uncertainty

16. Parameter Uncertainty Matters

Model-based decisions are never based on perfectly known parameters. Estimates of clearance, exposure-response parameters, treatment effects, and safety relationships all contain uncertainty.

For example, suppose an estimated exposure-response parameter is:

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

That point estimate does not imply that the true value is exactly 50 mg/L. The uncertainty interval around the estimate may be wide, particularly when exposure-response information is limited.

Simulation can propagate parameter uncertainty through the complete model so that the resulting dose-selection analysis reflects uncertainty in both the underlying parameters and future patient variability.

This distinction is important:

Concept Meaning
Between-subject variability Real differences in PK or response among patients.
Residual variability Unexplained variation between observations and model predictions.
Parameter uncertainty Uncertainty about the values of the model parameters themselves.
Model uncertainty Uncertainty about whether the chosen model adequately represents the system.
17 · Model evaluation

17. How Should the Model Be Evaluated?

A model used for dose selection should undergo appropriate evaluation before its predictions are relied upon for a development decision.

Important considerations can include:

  1. Goodness of fit: Does the model adequately reproduce observed data?
  2. Residual diagnostics: Are systematic patterns remaining in the residuals?
  3. Parameter plausibility: Are the estimated parameters scientifically reasonable?
  4. Precision: How much uncertainty surrounds key parameter estimates?
  5. Predictive performance: Does the model predict observations not directly used for fitting?
  6. Visual predictive checks: Does simulated behavior resemble the observed distribution?
  7. Covariate evaluation: Are important sources of variability represented appropriately?
  8. Sensitivity analysis: Do reasonable alternative assumptions materially change the dose-selection conclusion?
Fit is not validation. A flexible model can fit observed data well while still making unreliable predictions. Dose-selection models should therefore be evaluated for their intended predictive purpose.
18 · The shape of the curve

18. Why the Shape of the Exposure-Response Curve Matters

The location of the selected dose depends strongly on the shape of the exposure-response relationship.

Consider three conceptual possibilities:

Relationship Potential implication
Approximately linear Increasing exposure may continue to produce increasing response over the observed range.
Saturating Additional exposure eventually produces smaller incremental benefit.
Steep relationship Small exposure differences can produce meaningful response differences.
Flat relationship Large exposure changes may produce relatively little additional response.

A saturating relationship is particularly important for dose selection. If efficacy approaches a plateau while toxicity continues to increase with exposure, increasing the dose beyond the efficacy plateau may provide limited additional benefit.

19 · Biomarkers

19. Where Do Pharmacodynamic Biomarkers Fit?

Pharmacodynamic biomarkers can provide an intermediate link between systemic exposure and clinical outcome.

$$\text{Dose}\rightarrow\text{PK}\rightarrow \text{Biomarker}\rightarrow\text{Clinical response}$$

A biomarker may respond more quickly or more directly to drug exposure than a clinical endpoint. This can help characterize whether the administered doses are producing the expected pharmacologic effect.

Biomarker modeling can therefore contribute to dose selection even when the ultimate clinical endpoint is noisy or slow to observe.

However, a biomarker is not automatically a surrogate for clinical benefit. Its role in dose selection depends on the strength and relevance of the relationship between the biomarker, exposure, and clinical outcome.

20 · Preparing for Phase 3

20. Connecting Phase 2 Dose Selection to Phase 3

One major purpose of Phase 2 dose selection is to provide a quantitative basis for the doses carried forward into later development.

A model-based analysis can support questions such as:

  • Which doses are expected to cover the relevant exposure range?
  • What proportion of future patients is expected to achieve target exposure?
  • How much variability in exposure is expected at each dose?
  • Is there evidence that additional exposure produces additional efficacy?
  • Does the exposure-safety relationship constrain the upper end of the dose range?
  • How sensitive is the dose-selection decision to uncertainty in the model?

Clinical trial simulation can then be used to evaluate whether the proposed Phase 3 dosing strategy has a reasonable probability of providing informative results under plausible assumptions.

21 · Updating evidence

21. Dose Selection Can Be an Iterative Process

Model-based development is not necessarily a one-time analysis. As new clinical data become available, PK and exposure-response models can be updated.

For example, a development sequence might look like:

Data PK + efficacy Model estimate Simulate candidate doses Trial new data Evidence is updated as the development program progresses

Model-based development is iterative: new data can update the model and refine subsequent dose and study decisions.

22 · Interpretation

22. What Model-Based Dose Selection Does Not Guarantee

A sophisticated model does not eliminate uncertainty or guarantee that the selected dose will succeed in a future trial.

  • The model can be wrong. Structural assumptions may not fully represent the biological system.
  • The data can be insufficient. Limited exposure coverage or sparse sampling can make key parameters weakly identifiable.
  • Extrapolation can be uncertain. Predictions beyond the observed exposure range depend strongly on model assumptions.
  • Population differences can matter. Future trial participants may differ from the population used to develop the model.
  • Clinical endpoints contain variability. Even a well-described exposure-response relationship may not predict every patient's outcome.
  • Model uncertainty should be considered. Different plausible models can sometimes imply different dose-selection conclusions.
Interpretation principle: model-based dose selection converts available evidence into quantitative predictions. It does not turn uncertain evidence into certainty.
23 · Practical workflow

23. A Practical Workflow for Model-Based Phase 2 Dose Selection

  1. Define the clinical decision. Identify what dose-selection question the model needs to answer.
  2. Assemble the available evidence. Integrate PK, PD, efficacy, safety, biomarker, and relevant prior information.
  3. Develop the PK model. Characterize dose-exposure relationships and important sources of variability.
  4. Develop exposure-response models. Characterize efficacy and, where possible, safety relationships with exposure.
  5. Evaluate the models. Examine goodness of fit, diagnostics, parameter uncertainty, predictive performance, and sensitivity to alternative assumptions.
  6. Define candidate doses. Map relevant exposure ranges back to practical dosing regimens.
  7. Simulate future trials. Propagate population variability and parameter uncertainty through the proposed study design.
  8. Evaluate decision characteristics. Determine how often candidate doses achieve the predefined efficacy and safety criteria under plausible scenarios.
  9. Select the development strategy. Use the quantitative evidence together with clinical, statistical, operational, and regulatory considerations.
  10. Update the model as evidence accumulates. Treat dose selection as an iterative learning process rather than a one-time calculation.

24. Key Takeaways

  • Model-based dose selection connects administered dose to exposure and then to efficacy and safety.
  • A dose is not the same thing as exposure. Patients receiving the same dose can experience substantially different concentrations.
  • Population PK models can characterize typical pharmacokinetics, variability, and clinically relevant covariate relationships.
  • Exposure-response models can provide a more informative basis for dose selection than dose-response comparisons alone when exposure varies substantially among patients.
  • Efficacy and safety should be considered together when selecting a dose for subsequent development.
  • Target exposure ranges can be translated back into practical dosing regimens using the PK model.
  • Simulation can propagate patient variability and parameter uncertainty through future clinical trial designs.
  • A useful dose-selection model should be evaluated for its intended predictive purpose rather than judged only by goodness of fit.
  • Model uncertainty, parameter uncertainty, and between-subject variability are distinct sources of uncertainty and should be considered separately.
  • Model-based dose selection supports clinical decision-making but does not replace clinical judgment or eliminate uncertainty.
  • The most useful model is not necessarily the most complex model. It is the model that is sufficiently informative, identifiable, interpretable, and predictive for the development decision.
Next step

Where to Go Next

A natural next step is to study exposure-response modeling in greater detail, including Emax, sigmoid Emax, linear, indirect-response, and longitudinal models.

From there, model-based development can be extended to clinical trial simulation, population PK/PD modeling, covariate modeling, exposure-response analysis, dose optimization, and model-informed drug development across Phase 1 through Phase 3.

These methods provide the quantitative foundation for moving from "What doses were tested?" to the more informative question: "What exposure range should the next clinical study target, and why?"

Related Tutorials

Clinical Biostats

← Back to Pharmacokinetics Tutorials