QSP Parameter Identifiability: Can the Data Determine the Parameters?

1. What Is Parameter Identifiability?

Parameter identifiability asks a fundamental question: given a mathematical model and a set of observations, can the unknown parameters be determined sufficiently well from the information available?

This question is especially important in quantitative systems pharmacology (QSP), where models can contain many biological states, feedback loops, binding processes, turnover mechanisms, and parameters assembled from heterogeneous experimental sources. A model may reproduce observed data while still containing parameters that the data cannot uniquely or precisely determine.

2. Why Is Identifiability So Important in QSP?

QSP models often attempt to represent biological mechanisms at a level of detail substantially greater than the number of directly observed quantities. This creates a potential mismatch between model complexity and information content in the data.

If several parameter combinations produce nearly indistinguishable predictions, fitting the model may yield a good numerical fit without producing uniquely determined biological parameter values. QSP methodology papers emphasize that identifiability analysis is important precisely because model complexity can exceed what the available observations can constrain.

3. Three Different Questions About Identifiability

It is useful to separate three related but distinct questions.

1. Is the model structurally identifiable? In an idealized setting with perfect, noise-free observations, can the parameters be uniquely recovered?
2. Is the model practically identifiable? Given the actual experimental data, noise, sampling schedule, and measurement limitations, are the parameters sufficiently constrained?
3. Are the parameters relevant to the scientific question? Even if an individual parameter is poorly identified, can the model still make the prediction or decision of interest reliably?

4. Structural Identifiability

Structural identifiability is primarily a property of the mathematical model, its parameterization, its inputs, and its assumed observables. It asks whether unique parameter values could theoretically be recovered if the relevant observations were perfect and sufficiently rich.

Consider the simple model

Suppose only \(y(t)\) and \(x(t)\) are observed. The model depends on the product \(ab\), not on \(a\) and \(b\) independently. If \(a\) is multiplied by a factor \(c\) while \(b\) is divided by the same factor, the prediction remains unchanged:

Consequently, the individual values of \(a\) and \(b\) cannot be uniquely recovered from this output. The product \(ab\) may be identifiable even though the individual parameters are not.

5. Practical Identifiability

Practical identifiability concerns the information content of the actual dataset. A model may be structurally identifiable but still practically non-identifiable because the available observations are too sparse, too noisy, insufficiently informative, or poorly positioned in time.

For example, imagine that a QSP model contains a biomarker production rate \(k_{\mathrm{prod}}\) and degradation rate \(k_{\mathrm{deg}}\). If the experiment measures the biomarker only after it has reached approximately steady state, the data may contain little information about the individual rates even though both parameters appear explicitly in the model.

6. Structural vs. Practical Identifiability

The distinction is crucial because collecting more data does not automatically solve structural non-identifiability. If two parameters are mathematically redundant under the chosen model and observables, increasing sample size or reducing measurement noise cannot create information that the model-output relationship does not contain.

7. Parameterization Can Create or Reveal Identifiability Problems

Identifiability is affected by how a model is parameterized. Consider a QSP pathway in which an effect depends on the product of two quantities:

If the available observations depend only on \(k_1k_2\), attempting to estimate both parameters independently introduces a redundancy.

In some cases, the scientifically meaningful quantity may actually be the product itself. In others, independent experimental measurements may be available for one of the two parameters. Fixing or informing one parameter from reliable external evidence can then allow the remaining parameter to be estimated from the model.

8. Parameter Sensitivity and Identifiability

Sensitivity analysis examines how much model outputs change when parameters change. A local sensitivity coefficient can be written as:

where \(y_i(t)\) is an observable and \(\theta_j\) is a model parameter.

If an output is nearly insensitive to a parameter across the relevant experimental conditions, the data may contain little information about that parameter.

However, sensitivity and identifiability are not identical concepts. Two parameters can each be highly influential while producing very similar changes in the output. In that situation, the model may be sensitive to both parameters but unable to distinguish them individually.

9. The Fisher Information Matrix

A common local tool for assessing practical identifiability is the Fisher information matrix (FIM). In simplified notation, it can be represented as:

where \(S\) is a sensitivity matrix and \(W\) reflects the assumed observation-error weighting.

The eigenvalues and eigenvectors of the information matrix describe directions in parameter space that are strongly or weakly informed by the data.

FIM-based methods are useful because they can identify poorly informed directions before or during model calibration. However, they are generally local approximations and should not automatically be treated as definitive evidence of global identifiability.

10. Profile Likelihood

Profile likelihood provides a more direct way to investigate how strongly the data constrain an individual parameter.

Suppose the parameter of interest is \(\theta_j\). Rather than estimating all parameters freely, fix \(\theta_j\) at a sequence of values and re-estimate the remaining parameters:

This produces a profile likelihood for \(\theta_j\). The remaining parameters are allowed to compensate for changes in the parameter being profiled.

Profile likelihood is particularly useful in QSP because nonlinear models can have parameter relationships that are not adequately summarized by a local covariance approximation. QSP workflow literature has specifically recommended profile likelihood as a complementary tool for identifying poorly constrained parameters and obtaining confidence intervals.

11. Parameter Correlation and Compensation

One of the most common practical identifiability problems is that parameters compensate for one another.

Suppose increasing parameter \(k_1\) increases an observable while increasing \(k_2\) decreases it. If their effects are sufficiently similar over the observed conditions, many combinations of \(k_1\) and \(k_2\) may produce almost the same model output.

The fitted parameters may therefore exhibit strong covariance or correlation. A parameter estimate can appear precise when viewed alone while the joint parameter distribution reveals a long, narrow ridge.

12. Why One Successful Optimization Is Not Enough

Nonlinear QSP models can have complicated objective surfaces. A single optimization run that converges successfully does not establish that the solution is unique.

A useful strategy is multistart parameter estimation: fit the same model repeatedly from many plausible initial parameter sets.

1. Define scientifically plausible parameter ranges.
2. Generate many initial parameter vectors.
3. Fit the model from each starting point.
4. Compare objective-function values.
5. Compare resulting parameter vectors.
6. Compare model predictions across the resulting solutions.

13. Identifiability and Parameter Uncertainty

Identifiability and uncertainty are closely related but should not be treated as identical concepts.

A parameter may have a narrow confidence interval because strong prior information has been supplied, even though the current experimental data contribute little information about that parameter. Conversely, a parameter can be structurally identifiable while having substantial uncertainty because the dataset is small or noisy.

Bayesian approaches can represent parameter uncertainty through posterior distributions. Markov chain Monte Carlo or other posterior-sampling approaches can reveal whether a parameter is tightly constrained or whether broad, correlated parameter combinations remain plausible.

14. The Observable Determines What Can Be Learned

A QSP model may contain dozens or hundreds of mechanistic quantities, but only some may be directly connected to the available observations.

For example, consider a biomarker \(B(t)\) governed by:

If only steady-state biomarker concentration is measured, then:

The steady-state observation informs the ratio of production to degradation, but it may not separately determine both rates.

Time-course measurements following a perturbation can provide additional information about the dynamics and may help distinguish the two processes.

15. Using Identifiability to Design Better Experiments

Identifiability analysis should not be viewed only as a post hoc diagnostic. It can also guide the design of experiments intended to make the model more informative.

Suppose two candidate parameters produce nearly identical predictions in the existing experiment. A useful next experiment is one in which their predicted effects diverge.

Potential design strategies

• Measure additional biomarkers or molecular species.
• Collect time-course observations rather than only steady-state measurements.
• Introduce perturbations that selectively affect competing mechanisms.
• Use multiple doses or exposure levels.
• Study multiple physiological or disease states.
• Measure downstream outputs that distinguish alternative pathways.
• Combine independent data sources that constrain different parameter groups.

16. Worked Example: A Simple Turnover Model

Consider a hypothetical biomarker governed by:

Assume the biomarker is initially at steady state and the observed baseline concentration is:

Step 1: Apply the steady-state condition

At steady state:

Therefore:

Step 2: Try two parameter sets

All three parameter sets produce the same baseline steady-state concentration. Therefore, baseline concentration alone cannot distinguish these parameter combinations.

Step 3: Add a dynamic observation

Now suppose the system is perturbed so that production is suddenly reduced to zero. The model becomes:

with solution:

The three candidate parameter sets now predict different time courses because \(k_{\mathrm{deg}}\) determines the rate of decline.

17. Why Identifiability Is Particularly Challenging in QSP

QSP models combine biological detail with mathematical dynamics. Several characteristics make identifiability especially challenging.

This creates an important QSP modeling tension: increasing mechanistic detail can make a model more biologically expressive, but the additional parameters may also become poorly constrained. The goal is therefore not simply to maximize model complexity.

18. Not Every Parameter Needs to Be Estimated From One Dataset

QSP models frequently integrate information from many sources. A parameter may be supported by an in vitro experiment, a preclinical study, literature evidence, a clinical PK analysis, or a mechanistic assumption.

Consequently, the practical question is not always: “Can this parameter be estimated from the clinical dataset?”

It may instead be: “What evidence constrains this parameter, and is that evidence sufficient for the intended use of the model?”

19. What Should You Do With Non-Identifiable Parameters?

Discovering non-identifiability does not automatically mean the entire QSP model must be discarded. Several responses may be appropriate depending on the scientific purpose.

1. Fix parameters using independent evidence. If reliable external information exists, a parameter may be fixed or constrained rather than freely estimated.
2. Collect additional data. New observables or perturbations may distinguish competing parameter sets.
3. Reparameterize the model. Replace redundant individual parameters with a combination that the data actually inform.
4. Reduce model complexity. Remove mechanisms that cannot be supported by the intended data and scientific question.
5. Retain parameter uncertainty explicitly. If the model is still useful for prediction, propagate the uncertainty rather than presenting a single parameter vector as if it were uniquely determined.
6. Focus on identifiable predictions. The scientific question may depend on a model output rather than on precise estimation of every mechanistic parameter.

20. Can a Non-Identifiable Model Still Make Useful Predictions?

Yes. This is one of the most important nuances in QSP modeling.

Imagine that many parameter vectors fit the available observations equally well. If all of those parameter vectors produce nearly identical predictions for a particular clinical endpoint under a proposed dosing regimen, then the endpoint may still be robust even though the individual parameters are not.

The relevant question becomes: which predictions are stable across the parameter sets that remain compatible with the evidence?

21. A Practical QSP Identifiability Workflow

1. Define the scientific question. Identify which outputs or decisions the model must support.
2. Document parameter provenance. Separate measured, literature-derived, assumed, and calibrated parameters.
3. Specify the observables. Determine exactly which model states or transformations are connected to experimental measurements.
4. Assess structural identifiability where feasible. Look for parameter redundancies before attempting extensive calibration.
5. Perform sensitivity analysis. Determine which parameters affect the outputs under the experimental conditions of interest.
6. Fit using multiple starting points. Examine whether substantially different parameter vectors achieve similar objective-function values.
7. Inspect parameter correlations and the information matrix. Look for weakly informed directions.
8. Use profile likelihood or posterior analysis. Investigate nonlinear uncertainty that local covariance approximations may miss.
9. Assess predictive uncertainty. Determine whether uncertainty in parameters materially affects the outputs needed for the scientific question.
10. Design additional experiments when needed. Seek observations or perturbations that separate competing mechanisms.
11. Document what is and is not identified. Clearly distinguish estimated parameters, externally constrained parameters, uncertain parameters, and identifiable predictions.

22. Common QSP Identifiability Mistakes

23. From Parameter Identifiability to Model Credibility

Identifiability is one part of a broader QSP model-development process. A model can be numerically calibrated yet still require additional assessment before it is used for extrapolation or decision making.

The order is not always strictly linear, but the conceptual relationship is important. If parameters are not constrained, calibration may be unstable. If calibration is uncertain, predictions may inherit that uncertainty. If predictions are sensitive to poorly informed mechanisms, additional data or model refinement may be necessary.

Where to Go Next

A natural next step is to study QSP Model Calibration: how to formulate an objective function, combine heterogeneous datasets, choose which parameters to estimate or fix, perform multistart optimization, assess convergence, and propagate parameter uncertainty into model predictions.

After calibration, useful companion topics include QSP Sensitivity Analysis, QSP Model Reduction, Virtual Population Generation, and QSP Experimental Design.

References

1. Sher A, Niederer SA, Mirams GR, et al. A Quantitative Systems Pharmacology Perspective on the Importance of Parameter Identifiability. Bulletin of Mathematical Biology. 2022;84:39.
   https://doi.org/10.1007/s11538-021-00982-5
2. Cucurull-Sanchez L, Chappell MJ, Cheung SYA, et al. Best Practices to Maximize the Use and Reuse of Quantitative and Systems Pharmacology Models: Recommendations From the UK Quantitative and Systems Pharmacology Network. CPT: Pharmacometrics & Systems Pharmacology. 2019;8:610–621.
   https://doi.org/10.1002/psp4.12381
3. Nguyen VK, et al. Methodologies for Quantitative Systems Pharmacology (QSP) Models: Design and Estimation. CPT: Pharmacometrics & Systems Pharmacology. 2017.
   https://doi.org/10.1002/psp4.12206
4. Miao H, Xia X, Perelson AS, Wu H. On Identifiability of Nonlinear ODE Models and Applications in Viral Dynamics. SIAM Review. 2011;53(1):3–39.
   https://doi.org/10.1137/090757009
5. Raue A, Kreutz C, Maiwald T, et al. Structural and Practical Identifiability Analysis of Partially Observed Dynamical Models by Exploiting the Profile Likelihood. Bioinformatics. 2009;25(15):1923–1929.
   https://doi.org/10.1093/bioinformatics/btp358
6. Open Systems Pharmacology. Model Evaluation: sensitivity analysis, uncertainty assessment, and identifiability considerations for mechanistic models.
   Open Systems Pharmacology — Model Evaluation