1. What Is Sensitivity Analysis?
Sensitivity analysis examines how changes in model inputs affect model outputs. In quantitative systems pharmacology (QSP), those inputs can include model parameters, initial conditions, dosing variables, biological assumptions, or other quantities that influence the simulated system.
A QSP model may contain dozens or hundreds of parameters. Not every parameter contributes equally to a particular prediction. Sensitivity analysis provides a systematic way to determine which inputs have the greatest influence on an output of interest.
Sensitivity analysis connects model inputs to predictions by quantifying how much an output changes when an input is varied.
2. Why Is Sensitivity Analysis Important in QSP?
QSP models are designed to integrate biological mechanisms across multiple levels of a system. This mechanistic richness is useful, but it also creates many opportunities for uncertainty and parameter interactions.
Sensitivity analysis helps connect model complexity to practical scientific questions.
| Question | Role of sensitivity analysis | Example |
|---|---|---|
| Which parameters influence a prediction? | Identifies influential model inputs | Which parameters determine predicted tumor response? |
| Which measurements would be valuable? | Supports experimental prioritization | Would measuring receptor abundance reduce uncertainty? |
| Which mechanisms matter? | Tests the contribution of pathways or processes | Does target-mediated internalization materially affect response? |
| Which parameters are relatively unimportant? | Helps identify low-influence parameters for a specific output | Does a downstream degradation rate affect the clinical endpoint? |
| How robust is a conclusion? | Tests whether predictions change substantially under plausible input variation | Does the predicted dose-response relationship remain stable? |
Sensitivity analysis is therefore not simply a model-diagnostics exercise. It can be used as a bridge between model development, experimental design, uncertainty analysis, and decision-focused simulation.
3. The Mathematical Idea Behind Sensitivity
Suppose a model produces an output \(Y\) as a function of parameters \(\theta_1,\theta_2,\ldots,\theta_p\):
A local sensitivity asks how much \(Y\) changes when one parameter changes slightly while the other parameters are held fixed.
This derivative provides the local slope of the model output with respect to parameter \(\theta_i\).
For example, if:
then:
A one-unit increase in \(\theta\) produces a ten-unit increase in \(Y\), locally.
4. Normalized or Dimensionless Sensitivity
To make sensitivities more comparable across parameters, it is common to consider a normalized sensitivity coefficient:
This can be interpreted approximately as the fractional change in output associated with a fractional change in the parameter, under a local approximation.
For example, suppose a 1% increase in a parameter produces approximately a 2% increase in an output. The corresponding normalized local sensitivity is approximately:
If the output decreases by approximately 2% for a 1% increase in the parameter, the sensitivity is approximately:
| Normalized sensitivity | Approximate interpretation |
|---|---|
| \(S^N>0\) | Increasing the parameter tends to increase the output locally |
| \(S^N<0\) | Increasing the parameter tends to decrease the output locally |
| \(|S^N|\) large | The output is relatively sensitive to the parameter locally |
| \(|S^N|\) small | The output is relatively insensitive to the parameter locally |
These interpretations are local. They should not automatically be extrapolated to large parameter changes or to other outputs.
5. Local Sensitivity Analysis
Local sensitivity analysis evaluates model behavior around a particular nominal parameter set.
For a QSP model represented by differential equations:
where \(x(t)\) is the vector of system states and \(\theta\) is the parameter vector, sensitivity equations can describe how each state changes with respect to each parameter.
For parameter \(\theta_i\), define:
The resulting sensitivity trajectories show how perturbations in parameters propagate through the dynamic system over time.
Local sensitivities can vary over time. A parameter may strongly influence an early biomarker response but have relatively little influence on a later endpoint.
Local sensitivity analysis can therefore be especially useful for dynamic QSP models where the scientific question concerns the timing and direction of parameter influence.
6. Global Sensitivity Analysis
Global sensitivity analysis explores model behavior over a range of plausible parameter values rather than around a single nominal parameter set.
This distinction is important for nonlinear QSP models. A parameter that appears weakly influential near the nominal parameter set may become important elsewhere in the parameter space.
Conceptually, global analysis asks:
and examines how variation across the joint parameter space affects the output.
Common global approaches include:
- Latin hypercube sampling.
- Monte Carlo sampling.
- Partial rank correlation coefficients.
- Standardized regression coefficients.
- Morris screening.
- Variance-based methods such as Sobol indices.
7. One-at-a-Time Sensitivity Analysis
A simple approach is to vary one parameter while keeping the others fixed. This is commonly called one-at-a-time (OAT) analysis.
For example, a parameter may be varied by ±10%, ±25%, or another scientifically justified range:
where \(\delta\) represents the relative perturbation.
The resulting change in an output can then be summarized as:
| Advantage | Limitation |
|---|---|
| Easy to implement and explain | May miss interactions among parameters |
| Useful for quick model exploration | Results depend on the chosen perturbation size |
| Can produce intuitive tornado plots | Does not explore the full parameter space |
| Useful for communication with non-modelers | Can be misleading for strongly nonlinear models |
OAT analysis can be valuable as an exploratory tool, but it should not automatically be interpreted as a complete global sensitivity analysis.
8. Morris Screening
The Morris method is a global screening approach designed to identify parameters that have potentially important effects without requiring the computational burden of evaluating a full variance decomposition.
The method evaluates elementary effects associated with changes in individual parameters at multiple locations in the parameter space.
Two useful summaries are commonly considered:
- \(\mu^*\): the mean absolute elementary effect, which reflects overall influence.
- \(\sigma\): the standard deviation of elementary effects, which can indicate nonlinear effects and/or interactions.
| Pattern | Possible interpretation |
|---|---|
| Low \(\mu^*\), low \(\sigma\) | Relatively low influence over the explored range |
| High \(\mu^*\), low \(\sigma\) | Important effect that is relatively consistent |
| High \(\mu^*\), high \(\sigma\) | Important parameter with possible nonlinearity or interactions |
| Low \(\mu^*\), high \(\sigma\) | Potentially context-dependent or interaction-driven behavior |
Morris screening is particularly useful when a QSP model contains a large number of parameters and the first objective is to identify which parameters deserve more detailed investigation.
9. Sobol Sensitivity Indices
Sobol sensitivity analysis is a variance-based global method. It partitions output variability into contributions associated with individual parameters and parameter interactions.
Suppose the model output is \(Y=f(\theta_1,\ldots,\theta_p)\). The first-order Sobol index for parameter \(i\) can be expressed conceptually as:
This represents the fraction of output variance attributable to parameter \(i\) alone, under the specified parameter distributions and model.
Total-effect indices include the parameter's interactions with other parameters:
where \(\theta_{-i}\) denotes all parameters other than \(\theta_i\).
10. Correlation and Regression-Based Sensitivity
Another family of methods relates sampled parameter values to sampled model outputs.
For example, a rank-based correlation can be used to determine whether increasing a parameter tends to be associated with increasing or decreasing output across the sampled parameter space.
Partial rank correlation coefficients (PRCCs) attempt to assess the relationship between one parameter and an output while accounting for the effects of the other sampled parameters.
| Method | Typical use | Important consideration |
|---|---|---|
| Correlation coefficient | Simple association between parameter and output | Can be affected by other parameters |
| PRCC | Rank-based multivariable sensitivity screening | Works best when relationships are reasonably monotonic |
| Standardized regression coefficient | Regression-based influence assessment | Depends on the regression model and sampled parameter distributions |
These methods can be computationally attractive for large QSP models, especially when the goal is to screen many parameters before applying more expensive methods.
11. Sensitivity Is Always Relative to an Output
There is no single universal sensitivity ranking for a QSP model.
Sensitivity is defined relative to a particular model output, time point, or summary metric.
For example, consider a model containing:
- Target expression.
- Drug binding affinity.
- Internalization rate.
- Target degradation rate.
- Downstream signaling rate.
- Cell proliferation rate.
The parameters that influence early target occupancy may differ from those that influence a late clinical endpoint.
| Output | Potentially relevant sensitivity question |
|---|---|
| Target occupancy at 1 hour | Which parameters control early target engagement? |
| Biomarker at 24 hours | Which mechanisms propagate target engagement to biomarker change? |
| Tumor volume at 28 days | Which parameters influence long-term disease dynamics? |
| Clinical response probability | Which uncertain mechanisms drive predicted response? |
12. Sensitivity Can Change Over Time
QSP models are dynamic systems, so parameter influence can change during a simulation.
Suppose parameter \(\theta_i\) has sensitivity:
The value of \(S_i(t)\) can be small early in the simulation and large later—or the reverse.
This occurs because different biological processes can dominate different phases of a response.
Two parameters can have very different temporal sensitivity profiles. The parameter that matters most for an early response need not dominate a later endpoint.
For this reason, sensitivity analysis of dynamic QSP models should often consider both time-resolved outputs and clinically meaningful summary measures.
13. Sensitivity Analysis and Uncertainty Analysis
Sensitivity analysis and uncertainty analysis are closely related but answer different questions.
| Concept | Question |
|---|---|
| Uncertainty analysis | How much can the model prediction vary because the inputs are uncertain? |
| Sensitivity analysis | Which uncertain inputs contribute most strongly to variation in the prediction? |
Suppose a QSP model predicts a treatment response of 60% but the model's uncertain inputs produce a wide range of possible predictions.
Uncertainty analysis characterizes the distribution of predicted responses. Sensitivity analysis can then identify which parameters are responsible for much of that prediction uncertainty.
Sensitivity analysis helps identify which parts of that pathway matter most.
14. Sensitivity and Parameter Identifiability
Sensitivity analysis is closely connected to parameter identifiability, but the concepts are not identical.
A parameter with very little influence on a particular observed output may be difficult to estimate from that output. Conversely, a parameter can have a substantial influence on an output but still be difficult to estimate precisely if it is correlated with other parameters.
| Concept | Core question |
|---|---|
| Sensitivity | How strongly does the output respond to the parameter? |
| Identifiability | Can the available data uniquely or sufficiently constrain the parameter? |
| Uncertainty | How uncertain is the parameter or prediction? |
| Correlation | How strongly are parameters or outputs statistically associated? |
15. Why Parameter Interactions Matter
QSP models frequently contain mechanisms that interact. Two parameters may have modest individual effects but a substantial combined effect.
Consider a simplified output:
The output depends on the product of the two parameters. Changing either parameter changes the influence of the other.
Global sensitivity methods can help identify these interactions.
| Observation | Possible interpretation |
|---|---|
| High first-order effect | Parameter independently explains substantial output variation |
| Low first-order but high total effect | Parameter may be important primarily through interactions |
| High interaction contribution | Model behavior depends on combinations of mechanisms |
This is one reason global sensitivity analysis can provide information that simple one-at-a-time perturbations cannot.
16. Using Sensitivity Analysis to Guide Experiments
One of the most useful applications of QSP sensitivity analysis is identifying which measurements may provide the greatest value for reducing uncertainty in a prediction.
Suppose a model prediction is highly sensitive to a biological parameter that has been estimated from limited experimental data. Additional experiments targeting that parameter may be informative.
A simplified workflow is:
- Identify the prediction of interest.
- Quantify uncertainty in the prediction.
- Perform sensitivity analysis.
- Identify influential uncertain parameters.
- Determine which parameters can realistically be measured or better characterized.
- Evaluate whether additional information would materially reduce predictive uncertainty.
This approach can help connect mechanistic modeling with experimental planning rather than treating the QSP model as an isolated computational exercise.
17. Worked Example: Sensitivity of a QSP Response
Consider a simplified QSP model that predicts a biomarker response \(Y\) based on three uncertain parameters:
- \(\theta_1\): target abundance
- \(\theta_2\): drug-target affinity parameter
- \(\theta_3\): downstream signaling rate
Suppose the nominal model predicts:
Step 1: Perturb target abundance
A 10% increase in \(\theta_1\) produces \(Y=44\).
The approximate normalized sensitivity is:
Step 2: Perturb affinity
A 10% increase in \(\theta_2\) produces \(Y=42\).
Step 3: Perturb downstream signaling
A 10% increase in \(\theta_3\) produces \(Y=36\).
| Parameter | Output after +10% | Approx. normalized sensitivity |
|---|---|---|
| Target abundance \(\theta_1\) | 44 | +1.0 |
| Affinity \(\theta_2\) | 42 | +0.5 |
| Signaling rate \(\theta_3\) | 36 | −1.0 |
The calculation indicates that the output has a stronger local response to target abundance and signaling rate than to affinity under this particular perturbation and nominal parameter set.
18. Tornado Plots and Sensitivity Rankings
A common way to communicate OAT sensitivity results is a tornado plot. Each bar represents the change in an output caused by varying one parameter over a specified range.
A conceptual tornado plot. The width of each bar reflects the change in a selected output over the specified parameter perturbation range.
Tornado plots are intuitive, but the ranking should always be accompanied by information about the parameter ranges, output definition, and analysis method.
19. Sensitivity of Dose-Response Predictions
QSP sensitivity analysis becomes particularly useful when the model is used to simulate treatment response under different doses or exposure levels.
For example, the output of interest might be:
- Target occupancy.
- Biomarker suppression.
- Pathway inhibition.
- Tumor growth inhibition.
- Probability of achieving a response threshold.
- Probability of exceeding a toxicity threshold.
The sensitivity analysis can identify which biological and pharmacological assumptions most influence the simulated dose-response relationship.
This can help distinguish uncertainty arising from drug exposure from uncertainty arising from disease biology, target biology, downstream signaling, or other model components.
20. Practical Best Practices for QSP Sensitivity Analysis
- Define the scientific question first. Sensitivity analysis should be driven by a prediction or decision of interest.
- Define the output explicitly. Specify the biomarker, endpoint, time point, or summary measure.
- Use scientifically justified parameter ranges. Sensitivity results are conditional on the parameter ranges explored.
- Distinguish local from global analysis. Do not interpret a local derivative as a global statement about the model.
- Consider nonlinearities. Large perturbations can produce behavior that is not captured by local approximations.
- Consider interactions. Global methods can reveal effects that OAT analysis may miss.
- Track parameter uncertainty. A parameter's sensitivity is most useful when considered together with how uncertain that parameter actually is.
- Examine multiple endpoints when appropriate. A parameter may be influential for one output but not another.
- Use sensitivity results to guide experiments. Influential uncertain parameters may identify opportunities for additional data collection.
- Document the analysis. Report the parameter distributions, perturbation ranges, outputs, methods, sampling strategy, and number of model evaluations.
21. Common Mistakes in QSP Sensitivity Analysis
| Mistake | Why it can be problematic |
|---|---|
| Reporting a single sensitivity ranking for the entire model | Sensitivity is generally output- and context-dependent |
| Using arbitrary parameter ranges | The resulting sensitivity may not reflect scientifically plausible uncertainty |
| Interpreting OAT results as global results | OAT analysis does not fully explore parameter interactions |
| Ignoring parameter correlations | Independent variation may produce biologically implausible combinations |
| Confusing sensitivity with identifiability | An influential parameter is not necessarily estimable from available data |
| Ignoring output timing | Dynamic sensitivity can change substantially over time |
| Focusing only on magnitude | The direction and biological meaning of the effect also matter |
| Using sensitivity as proof of mechanism | Sensitivity describes model behavior; it does not by itself establish biological causality |
22. A Practical QSP Sensitivity Analysis Workflow
- Define the scientific objective. Identify the prediction, endpoint, or decision of interest.
- Define the model inputs. Identify parameters, initial conditions, dosing variables, and other uncertain quantities.
- Specify plausible ranges or distributions. Base these on experimental data, literature, prior knowledge, or justified assumptions.
- Select the sensitivity method. Choose local, OAT, Morris, PRCC, regression-based, Sobol, or another method according to the question and computational budget.
- Generate model simulations. Vary inputs according to the selected analysis design.
- Calculate sensitivity measures. Quantify the influence of parameters on the specified outputs.
- Visualize the results. Use sensitivity trajectories, tornado plots, scatter plots, rankings, or variance-decomposition summaries as appropriate.
- Investigate influential parameters. Determine whether they represent biological uncertainty, pharmacological uncertainty, structural uncertainty, or parameter interactions.
- Connect results to data. Identify whether additional experiments or measurements could reduce important sources of uncertainty.
- Document assumptions. Record parameter ranges, distributions, outputs, sampling design, and computational settings.
23. Key Takeaways
- Sensitivity analysis quantifies how changes in model inputs influence QSP predictions.
- Sensitivity is always relative to a particular output, endpoint, time point, or summary measure.
- Local sensitivity examines model behavior around a nominal parameter set, while global sensitivity explores a broader parameter space.
- Normalized sensitivities make parameter effects more comparable across different scales and units.
- One-at-a-time analysis is intuitive but does not fully capture parameter interactions.
- Morris screening can efficiently identify influential parameters and potential nonlinear or interaction effects.
- Sobol methods can partition output variance into first-order and total parameter contributions.
- Correlation- and regression-based methods can provide efficient global screening when their assumptions are appropriate.
- Sensitivity can vary substantially over time in dynamic QSP models.
- Sensitivity and uncertainty analysis are related but answer different questions: uncertainty describes how much predictions vary, while sensitivity helps identify which inputs drive that variation.
- Sensitivity is not the same as parameter identifiability. A parameter can strongly affect a prediction while remaining difficult to estimate from available data.
- Global sensitivity analysis can reveal important parameter interactions that simple perturbation analyses may miss.
- Sensitivity results can help prioritize experiments and identify measurements that may reduce uncertainty in important predictions.
- The most useful sensitivity analysis is tied to a clearly defined scientific question, clinically relevant output, and scientifically justified parameter uncertainty.
Where to Go Next
A natural progression after QSP sensitivity analysis is to study parameter uncertainty and identifiability, followed by structural versus practical identifiability, uncertainty propagation, parameter estimation, model calibration, and virtual population generation.
These topics build on sensitivity analysis by asking not only which parameters influence a prediction, but also whether available data can constrain those parameters and how uncertainty in them propagates into clinically relevant model predictions.