Why Optimize a Virtual Population?
A candidate virtual population is a finite collection of mechanistic parameter vectors. Even after applying admissible-range and physiological-feasibility rules, its simulated summaries may differ from the targets derived from observed data or prior knowledge. Optimization-based selection provides a systematic way to choose a population whose simulated summaries have small, explicitly defined discrepancies from those targets.
The result depends on the objective, the candidate set, the constraints, and the search procedure. A small objective value means good agreement on the summaries included in that objective; it does not establish that the selected population is unique, representative in every relevant respect, or biologically correct. Those distinctions matter when a population is used for QSP predictions beyond the summaries used to select it.
Specify the Objective and Constraints First
For \(K\) target summaries, a common objective is a weighted sum of squared, scaled residuals:
Here \(g_k(V)\) is the simulated summary, \(t_k\) is its target, \(s_k>0\) is a scale used to put discrepancies on comparable footing, and \(w_k\geq 0\) is a prespecified importance weight. The residual \(g_k(V)-t_k\) has the units of summary \(k\); division by \(s_k\) makes that component dimensionless. Weights express chosen priorities, not a guarantee that one target is scientifically more certain. If the target summaries have known sampling covariance, a covariance-weighted discrepancy may be more appropriate, but the covariance and the objective must still be specified transparently.
The optimization problem is not fully defined until feasibility is stated. Write \(V\in\mathcal{F}\) when the population satisfies all hard requirements, such as parameter bounds, physiological rules, population size, and any required representation of strata. Then the problem is
A hard constraint excludes a population from consideration; it should not be treated as a soft preference merely because that is convenient computationally. A soft preference can instead be represented by an explicit penalty in \(J(V)\), with its scale and weight justified. These choices can produce different selected populations.
Worked Example: Checking Seven Feasible Swaps
Worked Example: Checking Seven Feasible Swaps
Suppose a two-member candidate population is initially \(\{8,13\}\), where each number denotes a candidate parameter vector. The selection objective is the squared discrepancy between a simulated endpoint summary and its target of \(0\). For this hypothetical example, the endpoint residuals for the allowed one-for-one replacements have been computed by the QSP model. A replacement retaining 13 may use 4, 11, or 16; retaining 8 may use 4, 11, 16, or 17. The replacement \(\{17,13\}\) is infeasible under a prespecified constraint.
For each feasible replacement, the population mean of the displayed parameter is \(\bar{x}=(x_1+x_2)/2\), and \(J=(r-0)^2\), where \(r\) is the simulated endpoint residual. The seven feasible results are:
| Replacement population | \(\bar{x}\) | Endpoint residual \(r\) | Objective \(J=r^2\) |
|---|---|---|---|
| \(\{4,13\}\) | 8.5 | 0.75 | 0.5625 |
| \(\{11,13\}\) | 12 | 1 | 1 |
| \(\{16,13\}\) | 14.5 | 2.25 | 5.0625 |
| \(\{4,8\}\) | 6 | 2 | 4 |
| \(\{11,8\}\) | 9.5 | 0.25 | 0.0625 |
| \(\{16,8\}\) | 12 | 1 | 1 |
| \(\{17,8\}\) | 12.5 | 1.25 | 1.5625 |
For example, \(\{11,8\}\) has mean \((11+8)/2=9.5\), and its objective is \((0.25-0)^2=0.0625\). The seven feasible one-for-one swaps therefore have means 8.5, 12, 14.5, 6, 9.5, 12, and 12.5, with objectives 0.5625, 1, 5.0625, 4, 0.0625, 1, and 1.5625, respectively. The two distinct populations \(\{11,13\}\) and \(\{16,8\}\) both have mean 12 and objective 1; equal summaries do not make them the same selected population. Among these seven feasible swaps, \(\{11,8\}\) gives the smallest objective. This is a best swap within the stated neighborhood, not yet proof of a global optimum over all possible populations.
Why Scaling and Weighting Change the Answer
A raw sum of squared discrepancies can be dominated by a summary with large numerical units. For instance, an error of 2 concentration units and an error of 0.05 in a proportion cannot be compared just by inspecting their raw magnitudes. Scales convert each error to a relative discrepancy, and weights determine its contribution after scaling. The scale may be a scientifically meaningful tolerance or another justified reference quantity. It should not be chosen after looking for a value that makes a preferred population win.
Worked Example: A Constrained Two-Summary Objective
Consider three hypothetical candidate populations, each with 40 virtual patients. The targets are a mean concentration of 20 mg/L and a response proportion of 0.30. Use scales \(s_1=2\) mg/L and \(s_2=0.05\), with weights \(w_1=1\) and \(w_2=4\). The objective is
Suppose a prespecified feasibility rule also requires the mean concentration to lie between 19 and 22 mg/L, inclusive. Candidate A has summaries \((22,\ 0.35)\). It is feasible, and
Candidate B has summaries \((21,\ 0.32)\). It is feasible, and
Candidate C has summaries \((19,\ 0.40)\). Its objective would be
C is feasible under the stated concentration rule, but its response proportion contributes a large discrepancy. Thus B is the best of these three candidates. The objective alone does not certify that B is best among candidates not listed, nor that these two summaries are sufficient for the intended application.
How the Search Works and When to Stop
When the candidate set is finite, the global solution is the feasible population with the smallest objective. Exhaustive enumeration establishes that solution if every feasible population has been evaluated. For larger candidate sets, enumeration may be too costly. A practical alternative is a local search: start from a feasible population, propose nearby changes such as one-for-one swaps, reject infeasible proposals, and retain a proposal only when it reduces the objective.
The reason this procedure can stop at a local rather than global solution is straightforward. If no allowed neighboring move lowers \(J\), then no population in the searched neighborhood improves the current one. That fact says nothing about a feasible population farther away, unless the search has examined the full feasible set or a result for the search method guarantees global optimality. Random restarts or several starting populations can reveal sensitivity to the starting point, but do not by themselves prove global optimality.
Worked Example: Applying a Stopping Criterion
Suppose a swap-based search has objective values 3.20, 1.40, and 0.62 at successive accepted populations. Set \(\epsilon=0.01\). After the last accepted population, the algorithm evaluates every feasible one-for-one swap and finds the smallest objective to be 0.62; all other feasible swaps have objective at least 0.62. The best decrease is therefore \(0.62-0.62=0\), which is not greater than \(\epsilon\). The search stops under the neighborhood criterion. It has found a local optimum for the tested swaps to the stated tolerance, but not necessarily the global optimum. If the algorithm instead stopped after 500 simulations without evaluating the full neighborhood, it would have stopped by budget and could not make even that local-optimum claim.
A short justification for the local-search claim follows directly from the stopping rule. Let \(V^\ast\) be the current population and \(\mathcal{N}(V^\ast)\) its declared feasible neighborhood. If every \(U\in\mathcal{N}(V^\ast)\) satisfies \(J(U)\geq J(V^\ast)-\epsilon\), then no tested neighboring move improves the objective by more than \(\epsilon\). This implication is complete for that neighborhood, but it does not extend to feasible populations outside \(\mathcal{N}(V^\ast)\). The distinction is why the search space and stopping criterion belong in the report alongside the selected population.
Fix candidate vectors, population size, target summaries, scales, weights, and hard constraints before searching.
Simulate each population using the same model settings and calculate the same summaries and objective.
Use exhaustive or declared approximate search; retain objective traces, constraint failures, and the reason the procedure ended.
Inspect target fit, feasibility, robustness, and relevant summaries that were not optimized.
Diagnostics, Alternatives, and Limitations
Optimization is going wrong, or answering a narrower question than intended, when the numerical score is treated as sufficient evidence of population quality. Examine each component residual, not only the total \(J\). A low total can hide a poor fit to a high-priority endpoint if weights or scales were poorly chosen. Check that every retained vector satisfies all hard constraints and that simulated summaries are reproducible to adequate precision. If simulation noise changes the ranking of close candidates, rerun or stabilize the evaluation before interpreting the selected solution.
Other warning signs include strong dependence on starting populations, a selected population concentrated on a few candidates, and poor agreement in outcomes not included in the objective. These do not necessarily invalidate the optimization; they limit what its result supports. Selection against a small set of targets can overfit those targets, especially when the candidate pool is large. Preserve and assess diversity when it is important for downstream predictions, and validate the selected population against summaries or perturbations not used to select it.
| Approach | What it does | Key distinction |
|---|---|---|
| Rejection sampling | Retains candidates meeting prespecified target ranges. | Does not necessarily rank retained candidates by a joint objective. |
| Weighted resampling | Changes selection probabilities using candidate weights. | Produces a probabilistic resampling scheme rather than necessarily minimizing one population-level objective. |
| Optimization-based selection | Searches for a feasible population with a low defined objective. | Depends explicitly on the objective and search; local search may not find a global solution. |
These alternatives answer related but different questions. Range filtering can be appropriate when targets are acceptable intervals and no ranking within them is needed. Weighted resampling can preserve a probabilistic interpretation of candidate weights. Optimization is useful when a specified population-level discrepancy is the decision criterion. None automatically guarantees diversity, joint-structure fidelity, or predictive adequacy. The method should therefore be chosen to match the selection goal, not simply to obtain the smallest available score.
Common Mistake
Check Your Understanding
Answer the questions, then compare your reasoning with the solutions below.
- Why should a concentration discrepancy and a proportion discrepancy generally be scaled before they are combined in a squared-error objective?
- A search stops because its simulation budget is exhausted. What can it claim about optimality?
- For target values 10 and 0.20, scales 2 and 0.05, and equal weights, calculate the objective for simulated summaries 12 and 0.25.
- Can a population with the smallest objective still be unsuitable for downstream QSP predictions? Give one reason.
Solutions. (1) They have different units and numerical scales; scaling makes each residual dimensionless before combination, while weights encode explicit relative importance. (2) It can report the best candidate evaluated, but the budget stop alone does not establish local or global optimality. (3) \(J=((12-10)/2)^2+((0.25-0.20)/0.05)^2=1^2+1^2=2\). (4) Yes. For example, it may fit the optimized summaries but have poor diversity or fail to reproduce an important outcome that was not included in the objective.