Who chooses the compartments?

A concentration curve looks deceptively simple: a set of points, a rising or falling line, and time along the bottom. The difficult editorial question is what commitments were made when somebody drew that line. Which explanations were permitted, which were excluded, and which discrepancies were accepted? A model-selection paper offers an unusually concrete way to inspect those choices.

A bounded search

Richardson and colleagues reported an automated population-pharmacokinetic workflow in July 2025. It uses pyDarwin, Bayesian optimisation with a random-forest surrogate, and local search to choose models for extravascular administration. A penalty discourages excessive complexity and implausible parameters. One synthetic and four clinical datasets were evaluated against expert-built models.¹

The authors report comparable structures in under 48 hours on average, using 40 CPUs and 40 GB of memory. That is a stated computing condition, not a promise of laptop-speed analysis.¹

What a search can explain

Our interpretation begins with a distinction between searching and asking. A search can explore the alternatives it is given; it cannot establish that those alternatives contain the right scientific explanation. We would therefore read the definition of the candidate models before celebrating the chosen winner. An impressive search through an unsuitable collection would still answer an unsuitable question.

Imagine two candidate descriptions that both follow the observed points. One makes additional assumptions about a hidden process; another uses fewer moving parts. Their similar fit does not tell us which hidden process actually exists. In our view, the useful next test would be a prediction they disagree about, followed by a measurement collected expressly to distinguish them. That turns model comparison into an experimental question instead of a contest for the prettiest curve.

Keep the rejected alternatives

For a research team adopting this approach, we would request a compact record of the alternatives considered, the penalties applied and the reasons a model failed. That record should travel with the final result. Changing the penalty is a change in the analyst's preferences, even when the ensuing search happens automatically.

We would also separate three deliverables: a description of the existing concentration observations, a prediction for observations withheld from development, and a justified decision about what to measure next. Each is useful. None should silently stand in for the others. In particular, a satisfactory account of exposure does not itself answer whether a biological response is beneficial.

This is the value we see in the work: an opportunity to make a consequential analytical choice inspectable and repeatable. We would judge a future implementation by whether another modeller can understand why it chose its answer, identify the assumptions most likely to matter, and challenge them with new evidence. The attractive endpoint is an accountable pharmacokinetic analysis, with the responsibility for interpreting it still clearly assigned.