Skip to content
Parth Ajmera

Population pharmacokinetics

#pk-pd

Give the same drug, at the same dose, to a hundred different people, and you'll get close to a hundred different responses. One person clears it in four hours; another takes twelve. One gets the intended effect; a third gets toxicity at the same dose that did nothing for the first two. Age, weight, kidney function, genetics, diet, other medications — all of it moves the number.

Classical pharmacokinetics was built to characterise one patient at a time, intensively: many blood draws, controlled conditions. That tells you almost nothing about the hundred-person spread above. Population pharmacokinetics (PopPK) flips the approach — sparse data (a few blood samples per person, at irregular real-world times) across many patients, fed into a statistical model that separates out three things at once: the typical drug behaviour across the population, how much individuals vary around that typical behaviour, and which patient characteristics — age, renal function, a genetic variant — actually explain that variation.

The mathematical engine behind this is nonlinear mixed-effects modelling (NLME); the software that's been the industry standard since the 1970s is called NONMEM.

Where it sits in development

PopPK is unusual among PK/PD subfields in that it's used across the entire pipeline, not just one stage — animal PopPK projecting human PK in preclinical work, sparse-sampling PopPK in Phase I and II identifying which covariates matter, large-scale PopPK in Phase III supporting the dose that actually gets approved, and a PopPK report as a standard, expected component of every NDA/BLA regulatory submission.

What's actually settled, and what isn't

The statistical methodology — NLME modelling, covariate analysis, the regulatory expectations around a PopPK report — is mature; it's been the standard approach for decades. What's still actively moving:

  • Biologics. Monoclonal antibodies, antibody-drug conjugates, and fusion proteins behave very differently from small molecules, and the models built for small-molecule PK don't extrapolate cleanly.
  • Novel modalities — mRNA and gene therapy PK — are close to uncharted; there isn't an established modelling tradition to lean on yet.
  • Machine learning integration — using ML to find which covariates matter in high-dimensional genomic or proteomic data, rather than testing them one at a time by hand — is an active area, not a settled one.
  • Pediatric and neonatal PopPK is a real regulatory gap: children aren't small adults, and developmental physiology changes the picture in ways adult-derived models don't capture.
  • Real-world-data PopPK, built from electronic health records instead of clean trial data, is genuinely harder — real-world data is messy in ways clinical trial data by design isn't — and still being worked out.

I don't have a reliable figure for how much of the field is "mature" versus "frontier" — any such split is someone's rough characterisation, mine included, not a measured statistic.