%0 Journal Article
%T What Machine Learning Adds to Pharmacometrics—and What Mechanistic Information It Risks Removing
%A Ying Li
%A Hui Wang
%A Xiaoming Chen
%A Jun Zhang
%J Pharmacophore
%@ 2229-5402
%D 2025
%V 16
%N 4
%R 10.51847/maxWWV33ys
%P 91-101
%X Machine learning is increasingly used alongside pharmacokinetic, pharmacodynamic, population, exposure–response, and systems-pharmacology models, but its contribution is frequently judged through predictive performance without equivalent attention to the scientific information altered by the modeling choice. This integrative critical review examines where machine learning adds genuine value to pharmacometrics, where it merely substitutes one representation for another, and where it may remove mechanistic constraints needed for explanation, extrapolation, or decision credibility. The selected literature was compared according to pharmaceutical task, data suitability, model structure, mechanistic information content, validation design, uncertainty treatment, interpretability, and intended context of use. The synthesis distinguishes mechanistic models, predominantly data-driven models, and hybrid or scientific machine-learning architectures. Machine learning can expand nonlinear function estimation, covariate discovery, concentration prediction, exposure–response characterization, and computational efficiency. These capabilities are most persuasive when the task is bounded, the prediction domain is adequately represented, and the consequences of error are limited or reversible. However, predictive flexibility can obscure compartmental meaning, physiological constraints, parameter interpretation, causal assumptions, and the basis for extrapolation. Hybridization may preserve selected mechanistic structures while learning unknown components, but the label “hybrid” does not itself establish mechanistic validity or regulatory credibility. The article proposes a comparative analytical framework based on an explicit mechanistic-information ledger and task-dependent evidence requirements. The available evidence remains limited by heterogeneous validation practices, restricted external evaluation, incomplete calibration assessment, and sparse prospective demonstration. Accountable integration therefore requires model choice to follow the pharmaceutical question, not technological preference, and requires claims to remain proportionate to the information retained and the validation actually performed.
%U https://pharmacophorejournal.com/article/what-machine-learning-adds-to-pharmacometricsand-what-mechanistic-information-it-risks-removing-eicwaryksd2qtdd