A Foundation Model for Pharmacokinetic and Pharmacodynamic Trajectories with Mechanistic Constraints, Missingness Awareness, and Calibrated Uncertainty

Clinical pharmacokinetic and pharmacodynamic modelling supports the interpretation of dose, exposure, biological response, and patient variability, but most models remain tied to a particular drug, population, sampling design, endpoint, or clinical task. This task-specific orientation limits reuse across evolving treatment contexts and does not inherently address informative observation processes, distribution shift, or uncertainty under transfer. This Original Foundation-Model Architecture Article proposes a trajectory foundation model that combines reusable longitudinal representations with explicit pharmacological structure. Patient timelines, administered doses, concentration measurements, pharmacodynamic observations, time-varying covariates, drug and formulation characteristics, and observation-process information are represented through distinct provenance-preserving encoders. These representations inform continuous-time latent pharmacokinetic, exposure, effect-site, and pharmacodynamic states governed by modular hard or soft mechanistic constraints. Missing values, irregular sampling, and measurement intensity are represented as features of the observation process rather than being treated only as data-cleaning problems. The architecture separates aleatoric, epistemic, transfer, and out-of-domain uncertainty and permits selective prediction or abstention when evidence is insufficient. Evaluation is proposed against classical population models, mechanistic PK/PD models, task-specific machine-learning systems, and unconstrained pretrained models across predictive accuracy, calibration, transfer, mechanism consistency, missingness robustness, and decision relevance. Interpretability is grounded in state definitions, provenance, constraint auditing, and traceable adaptation rather than post hoc explanation alone. The contribution is an original conceptual synthesis, not an empirically validated system. Computational verification, independent external evaluation, prospective assessment, uncertainty validation, and context-specific clinical and implementation studies would be required before any decision-facing use.