Physiologically Constrained Deep Learning in Physiologically Based Pharmacokinetic Modeling: Virtual Patient Generation, Machine-Learned Scaling Laws, and Pediatric Precision Dosing

Physiologically based pharmacokinetic (PBPK) modeling is now central to pediatric dose selection, because clinical sampling in children is ethically constrained and simple weight-based scaling is unreliable while organ function and enzyme activity are still maturing. Three bottlenecks limit classical PBPK in this setting, namely building physiologically coherent virtual patient populations, extrapolating drug disposition across species and age, and estimating many mechanism-linked parameters from sparse data. Over the past five years, deep learning has begun to address each. Generative models constrained by physiological knowledge create virtual patients that preserve realistic correlations and whole-body coherence. Graph networks and neural ordinary differential equations learn continuous scaling and maturation relationships instead of relying on fixed exponents. Physics-informed networks and surrogate emulators estimate mechanistic parameters and quantify uncertainty in data-poor pediatric settings. This review surveys these developments, organizes them by the PBPK workflow they serve, and links them to precision dosing of three antiepileptic drugs, lacosamide, lamotrigine, and oxcarbazepine. Unlike earlier surveys focused on concentration prediction and therapeutic drug monitoring, we keep the mechanistic model in the foreground, treating the virtual patient as the unit of simulation. We also discuss open problems, including validation standards for generated populations, the regulatory qualification of learning-enhanced PBPK, uncertainty reporting, and the prospect of foundation models of pharmacokinetics. Throughout, we distinguish methodological maturity from demonstrated clinical benefit.