PK/PD Modeling
AI-driven parameter estimation and simulation
Intended for the following roles:
Pharmacokineticists /Pharmacometricians: use AI to accelerate simulations and explore covariate relationships
Preclinical modelers /Data scientists: build hybrid mechanistic + ML models
Study directors /translational scientists: evaluate model outputs for dose-selection or translational predictions
Regulatory operations /QA: validate model pipelines for GLP and submission documentation
ML methods are being investigated as complements to mechanistic pharmacokinetic and pharmacodynamic (PK/PD) models for parameter estimation, covariate discovery, surrogate modeling, and acceleration of simulation tasks. The literature emphasizes hybrid approaches that preserve mechanistic interpretability while leveraging ML for parts of the pipeline.
Challenges
Traditional population PK/PD and physiologically-based PK models can be computationally intensive, require manual feature engineering, and sometimes fail to capture unmodeled sources of variability. ML can assist in identifying nonlinear covariate relationships, approximating expensive model evaluations, and improving predictive accuracy in well-defined regimes. However, ML-only models lack mechanistic guarantees and interpretability required for regulatory decision making.
AI Techniques
Hybrid frameworks combining mechanistic models (PBPK or compartmental models) with ML components for parameter priors, surrogate models, or residual learning. Supervised learning for parameter prediction from chemical or physiological descriptors and probabilistic ML for uncertainty quantification are common motifs.
Published examples and programs
Reviews on machine learning in pharmacometrics outline opportunities and caution about replacing established population methods. BPS Publications
Studies demonstrating ML as surrogate models or in combination with PBPK to accelerate simulation and support covariate discovery. PubMed Central, ACS Publications
Before adopting these methods…
Treat ML outputs as adjuncts to, not replacements for, validated mechanistic models in regulatory contexts. Evidence and uncertainty quantification must be explicit
Randomized or stratified splitting and external validation are necessary to avoid overfitting when predicting PK/PD parameters
Document training data provenance, covariate distributions, and the limits of interpolation versus extrapolation
Pilot Validation Checklist
Define intended use: parameter estimation, surrogate simulation, or covariate discovery
Use historical studies to train and reserve independent trials for external validation
Compare ML-augmented workflows to established population PK/PD benchmarks and quantify gains and failure modes
Provide uncertainty estimates and stress-test models under covariate shifts
Keep mechanistic model available for counterfactual and safety assessments
Experiment design patterns
Start by using ML to accelerate non-safety-critical simulations or to prioritize covariates for mechanistic modeling
Use ML surrogates for exploratory analyses, and retrain/verify with prospective data before operationalizing
Resources
McComb et al., Machine learning in pharmacometrics: opportunities and challenges. accp1.org, BPS Publications
Hybrid ML and mechanistic modeling examples in the literature. PubMed Central, ACS Publications
