Hybrid mechanistic-machine learning PK/PD models with digital biomarkers: from cage to clinic.

Background: Pharmacometric PK/PD models remain central to dose selection for first-in-human studies, therapeutic drug monitoring, and animal-to-human extrapolation of exposure-response relationships, but modern datasets, including digital biomarkers from continuous monitoring, omics, and imaging, challenge traditional modeling assumptions. Hybrid mechanistic-machine learning (ML) approaches offer a structured way to combine causal pharmacological frameworks with data-driven flexibility. Methods: This review surveys the emerging landscape of hybrid mechanistic-ML PK/PD modeling with a focus on digital biomarker integration. We examine model architectures, covariate discovery methods, cross-species scaling strategies, and validation practices relevant to preclinical-to-clinical extrapolation. Literature was identified through PubMed, Scopus, and Web of Science using search terms combining pharmacokinetics, pharmacodynamics, machine learning, hybrid modeling, digital biomarkers, and cross-species pharmacology. Studies published between 2015 and 2025 were prioritized, with inclusion of foundational earlier work where necessary. Results: Hybrid approaches improve individual clearance estimation, covariate discovery, and cross-species PK scaling in defined settings, particularly when dense time-series data are available and the mechanistic model is structurally sound but parametrically under-identified. Integration of multimodal datasets introduces practical challenges around missingness, device drift, and data leakage that require explicit mitigation. Application examples span rodents, dogs, non-human primates, and minipigs, where continuous digital measures strengthen exposure-response inference. Discussion: The credibility of hybrid models depends on validation rigor, interpretability, and regulatory alignment rather than on algorithmic novelty. When mechanistic models fit well and sample sizes are small, adding an ML layer risks overfitting without measurable gain. Models that are prospectively tested, transparently documented, and fitted to a defined context of use will be more influential than those that are merely complex. We outline reporting and audit practices to support robustness, reproducibility, and regulatory review.