A machine learning approach to population pharmacokinetic modelling automation.

Background: Population pharmacokinetic (PopPK) models are crucial for understanding drug behaviour across populations, yet traditional development is often labour-intensive and slow. This study demonstrates an automated, out-of-the-box approach for popPK model development, leveraging optimisation algorithms implemented using pyDarwin to efficiently handle a diverse range of extravascular drugs.

Methods: We proposed a generic model search space for drugs with extravascular administration and developed a penalty function to discourage over-parameterisation whilst ensuring plausible parameter values. Optimisation within the model search space was conducted using pyDarwin, employing Bayesian optimisation with a random forest surrogate combined with exhaustive local search. This approach was evaluated on one synthetic and four clinical datasets, with results compared to manually developed models.

Results: Here we show that the automated approach reliably identifies model structures comparable to manually developed expert models in less than 48 h on average (40-CPU, 40 GB environment) while evaluating fewer than 2.6% of the models in the search space. Ablation experiments demonstrate the importance of our penalty function in selecting plausible models, and the benefit of global search algorithms in avoiding local minima.

Conclusions: These results demonstrate that a single penalty function and model space can be used within the pyDarwin framework to automatically identify model structures for a diverse range of drugs. By reducing the configuration required at search setup, this approach simplifies the process, potentially making the technology more accessible to users. Adoption of automatic model search can accelerate popPK analysis, improve model quality, increase reproducibility, and reduce manual effort for modellers.