Stochastic Gates for Covariate Selection in Population Pharmacokinetics Modeling

Area of Science:

• Pharmacometrics
• Machine Learning
• Computational Biology

Background:

• Covariate selection is critical in population pharmacokinetics (PPK) for understanding drug variability and optimizing dosage.
• Traditional methods like stepwise covariate modeling are time-consuming and can be inefficient.
• Machine learning offers potential for automated and more efficient covariate selection.

Purpose of the Study:

• To investigate the efficacy of neural networks with stochastic gates for automated covariate selection in PPK.
• To assess the method's performance in identifying relevant covariates while preventing overfitting.
• To evaluate its robustness across synthetic and real-world clinical datasets.

Main Methods:

• Development and application of a neural network model with stochastic gates for covariate selection.
• Testing on diverse synthetic datasets with varying complexities (e.g., correlations, low frequencies, high variability).
• Validation using real clinical data from monalizumab and tixagevimab/cilgavimab studies.

Main Results:

• The neural network approach demonstrated robustness in identifying important covariates on synthetic data, handling complex dependencies.
• It successfully identified expert-validated covariates in the monalizumab clinical dataset.
• For tixagevimab/cilgavimab, it identified a broader set of covariates, suggesting potential for further refinement.

Conclusions:

• Machine learning, specifically neural networks with stochastic gates, significantly enhances the covariate preselection process in PPK.
• This automated method offers substantial time savings and improved efficiency, even with challenging data.
• The approach shows promise for streamlining population pharmacokinetic model development.