Merck & Co., Inc. Norristown, Pennsylvania, United States
Background:
Background: Pharmacokinetic (PK) samples play a critical role in clinical studies providing valuable insights into the drug's fate. Demographics and covariates collected in clinical studies contribute to the understanding of variability in PK. We aimed to develop an algorithm to impute missing PK concentrations utilizing plasma concentrations and covariate data. Methods:
Methods: We simulated PK profiles for 400 subjects using mrgsolve, incorporating interindividual variabilities, random errors, and effects of significant covariates. We used gradient boosting machine learning to develop a model for PK concentration prediction which underwent tenfold cross-validation and grid search optimization of hyperparameter combinations. The model's performance was evaluated on a test dataset through dose and covariate inputs. Shapley tool was used to explain the relationship between covariates and feature importance. Results:
Results: Our machine learning model achieved an RMSE of 0.49 and an R2 of 0.91 on the training dataset and RMSE of 0.64 and R2 of 0.83 on the testing dataset. Shapley plots showed covariates such as body weight, race, gender, steady states, and albumin levels had significant impacts on PK profiles. Conclusion:
Conclusion: Our gradient boosting-based ML algorithm can predict drug PK profiles with reasonable accuracy for a simulated dataset. Future work will involve comparing its predictions with predictions from classic POPPK models.
