Prediction of Human Clearance Using In Silico Models with Reduced Bias.

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry and Cheminformatics
• Toxicology and Preclinical Studies

Background:

• Accurate prediction of human clearance is crucial for efficient drug development, reducing costs and animal testing.
• Existing in silico methods often struggle with general applicability due to structural similarities in training data.
• A large dataset of human pharmacokinetic (PK) data is essential for robust model development.

Purpose of the Study:

• To develop and validate robust random forest (RF) models for predicting human clearance using a large PK dataset.
• To assess the general applicability of RF models by excluding structurally similar compounds from training sets.
• To evaluate the predictive performance and confidence intervals using conformal prediction.

Main Methods:

• Development of RF models using 1340 compounds with human intravenous PK data.
• Systematic removal of structural analogues and similar compounds from training sets to test generalizability.
• Validation on a quasi-prospective test set of 343 compounds.
• Application of conformal prediction for model applicability assessment and confidence interval determination.

Main Results:

• RF models, even when excluding structurally similar compounds, predicted human clearance with a geometric mean fold error (GMFE) of 3.3.
• Predictive accuracy was higher for renally cleared compounds (GMFE of 2.3) when a classification model was applied.
• The developed in silico RF models demonstrated performance comparable to scaling from human hepatocytes or preclinical in vivo data.

Conclusions:

• Random forest models provide a realistic and broadly applicable approach for predicting human clearance.
• Excluding structural analogues from training data yields more reliable prospective predictions.
• Conformal prediction enhances model applicability assessment and provides crucial confidence intervals for predictions.