Multispecies Machine Learning Predictions of In Vitro Intrinsic Clearance with Uncertainty Quantification Analyses.

Area of Science:

• Computational chemistry and drug discovery
• Pharmacokinetics and drug metabolism

Background:

• Optimizing metabolic stability is crucial in pharmaceutical research to prevent rapid drug elimination.
• Intrinsic clearance (CLint) in liver microsomes or hepatocytes is a key metric for lead optimization.

Purpose of the Study:

• To develop machine learning models for predicting intrinsic clearance (CLint) and metabolic stability from compound structures.
• To create a multispecies machine learning model for simultaneous CLint prediction across six species.

Main Methods:

• A multitask (MT) learning architecture, specifically a graph neural network (MT-GNN), was employed.
• An ensemble of 10 MT-GNN models was developed and prospectively evaluated.
• Systematic quantification of uncertainty in experimental values and model predictions was performed.

Main Results:

• The MT-GNN model achieved geometric mean fold errors consistently below 2-fold.
• High precision was achieved in predicting high (>300 μL/min/mg) and low (<100 μL/min/mg) CLint compounds (75-97% precision).
• MT-GNN model performance approached the experimental variability of in vitro clearance assays.

Conclusions:

• The developed multispecies deep learning model significantly improves CLint predictions.
• Uncertainty quantification aids in identifying reliable predictions and facilitates informed decisions in early drug development.
• This approach supports efficient compound prioritization by accurately assessing metabolic stability.