Multi-Task ADME/PK prediction at industrial scale: leveraging large and diverse experimental datasets

Area of Science:

• Drug discovery and development
• Computational chemistry
• Pharmacokinetics

Background:

• Absorption, Distribution, Metabolism, and Excretion (ADME) properties are crucial for drug candidate success.
• Pharmacokinetic (PK) profiles determine a drug's behavior in the body.
• Predicting ADME and PK properties early is vital for efficient drug development.

Purpose of the Study:

• To evaluate multi-task machine learning (ML) models for predicting ADME and animal PK endpoints.
• To compare the performance of multi-task models against single-task models.
• To understand how leveraging in-house data across multiple endpoints impacts predictive accuracy.

Main Methods:

• Development and application of multi-task graph-based neural network models.
• Training models on Boehringer Ingelheim's internal ADME/PK data.
• Evaluation of models using realistic time-splits, considering both design and testing stages.

Main Results:

• Multi-task ML models demonstrated superior performance compared to single-task models.
• The benefit of multi-task learning was more pronounced when experimental data from earlier assays was available.
• Endpoints with more extensive data, such as physicochemical properties and microsomal clearance, contributed to improved predictivity of complex ADME/PK endpoints.

Conclusions:

• Multi-task learning effectively leverages data for multiple ADME/PK endpoints in pharmaceutical settings.
• Graph-based neural networks show strong potential for enhancing ADME/PK predictions.
• Optimizing the use of internal data across assays can significantly boost the predictivity of ML models in drug discovery.