Multitask Deep Learning Models of Combined Industrial Absorption, Distribution, Metabolism, and Excretion Datasets to...

Area of Science:

• Drug discovery and development
• Computational chemistry
• Pharmacokinetics

Background:

• Optimization of absorption, distribution, metabolism, and excretion (ADME) profiles is crucial for successful drug discovery.
• Machine learning (ML) models are vital for prioritizing compound design, but their effectiveness relies on diverse, high-quality experimental data.
• Expanding the chemical space explored by ML models is essential for identifying novel drug candidates.

Purpose of the Study:

• To evaluate the impact of expanding chemical space on ML model performance for ADME profiling.
• To assess the utility of combining large-scale, historical ADME datasets from different sources.
• To investigate the generalization capacity of multitask (MT) neural networks across disparate data sources.

Main Methods:

• Combined ADME datasets from Genentech and Roche, comprising over 1 million measurements across 11 endpoints.
• Utilized a multitask (MT) neural network architecture for simultaneous modeling of multiple ADME endpoints.
• Trained and compared single-site, single-task baseline models against cross-site MT models, treating data from different sites as separate tasks.
• Evaluated model performance on cluster-based, temporal, and external test sets to assess generalization.

Main Results:

• Cross-site MT models demonstrated superior generalization capacity compared to single-site models.
• Performance improvements were more significant for distant test sets (external and temporal), indicating an expanded applicability domain.
• The study validated the value of leveraging ADME data from multiple sources without direct aggregation, even with disparate experimental methods.

Conclusions:

• Combining ADME data from multiple sources enhances the predictive power and generalization of ML models.
• Cross-site multitask learning effectively expands the applicability domain of ADME models.
• This approach facilitates more robust compound prioritization in drug discovery by learning from diverse chemical spaces.