Improving Absorption, Distribution, Metabolism, and Excretion Property Predictions by Integrating Public and...

Area of Science:

• Computational chemistry
• Pharmacokinetics
• Drug discovery

Background:

• Accurate prediction of compound properties is vital for efficient drug discovery.
• Pharmacokinetic (ADME) properties are key determinants of drug efficacy and safety.

Purpose of the Study:

• To assess data-integration strategies for improving in silico ADME predictions.
• To evaluate the performance of single-source, pooled single-task, and multitask learning models for ADME prediction.

Main Methods:

• Leveraged internal and public high-quality ADME data.
• Developed and compared single-source, pooled single-task, and multitask learning models.
• Conducted applicability domain and prediction uncertainty analyses.

Main Results:

• Combined data, particularly with multitask learning, generally outperformed single-source models.
• Multitask learning demonstrated improved generalization and reduced errors for similar compounds.
• Proportional balancing of public and in-house data enhanced model performance.

Conclusions:

• Curated integration of public and proprietary ADME data yields more accurate in silico models.
• Multitask learning enhances the predictive power and applicability domain of ADME models.
• Improved ADME models support more effective computational compound design in drug discovery.