Predictive Multitask Deep Neural Network Models for ADME-Tox Properties: Learning from Large Data Sets.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Pharmacokinetics and pharmacodynamics modeling

Background:

• Drug discovery necessitates optimizing compound properties for pharmacokinetics, pharmacodynamics, and safety.
• Deep neural networks (DNNs) offer advanced capabilities for analyzing complex ADME-Tox data.
• Existing ADME-Tox databases are expanding, requiring sophisticated analytical tools.

Purpose of the Study:

• To industrialize and optimize the application of DNNs for modeling ADME-Tox properties.
• To investigate the impact of hyperparameters and molecular descriptors on DNN model performance.
• To develop and validate a novel multitask DNN approach for predicting ADME-Tox parameters.

Main Methods:

• Development of a fully industrialized approach for DNN setup, training, and interpretation.
• Utilizing large datasets (up to 50,000 compounds) from public and corporate ADME-Tox databases.
• Implementing both single-task and multitask DNN architectures for property prediction.

Main Results:

• Multitask DNN models demonstrated statistically superior performance compared to single-task models across various ADME-Tox datasets.
• Predictive accuracy for human metabolic lability improved from R² of 0.6 to 0.7 using multitask DNNs.
• A novel visualization technique, 'response map,' was introduced for interpreting DNN models and guiding drug design.

Conclusions:

• A scalable, industrialized DNN approach, particularly multitask learning, significantly enhances ADME-Tox property prediction.
• The 'response map' visualization tool aids in understanding structure-property relationships for rational drug design.
• This methodology provides a powerful framework for optimizing compound properties in drug discovery programs.