Deep-Learning: An Emerging Tool to Support Model-Informed Drug Development

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Traditional Model-Informed Drug Development (MIDD) utilizes hypothesis-driven (HD) mechanistic models.
• These models, while valuable, may struggle with complex nonlinear relationships.
• Deep learning (DL) offers a data-driven alternative for capturing intricate patterns.

Purpose of the Study:

• To compare the performance of DL and HD models in drug development.
• To evaluate if DL can augment, not replace, conventional mechanistic modeling.
• To assess the integration of DL and HD within the MIDD framework.

Main Methods:

• Implemented DL using artificial neural networks based on the universal approximation theorem.
• Utilized two independent datasets for comparative analysis.
• Performed rigorous validation including train/test splits, bootstrap analyses, and diagnostic evaluations (e.g., residual analyses, visual predictive checks).

Main Results:

• DL models demonstrated robust predictive performance and reliability comparable to HD models.
• DL approaches required fewer a priori assumptions than traditional HD models.
• The study confirmed that DL models can effectively complement established HD methods.

Conclusions:

• Mechanistic HD models and data-driven DL models are complementary tools in drug development.
• Integrating DL and HD within MIDD leverages the strengths of both approaches.
• This integration enhances predictive flexibility and biological interpretability in drug development.