Adapting Deep Learning QSPR Models to Specific Drug Discovery Projects

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Machine learning (ML) models predict molecular properties from structure, aiding drug design.
• Global ML models trained on diverse data are widely used but may lack specificity for focused drug discovery projects.
• Local ML models offer project-specific predictions but require substantial data.

Purpose of the Study:

• To benchmark global, local, and hybrid machine learning strategies for predicting molecular properties in drug discovery.
• To develop and evaluate hybrid global-local strategies using transfer learning to adapt model predictions.
• To assess the performance of these strategies across numerous drug discovery projects and ADME assays.

Main Methods:

• Benchmarking of ML-based global, local, and hybrid models.
• Development of hybrid strategies using transfer learning to combine historical (global) and project-specific (local) data.
• Fine-tuning pretrained global ML models for weight initialization (WI) as a key hybrid approach.

Main Results:

• Hybrid global-local strategies, particularly fine-tuning pretrained models (WI), outperformed standalone global and local models.
• Average improvements in mean absolute error were 16% for global and 27% for local models.
• Weight initialization fine-tuning demonstrated effectiveness even in low-data scenarios (approximately 10 molecules per project).

Conclusions:

• Domain adaptation through hybrid ML strategies significantly refines molecular property predictions.
• Fine-tuning pretrained models offers a powerful approach to leverage existing knowledge for new drug discovery projects.
• This method enhances predictive accuracy and is robust even with limited project-specific data.