Benefit of Retraining pKa Models Studied Using Internally Measured Data

Area of Science:

• Computational chemistry
• Drug discovery and development
• Medicinal chemistry

Background:

• Drug ionization state critically impacts pharmaceutical properties including solubility, permeability, and biological activity.
• Understanding the structure-property relationship for the acid-base dissociation constant (pKa) is vital for informed drug design during lead optimization.
• Computational tools like MoKa aid in pKa prediction, but often exhibit significant errors, particularly for proprietary compounds.

Purpose of the Study:

• To investigate the impact of model retraining on improving prediction accuracy for drug pKa.
• To assess computational model degradation over time in a real-world drug discovery setting.
• To establish the necessity of regular retraining to maintain prediction accuracy.

Main Methods:

• A longitudinal study spanning 15 years of data from a drug discovery environment.
• Utilizing the MoKa software for pKa prediction and analysis.
• Evaluating prediction errors and model performance over time with and without retraining.

Main Results:

• Retraining computational models substantially reduces prediction errors for drug pKa.
• Model performance degrades over a period of six to nine months due to evolving chemical space.
• Regular retraining is demonstrated to be essential for maintaining high prediction accuracy.

Conclusions:

• Computational pKa prediction models require regular retraining to counteract degradation caused by changes in chemical space.
• Proactive model maintenance through retraining is key to reliable predictions in drug discovery.
• Implementing a retraining strategy enhances the utility of computational tools for optimizing drug design.