Comparison of logP and logD correction models trained with public and proprietary data sets

Area of Science:

• Computational chemistry
• Drug discovery and development
• Quantitative Structure-Activity Relationship (QSAR) modeling

Background:

• Lipophilicity metrics like logP and logD are crucial for assessing drug candidates' bioactivity and bioavailability.
• Existing methods for logP and logD prediction have limitations, including systematic errors and difficulties with pKa estimation for ionizable compounds.

Purpose of the Study:

• To develop an improved machine learning QSAR model for predicting logD, overcoming limitations of current commercial software.
• To enhance the accuracy and applicability domain of lipophilicity predictions in drug discovery.

Main Methods:

• An integrated machine learning QSAR approach was employed, training models with experimental logD data.
• ClogP and pKa values predicted by commercial software were used as descriptors.
• A correction model was built by optimizing the loss function for software-calculated logD, incorporating both software descriptors and experimental logD data.

Main Results:

• The proposed approach demonstrated improved logD and logP predictions compared to commercial software.
• Models trained on publicly available data showed enhanced performance when applied to other datasets.
• The method extended the domain of applicability for lipophilicity predictions.

Conclusions:

• The integrated machine learning QSAR approach offers a significant improvement for predicting molecular lipophilicity (logD and logP).
• This method enhances the reliability of computational predictions, supporting more effective drug discovery pipelines.
• The approach provides a valuable tool for researchers seeking accurate and broadly applicable lipophilicity assessments.