Prediction of pK_a Using Machine Learning Methods with Rooted Topological Torsion Fingerprints: Application to...

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• The acid-base dissociation constant (pKa) is crucial for determining a compound's ionization state.
• Ionization state significantly impacts a molecule's biopharmaceutical properties and behavior.
• Accurate pKa prediction is essential for drug design and development.

Purpose of the Study:

• To develop a novel computational approach for predicting pKa values.
• To utilize rooted topological torsion fingerprints and machine learning (ML) for pKa prediction.
• To create accurate predictive models for aliphatic amines.

Main Methods:

• Employed rooted topological torsion fingerprints as molecular descriptors.
• Applied five machine learning algorithms: random forest, partial least squares, extreme gradient boosting, lasso regression, and support vector regression.
• Trained and validated models using a large dataset of 14,499 experimental pKa values and an external test set of 726 values.

Main Results:

• Achieved consistently good prediction statistics across multiple ML models.
• The top-performing model demonstrated high accuracy with RMSE of 0.45, MAE of 0.33, and R² of 0.84 on an external test set.
• Successfully generated accurate prospective pKa predictions for aliphatic amines.
• Identified and assessed factors influencing prediction accuracy and model applicability.

Conclusions:

• Rooted topological torsion fingerprints combined with ML methods offer a robust and promising strategy for accurate pKa prediction.
• This approach can aid in the rational design of drug candidates by predicting key physicochemical properties.
• The developed models provide valuable tools for computational drug discovery and development.