ClassicalGSG: Prediction of log P using classical molecular force fields and geometric scattering for graphs

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Predicting the partition coefficient (log P) is crucial for drug development and chemical property assessment.
• Traditional methods often rely on complex simulations or less informative molecular descriptors.

Purpose of the Study:

• To evaluate the performance of a new method, ClassicalGSG, for predicting log P values of small molecules.
• To compare ClassicalGSG against established graph convolutional network (GCN) models.

Main Methods:

• Utilized atomic attributes (radius, partial charge) from force fields as input.
• Transformed atomic attributes into index-invariant molecular features using geometric scattering for graphs (GSG).
• Trained neural network models on the OpenChem dataset (10,722 molecules) and validated on four independent test sets.

Main Results:

• ClassicalGSG demonstrated strong predictive performance for log P.
• Optimal results were achieved using atomic attributes from the CHARMM generalized force field.
• Employing 2D molecular structures alongside atomic attributes further enhanced prediction accuracy.

Conclusions:

• ClassicalGSG offers a robust and accurate approach for log P prediction.
• The method's reliance on fundamental atomic properties and 2D structures makes it a valuable tool in cheminformatics.
• This approach shows promise for accelerating drug discovery and chemical property prediction.