Machine learning to predict the specific optical rotations of chiral fluorinated molecules

Area of Science:

• Computational chemistry
• Chemoinformatics
• Stereochemistry

Background:

• Chiral molecules exhibit optical activity, crucial for drug efficacy and material properties.
• Accurate prediction of optical rotation is essential for characterizing enantiomers.
• Existing methods for predicting optical rotation have limitations.

Purpose of the Study:

• To develop and validate a chemoinformatics method for assigning absolute configurations.
• To quantitatively predict specific optical rotations of chiral fluorinated molecules.
• To compare the predictive performance against quantum chemistry calculations.

Main Methods:

• Utilized a dataset of 88 chiral fluorinated molecules (44 enantiomeric pairs).
• Employed counterpropagation neural networks for enantiomer classification (dextrorotatory/levorotatory).
• Trained regression models (MLP, RF, MLR) using physicochemical atomic stereo (PAS) and common physicochemical atomic stereo (cPAS) descriptors.

Main Results:

• Random forests (RF) with cPAS descriptors achieved high prediction accuracy (R=0.964, MAE=9.8°, RMSE=12.5°) in cross-validation.
• Predictions for compounds in chloroform showed excellent correlation (R=0.971, MAE=9.1°, RMSE=12.5°).
• The chemoinformatics approach demonstrated comparable or superior performance to quantum chemistry methods.

Conclusions:

• The developed chemoinformatics method is effective for predicting optical rotations of chiral fluorinated compounds.
• Physicochemical atomic stereo (PAS) and cPAS descriptors provide valuable information for stereochemical predictions.
• This computational approach offers a viable alternative to experimental measurements and complex quantum calculations.