Molecular partition coefficient from machine learning with polarization and entropy embedded atom-centered symmetry...

Area of Science:

• Computational chemistry
• Physical chemistry

Background:

• Predicting partition coefficients (log P) is crucial for drug and materials design.
• Accurate log P prediction can significantly reduce development time and costs.

Purpose of the Study:

• To develop a novel descriptor, 〈q - ACSFs〉conf, for efficient and accurate prediction of partition coefficients.
• To incorporate explicit polarization effects and conformational entropy into the prediction model.

Main Methods:

• Developed the 〈q - ACSFs〉conf descriptor by embedding partial charges into atom-centered symmetry functions (ACSFs).
• Accounted for entropic effects by averaging over conformations weighted by their Boltzmann distribution.
• Trained a high-dimensional neural network (HDNN) model on the PhysProp dataset (41,039 samples).

Main Results:

• Achieved satisfactory log P prediction performance on independent datasets: Martel (707 molecules), Star & Non-Star (266), and Huuskonen (1870).
• Demonstrated applicability to n-carboxylic acids (C2-C14) and 54 organic solvents.
• The model provides transferable, atom-based partition coefficients for various systems.

Conclusions:

• The 〈q - ACSFs〉conf descriptor effectively captures polarization and entropy, leading to accurate log P predictions.
• This method offers a transferable and efficient approach for predicting partition coefficients in diverse chemical systems.
• The developed model has the potential to accelerate the design cycle for drugs and materials.