Integrating multiscale and machine learning approaches towards the SAMPL9 log P challenge

Area of Science:

• Computational chemistry
• Drug discovery and development

Background:

• The partition coefficient (log P) is a key physicochemical property influencing pharmacokinetics, toxicity, and bioavailability.
• Accurate prediction of log P can significantly accelerate the drug design process.

Purpose of the Study:

• To evaluate current computational methods for predicting the toluene-water partition coefficient (log P_tol/w).
• To compare the performance of quantum mechanics (QM), molecular mechanics (MM), and machine learning (ML) approaches in a blind prediction challenge.

Main Methods:

• Three distinct computational approaches were employed: QM, MM, and ML.
• The study involved a blind prediction challenge (SAMPL-9) focused on calculating the log P_tol/w for sixteen drug molecules.

Main Results:

• Initial blind submissions showed mean unsigned errors (MUE) between 1.53-2.93 log P_tol/w units.
• QM methods, particularly DFT with a triple-ζ basis set, achieved a reduced MUE of 1.00 log P_tol/w.
• MM and ML methods performed well on smaller molecules but struggled with larger ones compared to DFT.

Conclusions:

• Density Functional Theory (DFT) coupled with a triple-ζ basis set emerged as the most effective and straightforward method for accurate partition coefficient prediction.
• The findings highlight the strengths and limitations of different computational approaches for log P calculations in drug discovery contexts.