Machine Learning for Fast, Quantum Mechanics-Based Approximation of Drug Lipophilicity.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Lipophilicity, measured by log P, is crucial for drug discovery but experimentally challenging for certain compounds.
• Existing in silico models struggle with data scarcity for these challenging compounds.
• Quantum mechanics (QM) methods offer accurate log P estimation but are computationally expensive.

Purpose of the Study:

• To develop computationally affordable machine learning models for estimating drug lipophilicity.
• To supplement or replace expensive QM calculations for log P prediction.
• To enable rapid, interactive ideation in drug discovery projects.

Main Methods:

• Trained various machine learning models (Random Forest, Lasso, XGBoost, Chemprop, Chemprop3D) on calculated log P values.
• Utilized both public and in-house datasets for model training and evaluation.
• Employed scaffold splitting for robust assessment of model generalizability.

Main Results:

• The message-passing neural network model, Chemprop, demonstrated superior performance.
• Chemprop achieved mean absolute errors of 0.44 and 0.34 log units on public and in-house test sets, respectively.
• Learning curve analysis indicated potential for further error reduction with increased training data.

Conclusions:

• Machine learning models, especially Chemprop, provide a cost-effective alternative for log P estimation.
• These models can handle compounds beyond experimental quantification limits.
• Potential applications include pre-screening large compound libraries and prioritizing candidates for QM calculations.