A NON-LINEAR STRUCTURE-PROPERTY MODEL FOR OCTANOL-WATER PARTITION COEFFICIENT

Area of Science:

• Computational chemistry
• Physical chemistry
• Drug discovery

Background:

• The octanol-water partition coefficient (Kow) is crucial for understanding chemical behavior in biological and environmental systems.
• Accurate Kow prediction is vital for pharmacology, pharmacokinetics, and environmental toxicology.
• Existing models may lack precision for novel molecular structures.

Purpose of the Study:

• To develop a novel, non-linear quantitative structure-property relationship (QSPR) model for predicting molecular Kow values.
• To utilize artificial neural networks (ANNs) and feature selection for enhanced prediction accuracy.
• To create a robust computational tool for in silico Kow determination.

Main Methods:

• Generated 823 molecular descriptors for 11,308 molecules from the PhysProp dataset.
• Employed a wrapper-based feature selection algorithm combining differential evolution and ANNs to optimize model inputs and architecture.
• Developed a neural network ensemble by averaging predictions from five ANNs with identical architecture but varied weights.

Main Results:

• Identified an optimal ANN architecture (50-33-35-1) yielding minimal root-mean-squared error (RMSE) on the training set.
• The neural network ensemble achieved an RMSE of 0.28 for the training set and 0.38 for internal validation.
• The ensemble model demonstrated competitive performance against established Kow prediction models on an external dataset.

Conclusions:

• The developed non-linear QSPR model, particularly the neural network ensemble, provides accurate in silico prediction of Kow values.
• This computational approach offers a valuable tool for researchers in pharmacology, environmental science, and chemical development.
• The study highlights the efficacy of combining advanced feature selection with ensemble neural networks for predicting physicochemical properties.