Exploring the octanol-water partition coefficient dataset using deep learning techniques and data augmentation.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Environmental Chemistry and Toxicology

Background:

• Increasing availability of large chemical datasets fuels the adoption of deep neural networks (DNNs).
• The octanol-water partition coefficient (log P) is a critical property in environmental chemistry, toxicology, and chemical analysis.
• Accurate prediction of log P from chemical structures is essential for various applications.

Purpose of the Study:

• To explore the potential of DNNs for predicting chemical properties, using log P as a case study.
• To develop and evaluate a DNN model for accurate log P prediction.
• To investigate the utility of DNNs in curating and improving existing log P datasets.

Main Methods:

• Development of a deep neural network (DNN) model for predicting the octanol-water partition coefficient (log P).
• Training the DNN using data augmentation, incorporating all potential tautomeric forms of chemical compounds.
• Evaluation of the DNN model's predictive performance on internal test datasets and an external dataset from the SAMPL6 challenge.

Main Results:

• The developed DNN achieved a root-mean-square error (RMSE) of 0.47 log units on the test dataset.
• The DNN demonstrated strong performance on an external dataset from the SAMPL6 challenge, with an RMSE of 0.33 log units.
• The study highlights the DNN model's capability to identify potential errors within log P datasets, aiding in data curation.

Conclusions:

• Deep neural networks are highly effective for predicting the octanol-water partition coefficient (log P) from chemical structures.
• The DNN approach, including data augmentation with tautomers, provides accurate and reliable log P predictions.
• DNN models offer a valuable tool for enhancing the quality and reliability of chemical property datasets.