Dimensionally reduced machine learning model for predicting single component octanol-water partition coefficients.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Environmental Science

Background:

• Octanol-water partition coefficients (LogP) are crucial for environmental fate and drug delivery predictions.
• Current LogP prediction methods rely on experimental data or complex structural information.
• Existing methods can be computationally intensive and require detailed molecular structures.

Purpose of the Study:

• To introduce MF-LOGP, a novel method for predicting single-component octanol-water partition coefficients.
• To develop a predictive model that utilizes only molecular formulas as input.
• To provide a computationally inexpensive and automatable alternative for LogP prediction.

Main Methods:

• MF-LOGP employs a random forest algorithm.
• The model is trained on 15,377 data points using 10 features derived from molecular formulas.
• Performance was validated on an independent set of 2,713 data points.

Main Results:

• MF-LOGP achieved an average LogP prediction accuracy of R² = 0.77, Q² = 0.52, and R²_pred = 0.83.
• The model's performance is comparable to existing, more complex methods.
• MF-LOGP requires minimal input features and no structural data.

Conclusions:

• MF-LOGP offers a practical and predictive tool for estimating octanol-water partition coefficients.
• The method is particularly useful when molecular structures are unknown or rapid predictions are needed.
• This work lays the foundation for advanced prediction models using big data analytics.