pH-dependent solubility prediction for optimized drug absorption and compound uptake by plants

Area of Science:

• Computational chemistry
• Pharmacokinetics
• Agrochemistry

Background:

• Aqueous solubility is critical for the bioavailability of drug and agrochemical candidates.
• Predicting solubility across various pH conditions (acidic, neutral, basic) is essential for understanding compound behavior in biological systems.
• Existing methods often lack direct prediction of pH-dependent solubility.

Purpose of the Study:

• To develop the first direct machine learning models for predicting pH-dependent aqueous solubility.
• To create a valuable tool for profiling compounds in pharmaceutical and agrochemical research.
• To improve the prediction accuracy of solubility across different pH environments relevant to biological systems.

Main Methods:

• Combined nearly 300,000 data points from 11 solubility assays and over one million data points from lipophilicity and melting point experiments.
• Utilized multi-task neural networks with ECFP-6 fingerprints, trained on data categorized into acidic, neutral, and basic pH classes.
• Incorporated five helper tasks alongside three primary solubility prediction tasks.

Main Results:

• Multi-task neural networks significantly outperformed baseline random forests and single-task networks.
• Achieved an overall root mean square error (RMSE) of 0.56 log units and a Spearman rank correlation of 0.83.
• The model demonstrated strong performance across pH classes: acidic (RMSE 0.61, Spearman 0.78), neutral (RMSE 0.52, Spearman 0.86), and basic (RMSE 0.54, Spearman 0.86).

Conclusions:

• The developed machine learning model provides accurate predictions of compound pH profiles.
• This tool is valuable for early-stage profiling of drug and agrochemical candidates.
• The model enables more efficient and effective research and development in the pharmaceutical and agrochemical industries.