Prediction of aqueous solubility from SCRATCH

Area of Science:

• Computational Chemistry
• Drug Discovery
• Physical Chemistry

Background:

• Accurate prediction of aqueous solubility is crucial for drug development and environmental assessments.
• Existing methods often rely on experimental data, which can be time-consuming and costly.
• Structure-based prediction models offer a valuable alternative for rapid solubility estimation.

Purpose of the Study:

• To introduce the SCRATCH model for predicting the aqueous solubility of compounds directly from their chemical structures.
• To evaluate the performance of the SCRATCH model against established methods like the General Solubility Equation (GSE).
• To demonstrate the utility of a purely structure-based approach for solubility prediction, eliminating the need for experimental measurements.

Main Methods:

• The SCRATCH model utilizes predicted melting points and aqueous activity coefficients.
• It incorporates additive (enthalpy of melting, aqueous activity coefficient) and non-additive (symmetry, flexibility) molecular descriptors.
• Model training involved datasets of over 2200 compounds for melting point prediction and approximately 1640 for activity coefficients, with validation via 10-fold cross-validation on 883 compounds for solubility.

Main Results:

• The SCRATCH model provides a rigorous and robust estimation of aqueous solubility based on molecular structure.
• While exhibiting a slightly higher average absolute error compared to the GSE, SCRATCH bypasses the requirement for experimental melting point data.
• The model successfully predicts aqueous solubility using only structural information and predicted parameters.

Conclusions:

• The SCRATCH model offers a viable, structure-only approach for predicting aqueous solubility.
• It overcomes the limitations of methods requiring experimental data, making it a valuable tool in computational chemistry and drug design.
• Further development could refine accuracy while maintaining the advantage of structure-based prediction.