Bridging the Gap between Differential Mobility, Log S, and Log P Using Machine Learning and SHAP Analysis.

Area of Science:

• Computational Chemistry and Cheminformatics
• Analytical Chemistry
• Machine Learning in Chemical Sciences

Background:

• Aqueous solubility (log S) and water-octanol partition coefficient (log P) are critical physicochemical properties for drug development and environmental fate assessment.
• Accurate prediction of these properties is essential for efficient screening of drug candidates and understanding chemical transport.
• Existing methods often rely on large datasets or purely structure-based models, which can have limitations.

Purpose of the Study:

• To develop and evaluate machine learning (ML) frameworks for predicting aqueous solubility (log S) and hydrophobicity (log P) using differential mobility spectrometry (DMS) data.
• To assess the explainability of the ML models using SHapley Additive exPlanations (SHAP) analysis.
• To compare the performance of DMS-based models with and without additional structural descriptors.

Main Methods:

• Differential mobility spectrometry (DMS) experiments were conducted in microsolvating environments to generate ion mobility/DMS data (e.g., collision cross-section, dispersion curves).
• Machine learning regressors and ensemble stacking were employed to build predictive models for log S and log P.
• The OPERA package was used to obtain reference log S and log P values for 333 analytes, and SHAP analysis was used for model interpretability.

Main Results:

• DMS-based regression models achieved R² = 0.67 for both log S and log P predictions with RMSE values of 1.03 ± 0.10 and 1.20 ± 0.10, respectively.
• SHAP analysis indicated that gas-phase clustering was a significant factor in log P predictions.
• Incorporating structural descriptors improved model performance, yielding RMSE = 0.84 and R² = 0.78 for log S, and RMSE = 0.83 and R² = 0.84 for log P.

Conclusions:

• Differential mobility spectrometry (DMS) data, when used with machine learning, provides a valuable and explainable approach for predicting key physicochemical properties like log S and log P.
• The inclusion of structural descriptors further enhances prediction accuracy, demonstrating a synergistic effect.
• This DMS-driven methodology offers a promising alternative to traditional structure-based models, especially when dealing with smaller datasets.