In silico Prediction of Aqueous Solubility: a Comparative Study of Local and Global Predictive Models

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Chemical informatics

Background:

• Accurate prediction of aqueous intrinsic solubility is crucial for drug development and chemical process design.
• Existing Quantitative Structure-Property Relationship (QSPR) models face challenges in balancing predictive accuracy with mechanistic interpretability.

Purpose of the Study:

• To develop and evaluate Quantitative Structure-Property Relationship (QSPR) models for predicting the aqueous intrinsic solubility of liquid and crystalline chemicals.
• To compare the performance and interpretability of global and local QSPR modeling approaches.

Main Methods:

• Construction of global QSPR models using Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Random Forest (RF).
• Development of local QSPR models utilizing k-nearest neighbour (kNN), Arithmetic Mean Property (AMP), and Local Regression Property (LoReP).
• Validation of models using datasets of 1022 liquid and 2615 crystalline compounds.

Main Results:

• Achieved optimal QSPR models with prediction RMSEs of 0.50-0.60 log units for liquids and 0.80-0.90 log units for crystalline compounds.
• Global models, while predictive, showed limited mechanistic interpretability due to numerous descriptors.
• Local models offered better mechanistic insights using few descriptors but required structurally similar neighbors.

Conclusions:

• A consensus approach combining global and local QSPR models is proposed for enhanced predictive accuracy and mechanistic understanding.
• This integrated strategy aims to provide stable and interpretable QSPR models for solubility prediction.
• The findings support the utility of QSPR in aiding medicinal chemists in solubility estimations.