SolPredictor: Predicting Solubility with Residual Gated Graph Neural Network.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Computational methods accelerate drug discovery by predicting compound properties.
• Machine and deep learning models are effective for in silico solubility prediction.
• Accurate solubility prediction aids formulation, lead optimization, and pharmacokinetic assessments, reducing costs and timelines.

Purpose of the Study:

• To develop an advanced computational model, SolPredictor, for precise molecular solubility prediction.
• To leverage residual graph neural network convolution (RGNN) for capturing complex molecular dependencies.

Main Methods:

• Utilized simplified molecular-input line-entry system (SMILES) representation.
• Compiled two large datasets and employed ten-fold cross-validation.
• Incorporated RGNNs designed for long-range dependency capture and residual connections for feature preservation.

Main Results:

• Achieved a Pearson correlation coefficient (R2) of 0.79±0.02 and a root mean square error (RMSE) of 1.03±0.04.
• Validated the model on five independent datasets.
• Conducted error, hyperparameter optimization, and explainability analyses to identify key predictive molecular features.

Conclusions:

• SolPredictor demonstrates high precision in in silico solubility prediction.
• The RGNN-based approach effectively captures essential molecular features for accurate predictions.
• This model offers significant potential for accelerating drug development and reducing associated costs.