Ensemble Geometric Deep Learning of Aqueous Solubility.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular property prediction

Background:

• Aqueous solubility is crucial for drug pharmacokinetics.
• Predicting solubility aids in developing effective drug candidates.
• Big data and deep learning offer new avenues for molecular property prediction.

Purpose of the Study:

• To develop and evaluate geometric deep learning models for predicting aqueous solubility.
• To compare the performance of spectral and spatial convolution-based graph neural networks.
• To assess the models' utility in drug discovery workflows.

Main Methods:

• Two graph neural network ensembles were constructed: SolNet-GCN (spectral convolution) and SolNet-GAT (spatial convolution).
• Models were pretrained and benchmarked against existing neural networks on a validation set of 207 molecules.
• Performance was evaluated using Root Mean Square Error (RMSE) and Pearson correlation coefficient (r²).

Main Results:

• SolNet-GCN and SolNet-GAT significantly outperformed existing neural networks.
• SolNet-GCN achieved the best performance with RMSE of 0.53 and 0.72 log molar units and r² of 0.95 and 0.75 on training and validation sets, respectively.
• Model ranking correlated well with a quantum mechanics-based thermodynamic cycle approach.

Conclusions:

• Geometric deep learning models, specifically SolNet-GCN, show high accuracy in predicting aqueous solubility.
• These models can effectively aid in the pharmacokinetic optimization of drug candidates.
• Incorporating atomic attributes related to hydrogen bonding and planarity can further enhance solubility prediction accuracy.