SolTranNet-A Machine Learning Tool for Fast Aqueous Solubility Prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Predicting aqueous solubility is crucial but challenging in drug discovery.
• Machine learning (ML) methods are increasingly used for solubility prediction.
• The Second Challenge to Predict Aqueous Solubility (SC2) highlighted the reliance on ML.

Purpose of the Study:

• To develop and present SolTranNet, a novel molecule attention transformer model.
• To predict aqueous solubility directly from a molecule's SMILES representation.
• To investigate the relationship between model size and performance in solubility prediction.

Main Methods:

• Developed SolTranNet, a molecule attention transformer architecture.
• Utilized SMILES strings as input for molecular representation.
• Performed 3-fold scaffold split cross-validation on the AqSolDB dataset.
• Evaluated performance using root-mean-square error (RMSE) and classification sensitivity.

Main Results:

• SolTranNet achieved an RMSE of 1.459 on AqSolDB and 1.711 on a withheld test set.
• Demonstrated that smaller models (SolTranNet: 3,393 parameters) outperform larger ones.
• Achieved 94.8% sensitivity in classifying insoluble compounds on the SC2 dataset.
• SolTranNet proved competitive with other methods in the SC2 challenge.

Conclusions:

• SolTranNet offers an accurate and efficient ML approach for aqueous solubility prediction.
• Model size is not directly correlated with improved performance for this task.
• SolTranNet can effectively filter insoluble compounds, supporting early-stage drug discovery.