SolTranNet

SolTranNet predicts aqueous solubility from molecular SMILES using a molecule attention transformer architecture to support solubility assessment in drug discovery.

Key Features:

• Model architecture: Uses a molecule attention transformer that processes molecular representations as SMILES strings.
• Input representation: Operates on SMILES string encodings of molecules.
• Model size: Implements a relatively small model comprising 3,393 parameters.
• Regression performance: Achieves a cross-validation RMSE of 1.459 on the AqSolDB dataset using a 3-fold scaffold split.
• Test set performance: Reports an RMSE of 1.711 on a withheld test set.
• Classification capability: Functions as a classifier that filters insoluble compounds with 94.8% sensitivity on the SC2 dataset.
• Task scope: Supports both regression and classification tasks for solubility prediction.

Scientific Applications:

• Aqueous solubility prediction: Predicts experimental solubility values for compounds to inform drug discovery decision-making.
• Compound filtering: Identifies and filters insoluble compounds during screening via classification sensitivity on SC2.
• Benchmarking: Evaluated on public datasets including AqSolDB (3-fold scaffold split) and SC2 for comparative performance assessment.

Methodology:

Processes SMILES with a molecule attention transformer model (3,393 parameters); evaluated via 3-fold scaffold split cross-validation on AqSolDB (RMSE 1.459), on a withheld test set (RMSE 1.711), and as a classifier reporting 94.8% sensitivity on SC2.