A Novel Molecular Representation Learning for Molecular Property Prediction with a Multiple SMILES-Based Augmentation

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Deep learning models are vital for molecular property prediction in drug discovery.
• SMILES strings, while common, suffer from non-uniqueness, hindering model performance.
• Overfitting is a challenge due to limited data in molecular property prediction datasets.

Purpose of the Study:

• To develop an improved molecular representation technique for deep learning.
• To address the non-unique SMILES string issue in molecular property prediction.
• To enhance the accuracy and robustness of molecular property prediction models.

Main Methods:

• Encoding multiple SMILES strings for each molecule.
• Utilizing automated data augmentation strategies.
• Applying deep neural network models inspired by natural language processing.

Main Results:

• Achieved superior molecular representation through multi-SMILES encoding.
• Demonstrated improved performance in molecular property prediction tasks.
• Successfully alleviated overfitting issues common in small molecular datasets.

Conclusions:

• Multi-SMILES encoding is an effective data augmentation technique for molecular property prediction.
• This method enhances deep learning model performance in drug discovery.
• The approach offers a promising solution for leveraging limited molecular data.