Going beyond SMILES enumeration for data augmentation in generative drug discovery

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Small molecular datasets limit generative deep learning models.
• Data augmentation, particularly SMILES enumeration, is crucial for de novo molecule design.
• Existing SMILES augmentation methods may not fully exploit chemical and linguistic information.

Purpose of the Study:

• To investigate novel SMILES augmentation techniques for enhancing de novo molecule design.
• To introduce and evaluate new strategies inspired by natural language processing and chemistry.
• To expand the toolkit for designing molecules with specific properties in data-scarce environments.

Main Methods:

• Developed four novel SMILES augmentation strategies: token deletion, atom masking, bioisosteric substitution, and self-training.
• Applied these methods to small molecular datasets for generative deep learning.
• Systematically analyzed the performance and advantages of each strategy.

Main Results:

• All introduced SMILES augmentation strategies demonstrated potential for improving de novo design.
• Atom masking proved effective for learning physico-chemical properties in low-data regimes.
• Token deletion showed promise in generating novel molecular scaffolds.

Conclusions:

• Rethinking SMILES augmentation techniques can significantly enhance generative deep learning for molecule design.
• The developed strategies offer distinct advantages for addressing limitations of small datasets.
• This expanded repertoire provides chemists with more powerful tools for designing molecules with bespoke properties.