Data augmentation in a triple transformer loop retrosynthesis model.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Data Science

Background:

• US Patent Office (USPTO) reaction datasets are biased, limiting their utility for computer-assisted synthesis planning (CASP).
• Existing datasets may not adequately represent the full scope of chemical reactions for training robust models.

Purpose of the Study:

• To generate an equilibrated and expanded dataset of chemical reactions for improved CASP.
• To overcome the limitations of biased reaction data in computational chemistry.

Main Methods:

• Applied retrosynthesis templates to USPTO molecules to generate starting materials (SM) from products (P).
• Utilized a triple transformer loop (TTL) model (T2) to predict reagents (R) for SM → P reactions.
• Validated predictions using TTL transformer T3, generating over 27.5 million fictive reactions.

Main Results:

• Created a large, validated dataset of 27.5 million fictive reactions, covering a wider chemical space than the original USPTO data.
• Demonstrated that a retrosynthesis model trained on a subset of this new dataset (1,097,374 reactions) outperforms models trained solely on USPTO data.
• Showcased improved performance in single-step retrosynthesis prediction.

Conclusions:

• The generated fictive reaction dataset significantly enhances the performance of retrosynthesis models for CASP.
• Template-equilibrated datasets are crucial for overcoming biases in existing chemical reaction databases.
• This approach provides a valuable resource for advancing computer-assisted chemical synthesis.