Enzyformer: a two-stage pretrained model for enzymatic retrosynthesis

Area of Science:

• Computational chemistry
• Drug discovery
• Biotechnology

Background:

• Enzymatic retrosynthesis planning accelerates drug synthesis but faces challenges in knowledge transfer and EC number prediction.
• Current models struggle with limited transferability across general chemical spaces.

Purpose of the Study:

• Introduce Enzyformer, a novel model for improved enzymatic retrosynthesis.
• Enhance the accuracy and interpretability of retrosynthesis prediction and EC number assignment.
• Reduce workload and costs in drug design and synthesis.

Main Methods:

• Enzyformer integrates molecule and reaction pretraining to learn SMILES syntax and organic reaction rules.
• A two-stage pretraining strategy is employed for enhanced retrosynthesis prediction.
• Contrastive learning is utilized for sequential EC number prediction on the ECREACT dataset.

Main Results:

• Enzyformer improved top-1 and top-10 retrosynthesis prediction accuracies by 7.5% and 11.7% over R-SMILES.
• The contrastive learning model achieved a top average F1-score of 0.924 for first-level EC number prediction.
• The model demonstrates enhanced accuracy and interpretability in enzymatic retrosynthesis.

Conclusions:

• Enzyformer provides a significant advancement in enzymatic retrosynthesis prediction accuracy and interpretability.
• The model has the potential to accelerate drug design, reduce costs, and minimize environmental impact.
• This approach offers a promising solution for challenges in computational drug discovery.