TL;DR
This paper introduces a data-driven approach for planning chemical syntheses using enzymes, extending existing AI models to predict both forward reactions and retrosynthetic pathways. By leveraging a new tokenization scheme based on enzyme classification numbers and a comprehensive dataset of biocatalytic reactions, they achieve significant progress in making biocatalysis more accessible for designing sustainable chemical processes.
What Problem Does This Paper Address?
Green chemistry initiatives increasingly rely on enzymes as catalysts for more sustainable and efficient chemical synthesis. However, integrating biocatalysis into retrosynthetic planning (the process of designing a synthesis route) has been challenging. This is primarily due to the difficulty in predicting enzyme activity for novel substrates and understanding enzyme-specific selectivity (stereo- and regioselectivity). Current methods are largely rule-based, and existing data-driven approaches are limited to predicting reaction products (forward prediction) rather than planning synthesis routes. This gap hinders the widespread adoption of biocatalysis in everyday laboratory work and industrial process design.
Methods
The core of this work is the adaptation of the Molecular Transformer architecture, a deep learning model previously successful in predicting organic reactions, to the domain of biocatalysis. Key methodological advancements include:
- Enzyme Commission (EC) Number Tokenization: Instead of using natural language enzyme names, the authors introduce a novel tokenization scheme that encodes enzyme information using their hierarchical EC classification numbers. This allows the model to learn catalysis patterns across enzymes with similar functions, even if their names differ.
- Template-Free Retrosynthesis: They extend existing template-free retrosynthetic planning methods to incorporate biocatalytic reactions. This means the model doesn't rely on pre-defined reaction templates but learns to propose disconnections and identify suitable enzymes based on the learned patterns.
- Multi-task Transfer Learning: The Molecular Transformer is trained using a combination of a large dataset of general organic reactions (USPTO) to learn fundamental chemical grammar and a specialized dataset of biocatalytic reactions (ECREACT) to learn enzyme-specific behavior.
Datasets
- ECREACT: A newly curated dataset of unique enzyme-catalyzed reactions, merged from public databases like Rhea, BRENDA, PathBank, and MetaNetX. This dataset is crucial for training and evaluating the biocatalytic models. It's made publicly available.
- USPTO: A large dataset of 1 million general organic chemical reactions, used for pre-training the Molecular Transformer to learn general chemical reaction rules and SMILES grammar.
- RetroBioCat: Used for evaluating retrosynthesis use-cases, comparing the model's predictions against existing biocatalytic retrosynthesis planning tools.
Key Insights and Ideas
- EC Numbers as Powerful Catalysis Signatures: Encoding enzymes via their hierarchical EC numbers proves more effective than using natural language names for learning generalizable biocatalytic patterns. This allows the model to group similar enzymes and predict reactions even for less common ones.
- Bridging Forward and Backward Prediction: The successful implementation of both forward and backward prediction models for biocatalysis is a significant step. This dual capability is essential for true retrosynthetic planning, enabling the design of synthesis routes from target molecules.
- Data Imbalance is a Major Hurdle: The performance of the models is heavily influenced by the availability of training data for specific enzyme classes. This highlights the need for more comprehensive and balanced biocatalytic datasets.
- Implicit vs. Explicit Features: The chemical space of substrates and products can provide implicit information about enzyme class. However, for classes with more diverse or less defined chemical spaces, explicit EC number tokens become critical for accurate prediction.
Key Results
- Forward Prediction: Achieved a top-1 accuracy of 49.6%.
- Retrosynthetic Pathway Prediction: Demonstrated a top-1 single-step round-trip accuracy of 39.6%.
- Model Performance: The EC3 token scheme (using EC levels 1-3) was identified as the best balance between specificity and performance for biocatalytic planning.
- Use-Cases: The model successfully proposed biocatalytic synthesis routes for several target molecules, offering alternatives to traditional methods that often require harsher conditions.
Limitations and Future Work
- Data Scarcity and Imbalance: The primary limitation acknowledged is the imbalanced nature of the ECREACT dataset, with some enzyme classes having significantly fewer samples than others (e.g., isomerases, translocases). This directly impacts prediction accuracy for those classes.
- Bias Towards Natural Products: Due to the sourcing of data from predominantly biosynthetic databases, the model tends to suggest natural products as substrates.
- Stereochemistry Prediction: While improved, predicting correct stereochemistry remains a challenge, particularly for isomerase-catalyzed reactions, due to limited data coverage.
- Future Directions: The authors suggest that as more biocatalytic data becomes available and experimental validation of proposed routes increases, models can be retrained for higher accuracy and broader scope. They also envision the model's use in narrowing or expanding substrate scope depending on specific use-cases. The development of template-free biocatalytic synthesis planning tools is a significant step towards integrating greener chemistry into daily laboratory practices.
Summary updated