ORDerly: Data Sets and Benchmarks for Chemical Reaction Data.

Area of Science:

• Chemistry
• Life Sciences
• Machine Learning

Background:

• Machine learning (ML) offers significant potential in life sciences for accelerating new molecule discovery and product development.
• High-quality, open-source chemical reaction datasets are scarce, hindering the training of robust ML models.
• Existing reaction data often requires extensive cleaning and standardization for reliable ML applications.

Purpose of the Study:

• To present ORDerly, an open-source Python package for preparing chemical reaction data.
• To enable customizable and reproducible data preparation adhering to the Open Reaction Database (ORD) schema.
• To generate benchmark datasets for reaction prediction tasks and evaluate ML model performance.

Main Methods:

• Developed ORDerly, a Python package for processing chemical reaction data.
• Utilized ORDerly to clean and structure United States patent data according to the ORD schema.
• Generated datasets for forward prediction, retrosynthesis, and reaction condition prediction.
• Trained neural networks and transformers for prediction tasks using generated datasets.

Main Results:

• Demonstrated that incomplete data cleaning can lead to inflated ML performance metrics.
• Established the first benchmark for reaction condition prediction.
• Showcased the utility of non-patent data for evaluating the generalization of forward and retrosynthesis prediction models.
• Highlighted the impact of data quality on ML model reliability in chemistry.

Conclusions:

• ORDERly provides a customizable, open-source solution for preparing large chemical reaction datasets.
• The package facilitates reproducible research and advances ML applications in chemistry.
• Ensuring rigorous data cleaning is essential for accurate ML model evaluation in chemical reaction prediction.