Data Checking of Asymmetric Catalysis Literature Using a Graph Neural Network Approach

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Data Science

Background:

• Chemical databases have grown substantially, but data quality is often compromised by human errors.
• Manual curation of large chemical datasets is time-consuming and inefficient.
• Automated tools are needed to improve the accuracy and reliability of chemical data.

Purpose of the Study:

• To develop and apply Graph Neural Networks (GNNs) for identifying stereochemical misassignments in asymmetric catalysis literature.
• To create an automated method for enhancing the quality control of chemical databases.

Main Methods:

• An ensemble of GNN models was employed to predict the stereoselectivity of reactions.
• Discrepancies between predicted and reported outcomes were flagged as potential stereochemical misassignments.
• The methodology was tested on databases of asymmetric 1,4-addition reactions using bisphosphine and diene ligands.

Main Results:

• The GNN approach successfully identified potential stereochemical misassignments in the studied databases.
• This automated method significantly reduced the need for extensive manual data checking.
• Expert review was reduced to 2.2% and 3.5% of the total data points in the investigated datasets.

Conclusions:

• Graph Neural Networks offer an effective and efficient solution for identifying stereochemical errors in chemical literature.
• This methodology accelerates database curation and improves overall data quality.
• The approach is valuable for maintaining the integrity of chemical information in scientific literature.