Extracting structured data from organic synthesis procedures using a fine-tuned large language model | JoVE Visualize

Area of Science:

Organic Chemistry
Computational Chemistry
Data Science

Background:

The increasing reliance on data-driven methods and machine learning (ML) in organic chemistry highlights the need for structured reaction data.
A significant portion of chemical data exists as unstructured text in literature, making manual data extraction a bottleneck for applications like reaction prediction.
Automated tools are crucial for efficiently converting unstructured text into structured formats to unlock the potential of chemical big data.

Purpose of the Study:

To fine-tune a large language model (LLM) for extracting structured reaction information from organic synthesis procedures.
To convert unstructured text into a standardized format adhering to the Open Reaction Database (ORD) schema.
To enable downstream applications by facilitating the creation of a comprehensive, machine-readable organic reaction database.

Main Methods:

Fine-tuning a large language model (LLM) on a dataset of organic synthesis procedure texts.
Developing a system to extract reaction details and map them to the Open Reaction Database (ORD) schema.
Evaluating the model's performance on accuracy, syntactic correctness, and its ability to recognize compound references and infer reaction roles.

Main Results:

The fine-tuned LLM successfully generated syntactically correct ORD records from unstructured text.
Achieved an average accuracy of 91.25% for ORD "messages" (e.g., compounds, workups, conditions).
Demonstrated high accuracy (92.25%) for individual data fields (e.g., compound identifiers, quantities) and recognized compound-referencing tokens.

Conclusions:

The developed LLM is effective for automated extraction of structured reaction data from organic chemistry literature.
This approach significantly enhances the efficiency of creating structured datasets for applications like reaction prediction and condition recommendation.
The model's ability to infer reaction roles and handle complex text structures paves the way for more advanced computational chemistry tools.