Fast Customization of Chemical Language Models to Out-of-Distribution Data Sets.

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Chemical Informatics

Background:

• Language models (LMs) are increasingly vital for industrial applications, offering predictive insights.
• In the chemical industry, LMs have been used since 2016 for tasks like reaction outcome prediction and retrosynthesis.
• Limited public datasets hinder LM performance, necessitating the use of proprietary data.

Purpose of the Study:

• To develop and validate a methodology for retraining LMs using proprietary, non-public chemical datasets.
• To enhance the accuracy of reaction outcome prediction and single-step retrosynthesis models.
• To establish guidelines for customizing chemical LMs in corporate environments.

Main Methods:

• Retraining of language models on proprietary, non-public datasets.
• Application of a multidomain learning formulation combining patent and proprietary data.
• Validation of the methodology on reaction outcome prediction and single-step retrosynthesis tasks.

Main Results:

• A significant improvement in model accuracy was achieved through retraining with proprietary data.
• Combining patent and proprietary data in a multidomain learning approach yielded considerable accuracy gains.
• The developed methodology demonstrates a successful approach to enhancing chemical LM performance.

Conclusions:

• Proprietary datasets are crucial for improving the performance of chemical language models.
• Multidomain learning offers a powerful strategy for leveraging diverse data sources.
• The study provides practical guidelines for the effective customization of chemical LMs in industry.