Recommending reaction conditions with label ranking.

Area of Science:

• Chemical synthesis
• Machine learning in chemistry
• Reaction optimization

Background:

• Identifying optimal reaction conditions is crucial but challenging in chemical synthesis.
• Current methods often focus on yield regression, which may not generalize well to new substrates.

Purpose of the Study:

• Introduce a new approach using label ranking models for prioritizing chemical reactions.
• Compare the performance of label ranking against traditional yield regression models.

Main Methods:

• Developed and evaluated label ranking models using only substrate features.
• Utilized Borda's method for ranking aggregation.
• Tested models on reaction condition selection scenarios with varying numbers of conditions and missing data.

Main Results:

• Label ranking models demonstrated superior generalization to new substrates compared to prior models.
• The approach proved effective in practical reaction condition selection tasks.
• Borda's method and model simplicity contributed to consistent high performance.

Conclusions:

• Label ranking offers a conceptually new and effective strategy for prioritizing experimental conditions in chemical synthesis.
• This method enhances predictive accuracy and experimental efficiency for new chemical reactions.