Yield Prediction of Organic Reactions in Biased Data Sets via Positive-Unlabeled Learning

Area of Science:

• Chemistry
• Machine Learning
• Data Science

Background:

• Scientific literature contains vast reaction data valuable for machine learning.
• This data is often imbalanced due to selection and reporting biases.
• Existing methods struggle with scarcity of fully labeled reaction datasets.

Purpose of the Study:

• To introduce a machine learning framework, Positivity is All You Need (PAYN), to address data scarcity in reaction prediction.
• To develop a method that learns effectively from biased, positive-only reaction data.
• To improve the scalability and accessibility of data-driven strategies for chemical discovery.

Main Methods:

• Developed PAYN, a framework utilizing a spy-based positive-unlabeled (PU) learning strategy.
• Treated reported high-yielding reactions as the 'positive' class and unexplored chemical space as 'unlabeled'.
• Simulated literature bias on high-throughput experimentation (HTE) datasets for validation.

Main Results:

• PAYN significantly improves model performance on biased datasets.
• The framework balances data by augmenting with generated negative data points.
• Validated on Ni-catalyzed borylations, Buchwald-Hartwig, and Suzuki-Miyaura couplings.

Conclusions:

• PAYN offers a robust strategy for leveraging biased reaction data.
• This approach enhances the utility of scientific literature for machine learning.
• Paves the way for accelerated synthesis design, optimization, and chemical discovery.