Negative chemical data boosts language models in reaction outcome prediction

Area of Science:

• Chemistry
• Machine Learning
• Artificial Intelligence

Background:

• Traditional chemistry research often discards negative experimental results, representing a missed opportunity for data enrichment.
• Underutilization of negative reaction data limits the performance of predictive models, particularly in data-scarce scenarios.

Purpose of the Study:

• To demonstrate the utility of negative chemical reaction data in enhancing reactivity-prediction models.
• To develop and evaluate a machine learning approach that leverages unsuccessful experimental outcomes.

Main Methods:

• Extended tuning of language models with reinforcement learning for chemical reaction prediction.
• Trained a transformer model incorporating negative data from both controlled and high-throughput screening datasets.
• Evaluated model performance on diverse datasets with varying ratios of positive to negative examples.

Main Results:

• Achieved state-of-the-art performance in chemical reaction prediction.
• Successfully leveraged minimal positive data (as few as 20 points) when supported by a significantly larger negative dataset (40x).
• Demonstrated consistent model improvement across controlled and high-throughput experimental data.

Conclusions:

• Negative experimental results are a valuable, underutilized resource for improving chemical reactivity prediction.
• Machine learning models, particularly transformer-based ones, can be effectively trained with limited positive data by incorporating negative outcomes.
• This approach offers significant advantages for accelerating chemical discovery and optimizing experimental design.