Machine Learning C-N Couplings: Obstacles for a General-Purpose Reaction Yield Prediction | JoVE Visualize

Area of Science:

Organic Chemistry
Computational Chemistry
Chemical Informatics

Background:

Palladium-catalyzed C-N couplings are vital in synthesis but optimizing conditions for high yields is labor-intensive.
Machine learning (ML) offers a promising approach to accelerate the selection of effective reaction conditions by predicting yields.
Predicting reaction outcomes from chemical reaction databases is an emerging area with potential to streamline chemical research.

Purpose of the Study:

To evaluate the feasibility of predicting palladium-catalyzed C-N coupling reaction yields using machine learning models.
To assess the generalizability and limitations of ML models when applied to new chemical reaction data.
To identify data requirements for developing robust, broadly applicable yield prediction models.

Main Methods:

Utilized databases of chemical reactions to train machine learning models for yield prediction.
Tested model performance on challenging data splits and a dedicated experimental dataset.
Analyzed model generalizability by assessing performance on reactions outside the training data's chemical space.

Main Results:

ML models demonstrated good performance when predicting yields within the chemical space covered by the training data.
Model generalizability was limited, with performance dropping significantly on reactions outside the training set's scope, even for similar reaction types.
Algorithmic approaches for yield prediction are viable, but practical application is constrained by the availability and diversity of training data.

Conclusions:

Accurate yield prediction for novel C-N coupling reactions requires significantly more diverse training data covering a wider range of reagents.
Current ML models for reaction yield prediction are susceptible to leaving their applicability domain, even with seemingly minor variations in reaction conditions.
Evaluating ML models solely on literature data, even with challenging splits, can be misleading regarding their real-world predictive power.