Intermediate knowledge enhanced the performance of the amide coupling yield prediction model.

Area of Science:

• Medicinal Chemistry
• Machine Learning
• Chemical Synthesis

Background:

• Amide coupling is crucial in medicinal chemistry, but recommending optimal reaction conditions is challenging due to the vast condition space.
• Accurate yield prediction is difficult due to complex relationships between reaction parameters and outcomes.
• Machine learning offers a promising approach for efficient condition recommendation and yield prediction in chemical reactions.

Purpose of the Study:

• To develop a robust machine learning strategy for accurate yield prediction in amide coupling reactions.
• To improve the efficiency and reliability of condition recommendation for chemical transformations.
• To demonstrate the feasibility of combining high-throughput experimentation (HTE) with knowledge-embedded models.

Main Methods:

• Dataset quality was ensured by unbiased machine-based sampling for diverse substrates and conditions.
• Experiments were performed using an in-house high-throughput experimentation (HTE) platform to minimize human error.
• An intermediate knowledge-embedded strategy was employed to enhance model robustness and predictive accuracy.

Main Results:

• The model achieved high performance on test datasets, with R² of 0.89, MAE of 6.1%, and RMSE of 8.0% under full substrate novelty.
• The strategy demonstrated good generalization on external literature datasets, yielding R² of 0.71, MAE of 7%, and RMSE of 10%.
• The model successfully recommended conditions to elevate reaction yields and identified higher-yielding reactions in pairs with reactivity cliffs.

Conclusions:

• Accurate yield prediction for amide coupling reactions is feasible through the integration of HTE and intermediate knowledge embedding in machine learning models.
• This combined approach significantly improves condition recommendation and facilitates related machine learning applications in chemistry.
• The developed strategy shows potential for broader application in accelerating chemical discovery and optimization.