Yield Smarter, Not Harder: Good Practices for Machine Learning of Reaction Outcomes

Area of Science:

• Chemical synthesis
• Machine learning
• Computational chemistry

Background:

• Predicting reaction yield is crucial for synthetic chemistry, impacting route planning and high-throughput experimentation (HTE).
• Current machine learning (ML) models often rely on complex descriptors and deep architectures, limiting interpretability and scalability.
• There is a need to evaluate simpler descriptors for reaction yield prediction.

Purpose of the Study:

• To assess the information content of simpler chemical descriptors for reaction yield prediction.
• To determine if increasing descriptor complexity improves ML model accuracy.
• To benchmark predictive performance using classical ML models and varying descriptor complexity.

Main Methods:

• Trained classical ML models on descriptors of varying complexity.
• Benchmarked performance on four public HTE datasets (Buchwald-Hartwig amination, Suzuki-Miyaura coupling, silicon-amine protocol).
• Evaluated generalization, robustness via external validation, and performance on asymmetric yield distributions.

Main Results:

• Simpler models with interpretable features achieved competitive performance under rigorous validation.
• Contrary to expectations, increased descriptor complexity did not consistently improve accuracy.
• Findings suggest simpler descriptors hold significant predictive power.

Conclusions:

• Simpler ML models and descriptors offer a viable and potentially superior alternative for reaction yield prediction.
• Established good practices for future ML studies in reaction yield prediction, emphasizing baseline model comparisons.
• Advocated for the use of interpretable models in HTE for better scalability and understanding.