Predicting Reactivity and Reaction Yields in Parallel Synthesis: Meeting of Expectations and Reality

This study focuses on the essential problem in modern machine learning in chemistry: data quality and its impact on ML predictions. We explored a dataset of Biginelli reactions, obtained via a parallel synthesis protocol at Enamine LTD over more than 20 years. Prediction models, both classification (feasibility) and regression (estimated yield), were developed. We observed a relatively uniform performance of different ML methods, representations, and descriptor types. The performance of regression models is limited by the inaccuracies in the reported yield values, which were explored in depth (model outlier identification, resynthesis, and structural analysis). Our findings highlight the potential of machine learning models not only for predictive tasks but also as tools for quality control and deeper insights into the synthetic process-particularly when results deviate from expectations. They also demonstrate the critical impact of input data quality on the excellence of prediction models. Our models were finally evaluated on a set of newly synthesized molecules, selected from Enamine's REAL database, and found to be of satisfactory quality.