Classification of organometallic reactions using machine learning

Abstract

Classifying organometallic reactions into distinct reaction types is fundamentally important for understanding mechanisms and predicting reactions, for synthesis optimisation. Fundamental to classification of organometal reactions is reaction representation, but current methods often fail to capture organometal chemical transformation adequately. The study has adopted a hybrid fingerprinting approach, whereby new fingerprints were concatenated with permutation important Morgan fingerprints to create 49 to 63 bits fingerprints. The hybrid fingerprints were used to build KMeans clustering, Guassian mixtures, Affinity propagation and Heirarchical clustering models for classification of organometal reactions. The models successfully classified reactions into 6-8 major organometal reaction types. Of note, the fingerprints consistently outperformed Morgan fingerprints across all clustering models with good Davies–Bouldin Index (DBI) ranging from 0.3 to 0.6 and Silhouette score from 0.3 to 0.8. Furthermore, the clustering models were visualized using Principal Component Analysis. Affinity Propagation and KMeans demonstrated superior performance over Hierarchical and GMM algorithms in distinguishing major reaction categories. In contrast, the Hierarchical model excelled at identifying sub-level classifications compared to the other methods. Consequently, Affinity Propagation and KMeans are recommended for broad reaction type classification, while the Hierarchical approach is better suited for resolving detailed subclass distinctions. The observed variability in model performance further highlighted the importance of feature selection and representation in clustering models. Future studies should look into improving the fingerprints to recognize subcategories.