Deep Learning‐Driven Quantum Chemistry Descriptors for Predicting Electrophilicity Index in Small Organic Molecules

In this work, we employ different machine learning (ML) methods, including feature-based traditional methods and graph-based deep learning methods, to predict the gas phase electrophilicity of small organic molecules. For this goal, atomic and molecular descriptors, and topological features based on graph theory were calculated for small organic molecules in the GDB-9-Ex dataset. SHAP analysis carried out on the traditional ML methods developed helped us identify the important molecular features relevant for the reactivity parameter predictions. We observe that the features identified by traditional ML methods are well captured by the embedding space of the GNNs, indicating that both methods capture similar information from the training data. Basic qualitative analysis of the ML models shows that the trained models are able to predict the electrophilicity values of amines and carbonyl groups, while struggling to predict the reactivity parameters of alcoholic groups. Further, we observe that the traditional ML models trained using molecular-level features struggle in the out-of-the-box molecule prediction task. We believe that the data, models, and results obtained from this study can act as a starting point for organic molecule design with tailored reactivity profiles.