Multi-Conformation Enhanced Equivariant Graph Neural Network: Advancing Melting Point Prediction Accuracy for Organic Small Molecules

The melting points of organic small molecules are governed by their three-dimensional (3D) conformations. We developed a novel framework termed 3D ^<i>x</i> G-MP (where <i>x</i> denotes the number of molecular conformations introduced into the model) to address the lack of 3D conformational information in existing melting point prediction approaches. This framework integrates a 3D equivariant graph neural network (EGNN) with 3D structural information and conformational data, and was trained on an extensive data set of 237,406 melting points. The 3D³G-MP model, which was augmented with three low-energy conformations per molecule, outperformed traditional methods (XGBoost, long short-term memory, two-dimensional (2D) graph attention networks) in melting point prediction, with mean absolute errors (MAEs) of 21.719, 23.014, and 22.956 °C for the training, validation, and test sets, respectively. Notably, the MAE was 10.04% lower than that of XGBoost for the melting point prediction of flexible molecules with ≥7 rotatable bonds. External validation with 88 independent molecules confirmed the reliability of the model, yielding a MAE of 22.471 °C. This study demonstrates that combining 3D structural information with conformational data significantly improves melting point prediction accuracy. The 3D ^<i>x</i> G-MP framework is promising for predicting the melting points of organic small molecules while offering extensibility to the prediction of other physicochemical properties.