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

Area of Science:

• Computational chemistry and cheminformatics.
• Machine learning applications in physical property prediction.

Background:

• Melting points of organic small molecules are critically dependent on their three-dimensional (3D) conformations.
• Existing melting point prediction models often lack sufficient 3D conformational information, limiting their accuracy.

Purpose of the Study:

• To develop a novel framework, 3D^xG-MP, that integrates 3D structural and conformational data for improved melting point prediction.
• To address the limitations of current methods by incorporating multiple low-energy molecular conformations.

Main Methods:

• Development of the 3D^xG-MP framework utilizing a 3D equivariant graph neural network (EGNN).
• Training the model on an extensive dataset of 237,406 melting points, incorporating three low-energy conformations per molecule for the 3D³G-MP variant.
• Comparison against traditional methods like XGBoost, long short-term memory, and 2D graph attention networks.

Main Results:

• The 3D³G-MP model achieved mean absolute errors (MAEs) of 21.719 °C (training), 23.014 °C (validation), and 22.956 °C (test).
• Demonstrated a 10.04% lower MAE compared to XGBoost for flexible molecules (≥7 rotatable bonds).
• External validation on 88 independent molecules yielded a MAE of 22.471 °C, confirming model reliability.

Conclusions:

• Integrating 3D structural information with conformational data significantly enhances the accuracy of melting point prediction for organic small molecules.
• The 3D^xG-MP framework shows promise for predicting melting points and can be extended to other physicochemical properties.