Q-GEM: Quantum Chemistry Knowledge Fusion Geometry-Enhanced Molecular Representation for Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning for drug discovery
• Molecular representation learning

Background:

• Existing self-supervised learning (SSL) methods using 3D graph neural networks (GNNs) for molecular representation often overlook crucial electronic structural information.
• This neglect limits their ability to accurately predict molecular properties influenced by electronic factors, such as reactivity and adsorption.

Purpose of the Study:

• To develop a novel molecular representation learning method, Q-GEM, that integrates both quantum chemical and 3D geometric structural information.
• To enhance the characterization of molecules for improved drug discovery and property prediction.

Main Methods:

• Q-GEM utilizes a GNN incorporating comprehensive 3D geometrical and electronic structural data.
• It employs multiscale self-supervised learning (SSL) tasks and leverages the quantum chemical property database QuanDB.
• The method focuses on enhancing molecular conformation prediction and discrimination.

Main Results:

• Q-GEM achieved state-of-the-art performance on 12 out of 13 MoleculeNet prediction tasks.
• It demonstrated an average performance improvement of 3.3% for classification and 2.0% for regression tasks.
• Significant improvements were also observed in predicting localized quantum chemical properties, highlighting its strength in distinguishing electronic structures.

Conclusions:

• Q-GEM represents a significant advancement in molecular representation learning by effectively combining quantum and geometric information.
• The method shows superior performance in predicting molecular properties and characterizing electronic structures.
• This breakthrough offers a powerful tool for accurate molecular property prediction, advancing drug discovery efforts.