Neural graph distance embedding for molecular geometry generation.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Chemistry

Background:

• Generating accurate 3D molecular geometries is crucial for drug discovery and materials science.
• Existing methods, such as distance geometry, face challenges with complex molecular structures like polycyclic systems.
• Graph-based representations offer a promising avenue for molecular structure prediction.

Purpose of the Study:

• To introduce Neural Graph Distance Embedding (nGDE), a novel method for predicting 3D molecular geometries.
• To demonstrate the effectiveness of machine learning, specifically graph neural networks, in predicting interatomic distances for geometry generation.
• To compare nGDE's performance against state-of-the-art methods, particularly for challenging molecular scaffolds.

Main Methods:

• Utilizing a graph neural network trained on the OE62 dataset to predict interatomic distances from molecular graphs.
• Employing multidimensional scaling with predicted distances to generate initial 3D molecular coordinates.
• Refining the generated 3D geometries using standard bioorganic force fields.

Main Results:

• nGDE successfully predicts interatomic distances, enabling the generation of 3D molecular geometries.
• The machine learning-based graph distances outperform conventional shortest path distances in graph drawing applications.
• Comparative analysis shows nGDE achieves competitive performance, outperforming existing methods for polycyclic molecules.

Conclusions:

• Neural Graph Distance Embedding (nGDE) offers a robust and accurate method for 3D molecular geometry generation.
• The approach demonstrates significant advantages over traditional methods, particularly for complex polycyclic structures.
• nGDE represents a promising advancement in applying machine learning to molecular structure prediction.