3D Conformational Generative Models for Biological Structures Using Graph Information-Embedded Relative Coordinates.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Artificial Intelligence in Chemistry

Background:

• Generating accurate 3D molecular conformations is crucial for drug discovery and materials science.
• Existing methods for 3D structure generation face challenges with coordinate representation and error sensitivity.
• Autoencoder-based generative models are emerging as powerful tools for molecular design.

Purpose of the Study:

• To develop an autoencoder-based generative model for direct 3D molecular conformation generation.
• To introduce and validate a novel encoding method, graph information embedded relative coordinate (GIE-RC), for molecular structures.
• To demonstrate the model's utility in generating diverse molecular conformations and augmenting training data for force fields.

Main Methods:

• An autoencoder architecture was employed for learning and generating molecular representations.
• Graph Information Embedded Relative Coordinate (GIE-RC) was utilized for encoding 3D molecular structures, ensuring translation and rotation invariance.
• The model was trained and tested on generating conformations for both ligands and peptides.

Main Results:

• The GIE-RC method proved less sensitive to errors during the decoding of latent variables compared to Cartesian and internal coordinates.
• The autoencoder model successfully generated 3D conformations for ligands and peptides.
• The model served as an effective method for conformation sampling to enhance datasets for neural network-based force fields.

Conclusions:

• The proposed GIE-RC based autoencoder model provides an effective and robust approach for 3D molecular conformation generation.
• This method simplifies complex 3D generation tasks into graph node feature generation.
• The model has significant potential for applications in drug discovery, materials science, and computational chemistry research.