Physics-Informed Generative Model for Drug-like Molecule Conformers.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Accurate molecular conformer generation is crucial for drug discovery and materials science.
• Classical force fields provide physically relevant representations but can be limited in scope.
• Deep learning offers powerful tools for inferring complex molecular properties.

Purpose of the Study:

• To develop a diffusion-based generative model for accurate conformer generation.
• To ensure the model generates physically relevant molecular structures.
• To leverage deep learning for inferring molecular parameters.

Main Methods:

• A diffusion-based generative model was developed for conformer generation.
• Deep learning techniques were employed to infer atom typing and geometric parameters.
• The model was trained on large, synthetic datasets of drug-like molecules optimized with the GFN2-xTB method.

Main Results:

• The model accurately reproduces bonded structures, a key aspect of molecular representation.
• High accuracy for bonded parameters was achieved, surpassing conventional knowledge-based methods.
• Generated conformers were compared against experimental data from the Protein Databank and Cambridge Structural Database.

Conclusions:

• Diffusion-based generative models show significant promise for accurate conformer generation.
• The developed model offers a physically relevant and accurate approach to molecular modeling.
• This method has the potential to advance computational drug design and molecular simulation.