Efficient 3D Molecular Design with an E(3) Invariant Transformer VAE.

Area of Science:

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

Background:

• Accurate prediction of molecular properties is crucial for drug discovery and materials science.
• Existing methods for molecular generation often struggle to simultaneously optimize 3D structure and quantum chemical properties.
• Developing efficient and accurate generative models for novel molecules remains a significant challenge.

Purpose of the Study:

• To introduce Vagrant, a 3D invariant graph-to-string transformer variational autoencoder (VAE).
• To enable the generation of molecules with accurate density functional theory (DFT)-level properties.
• To improve the quality and computational efficiency of molecular generation compared to existing methods.

Main Methods:

• Encoding 3D molecular structures into a 3D-aware latent space using a graph-to-string transformer VAE.
• Directed navigation within the latent space to implicitly optimize 3D molecular structures.
• Conditioning a generative transformer with latent embeddings to predict 1D molecular sequences.
• Introducing novel sampling methods tailored for VAEs to enhance performance.

Main Results:

• Vagrant successfully models the joint probability distribution of 3D molecular structures and their properties.
• The method demonstrates superior performance over comparable 3D autoregressive and diffusion models.
• Achieved higher sample quality and computational efficiency in predicting quantum chemical properties of novel molecules.

Conclusions:

• Vagrant offers a powerful new approach for generating molecules with high-fidelity 3D structures and accurate quantum chemical properties.
• The developed VAE framework and sampling strategies significantly advance the field of generative chemistry.
• This work paves the way for accelerated discovery of novel molecules with desired properties.