ProxelGen: Generating Proteins as 3D Densities.

Area of Science:

• Computational biology
• Structural bioinformatics
• Artificial intelligence in drug discovery

Background:

• Current protein structure generation models predominantly use 3D point cloud representations.
• Limitations exist in the flexibility and conditioning capabilities of existing generative models.

Purpose of the Study:

• To develop ProxelGen, a new generative model for protein structures based on 3D densities (proxels).
• To explore the advantages of density-based representations for protein generation tasks.
• To enable more flexible shape conditioning in protein design.

Main Methods:

• Developed ProxelGen, a generative model operating on voxelized 3D densities (proxels).
• Utilized a 3D Convolutional Neural Network (CNN)-based Variational Autoencoder (VAE).
• Incorporated a diffusion model operating in the latent space of the VAE.

Main Results:

• ProxelGen samples exhibit higher novelty and improved FID scores compared to state-of-the-art models.
• Achieved designability comparable to the training set.
• Demonstrated superior performance in a motif scaffolding benchmark.
• Showcased enhanced flexibility in shape conditioning through 3D density generation.

Conclusions:

• Protein structure generation using 3D densities (proxels) offers significant advantages over point cloud methods.
• ProxelGen provides a powerful new tool for de novo protein design with improved novelty and control.
• Density-based generation opens new avenues for flexible and conditionable protein structure modeling.