Ig-VAE: Generative modeling of protein structure by direct 3D coordinate generation

Area of Science:

• Protein science
• Computational biology
• Structural biology

Background:

• Deep learning is increasingly used in protein science.
• Protein backbone generation is crucial for structure-based design but underexplored by deep learning.
• Existing methods lack efficient backbone generation capabilities.

Purpose of the Study:

• To present a novel deep learning approach for generating class-specific protein backbones.
• To apply variational auto-encoders for direct 3D coordinate generation of immunoglobulins.
• To demonstrate the utility of generative models in computational protein design.

Main Methods:

• Developed a torsion- and distance-aware variational auto-encoder (Ig-VAE) for immunoglobulin backbone generation.
• Learned a high-resolution embedding of protein structure data.
• Integrated the Ig-VAE with Rosetta for computational design tasks.

Main Results:

• Generated novel, high-quality immunoglobulin backbones compatible with existing design tools.
• Successfully created a computational model of a SARS-CoV-2 RBD binder using latent space sampling.
• Demonstrated the effectiveness of the generative prior for guiding protein design.

Conclusions:

• The Ig-VAE offers a new method for protein backbone generation.
• Generative models provide a powerful prior for computational protein design.
• This work suggests a new paradigm for backbone design as a constrained optimization problem in latent space.