ProT-VAE: Protein Transformer Variational AutoEncoder for functional protein design

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Deep generative models excel at predicting protein sequence-function relationships.
• Designing synthetic proteins with novel functionalities requires advanced computational tools.

Purpose of the Study:

• Introduce the Protein Transformer Variational AutoEncoder (ProT-VAE) for data-driven protein design.
• Leverage variational autoencoders and transformer-based models for enhanced protein sequence design.
• Validate ProT-VAE's performance in functional prediction and design.

Main Methods:

• Implemented ProT-VAE using NVIDIA's BioNeMo framework.
• Combined variational autoencoders for latent embeddings with transformer models for featurization.
• Validated the model through retrospective functional prediction and prospective functional design.

Main Results:

• Identified a phenylalanine hydroxylase with 2.5x wild-type catalytic activity.
• Engineered a gamma-carbonic anhydrase with a +61°C melting temperature elevation.
• Demonstrated enzyme activity under industrially relevant conditions for carbon capture.

Conclusions:

• ProT-VAE is an accurate, fast, and transferable model for protein design.
• The model enables the discovery of synthetic proteins with engineered functions.
• ProT-VAE serves as a powerful platform for machine learning-guided directed evolution.