ProT-VAE: Protein Transformer Variational AutoEncoder for functional protein design
- Emre Sevgen
- Joshua Moller
- Adrian Lange
- John Parker
- Sean Quigley
- Jeff Mayer
- Poonam Srivastava
- Sitaram Gayatri
- David Hosfield
- Clayton Dilks
- Claire Buchanan
- Thomas Speltz
- Maria Korshunova
- Micha Livne
- Michelle Gill
- Rama Ranganathan
- Anthony B. Costa
- Andrew L. Ferguson
2025-10-06
Deep generative models have demonstrated success in learning the protein sequence to function relationship and designing synthetic sequences with engineered functionality. We introduce the Protein Transformer Variational AutoEncoder (ProT-VAE) as an accurate, generative, fast, and transferable model for data-driven protein design that blends the merits of variational autoencoders to learn interpretable, low-dimensional latent embeddings for conditional sequence design with the expressive, alignment-free featurization offered by transformer-based protein language models. We implement the model using NVIDIA’s BioNeMo framework and validate its performance in retrospective functional prediction and prospective functional design. The model identifies a phenylalanine hydroxylase enzyme with 2.5× catalytic activity over wild-type, and a γ -carbonic anhydrase enzyme with a melting temperature elevation of ΔT m = +61 ° C relative to the most thermostable sequence reported to date and activity in 23% v/v methyl diethanolamine at pH 11.25 and 93 ° C corresponding to industrially relevant conditions for enzymatic carbon capture technologies. The ProT-VAE model presents a powerful and experimentally validated platform for machine learning-guided directed evolution campaigns to discover synthetic proteins with engineered function.