ProtWave-VAE: Integrating Autoregressive Sampling with Latent-Based Inference for Data-Driven Protein Design.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning

Background:

• Deep generative models (DGMs) like VAEs and AR models are used for protein design.
• VAEs learn from aligned sequences (MSAs) and create latent spaces.
• AR models handle unaligned sequences and variable lengths but lack latent spaces.

Purpose of the Study:

• Introduce ProtWave-VAE, a novel DGM combining VAE and AR strengths.
• Enable training on unaligned sequences and design variable-length proteins.
• Generate interpretable latent spaces for protein design.

Main Methods:

• Developed ProtWave-VAE using an information-maximizing VAE with a dilated convolution encoder and WaveNet decoder.
• Trained and evaluated the model on alignment-free homologous protein families.
• Applied the model to design synthetic proteins and engineer functions.

Main Results:

• ProtWave-VAE successfully inferred functional and phylogenetic patterns in latent spaces.
• The model achieved high accuracy in semisupervised fitness prediction tasks.
• Experimental validation confirmed the design of functional synthetic proteins, including engineered osmosensing in SH3 domains.

Conclusions:

• ProtWave-VAE effectively blends VAE and AR model advantages for protein design.
• The model facilitates the design of novel, functional proteins from unaligned sequence data.
• ProtWave-VAE enables conditional design and functional engineering of proteins.