Generating functional protein variants with variational autoencoders.

Area of Science:

• Computational Biology
• Protein Engineering
• Machine Learning

Background:

• Protein sequence databases are expanding, offering opportunities for new protein design methods.
• Deep generative models learn from sequence data but their use in novel protein design is underexplored.

Purpose of the Study:

• To investigate the use of variational autoencoders (VAEs) for designing novel, functional protein variants.
• To compare VAEs using aligned (MSA VAE) versus raw sequence input (AR-VAE).

Main Methods:

• Trained VAE models on a dataset of ~70,000 luciferase-like oxidoreductases.
• Generated novel variants of the luxA bacterial luciferase using unconditional and conditional VAEs.
• Experimentally validated the luminescence activity and solubility of generated variants.

Main Results:

• Both MSA VAE and AR-VAE captured amino acid usage patterns; MSA VAE better reflected 3D structure influences.
• Multiple generated variants retained measurable luminescence, with some differing significantly from training sequences.
• Conditional VAEs successfully increased luxA solubility without compromising function.

Conclusions:

• Deep generative models are feasible for exploring protein sequence space and generating useful variants.
• This approach complements existing rational design and directed evolution methods.
• The study demonstrates successful experimental validation of VAE-generated functional proteins.