De novo protein design by deep network hallucination

Area of Science:

• Computational biology
• Protein engineering
• Deep learning

Background:

• Deep neural networks have advanced protein structure prediction by analyzing amino acid sequences.
• The potential of these networks to generate entirely new protein sequences and structures remains largely unexplored.

Purpose of the Study:

• To investigate if deep neural networks trained for protein structure prediction can be used to design novel, functional proteins.
• To explore the de novo design of proteins with sequences unrelated to known natural proteins.

Main Methods:

• Utilized the trRosetta network to predict inter-residue distance maps from random amino acid sequences.
• Employed Monte Carlo sequence space sampling to optimize predicted distance distributions against background averages.
• Synthesized genes for designed sequences, expressed proteins in E. coli, and analyzed their structures.

Main Results:

• Generated novel protein sequences with diverse predicted structures.
• Successfully expressed and purified 27 of 129 designed proteins, yielding monodisperse samples.
• Determined the 3D structures of three designed proteins (via X-ray crystallography and NMR), which closely matched computational models.

Conclusions:

• Deep neural networks trained for structure prediction can be inverted for de novo protein design.
• This approach offers a powerful complement to traditional physics-based methods for creating proteins with new functions.