Engineering Proteins with 3D Convolutional Neural Networks

Tuesday June 21th, 4-5pm EST | Danny Diaz, University of Texas at Austin

(This talk was previously titled: Machine learning-aided engineering of hydrolases for PET depolymerization)

An extremely important task in biotechnology is the ability to engineer proteins by introducing mutations into their sequences, which ultimately alters their folded structure and function. We have trained self-supervised 3D convolutional neural networks (http://www.mutcompute.com) over the entire Protein Data Bank (PDB) to learn the optimal chemical pockets for a given amino acid within a given protein are or should be. In most cases (ca. 80%), MutCompute predicts the wild-type (natural) amino acid, but in some cases, it uses its broad computer vision to adjudge that a different amino acid might be a better fit for a given pocket. In these cases, have successfully introduced mutations to improve function, basically anticipating what evolution might have done on a longer timescale and/or under stronger selection pressure.  We have used this insight to engineer a wide variety of biotechnologically important proteins. So far, we have improved the fluorescence of blue-fluorescent protein (BFP) by 10-fold, increased the stability of a plastic-eating enzyme by over 10 degrees Celsius (to the point where it can completely degrade plastic packaging from Walmart within 48 hours), stabilized a polymerase for isothermal COVID19 diagnostics, and, most recently, improved the expression of a potential cancer therapeutic enzyme. Additionally, we have rebuilt the data engineering stack to extend MutCompute’s capabilities to non-amino acid chemistries, such as protein-DNA or protein-ligand interactions, and are currently collaborating with protein engineers around the world to validate these new capabilities.

Paper: https://www.nature.com/articles/s41586-022-04599-z

Recording link: https://youtu.be/Gaoeipwx5p4