Beyond AlphaFold: Artificial intelligence creates new proteins

Over the past two years, AI/machine learning tools have stunned the world with the accuracy of protein structure predictions and have led the research field to produce some meaningful results

Proteins are often referred to as "the building blocks of life" because they are necessary for the structure and function of

The problem of protein sequence design is to find an amino acid sequence

"Proteins are the foundation of the whole biology, but we know that all proteins found in every plant, animal and microbe are far less than the possible 1 percent

To go beyond the proteins found in nature, Baker's team members broke down the challenges of protein design into three parts and used new software solutions

First, an expected new protein sequence must be constructed

In a paper published July 21 in the journal Science, the team showed that AI can design proteins

Then, to speed up the process, the team designed a deep learning-based protein sequence design method, ProteinMPNN, that is broadly suited to the design of monomers, cyclic oligomers, protein nanoparticles, and protein-protein interfaces, with excellent performance in both computer and experimental tests

"If you have a lot of data, neural networks are easy to train, but for proteins, there aren't as many examples

The team then used AlphaFold (a tool developed by Alphabet's DeepMind) to independently assess whether the amino acid sequences they designed were likely to fold into the intended shape structure

"Software that predicts protein structure is part of the solution, but it can't come up with anything new on its own," Dauparas explains

In another paper, published Sept.

"We found that proteins created using ProteinMPNN were more likely to fold as expected, and we can use these methods to create very complex combinations of