Gradient Descent Approach to Optimizing AlphaFold2 Protein Predictions

Proteins are essential to human bodies as building blocks, transport systems, enzymes, and antibodies. As a result, scientists are attempting to replicate them or produce so-called de novo proteins, which are not found in nature. For instance, these synthetic proteins are made to deliver medications or attach to specific viruses.

Machine learning is being used more and more by scientists to create them. The Nobel Prize in Chemistry was recently awarded to recognize advancements in this discipline. This year's winners were software developers Demis Hassabis and John Jumper, who created Alphafold2, and David Baker, who pioneered de novo protein creation. Protein structures can be accurately predicted on a computer thanks to this software.

Now, an international team led by Sergey Ovchinnikov, a Biology Professor at MIT, and Hendrik Dietz, Professor of Biomolecular Nanotechnology at the Technical University of Munich (TUM), has created a technique that combines the precise structure prediction of Alphafold2 with a so-called gradient descent approach for effective protein design. The study was published in the journal Science.

One popular technique for model optimization is gradient descent. It can be applied methodically to detect departures from the intended goal function and modify the parameters until the best outcome is obtained.

The structure of new proteins predicted by AlphaFold2 can be compared with the intended protein structure in protein design using gradient descent. This enables researchers to refine the structure and newly created amino acid chain further. The latter relies on minor energy interactions and plays a major role in determining the protein's stability and function.

Virtual Superposition of the Building Blocks

Large new proteins can now be better designed and tailored to have certain functions, such as accurately binding to other proteins, thanks to the new technique. Their design methodology is different from earlier methods in several respects.

We have designed the process for new proteins so that we initially ignore the limits of what is physically possible. Usually, only one of the 20 possible building blocks is assumed at each point of the amino acid chain. Instead, we use a variant in which all possibilities are virtually superimposed.”

Christopher Frank, Doctoral Candidate and Study First Author of the study, Chair of Biomolecular Nanotechnology, Technical University of Munich

It is not possible to immediately transform this virtual superposition into a protein that can be produced. However, it enables iterative protein optimization.

We improve the arrangement of the amino acids in several iterations until the new protein is very close to the desired structure.”

Christopher Frank, Doctoral Candidate and Study First Author of the study, Chair of Biomolecular Nanotechnology, Technical University of Munich

Using this optimized structure, the amino acid sequence that can be assembled to form a protein in the lab is then ascertained.

The Crucial Test: How Do the Predictions Hold Up in Real Life?

Does the actual structure match the expected architecture and the intended function? This is the ultimate test for all freshly developed proteins. The scientists used the novel technique to design over 100 proteins virtually, synthesize them in the lab, and conduct experimental testing.

“We were able to show that the structures that we designed are very close to the structures that are actually produced,” said Christopher Frank.

They were able to create proteins with up to 1000 amino acids using their novel technique.

This brings us closer to the size of antibodies, and just as with antibodies, we can then integrate several desired functions into such a protein. These could, for example, be motifs for recognizing and suppressing pathogens.”

Hendrik Dietz, Professor, Biomolecular Nanotechnology, Technical University of Munich