In June, South Korean authorities approved the production of the first drug, the COVID-19 vaccine, using a new human-created protein. The vaccine is based on a spherical-shaped protein "nanoparticle" developed by scientists over a decade ago through a labor-intensive trial and error process.
Now, a team led by David Baker, a biochemist at the University of Washington (UW) in Seattle, reveals in the journal Science that they can build such molecules in seconds instead of months, due to tremendous advances in artificial intelligence (AI).
Such initiatives are part of a larger scientific shift as life scientists embrace artificial intelligence technologies such as DeepMind's protein structure prediction software AlphaFold. The latest iteration of AlphaFold has revealed that it predicts the structures of every protein known to science, DeepMind announced in July. And in recent months, the number of AI tools that can rapidly imagine entirely new proteins has grown dramatically. Some of these tools are based on AlphaFold. This was previously a painstaking endeavor with a high failure rate.
According to computational biologist Noelia Ferruz of the University of Girona in Spain, “since AlphaFold, there has been a revolution in the way we work with protein design.” We live in extremely exciting times.
Much of the effort has focused on developing methods for creating new proteins unlike anything found in nature, with little attention paid to the potential of these molecules. But scientists want to develop proteins that can perform useful tasks, from cleaning up toxic waste to treating disease, and a growing number of businesses are doing just that. DeepMind in London and Meta (formerly Facebook) organizations in Menlo Park, California are among those pursuing this goal.
The techniques are already very effective. Baker predicts they will grow stronger. What issues you will address with them is the question.
For the past three decades Baker's lab has been producing new proteins. The procedure is segmented by software called Rosetta, which his lab started building in the 1990s. The researchers first imagined a structure for a unique protein, often by putting other protein parts together, and the software then determined an amino acid sequence that fit that shape.
But when created in the lab, these "first draft" proteins were rarely folded into the required form and instead were trapped in various ways. Therefore, modifying the protein sequence so that it simply folds into a desired structure required another step.
According to Sergey Ovchinnikov, an evolutionary biologist at Harvard University in Cambridge, Massachusetts, who once worked in Baker's group, this step was computationally expensive, requiring simulating all possible ways that certain sequences might fold. You really need to run 10.000 computers for weeks on end to complete this.
Ovchinnikov claims that this time-consuming phase has been accelerated by replacing AlphaFold and other AI programs. Hallucination is a method created by Baker's team, in which random amino acid sequences are fed into a structure-prediction network and the structure is modified to become increasingly protein-like, as determined by the network's predictions. In a 2021 study, Baker's group reported evidence that about one-fifth of the small, "hallucinatory" proteins they produced in the lab resembled the predicted form.
AlphaFold and RoseTTAFold, a related technology created by Baker's lab, were trained to predict the structure of specific protein chains. However, it wasn't long before researchers realized that these networks could also represent an assembly of multiple interacting proteins. In light of this, Baker and his team were confident they could imagine proteins that would self-assemble into nanoparticles of various sizes and shapes; they will consist of several copies of a single protein and be comparable to that on which the COVID-19 vaccine is based.