In 2020, Google DeepMind researchers presented AlphaFold2, an artificial intelligence model that tackled a major scientific challenge: Is it possible to determine the structure of a protein from its components alone? The ability to do this without slow and expensive laboratory experiments had the potential to revolutionize our understanding of biology and accelerate drug discovery.
In an annual competition of scientists trying to solve this puzzle, AlphaFold2 became the first artificial intelligence model to perform as well as laboratory experiments. Over the next few years, he predicted structures for more than 200 million different proteins, a 1,500-fold increase over proteins previously characterized in decades of laboratory work. In 2024, the main developers of the model received the Nobel Prize in Chemistry.
AlphaFold2 and related models have been “extremely exciting” for science, he says Ryan Hillassistant professor of strategy at Kellogg. With collaborator Carolyn Stein of UC Berkeley, Hill set out to measure exactly how AlphaFold has changed the field of biology, from the pace and direction of discovery to the nature of scientific work.
Their study provides a preview of what happens when artificial intelligence transforms a niche field overnight—a level of disruption that may become increasingly familiar in the future as the world enters the age of artificial intelligence.
“Economists are grappling with many questions about the role of artificial intelligence in our lives and in the economy,” says Hill. “This is an interesting microcosm for studying these forces because it’s a powerful artificial intelligence tool that does a specific job that we can observe and measure.”
A transformation overnight
While there are over 200 million known proteins, they are all made from the same 20 organic building blocks, called amino acids. Because these components can fold into incredibly complex structures, simply figuring out which ones make up a protein is like imagining a car based on a box of loose parts. To then discover the three-dimensional structure of this protein – and better understand how it works and how it can be controlled – requires even more work. Or at least it used to be.
Using experimental methods to determine the three-dimensional structure of proteins takes years to complete and costs about $100,000 per protein solved. As a result, less than 0.1 percent of known proteins had solved structures in 2020.
AlphaFold2 changed that almost overnight, providing millions of structures with similar accuracy to those costly experiments.
“There were a lot of people who had been trained for many years in the methods of solving experimental structures,” says Hill. “And one day, Google DeepMind just ran its algorithm on every known protein and published it on the Internet.”
Enhancement, not replacement
Hill and Stein investigated the impact of AlphaFold’s sudden discovery on the world of science by combining multiple scientific databases for protein research. These sources contain information about what scientists have studied and published about each known protein, a data trail that stretches back decades.
This allowed the researchers to ask how the seismic shock of AlphaFold2 affected the field of structural biology. On the surface, the answer appeared to be, surprisingly, not much. In the years since the introduction of AlphaFold2, the number of journal articles using the time-consuming experimental methods to determine protein structure has not decreased.
“Structural biologists are still doing what they’ve been doing in a lot of ways,” says Hill. “They are publishing the same number of papers as they used to, and surprisingly to us, they are still publishing in the best journals in science – even though in some ways their work can now be replaced by artificial intelligence.”
But looking deeper, the researchers found signs that AlphaFold2 is extending the experimenters’ abilities rather than replacing them. Structural biologists have been using the AI model to enhance their skills, leading to faster, more accurate discoveries.
“There’s often a lot of complementary insight,” says Hill. “Artificial intelligence is not perfect. Sometimes there are variations of the protein or parts of the structure that are harder for AI tools to predict. Experimental methods can also have quality issues. Combining information from experiments and AI gives us more confidence that we have the right protein structure in a way that can matter for downstream research.”
A spotlight on science
In addition, the release of AlphaFold2 appeared to increase the number of proteins the scientists investigated. Many of the proteins that did not have structures before the appearance of AlphaFold2 were neglected due to lack of interest. some may simply have been impractical for experimental methods, or there simply may not have been enough structural biologists to keep up with the demand.
Hill and Stein found that AlphaFold2 rapidly expanded the number of proteins the field examined, which they describe as a “spotlight” effect.
For example, scientists studying reproduction in zebrafish had identified a key protein, but their lab lacked the expertise to determine its structure.
“These scientists would just have to wait and hope that someone else would make a breakthrough, and then they could build on that,” says Hill. “And that’s not unusual in many parts of science.”
After the release of AlphaFold2, researchers were able to obtain an AI prediction of the protein’s structure and then use that information to inform further experiments about its function. Their research was eventually published in Cellone of the leading biology journals.
It’s a pattern Hill and Stein saw on the field.
“Within a few years of launching AlphaFold, we’re seeing very large increases in activity in these previously unsolved proteins,” says Hill. “This often happens with technological change. If a job becomes very cheap through automation, it allows people to do a variety of new jobs that they couldn’t do before.”
One traffic jam after another
Creating cool 3D images of complex protein structures wasn’t the whole point of the protein folding challenge. With these structures, scientists now hope to better understand how the important proteins work and eventually influence their activity with new drugs that treat disease, slow aging or otherwise improve human health.
Hill and Stein looked for evidence that AlphaFold2 accelerated these downstream discoveries as well. But they didn’t see a significant impact of introducing the AI model on drug development — at least not yet.
“Even with a very capable machine learning tool for these structural steps, it’s only one piece of a very large puzzle,” says Hill. “Because there are so many bottlenecks, there is no task that you could fully automate that would have a meaningful impact on the rate of drug discovery.”
“The good news is that it opens up opportunities, because we can divert our human efforts towards some of these bottlenecks, which hopefully will make us more productive,” he continues. “It will have benefits, but I don’t expect it to happen overnight.”
A new partner
The release of AlphaFold2 came at a very different time for AI – two years before the revelation of ChatGPT sparked a huge wave of interest in genetic AI models and their potential to change the way we work. As these models gain new capabilities, anxiety has grown about their ability to replace even highly skilled humans.
But there are few professions more specialized than the structural biologist. So the AlphaFold2 story may provide some reassurance to skilled workers in other fields, Hill says.
“There have been a lot of new technologies that have done a lot of work that workers did in the past, so there’s always the concern that people will be adversely affected,” he says. “On the other hand, in most past cases of automation, it’s also opened up a lot of new opportunities in the economy, new jobs that appear or changes in the types of jobs that are available. These types of tools usually make people more productive.”
A disruptive technology like AlphaFold2 that extends rather than displaces the capabilities of human experts provides early evidence for an optimistic vision of human-AI collaboration, both in science and beyond.
“There’s a sense that if AI can get more involved and make scientists super productive or even generate ideas, that could have a huge ripple effect on the economy, making more people and more processes efficient,” says Hill. “I find that exciting. This means that there could be new discoveries that would be impossible without the help of AI tools.”
