In 2020, Google DeepMind unveiled AlphaFold2, an AI model that tackled one of biology's most fundamental challenges: determining a protein's three-dimensional structure from its amino acid sequence alone. The breakthrough was stunning. AlphaFold2 became the first AI system to match the accuracy of laboratory experiments, a feat that had eluded researchers for decades. In subsequent years, the model predicted structures for more than 200 million different proteins—a 1,500-fold increase over the proteins characterized through traditional laboratory methods in decades of work. In 2024, the model's lead developers received the Nobel Prize in Chemistry.

Yet six years after AlphaFold2's release, something unexpected has happened. The number of journal articles using traditional, time-intensive experimental methods to determine protein structure has not declined. Structural biologists are still publishing the same number of papers as they did before AlphaFold2. They are still publishing in the best journals in science. In many ways, they are doing what they were doing before.

This apparent paradox offers a crucial lesson about how AI actually transforms specialized fields. It is not through wholesale replacement of human expertise, but through augmentation and expansion of human capability. A new study by researchers at Northwestern University's Kellogg School of Management, conducted in collaboration with UC Berkeley, provides the first rigorous analysis of how AlphaFold2 has actually changed the field of structural biology. The findings challenge both utopian and dystopian narratives about AI's impact on specialized work.

The Floodlight Effect

The traditional approach to determining protein structure is extraordinarily expensive and time-consuming. Using experimental methods like X-ray crystallography or cryo-electron microscopy, scientists can determine a protein's structure, but the process takes years and costs an estimated $100,000 per solved protein. As a result, less than 0.1 percent of known proteins had solved structures in 2020. The vast majority of proteins remained structurally uncharacterized, a limitation that constrained biological research and drug discovery.

AlphaFold2 changed this almost overnight. Google DeepMind ran its algorithm on every known protein and posted the results to the internet. Suddenly, researchers had access to predicted structures for 200 million proteins. Yet the response from structural biologists was not what many observers expected. Rather than abandoning experimental methods and relying entirely on AI predictions, researchers began using AlphaFold2 to augment their experimental work.

In one striking example, scientists studying reproduction in zebrafish had identified a key protein but lacked the expertise to determine its structure experimentally. Before AlphaFold2, they would have had to wait and hope that another lab would solve the problem. Instead, they used AlphaFold2's prediction as a starting point, conducted targeted experiments to validate and refine the prediction, and published their findings in Cell, one of the leading biology journals.

"There's often a lot of complementary insight. The AI is not perfect. There are sometimes variations of the protein or pieces of the structure that are more difficult for the AI tools to predict. Combining insights from the experiments and the AI gives us more confidence that we have the correct protein structure in a way that might matter for downstream research."

— Ryan Hill, Assistant Professor of Strategy, Kellogg School of Management