A protein begins as a chain of amino acids. What that chain does depends in large part on how it folds into a three-dimensional shape: a form that can help it bind to another molecule, speed up a chemical reaction or carry out another task in a cell. For decades, finding a protein’s structure could require laborious experiments. In 2024, the Nobel Prize in Chemistry recognized work that changed the scale and ambition of the problem.
The prize highlighted two related but distinct achievements. David Baker was honored for computational protein design: using computers to help create proteins that do not exist in nature. Demis Hassabis and John Jumper were recognized for developing AlphaFold2, an artificial-intelligence system that predicts protein structures from their amino-acid sequences.
Together, the advances offer researchers new ways to ask questions about molecular shape. They do not make laboratory experiments unnecessary, and they are not the same task. One aims to predict what a natural protein may look like; the other aims to design a new protein with a chosen structure or purpose.
Why a protein’s shape matters
Amino acids are linked into a chain, and interactions among parts of that chain help it fold. The resulting structure is not just a decorative arrangement. Its surfaces and internal features affect how it interacts with other molecules. A small change in sequence can alter folding or function, which is one reason it can be difficult to infer a protein’s behavior from its sequence alone.
Scientists can determine structures experimentally, including through methods such as X-ray crystallography and cryo-electron microscopy. These approaches have been essential to structural biology, but they take time and do not work equally well for every protein. A reliable computational prediction can help researchers decide what to investigate, suggest how a molecule might work, or guide the next experiment.
AlphaFold2 made prediction a practical tool at scale
AlphaFold2 drew wide attention after its performance in the Critical Assessment of protein Structure Prediction, a community-wide contest that compares predictions with experimentally determined structures. In a 2021 paper in Nature, the developers described a system that could produce highly accurate predictions for many proteins. The achievement addressed a long-standing challenge: estimating a protein’s three-dimensional structure from its sequence.
The practical reach grew when predicted structures were made available through the AlphaFold Protein Structure Database, operated by EMBL’s European Bioinformatics Institute in partnership with Google DeepMind. Instead of requiring each research group to run a prediction from scratch, scientists could search a large public resource and inspect predictions for proteins of interest. That access can make structural clues available early in a project, including for proteins that have not yet been studied experimentally in detail.
A prediction is still a model, not a laboratory observation. Researchers need to consider how confident a model is in different regions and whether the predicted shape fits the biological conditions they are investigating. Some proteins change shape, form complexes with other molecules or contain flexible regions. A structure prediction may not capture every state or interaction. Experimental evidence remains important for checking what a protein does and how it behaves.
Designing proteins is the reverse challenge
Baker’s work tackles a different question: rather than starting with a natural sequence and predicting its shape, can researchers start with a desired shape or function and find an amino-acid sequence that will produce it? That is computational protein design. It could, in principle, give scientists molecular tools with properties that natural proteins do not provide.
Designing a protein is not simply drawing a shape on a screen. A proposed sequence must fold into a stable structure and, if it is meant to perform a task, interact as intended. Computational methods can narrow the search and generate candidates, but researchers then have to make and test those candidates in the laboratory. The experimental step is crucial: a promising design is a hypothesis until its behavior is checked.
One example of the direction of travel is RFdiffusion, a system described by Baker’s group and collaborators in Nature in 2023. The researchers used the method to generate protein structures and designs, then assessed selected candidates experimentally. Work of this kind illustrates how computational design can help explore possibilities that would be difficult to find by trial and error alone. It does not mean that every generated protein will work, or that a computer can specify a useful biological effect without testing.
A faster route from question to experiment
The importance of these methods is less about replacing structural biology than about changing its starting point. A scientist studying a disease-related protein might use a prediction to identify a region worth examining. A researcher developing a molecular tool might use design software to propose candidate proteins, then test which ones fold and function as hoped. In both cases, computation can help prioritize work, while experiments establish whether the ideas hold up.
The two approaches also reinforce a broader shift in chemistry and biology: researchers can increasingly use digital models to explore molecular possibilities before committing to laboratory work. Prediction and design provide different kinds of leverage. One helps interpret the proteins nature has produced; the other explores what proteins might be made to do.
The Nobel-recognized advances mark a substantial change in what scientists can attempt, not a final answer to every question about proteins. A predicted structure is not a complete account of a living molecule, and a designed sequence is not a guaranteed working tool. But when computer models are paired with careful experiments, they can make the search for molecular understanding more focused—and open paths to proteins that researchers would not have known how to look for before.
