John Jumper on AlphaFold, Protein Design and Experimental Science

Hannah Fry interviews John Jumper about AlphaFold, his 2024 Nobel Prize and the scientific work enabled by protein-structure predictions. They discuss how predictions generate hypotheses that researchers must still test.

Jumper recalls waiting anxiously on the morning of the Nobel announcement and nearly concluding that the call would not come. He describes the surprise and celebration after receiving the news.

His route into computational biology was not a straight line. Earlier research and work on hardware for protein simulation eventually led him toward AI and a role that combined several of his interests.

The prize gave him a public platform as well as new responsibilities. Jumper reflects on how he might use that position to support public understanding and funding of science.

The team knew AlphaFold 2 had performed well in the CASP structure-prediction assessment, but did not know how rapidly the wider scientific community would adopt it. Jumper says that uptake exceeded his expectations.

A useful prediction can shorten the path from a biological question to a plausible structure and help researchers prioritize experiments. It does not mean that all questions about protein behavior or disease are solved.

After the code and structure database were released, researchers rapidly investigated both capabilities and limitations. Jumper recalls scientists recognizing that some predictions closely resembled structures they had been studying.

The community also learned where caution was needed, including AlphaFold 2’s limited sensitivity to some single-amino-acid changes. Scientists incorporated predictions and confidence estimates into workflows rather than treating every result as certain.

Fry asks about unexpected applications. Jumper mentions work on honeybee biology as an example of a general tool reaching problems the team had not specifically anticipated.

He also describes researchers using structure predictions to narrow a search for proteins involved in fertilization. They then altered or removed a candidate protein to test its role experimentally.

Jumper summarizes the workflow as moving from a broad biological hypothesis to a computational prediction and then to laboratory confirmation. The prediction helps focus the search; it is not the final experiment.

AlphaFold 3 extends the problem from individual proteins to complexes containing proteins and other molecular types, including nucleic acids and small molecules. Accuracy varies by case and model limitations remain.

The team’s “whole PDB” ambition was to handle a broader range of structures in the Protein Data Bank. Some design choices that worked well for proteins in AlphaFold 2 were difficult to extend to other molecules.

Jumper explains that evolutionary information was valuable for protein prediction but was not available in the same form for every molecular component. AlphaFold 3 reduced its centrality in much of the network and emphasized geometric relationships.

Fry uses a wedding seating-plan analogy: deciding where one person sits depends on relationships with others. Jumper agrees that jointly resolving constraints is a useful intuition for the structural problem.

AlphaFold 3 uses a diffusion-based approach that refines an initial noisy arrangement into a predicted structure. Jumper says this provides a more flexible way to handle different molecular types.

That flexibility introduces a risk of plausible-looking but invented structure, particularly in disordered regions. The [AlphaFold 3 paper](https://www.nature.com/articles/s41586-024-07487-w) also lists limitations involving stereochemistry, dynamics and particular targets.

Fry asks how users should respond to hallucinations. Jumper says researchers can inspect confidence measures and use predictions to formulate hypotheses for experiments, while recognizing that no confidence measure is perfect.

On interpretability, Jumper distinguishes understanding a tool’s internals from determining whether it is useful and reliable for a task. He welcomes more interpretability research but does not treat complete internal understanding as the only route to scientific use.

He compares AI development with engineering based partly on practical intuition and testing. For downstream scientists, a computational tool can be useful when its behavior and limits are understood well enough for the experiment.

The conversation turns to disease and drug discovery. Jumper emphasizes that protein structure is only one piece: researchers still need to understand biology, molecular effects, metabolism and many other properties.

He notes the high attrition of drug candidates in clinical development and the difficulty of establishing disease mechanisms. Better structures can help research without guaranteeing a successful therapy.

An example involving a lipid-transport protein illustrates why structure can suggest where a molecule might bind while also revealing a biological function that should not casually be disrupted. The next step is to test the hypothesis.

Fry asks about designing new proteins. Jumper says prediction tools can help evaluate many proposed designs and prioritize those most likely to have the desired structure or interaction.

Researchers have used AlphaFold in protein-design workflows in ways the original team did not fully anticipate. Jumper calls this an example of a community learning how to repurpose a powerful predictive tool.

The discussion mentions AlphaProteo and designed binding proteins. Jumper stresses that designed proteins must still be made and tested in the lab; a computational design is a candidate, not a confirmed product.

Potential applications include tools for studying cells and enzymes for environmental problems. Practical deployment requires properties beyond a single successful reaction, such as stability and repeated activity.

Design and prediction pose different challenges: recognizing a plausible structure is not the same as engineering a molecule that reliably performs a desired function. Jumper expects progress but frames his timeline as a forecast.

Asked whether AI “thinks” or understands biology, Jumper is cautious about the terminology. He prefers to ask whether a system is reliable enough to help solve a scientific problem.

Fry asks about a simulated cell. Jumper notes that biology does not yet have a complete parts list and suggests combining specialized models with systems that can use broader scientific information.

He imagines deeper integration than simply calling AlphaFold as an external tool: a system might need to reason about three-dimensional structures alongside the biological literature. This remains a research goal.

Some questions, such as the origin of life, have limited direct data and may remain difficult even with better models. Computational tools can narrow possibilities, while experiments help decide among them.

Fry concludes that AlphaFold is valuable as a tool that accelerates scientific work. The discussion leaves open the philosophical question of machine understanding while emphasizing practical utility and experimental validation.