A larger map

More than 200 million predicted protein structures are available in the public AlphaFold database. A researcher can look up a model of a protein’s three-dimensional shape without first commissioning an experiment to determine it. The system, developed by Google DeepMind, makes its predictions from the protein’s amino-acid sequence; EMBL-EBI and DeepMind make the database available together. AlphaFold DB

The obvious question is how much experimental work this saves. Another question may prove just as consequential: which experiments does it make worth starting?

A scientist choosing a problem needs a plausible way to reach an answer. An important question with no useful starting point can lose out to a less important question surrounded by tools, data and previous work. Improve the starting points, and researchers may choose different questions even if their laboratories do not become faster.

A prediction can be enough to start

A protein’s structure helps researchers reason about its function. But a predicted structure does not have to be a final, authoritative description to help with the next experiment. It can suggest what to investigate, supply a starting model, or make a previously obscure protein easier to compare with something better understood.

The distinction is supported by experimental work. A study in Nature Methods compared AlphaFold predictions with experimental structures and found that the predictions could accelerate structure determination while still containing errors that mattered. The authors treated them as hypotheses to test. That is a useful way to think about their scientific value: an imperfect model can help organise an investigation without settling its outcome. Terwilliger and colleagues

The economic effect depends on what the preliminary answer unlocks. If researchers were already committed to the same programme, it may save time within that programme. If the lack of a structural model was keeping them away from a protein altogether, it can make a different programme worth attempting.

Ryan Hill and Carolyn Stein find that research shifted toward proteins without prior experimental structures, while experimental structure determination continued at almost the same rate. Their observational comparison assumes the groups would otherwise have followed comparable research trends. The chart shows both research and medical-patent estimates. Working paper

The combination is revealing. A technology can change the direction of inquiry while leaving a familiar measure of laboratory output looking surprisingly stable. Measuring only how much old work disappears would miss that change.

The advantage of a well-lit problem

A well-studied subject offers more tools, more people to ask and more results to build on. Those advantages make the next project easier to justify. An obscure subject can remain obscure partly because nobody has paid the cost of making it approachable.

A broadly available predictive resource could weaken that loop. Instead of producing preliminary knowledge only when someone can already justify the next project, it supplies starting points across a much larger range of possible projects. Researchers can then bring questions to that resource that its creators never thought to ask.

This is the part of AI in science I find most consequential. Time saved on an established project is valuable. Making an unfamiliar problem approachable can also change what enters the research agenda. It may help a field discover where it has been overlooking opportunities.

There is no guarantee that the newly approachable questions are the most valuable ones. Ease can pull research toward attractive-looking predictions just as older tools pulled it toward familiar subjects. Nor does more activity prove that the work will lead to a useful application. The harder test is whether these projects produce findings that survive experimental scrutiny and open further lines of work.

Removing one obstacle need not remove the next. Understanding a protein well enough to study it is different from knowing how to intervene usefully in a biological system. A larger research agenda can coexist with stubborn difficulty further downstream.

Who gets to explore?

The database’s public availability is an institutional choice with scientific consequences. EMBL-EBI describes the resource as freely and openly available to researchers across public and private organisations. Access to the predictions does not require belonging to the organisation that trained the model. EMBL-EBI on open access

This separates the expensive creation of a capability from the decision about what to investigate with it. A central team builds the resource; many other people choose the questions. The potential value comes partly from that division. No single institution has to recognise every worthwhile use in advance.

Open predictions do not equal open science in every other respect. A researcher still needs the skills to interpret them, access to suitable experiments and enough support to pursue a result that may fail. A prediction can be universally available while the opportunity to test it remains concentrated. If those complementary resources become the decisive constraint, improving the model alone may do little to widen participation.

The practical implication is that investment in AI for science should be judged partly by what happens around the model. Can researchers retrieve and compare its outputs? Can they assess uncertainty? Can an unfamiliar lab obtain the experiment needed to challenge a promising prediction? These conditions determine whether a computational advance becomes a shared scientific resource or an impressive capability that only a few institutions can use fully.

AlphaFold does not establish that AI will automate discovery. It gives us a more specific possibility to take seriously: we can make enough preliminary knowledge available that people reconsider what is worth investigating. The transformation may first appear in the questions researchers choose, long before it appears in the products the rest of us can buy.