A molecular biologist deciding which compound to order and a weather forecaster watching a cyclone face a similar problem. Each can ask a computer for thousands of plausible answers. Each must choose what to do before the evidence is complete. The future of AI in science depends partly on what happens in that interval. Two studies show how much a prediction can buy, and how readily its value can disappear when the question changes.
Where does the molecule actually sit?
Proteins have surfaces with dips, ridges and pockets. A small molecule that settles in a pocket may change what the protein does. Knowing roughly where the two meet gives a chemist somewhere to begin. Which part of the compound could be made larger? Which part might already press against the pocket wall? A wrong turn at this stage can send a team away to make a series of compounds that all share the same failing.
In 2024, Josh Abramson and colleagues described AlphaFold 3, a system that predicts how combinations of proteins, small molecules and nucleic acids such as DNA fit together. The team compared its predicted arrangements with structures determined experimentally. For several classes of interaction, including proteins with small molecules, it improved on the methods used for comparison. A chemist can inspect the resulting map atom by atom, including the model’s confidence around the proposed contact.
One failure in the paper is particularly instructive. A group of proteins called E3 ubiquitin ligases can adopt an open shape when a binding partner is absent and a closed one when it is present. In the example reported, the model supplied the closed shape in both cases. A convincing picture of the closed protein would offer little help to a researcher looking for the pocket exposed when it opens.
Putting a shortlist on the laboratory bench
A separate study illustrates what a predicted shape can still accomplish. Researchers working with the earlier AlphaFold2 set out to find molecules that bind to two receptors, including one involved in serotonin signalling. They screened virtual libraries against predicted receptor structures and compared the results with screens using experimentally determined structures. Then they tested hundreds of candidate molecules. Both routes produced similar proportions of hits in this prospective ligand discovery study.
The interesting detail is that the two routes suggested different compounds. For one promising serotonin receptor candidate, the team used cryogenic electron microscopy to examine how the compound sat in its pocket. Parts of the measured pocket resembled the prediction. A second route into the same target had opened up, and there was a physical specimen against which to check the proposed explanation. A binding test alone leaves the compound’s effects inside cells unresolved.
For a future project, that result suggests a specific trial. Give two teams the same budget to buy and test compounds. Let one select them using the usual method and the other add predicted structures. Count confirmed hits, but also ask how many genuinely different chemical families each team found. Ten close relatives of the same compound leave a chemist with fewer directions to explore if a problem later emerges across that family.
Why a forecast needs several futures
In weather forecasting, the uncertainty starts before the first calculation. Observations reach meteorologists from scattered instruments at different times. A forecast begins with an estimate of the atmosphere assembled from those measurements. Small differences in the starting conditions can grow as the weather develops. For a harbour expecting high winds next week, several possible storm tracks convey something the single likeliest track cannot.
Ilan Price and colleagues built GenCast to generate an ensemble, a collection of possible weather sequences. In a comparison using 2019 forecasts, it outperformed the European Centre for Medium-Range Weather Forecasts’ operational ensemble of that period on 97.2 per cent of 1,320 targets. The targets combined weather variables, atmospheric levels and forecast lead times. That percentage describes wins across the chosen comparisons. A harbourmaster needs a different calculation to judge the chance of dangerous winds on Tuesday.
Consider a harbour that suspends crane operations above a particular wind speed. A forecast could get the weekly average almost right while missing the six hours that matter. To test its value, the harbour would compare forecasts issued before each shift with local wind measurements. Among shifts assigned a one in five chance of crossing the threshold, how often did that happen? How many dangerous shifts received a reassuring forecast? Both questions matter when deciding when to secure the cranes before the storm.
Some of this work has already reached the forecast desk. ECMWF put its own AI ensemble into operational service in July 2025 and upgraded it in May 2026. A centre running several models can investigate days on which their forecasts diverge. For the harbour, a disagreement about when the wind will peak might matter more than a small difference in its predicted strength. Comparing both timings with the local wind readings would reveal which forecast gave the crew enough time to finish unloading safely.
Storms can hide below the grid
A cyclone provides a physical reason for caution. The region containing its strongest winds can be small compared with the spacing of points on a global weather model’s grid. The forecast may locate the storm centre well while smoothing away the peak winds. In August 2026, ECMWF and Cambridge researchers reported testing a correction for cyclone intensity using storms from 2025, a season held back from training. The correction reduced the model’s average underestimate of wind speed from nearly 29 knots to about two.
An average bias near zero leaves room for substantial misses on individual storms. An overestimate for one system can cancel an underestimate for another. The next question concerns the worst misses, especially storms that strengthen rapidly close to land. New observations of ocean conditions or winds might help, provided researchers can show that a forecaster would actually have received those observations before issuing the forecast.
Tests worth keeping secret
Structural biologists have built an institution around withholding the answer. In the CASP protein prediction exercise, participating laboratories keep the measured structure confidential while modellers submit predictions. The 2026 call specifically sought difficult examples, including membrane proteins and complexes with limited evolutionary information. Each submission leaves a dated record of what its authors could infer before anyone revealed the experimental coordinates.
Weather researchers can borrow that discipline by archiving every forecast as it is issued, including the uneventful days. A model that repeatedly predicts a rare storm will eventually be able to point to an impressive success. The quiet days supply the cost of that success in false alarms. Recording the whole series also allows a second team to check how a method behaved after a change of instruments or during an unusual season.
For the biologist, the most valuable next prediction may identify two experiments whose results would distinguish an open protein from a closed one. For the forecaster, it may identify the incoming observation most likely to change a warning. If models can help researchers spend a scarce instrument hour on the measurement that settles a disagreement, their contribution will be visible in the next sample tube or the next set of observations.