AI in science: proteins, weather and more

ProteinsCopy Anchor Link

A protein is a chain of amino acids that folds into a 3D shape, and the shape determines what it does. Predicting the shape from the chain was a 50-year problem. AlphaFold 2 (DeepMind, 2020 to 2021) reached near-experimental accuracy at the CASP14 competition. DeepMind released predicted structures for hundreds of millions of proteins in a free database. In 2024 Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry with David Baker for protein structure prediction and design.

Other areasCopy Anchor Link

Weather forecasting models trained on decades of data now match or beat traditional physics simulations on many measures at a fraction of the compute. Models help search for new materials and drug candidates, and they help read medical images. In math, systems solved olympiad-level problems, with some reaching gold-medal level at the International Mathematical Olympiad in 2025.

CaveatCopy Anchor Link

These systems are tools in a pipeline. Predictions need lab or expert verification, and models can be confidently wrong outside the data they were trained on.

A worked example: why folding is hardCopy Anchor Link

A short protein of 100 building blocks has an astronomical number of possible shapes. Trying them all is impossible. AlphaFold learned from about 170,000 known structures which blocks tend to sit near each other, and predicts the shape directly. The result is not magic. It is pattern knowledge from known proteins, applied to a new one.

How the prediction worksCopy Anchor Link

Related proteins from different species are lined up next to each other. If two positions tend to change together across species, they are probably close in 3D, because a change at one spot needs a matching change at the other to keep the protein stable. The network learns to read these paired signals and turns them into a predicted shape. This is a simplified picture, but it shows why the method needs many known proteins to learn from.

A second example: weather as repeated guessingCopy Anchor Link

A learned weather model takes the state of the atmosphere now and predicts the state a short time ahead. To forecast further, it feeds its own output back in and repeats. If each step covers 6 hours, a 10-day forecast takes 10 × 24 ÷ 6 = 40 steps (invented step size). A small error in one step is carried into the next, so errors can grow. A physics simulation solves equations instead. A learned model copies patterns, which is why it is fast once trained but can struggle with situations that rarely appeared in its data.

Where it breaksCopy Anchor Link

The symptom is a smooth, confident answer that is wrong for an unusual case, such as a rare extreme storm or a protein unlike any known one. Flexible protein regions that have no single fixed shape are another weak spot. To check, test the model on past events it did not train on, and report the error on the extreme cases separately from the average. A good average can hide poor results on the cases that matter most.

Try it yourself: read a confidence mapCopy Anchor Link

Open the AlphaFold Protein Structure Database and search for a protein you have heard of, such as insulin. The viewer colors each region by how confident the model is. Find a region with low confidence and note where it sits. Many readers see that flexible ends and loose loops are less certain than the core. This shows that the tool reports its own doubt, and that you should read it.

Common mistakeCopy Anchor Link

Treating a predicted structure as a measured one. It is a strong hypothesis. Check with experiments before relying on it for a drug.