models/Chai-1
Chai-1
Chai-1 is a computer model that takes the sequence of a protein and works out the shape it folds into, together with the shapes of the molecules that stick to it.
What it is
Proteins are long chains built from a small alphabet of chemical units called amino acids. Each chain folds up into a particular three-dimensional shape, and that shape decides what the protein does. Chai-1 is a trained model that reads the sequence of units in the chain and predicts where every atom sits. It was trained on the many protein shapes that experiments have already measured, learning the patterns that link a sequence to a fold.
It also handles partners: the smaller molecules, short protein fragments and other chains that attach to a protein. Given the sequences of several pieces at once, it predicts how they fit together, and it reports its own confidence, a score saying how sure it is about each part of the answer.
Why scientists use it
Measuring a protein's shape in the laboratory is slow and sometimes impossible, while sequences are cheap and plentiful. A predictor stands in for that measurement. It lets researchers look at a shape for almost any sequence, including ones nobody has crystallised, and it can be run over thousands of candidates in the time one experiment would take.
The partners matter as much as the protein. Drugs, hormones and signals all work by binding, so a model that places the partner as well as the protein can suggest which molecules will stick and where they touch. Because such models are released ready-trained, other groups use them as finished tools and test how far the answers can be trusted.
In the record
Limits
A prediction is a guess shaped by the structures the model was trained on. Where little similar data existed, or where a protein is flexible and has no single shape, the answer can be confidently wrong. Confidence scores are the model's own opinion, not a measurement, and they can be high for a pose that experiments do not support. Placing a small partner correctly is harder than folding the protein itself. Predictions suggest what to test; they do not settle it.
Written by AIxSci: the background from general knowledge, the examples from checked records.