AI glossary · Advanced · Science & bio · Healthcare & bio
Protein structure prediction
Protein structure prediction means working out computationally, from only the sequence of amino acids that make up a protein, what three-dimensional shape the protein actually folds into.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Industry AI
- Fields
- Science & bio
- Industries
- Healthcare & bio
- Added
- 2026-10-10
How it works
A protein is a molecule made of 20 kinds of amino acids linked in a chain, and this chain folds into a specific three-dimensional shape. In the past, shapes were determined by experiments such as X-ray crystallography or cryo-electron microscopy, which took a lot of time and money. The AI approach trains a deep learning model on data from many proteins whose shapes are already known. The model uses clues such as pairs of amino acids that have changed together during evolution, along with attention-based architectures, to predict the 3D positions of atoms, and it also outputs a score for how reliable the prediction is.
From sequence to structure
- 1Amino acid sequence inputInformation about a single linked chain
- 2Gathering cluesAmino acid pairs that changed together, etc.
- 3Deep learning predictionComputing 3D positions of atoms
- 4Confidence scoreHow reliable each region is
- 5Experimental verificationConfirm before important decisions
Why it matters
A protein's shape largely determines what job it does in the body. Drugs usually work by fitting snugly into a pocket on a specific protein, so knowing the shape makes it much faster to narrow down which drug candidates to make. AI prediction has made it possible to obtain structures quickly that would have taken a long time experimentally, and predicted structures for many proteins are being shared in public databases. That is why it is considered a foundational technology for AI drug discovery, enzyme design and disease research.
Things to know
A prediction is still only a prediction, so important decisions still require experimental confirmation. Proteins actually move and change shape, but prediction results mainly show a single static shape. Floppy regions without a fixed shape, and shapes that change when binding to other molecules, are harder to predict. Recently, research has been expanding beyond single proteins to predicting binding with other proteins, DNA and drug molecules.
Example
The best-known example is Google DeepMind's AlphaFold. When researchers enter the amino acid sequence of a protein they are interested in, they get a predicted 3D structure and per-region confidence, and precomputed predicted structures can also be downloaded from a public database. Pharmaceutical companies and research labs use these predicted structures as a starting point to find where a drug might bind, select candidate compounds and then verify them experimentally.