# How does AlphaFold predict protein structure?

> Written with AI · [Biotechnology: bioinformatics, biological data and the computer](https://lessoons.com/en/bioinformatics) · Grade 11 · 2026-11-13 · https://lessoons.com/en/bioinformatics/how-does-alphafold-predict-protein-structure

AlphaFold predicts a protein's three-dimensional structure from its amino acid sequence with a neural network trained on structures measured in the laboratory. It lines up the sequence with related sequences, learns which amino acids sit close together, and builds a model with a confidence score for every amino acid. Its 2020 version came close to experimental accuracy and shared the 2024 Nobel Prize in Chemistry.

## The topic explained

### Shape decides the job

A protein is a chain of amino acids that folds into one precise shape, and the shape decides what it does: an enzyme's active site, an antibody's binding pocket. In 1961 Christian Anfinsen unfolded the enzyme ribonuclease with chemicals and watched it fold back and work again once they were removed. The conclusion, that the amino acid sequence holds all the information for the shape, earned him a share of the 1972 Nobel Prize in Chemistry.

Measuring a shape is slow. X-ray crystallography, NMR and cryo-electron microscopy can take months or years for one protein, and the Protein Data Bank, begun in 1971, holds the results. Sequences, by contrast, are read by the million. Predicting the shape from the sequence, the protein folding problem, stayed largely unsolved for some fifty years, because a chain can in principle take an astronomical number of shapes.

### How AlphaFold makes its prediction

AlphaFold, built at Google DeepMind by a team led by Demis Hassabis and John Jumper, is a neural network, a computer model that learns patterns from examples, trained on the structures in the Protein Data Bank. It starts by lining up the query sequence with related sequences from many species. Where two positions change together through evolution, they are probably touching in the folded protein, since a change in one must be balanced by the other. From these clues and what it learned from known structures, the network works out the distances between amino acids and builds a three-dimensional model. For a protein with few known relatives the clues are weaker, and the prediction is usually less confident.

Progress is judged at CASP, a blind contest held every two years in which teams predict structures that have been measured but not yet published. At CASP14 in 2020, AlphaFold2 reached a median score of 92.4 out of 100, around the level usually treated as matching experiment. A common misreading is that AlphaFold simulates the physics of folding step by step. It does not: it predicts the final shape from patterns learned from sequences and structures.

### Reading the confidence score

Every amino acid in a model carries a confidence score, pLDDT, from 0 to 100. At 90 and above the prediction is usually reliable even in fine detail; from 70 to 90 the backbone of the chain is likely right; from 50 to 70 it should be treated with caution; below 50 it is very uncertain, and such stretches are often disordered regions that have no fixed shape in the cell. A student who shows a low-confidence region as if it were measured has misread the model. In the public database the scores are painted onto the model as colors, dark blue for very high confidence, light blue, yellow, and orange for very low, so one glance shows which parts to trust.

### What changed, and what remains

The team has since predicted structures for almost all of the roughly 200 million known proteins, and the 2024 Nobel Prize in Chemistry went half to Hassabis and Jumper, and half to David Baker for designing new proteins by computer. Researchers now start many projects from a predicted model, to plan a mutation or find where a drug might bind. Yet a prediction is not a measurement: it shows one static shape, may miss how a protein changes when it binds a partner, and is checked by experiment when the details matter. Here too bioinformatics and the laboratory work together: the model offers a detailed hypothesis, and the experiment decides.

## Practice: predicting protein structure

For 11th grade biotechnology students. Answer each question from the sheet, then check your answers.

1. **What did Anfinsen's ribonuclease experiment show, and why does it make structure prediction possible in principle?** — That an unfolded protein folds back into its working shape by itself, so the amino acid sequence holds all the information for the shape. If the sequence determines the shape, the shape can in principle be calculated from it.
2. **Fill in the missing step and explain what AlphaFold learns from it.** — Related sequences, aligned with the query. Positions that change together through evolution are probably in contact in the folded protein, a clue to the shape.
3. **Why was predicting structure from sequence so hard before AlphaFold?** — A chain of hundreds of amino acids can in principle take an astronomical number of shapes, and measuring shapes by experiment is slow, so there was no fast way to find the right one.
4. **Complete the hidden cells: how should regions scoring 50 to 70, and below 50, be used?** — 50 to 70: low confidence, treat with caution. Below 50: very low confidence, often a disordered region with no fixed shape.
5. **Fill in the event of 2020 on the timeline, and say what CASP is.** — AlphaFold2 at CASP14. CASP is a blind contest held every two years in which teams predict structures already measured but not yet published, so predictions can be checked fairly.
6. **A researcher's model has pLDDT above 90 for most of the protein but below 50 for a long loop. What should she conclude about the loop?** — The loop's shape is very uncertain; it may be disordered in the cell. She should not draw conclusions from its predicted position without experimental evidence.
7. **Give two limits of predicted structures.** — For example: a prediction is not a measurement; it shows one static shape; it may miss how a protein changes on binding a partner; low-confidence regions are unreliable.

## Common questions

### Who won the 2024 Nobel Prize in Chemistry?

Half went to David Baker for computational protein design, and half was shared by Demis Hassabis and John Jumper of Google DeepMind for predicting protein structures with AlphaFold.

### Is AlphaFold free to use?

Its database of predicted structures can be searched and downloaded freely, and versions of the program are available to researchers, so a model of most known proteins can be viewed in a browser.

### What is the protein folding problem?

The question of how to predict a protein's three-dimensional shape from its amino acid sequence alone, open for about fifty years until AlphaFold2 came close to experimental accuracy in 2020.

### What does pLDDT mean?

Predicted local distance difference test: AlphaFold's confidence score, from 0 to 100, for each amino acid of a model. Above 90 is very high confidence; below 50 often marks a disordered region.

### Does AlphaFold make experiments unnecessary?

No. It gives a strong starting model, but experiments still measure shapes, check predictions where details matter and show how proteins move and bind partners.
