AI for scienceMiddleLesson 2525 min read

Protein structure and AlphaFold

Proteins fold into 3D shapes, and shape decides what they do. AlphaFold predicts those shapes from the sequence. Learn what it does, what it does not do, and how to explore its public database.

✦ Ask the tutor 🧪 Lab · 25 min 📝 Lesson test

Speed

Lesson in motion

In 60 seconds

Protein structure and AlphaFold

Proteins fold into 3D shapes, and shape decides what they do. AlphaFold predicts those shapes from the sequence. Learn what it does, what it does not do, and how to explore its…

1/4

In simple words

Take a long necklace of beads and drop it in a box. It crumples into a shape. Now imagine the beads are sticky in special ways, so the necklace always folds into the same shape. Proteins are like that necklace. Their shape decides their job in your body.

What protein folding is

A protein is a long chain made from small building blocks called amino acids. There are about twenty common kinds. The order of amino acids is the protein's sequence, and it is written in the cell's DNA.

After the chain is made, it folds into a 3D shape. Parts of the chain attract or push away from each other and from water. The final shape is called the protein's structure.

Why structure matters

Shape decides function. An enzyme has a pocket that fits one kind of molecule, like a lock and key.
Many medicines work by fitting into a pocket on a protein. Knowing the shape helps chemists design them.
Some diseases are linked to proteins that fold the wrong way.

Scientists find structures with experiments such as X-ray crystallography, cryo-electron microscopy and NMR. These methods are powerful but can take a lot of time, skill and money. Many solved structures are shared in a public archive called the Protein Data Bank (PDB).

What AlphaFold does

AlphaFold is an AI system from DeepMind. You give it a protein sequence, and it predicts the 3D structure. It learned from the many structures that scientists had already solved by experiment.

There is a regular blind competition for structure prediction called CASP. Teams predict structures that have been solved but not yet made public, and then predictions are compared with the real ones. AlphaFold 2 performed strongly at CASP14 in 2020, and this was widely seen as a major step forward.

The 2024 Nobel Prize in Chemistry was shared by David Baker, Demis Hassabis and John Jumper. Hassabis and Jumper were recognised for protein structure prediction with AlphaFold. Baker was recognised for his work on computational protein design: designing new proteins.

The public AlphaFold database

DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI) run a free public database of predicted structures at alphafold.ebi.ac.uk. Anyone can search it by protein name, view the 3D model and download files.

Each prediction comes with a confidence score for every part of the chain, shown with colours. High confidence usually means the model is likely close to the real shape there. Low confidence can mean the model is unsure, or that this part of the protein is floppy and has no single fixed shape. Always read the colour legend on the page.

What AlphaFold does not do

AlphaFold gives you	It does not give you
A predicted structure, usually one main shape	How the protein moves and changes shape over time (its dynamics)
A confidence estimate for each part	Certainty. A prediction is not an experiment.
A strong starting point for research	The protein's full function. Shape is a clue, not the whole story.
Structures for many known sequences	All the details of how it binds to drugs, DNA or other proteins in the cell

Note

Newer versions of AlphaFold and other tools try to predict how proteins fit together with other molecules. What each version can do changes, so check the current documentation before you rely on it.

Overclaiming

"AlphaFold solved biology."
Using a low-confidence region as if it were certain
Skipping experiments because a prediction exists

Accurate

"AlphaFold made structure prediction much better and faster."
Reading the confidence colours before drawing conclusions
Using predictions to plan better experiments

Real example

A student in Hyderabad studying a plant enzyme may have no structure from experiments. With a predicted structure, she can guess where the active pocket is and plan which amino acids to change in the lab. The prediction does not answer her question. It helps her ask a sharper one.

Watch and read more

AlphaFold Protein Structure DatabaseEMBL-EBI · reading

Lab

A short protein profile made from exploring one entry in the public AlphaFold database.

~25 min

The problem

Open alphafold.ebi.ac.uk. No login is needed.

Search for a protein you have heard of, for example insulin or haemoglobin. Pick the entry for the human protein (check the organism name).
Write down: the protein name, the organism, and what the page says the protein does (if it gives a summary).
Rotate the 3D model. Describe the shape in two sentences: coils, flat sheets, loose loops?
Look at the confidence colours and legend. Which parts have high confidence and which have low confidence?
Write three sentences: what this prediction can help a scientist do, and one thing it cannot tell you.

You are done when

You picked a real entry and recorded the protein and organism correctly
You described the shape in your own words
You used the confidence colours and legend to describe which parts are reliable
You named one clear limit, such as movement, binding or full function

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Your protein has a long low-confidence tail at one end. A friend says, "The model failed there." Is that the only explanation?Reveal

No. Some parts of real proteins are intrinsically disordered: they do not have one fixed shape and may only fold when they bind to a partner. Low confidence there can reflect real flexibility, not only a model failure. You would need other evidence, such as experiments, to tell the difference.

Q2AlphaFold learned from structures in the Protein Data Bank. Why might it be less reliable for some kinds of proteins?Reveal

A model is strongest where its training data is rich. Proteins that are unusual, poorly represented in known structures, or that change shape a lot depending on their partners may be harder to predict. This is a general rule for all machine learning: be most careful far from the training data.

Please sign in to continue.

Questions people ask

Did AlphaFold make lab experiments unnecessary?

No. Experiments are still needed to confirm structures, to see how proteins move, and to study how they bind to other molecules. Predictions help scientists decide which experiments to do, and sometimes help them make sense of experimental data.

Why does the order of amino acids decide the shape?

Each amino acid has different chemical properties. Some like water, some avoid it, some carry charge. As the chain folds, these properties push it towards a shape where the parts fit together well. So the sequence carries the instructions for the fold.

What does a low-confidence region mean?

The model is not sure about the shape there. Sometimes this is because that part of the real protein is flexible and has no single fixed shape. Either way, do not treat that region as reliable without other evidence.

Can I use the database for a school project?

Yes. It is free to search and view. Check the licence and citation guidance on the website before you reuse images or data in a report, and say that the structures are predictions.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.