Structure Prediction: AlphaFold and Its Limits

What structure prediction does

Proteins are long chains of amino acids that fold into three-dimensional shapes, and the shape helps decide what a protein does. Measuring a shape in the lab can take months to years for a single protein, and the structures of around 100,000 unique proteins had been determined by 2021, a small fraction of the billions of known protein sequences. [1] Structure prediction tools use artificial intelligence to estimate a shape from the sequence alone. This page is information only.

AlphaFold

In 2021 the AlphaFold team reported a method that regularly predicted structures with atomic accuracy in many cases, including when no similar structure was known, and showed accuracy competitive with experimental structures in a majority of cases in a blind assessment called CASP14. [1] The authors filed patent applications relating to the method. [1] In 2024 the Nobel Prize in Chemistry recognized Demis Hassabis and John Jumper for protein structure prediction, and David Baker for computational protein design. [2] That recognition is for the method, not for any medicine.

What predictions do not capture

An independent test compared AlphaFold predictions with experimental crystallography maps. In many cases the match was close. In other cases even very high-confidence predictions differed from the experimental map in overall distortion, domain orientation and local detail. [3] The authors also note that predictions do not account for ligands (molecules that attach to a protein), covalent modifications or other environmental factors. [3] They recommend treating predictions as exceptionally useful hypotheses and confirming structural details by experiment. [3]

A 2025 review of AlphaFold3 found it falls short in reliably predicting RNA and protein-RNA complex structures, especially for non-canonical interactions where training data are scarce. [4] Users are advised to treat such predictions as hypothesis generators. [4]

What a predicted shape can and cannot establish

A tool that predicts a shape says nothing about dose, side effects, interactions or whether changing the protein helps a person. Those are questions for cells, animals and, above all, human studies. See mechanism vs demonstrated benefit and the translational pipeline.

How it connects to other pages

Predicted shapes feed into molecular modeling and docking and into AI-assisted drug discovery. For peptide products, see peptide research methods. For background on shape and function, see why molecular structure matters.

Reading a claim that cites AlphaFold

• Was the structure measured, or predicted? Look for the words "predicted" or "model".
• Did the study test the idea in cells or people, or stop at the picture?
• Does the product claim go beyond the structure, for example to a benefit?
• Remember that a public database of predictions is a catalogue of hypotheses. I could not read the AlphaFold Database help pages in this session, so nothing here describes that database.