Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How AI predicts protein structures

    Proteins work because of their 3D shape, and measuring that shape in the lab takes months to years.[1] Deep-learning models such as AlphaFold now predict many structures from sequence alone with accuracy competitive with experiment, work that shared the 2024 Nobel Prize in Chemistry.[2][3]

    Editor reviewedUpdated AI for scienceLife sciencesArtificial intelligenceScience

    Proteins are chains of amino acids that fold into a 3D shape, and that shape decides what they do. Experimental methods had determined structures for about 100,000 unique proteins by 2021, a small fraction of the billions of known sequences, because each structure can take months to years.[1] Predicting the shape from the sequence alone is what the Nobel committee called a 50-year-old problem.[4]

    Why the shape matters

    Think of a protein as a string of beads that crumples into a precise knot. The knot’s shape lets it grab another molecule, cut it, or carry a signal. If you know the shape, you can guess how the protein works and how a drug might block it. Scientists used to measure each knot by hand, slowly.[1]

    Structure links sequence to mechanism: binding sites, catalytic residues and interfaces are only interpretable in 3D. Experimental structures are deposited in the protein-data-bank, the single global archive for 3D structures of proteins, nucleic acids and complexes since 1971.[5] As of October 2026 it held 260,626 experimental structures.[6]

    How the models work

    An AI model studies a large archive of solved structures and learns the patterns that link a sequence to its shape. It also compares the protein with its evolutionary relatives: if two positions always change together across species, they are probably touching in 3D. alphafold learned from the Protein Data Bank in this way.[7][8]

    AlphaFold2 combined a deep network with physical and biological priors and used multiple sequence alignments (MSAs) as a key input.[8] It was trained on a PDB snapshot from August 2019.[7] AlphaFold 3 replaced much of that architecture with a diffusion-based module that generates atom coordinates for whole complexes, covering proteins, nucleic acids, small molecules, ions and modified residues.[9]

    How good is it?

    The test that matters is blind prediction. In the CASP14 assessment, AlphaFold2 reached accuracy competitive with experimental structures in a majority of cases.[2] AlphaFold 3 reported far greater accuracy for protein–ligand interactions than standard docking tools and better antibody–antigen predictions than its predecessor.[10] The Nobel committee said AlphaFold2 has been used to predict structures for virtually all 200 million known proteins.[11]

    What is still hard

    Some proteins change shape, work in large groups, or bind to antibodies in ways the models still miss. The scientists who run the main blind test now aim it at these harder cases.[12]

    CASP17, whose prediction season closed in September 2026, targets areas where deep learning has yet to deliver: immune complexes, protein–ligand complexes, nucleic acids, conformational ensembles and difficult targets.[12][13] A prediction is a model, not a measurement. Predicted structures sit alongside experimental ones, which is why the RCSB PDB lists more than a million computed models separately from its experimental entries.[6]

    Who can use these models

    Anyone can look up a predicted structure for free: the AlphaFold Database offers over 260 million of them.[14] Running the newest models yourself is less simple. AlphaFold 3 is free on a web server for non-commercial work, but its full model is available only on Google’s terms.[15]

    AlphaFold 3’s inference code is public, but its parameters may only be used if obtained directly from Google.[15] Open reimplementations fill the gap: Boltz-1 is described by its developers as the first fully open model to approach AlphaFold 3 accuracy under an MIT licence, and OpenFold3-preview aims at a bitwise reproduction under Apache 2.0.[16][17] Boltz-2 adds binding-affinity prediction, a step toward drug design.[18]

    From prediction to design

    The same networks can be run in reverse to invent proteins. rfdiffusion was built by fine-tuning a structure predictor, RoseTTAFold, to generate new protein backbones.[19] That step, from reading nature’s proteins to writing new ones, is covered in the design pages of this course.

    Questions readers ask

    How many protein structures had been solved before AlphaFold?

    Around 100,000 unique proteins had experimentally determined structures, a small fraction of the billions of known sequences.[1]

    Where does AlphaFold's training data come from?

    AlphaFold2 was trained on a copy of the Protein Data Bank, the open archive of experimentally determined structures.[7][5]

    Can these models predict how proteins bind drugs or DNA?

    AlphaFold 3 predicts complexes of proteins with nucleic acids, small molecules and ions, and reported much better protein–ligand accuracy than docking tools.[9][10]

    Is protein structure prediction a solved problem?

    Not entirely. The 2026 CASP17 assessment focuses on areas where deep learning has yet to deliver, such as immune complexes and conformational ensembles.[12]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      Before AlphaFold2, experiments had determined structures for around 100,000 unique proteins, a small fraction of the billions of known protein sequences, because each structure takes months to years of work. confirmedas of 2021-07-15

    2. [2]

      AlphaFold2, published in Nature in July 2021, was validated in the CASP14 blind assessment and predicted structures with accuracy competitive with experiment in a majority of cases. confirmedas of 2021-07-15

    3. [3]

      The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for protein structure prediction. confirmedas of 2024-10-09

    4. [4]

      The Nobel committee described protein structure prediction as a 50-year-old problem that the AlphaFold AI model solved. confirmedas of 2024-10-09

    5. [5]

      Since 1971 the Protein Data Bank has been the single repository of 3D structures of proteins, nucleic acids and complex assemblies, managed by the Worldwide PDB partnership. confirmedas of 2026-10-10

    6. [6]

      As of 10 October 2026 the RCSB PDB listed 260,626 experimental structures and 1,062,058 computed structure models from AlphaFold DB and ModelArchive. confirmedas of 2026-10-10

    7. [7]

      AlphaFold2 was trained on a copy of the Protein Data Bank downloaded in August 2019. confirmedas of 2021-07-15

    8. [8]

      AlphaFold2 combines a deep-learning architecture with physical and biological knowledge about protein structure and uses multiple sequence alignments of related proteins. confirmedas of 2021-07-15

    9. [9]

      AlphaFold 3, published in Nature in May 2024, uses a diffusion-based architecture to predict joint structures of complexes containing proteins, nucleic acids, small molecules, ions and modified residues. confirmedas of 2024-05-08

    10. [10]

      The AlphaFold 3 paper reported far greater accuracy for protein–ligand interactions than state-of-the-art docking tools and higher antibody–antigen accuracy than AlphaFold-Multimer. confirmedas of 2024-05-08

    11. [11]

      According to the Nobel committee, AlphaFold2 has been used to predict the structure of virtually all 200 million known proteins and has been used by more than two million people in 190 countries. confirmedas of 2024-10-09

    12. [12]

      CASP17 emphasises areas where deep learning has yet to deliver, including immune complexes, protein–ligand complexes, nucleic acids and conformational ensembles. confirmedas of 2026-10-10

      • CASP17 home page · Protein Structure Prediction Center (UC Davis) (retrieved 2026-10-10)
    13. [13]

      The CASP17 prediction season ended on 11 September 2026, with assessment results due on 30 November 2026, a day before the CASP17 conference in Rome. confirmedas of 2026-10-10

      • CASP17 home page · Protein Structure Prediction Center (UC Davis) (retrieved 2026-10-10)
    14. [14]

      As of October 2026 the AlphaFold Protein Structure Database, developed by Google DeepMind and EMBL-EBI, said it provides open access to over 260 million protein structure predictions. confirmedas of 2026-10-10

    15. [15]

      AlphaFold 3's inference code is public on GitHub, but its model parameters may only be used if obtained directly from Google under terms of use; the model is also offered on AlphaFold Server for non-commercial use, with a more limited set of ligands. confirmedas of 2026-10-10

    16. [16]

      The Boltz team describes Boltz-1 as the first fully open-source model to approach AlphaFold 3 accuracy, and releases Boltz code and weights under the MIT licence for academic and commercial use. confirmedas of 2026-10-10

    17. [17]

      OpenFold3-preview, from Columbia University's AlQuraishi Lab and the OpenFold consortium, aims to be a bitwise reproduction of AlphaFold 3 and is available under the Apache 2.0 licence for academic and commercial use. confirmedas of 2026-10-10

    18. [18]

      Boltz-2, described in a June 2025 preprint by MIT and Recursion researchers, predicts both complex structures and small molecule–protein binding affinity, and its authors say it is the first AI model to approach free-energy perturbation accuracy while being at least 1,000 times more computationally efficient. confirmedas of 2025-06-18

    19. [19]

      RFdiffusion, published in Nature in July 2023, was made by fine-tuning the RoseTTAFold structure prediction network on structure denoising, producing a generative model of protein backbones. confirmedas of 2023-07-11

    Revision history (2)
    1. Page created.
    2. Added a section on who can use these models: AlphaFold 3 access terms and open alternatives (Boltz, OpenFold3).

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "How AI predicts protein structures." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/how-ai-predicts-protein-structures

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents