Protein structures used to be measured in careers, not minutes. Months at the low end, years at the high end, all to pin down the shape of one molecule. Then in November 2020 the organisers of the fourteenth Critical Assessment of Structure Prediction announced that a DeepMind entry called AlphaFold2 had outscored every other method by a margin nobody in the field had seen before.

Across all targets it posted a median Global Distance Test score of 92.4, and 87.0 in the free-modelling category, where no closely related structure exists to work from. The next best methods managed 72.8 and 61.0.

Those numbers come from the assessment itself and are reproduced in the AlphaFold Protein Structure Database FAQ. AlphaFold2 was described the following July in Nature, in a paper led by John Jumper. Two years after that, the predicted structures of roughly 200 million proteins were sitting on a public server, downloadable by anyone, at no cost.

Why the old way took so long

Jumper and his co-authors open by naming the bottleneck the system was built to address. At the time of writing, the structures of about 100,000 unique proteins had been determined experimentally, against billions of known sequences, and they attribute the gap to the “months to years of painstaking effort” a single structure demanded. X-ray crystallography requires coaxing a protein into a crystal, which some proteins simply refuse to do. Nuclear magnetic resonance and cryo-electron microscopy have their own constraints.

The Protein Data Bank, the archive where those experimental results are deposited, now holds more than 240,000 structures, according to the RCSB PDB team’s 2026 update in Nucleic Acids Research.

That is the accumulated output of half a century of laboratory work across thousands of groups.

AlphaFold2 predicts a structure from the amino acid sequence alone. Inference takes something in the region of a GPU-minute for a 384-residue protein. In practice a user also has to build the multiple sequence alignments the model draws on, which takes additional time, and this is precisely what later tools such as ColabFold were written to speed up.

From 350,000 to 214 million

DeepMind and EMBL’s European Bioinformatics Institute launched the database on 22 July 2021 with roughly 350,000 entries: the human proteome plus twenty other model organisms, among them E. coli, the fruit fly, the mouse and the malaria parasite. A year later, in July 2022, the pair expanded it to beyond 200 million predictions, covering close to the whole of UniProt. That release predicted the shape of nearly every protein catalogued by science.

Mihaly Varadi and colleagues at EMBL-EBI and Google DeepMind put the running total at 214 million in their 2024 database paper for Nucleic Acids Research. They also document a numerical bug in the July 2022 release that produced low-accuracy predictions for about four per cent of entries, corrected that November. AlphaFold DB now runs a public changelog and versioned coordinate files, which is the kind of housekeeping that distinguishes a maintained scientific resource from a data dump.

Access is genuinely open. The data carries a CC-BY-4.0 licence covering both academic and commercial use, with attribution expected. Writing in Fortune in February 2026, Demis Hassabis and James Manyika put usage at more than three million researchers across upwards of 190 countries, with over a third of them in low- and middle-income countries. Those are company figures rather than independently audited ones, though the integration of AlphaFold predictions into UniProt, Ensembl, InterPro and the PDB itself is consistent with uptake on that scale.

What the confidence score is doing

Every residue in every entry carries a pLDDT score from 0 to 100, the model’s own estimate of how well it has done. Varadi and co-authors set out the interpretation: above 90 is modelled to high accuracy and suitable for work that depends on it, such as characterising binding sites; between 70 and 90 gives a reliable backbone; between 50 and 70 warrants caution.

Applied to the full 214 million, EMBL-EBI’s assessment is that about 35 per cent of predictions are highly accurate, comparable to experimental structures, with a further 45 per cent confident enough to rely on for many purposes. That leaves a fifth of the library where the model is telling users, in effect, that it is unsure.

Summaries tend to drop that last number, which gets the resource backwards: sorting by confidence is what makes a library this size usable at all, since it lets a researcher separate the entries they can build on from the ones they cannot.

A prediction is not an experiment

Low pLDDT regions are not simply failures. Apolline Bruley and colleagues, writing in Biomolecules in 2022 under the title “Digging into the 3D Structure Predictions of AlphaFold2 with Low Confidence: Disorder and Beyond,” found that low-confidence regions frequently correspond to intrinsically disordered stretches of protein, where there is no single stable shape to predict. Frequently, not always. Their own framing leaves room for other causes, and a low score is a flag to investigate rather than an explanation in itself.

A second limit runs alongside it. AlphaFold2 returns a structure, singular, and many proteins move between conformations as part of how they work. Hannah Wayment-Steele and co-authors proposed a way to coax multiple conformations out of the model by clustering input sequences, a method called AF-Cluster, in a 2024 Nature paper.

That method has since been contested. A Matters Arising from February 2025, from a team at the US National Institutes of Health led by Joseph Schafer and Lauren Porter, reported that AF-Cluster underperforms random sequence sampling on the same targets. Its confidence scores also failed to separate correct predictions from incorrect ones without prior knowledge of the answer. Both papers stand, but the direction of the dispute matters: predicting how a protein moves remains open in a way that predicting its dominant fold no longer is.

RCSB.org handles the underlying distinction structurally. It serves upwards of a million computed structure models alongside its experimental entries, but flags their provenance clearly and excludes them from searches unless a user opts in.

The giveaway had limits

The 2024 Nobel Prize in Chemistry, announced on 9 October, went half to David Baker for computational protein design and half jointly to Hassabis and Jumper for protein structure prediction. By then the openness question had already turned.

AlphaFold 3 arrived in May 2024, in a Nature paper with Josh Abramson as lead author, extending prediction to proteins interacting with DNA, RNA and small molecules. It came without its code. An open letter from researchers objecting to publication on those terms gathered more than a thousand signatures, and DeepMind released the code on 11 November 2024 for non-commercial use, with model weights available to academics on request. Nature’s own news coverage of the release carried a correction three days later clarifying that the tool is not, in fact, open source.

Those 214 million entries remain free. Its successor does not sit under the same terms, and the commercial interest is explicit: Isomorphic Labs, the drug discovery company spun out of DeepMind, was named in the reasoning for the initial restriction.

What is worth watching is whether the free tier holds as the capability moves towards molecular interaction, which is where the money in drug discovery actually is. A 2025 update aligned coverage with the UniProt 2025_03 release and added isoforms and the underlying sequence alignments. That is a resource still being fed. Whether the same institutions keep giving away the more commercially valuable version is an open question, and not one the licence file answers.