# What is AlphaFold: DeepMind's AI for predicting protein structures

> How AlphaFold 2 solved protein folding, what AlphaFold 3 added, how scientists can use it for free, where it fails, and why Isomorphic Labs just raised $5 billion on it.

URL: https://claudegpt.si/news/what-is-alphafold  
Published: 2026-11-17T15:04:44.349Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: AlphaFold, Google DeepMind, Isomorphic Labs, Biology, Drug discovery  
Publisher: ClaudeGPT Si

## Key takeaways

- AlphaFold is a Google DeepMind AI system that predicts a protein's 3D shape from its amino acid sequence, and AlphaFold 3 extends this to DNA, RNA and drug-like molecules.
- Its database of more than 200 million predicted structures is free for anyone, while AlphaFold Server and the AlphaFold 3 weights are free only for non-commercial research.
- As of November 17, 2026 there is no public AlphaFold 4; Isomorphic Labs keeps its newer IsoDDE engine in-house.

What is AlphaFold, in short: an AI system from Google DeepMind that predicts the 3D shape of a protein from its sequence of amino acids, the chemical letters that make it up. Its newest public version, AlphaFold 3, also predicts how proteins fit together with DNA, RNA and drug-like molecules, and is free for non-commercial research through the [AlphaFold Server](https://alphafoldserver.com).

The name is back in the news because Isomorphic Labs, the drug discovery company DeepMind spun out in 2021 to build on AlphaFold, [announced a $5 billion Series C](https://www.isomorphiclabs.com/articles/isomorphic-labs-announces-series-c-investment-round) on November 17, 2026. Our [report on the round](https://claudegpt.si/news/isomorphic-labs-series-c-5-billion) covers who invested.

## The protein folding problem in plain words

A protein starts as a long chain of amino acids. It then folds into a precise 3D shape, and that shape decides what it does: carry oxygen, fight an infection, copy DNA. Knowing the shape shows where a drug might attach.

The shape is the hard part. Working it out in the lab, with methods such as X-ray crystallography, could take years and hundreds of thousands of dollars per protein, [according to DeepMind](https://deepmind.google/science/alphafold/). Predicting the shape from the sequence alone had been "an important open research problem for more than 50 years," as the [AlphaFold 2 paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC8371605/) puts it. AlphaFold does it in minutes.

## AlphaFold 2: the version that cracked it

The breakthrough came at CASP14, a blind contest for structure prediction, where DeepMind's AlphaFold 2 results were announced on November 30, 2020.

Its predictions had a median backbone error of 0.96 ångströms, against 2.8 ångströms for the next best method. An ångström is a ten-billionth of a meter, about the width of an atom.

The model reads a protein's sequence plus a multiple sequence alignment: the same protein's versions in many other species, which hint at which parts sit close together in 3D. DeepMind published the method in [Nature on July 15, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8371605/) and released the code. A later version, AlphaFold-Multimer, extended it to complexes made of several protein chains.

## The AlphaFold Protein Structure Database

DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI) ran the model on known proteins and published the results.

- The [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk/) opened on July 22, 2021 with more than 350,000 structures.
- On July 28, 2022 it grew to more than 200 million, covering nearly every catalogued protein known to science. The database homepage now lists more than 260 million predictions.
- The data is free under a CC-BY-4.0 licence, which allows academic and commercial use with attribution.
- In October 2026 it added links from individual amino acids in a structure to the papers that discuss them.

DeepMind says AlphaFold is now used by more than 3 million researchers in over 190 countries.

## AlphaFold 3: beyond proteins

Proteins rarely work alone, and most drugs work by binding to one. AlphaFold 3, [published in Nature on May 8, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11168924/) by Google DeepMind and Isomorphic Labs, predicts the joint structure of all of these together: proteins, DNA, RNA, ligands (small molecules, including many drugs), ions and chemical modifications.

AlphaFold 3 also adds a diffusion network, the technique behind AI image generators. It starts from a cloud of atoms and refines it step by step into a final structure.

The [companies claim](https://www.isomorphiclabs.com/articles/alphafold-3-predicts-the-structure-and-interactions-of-all-of-lifes-molecules) at least a 50% improvement over existing methods for proteins interacting with other molecule types, and double the accuracy for some important categories. On the PoseBusters benchmark of protein-ligand complexes, they say it is 50% more accurate than the best traditional docking tools, without being given any structural input.

## How scientists can use AlphaFold today

There are four routes, with different rules:

| Route | What you get | Who can use it |
| --- | --- | --- |
| [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk/) | Ready-made predictions for over 200 million proteins | Anyone, including companies (CC-BY-4.0) |
| [AlphaFold Server](https://alphafoldserver.com) | AlphaFold 3 in a browser, no GPU needed, with fewer ligand types | Free, non-commercial research only, with a daily job quota |
| [AlphaFold 3 code and weights](https://github.com/google-deepmind/alphafold3) | Run the full model on your own hardware | Code: Apache 2.0. Weights: non-commercial organizations only |
| [AlphaFold 3 on Google Cloud](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/alphafold-3) | A managed deployment | Commercial users, through an allowlist and a Google Cloud subscription |

DeepMind released the AlphaFold 3 code and weights for academic use on November 11, 2024. The code has since moved from a CC BY-NC-SA licence to Apache 2.0, a permissive open-source licence. The weights are a different matter. Under the [weights terms of use](https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md), they and their outputs are only for non-commercial use by universities, non-profits, research institutes, journalists and government bodies. You may not use them for research on behalf of a company, may not share the weights, and may not train a competing structure prediction model on the outputs.

The [latest code release](https://github.com/google-deepmind/alphafold3/releases), v3.0.5 from October 30, is mostly speed work. The one before, v3.0.4, made AlphaFold 3 run on CPU-only machines, slowly, and on Apple Silicon Macs, using the Mac's GPU.

## What AlphaFold gets wrong

The AlphaFold 3 paper lists its own limits, and they matter if you plan to rely on a prediction:

- **It predicts one still picture.** Proteins move. The model generally predicts static structures like those in the Protein Data Bank, not how a molecule behaves in solution.
- **It can pick the wrong shape.** For a protein called cereblon, which is open on its own and closes when a drug binds, AlphaFold 3 predicted the closed shape in both cases.
- **It can invent structure.** Because it is a generative model, it sometimes adds false order to floppy, disordered regions. These usually come with very low confidence scores.
- **Chemistry errors.** It sometimes gets chirality (the handedness of a molecule) wrong, at a rate of 4.4% on PoseBusters, and sometimes overlaps atoms, mostly in large protein-nucleic acid complexes.

A prediction is a strong hypothesis, not a measurement.

## The 2024 Nobel Prize in Chemistry

On October 9, 2024, the Royal Swedish Academy of Sciences [awarded the Nobel Prize in Chemistry](https://www.nobelprize.org/prizes/chemistry/2024/press-release/) half to David Baker "for computational protein design", and the other half jointly to DeepMind's Demis Hassabis and John Jumper "for protein structure prediction". The Academy cited AlphaFold 2, which it said had predicted the structure of virtually all of the roughly 200 million proteins researchers have identified.

## Is there an AlphaFold 4

Not from DeepMind, as of November 17, 2026. AlphaFold 3 is still the newest public version, and DeepMind's AlphaFold page lists no newer model.

The next step has happened inside Isomorphic Labs instead, and it is not public. On February 10, 2026, the company [described IsoDDE](https://isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier), its drug design engine, which it says goes beyond AlphaFold 3. Isomorphic claims it more than doubles AlphaFold 3's accuracy on a protein-ligand benchmark, in the cases least similar to what the models saw in training, and gets 2.3 times as many highly accurate predictions as AlphaFold 3 on a set of 334 antibody-antigen complexes. These are the company's own numbers, and outside researchers cannot run the model to check them.

The AlphaFold 2 breakthrough was published and free for anyone to use. The newest work is being kept private and turned into a business, and the $5 billion Isomorphic just raised is a bet that it will produce drugs. The company has not yet put one into human trials.

## Frequently asked questions

### Is AlphaFold free for commercial use?

The AlphaFold Protein Structure Database is free for commercial use under CC-BY-4.0. AlphaFold Server and the AlphaFold 3 model weights are for non-commercial research only; companies can get AlphaFold 3 through an allowlisted Google Cloud subscription.

### Is AlphaFold 3 open source?

Its code is under the Apache 2.0 licence on GitHub. The model weights are not open source: they are free only for non-commercial organizations under separate terms of use, which also bar sharing them or training rival structure models on their outputs.

### Who won the Nobel Prize for AlphaFold?

Demis Hassabis and John Jumper of Google DeepMind shared half of the 2024 Nobel Prize in Chemistry "for protein structure prediction". The other half went to David Baker for computational protein design.

### Is there an AlphaFold 4?

Not as of November 17, 2026. AlphaFold 3 is the newest public version. Isomorphic Labs says its proprietary IsoDDE engine beats AlphaFold 3 on several benchmarks, but it has not been released.

## Sources

- [AlphaFold](https://deepmind.google/science/alphafold/), Google DeepMind
- [Highly accurate protein structure prediction with AlphaFold (Jumper et al., Nature 2021)](https://pmc.ncbi.nlm.nih.gov/articles/PMC8371605/), Nature
- [Accurate structure prediction of biomolecular interactions with AlphaFold 3 (Abramson et al., Nature 2024)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11168924/), Nature
- [AlphaFold 3 predicts the structure and interactions of all of life's molecules](https://www.isomorphiclabs.com/articles/alphafold-3-predicts-the-structure-and-interactions-of-all-of-lifes-molecules), Isomorphic Labs
- [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk/), EMBL-EBI
- [AlphaFold Server](https://deepmind.google/science/alphafold/alphafold-server), Google DeepMind
- [AlphaFold 3 repository](https://github.com/google-deepmind/alphafold3), Google DeepMind
- [AlphaFold 3 model parameters terms of use](https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md), Google DeepMind
- [AlphaFold 3 releases](https://github.com/google-deepmind/alphafold3/releases), Google DeepMind
- [AlphaFold 3 on Gemini Enterprise Agent Platform](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/alphafold-3), Google Cloud
- [The Nobel Prize in Chemistry 2024: press release](https://www.nobelprize.org/prizes/chemistry/2024/press-release/), The Royal Swedish Academy of Sciences
- [Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry](https://deepmind.google/discover/blog/demis-hassabis-john-jumper-awarded-nobel-prize-in-chemistry/), Google DeepMind
- [A glimpse of the next generation of AlphaFold](https://www.isomorphiclabs.com/articles/a-glimpse-of-the-next-generation-of-alphafold), Isomorphic Labs
- [The Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold](https://isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier), Isomorphic Labs
- [Isomorphic Labs announces Series C investment round](https://www.isomorphiclabs.com/articles/isomorphic-labs-announces-series-c-investment-round), Isomorphic Labs
