AI protein structure prediction
DeepMind's AlphaFold predicts a protein's 3D shape from its amino acid sequence, nearly as well as experiments.
Open in the interactive tree →AlphaFold 2 won the CASP14 competition in late 2020 with accuracy close to experiments, solving a 50-year-old problem. The open AlphaFold database followed in 2021, and AlphaFold 3 (May 2024) added DNA, RNA and drug molecules. David Baker (protein design) and Demis Hassabis and John Jumper (AlphaFold) received the 2024 Nobel Prize in Chemistry.
As of October 2026
The AlphaFold Protein Structure Database now holds more than 260 million predicted structures; on 24 September 2026 it added viral protein complexes, and in October 2026 AI-derived residue-level annotations from the scientific literature. AlphaFold 3's code was released for academic, non-commercial use in November 2024. Predictions of how proteins move, bind several partners or change shape remain less reliable than predictions of static shapes.
Open steps
- Protein motion, not one shape High AI leverageProteins flex between states; models must predict full ensembles and rare states that matter for drug binding, not a single snapshot.
- Reliable binding strength High AI leveragePredicting how tightly a molecule binds, with lab-level accuracy and fast enough to rank millions of candidates, is not yet routine.
- Disordered proteins and condensates Medium AI leverageMany protein regions have no fixed shape; predicting their behaviour, partners and phase separation needs new data and methods.
- Complexes, modifications and membranes Medium AI leverageLarge assemblies, chemically modified residues and membrane proteins are predicted less well than single soluble chains, yet many drug targets sit there.
Where AI could help
High AI leverage. The open gaps (motion, binding, complexes) are prediction tasks, but they lack measured training data.
- Generative models of protein conformational ensembles instead of single shapes
- Fast binding-affinity prediction to rank drug candidates before synthesis
- Complex and protein-ligand prediction at database scale
- Literature mining to annotate residues and functions
Shown so far
- In July 2025 Science published BioEmu, a Microsoft generative model of protein conformations that samples thousands of independent structures per hour on one GPU and was trained on over 200 milliseconds of simulation data. source
- In 2025 the open-source Boltz-2 model reported (preprint, according to its developers) a Pearson correlation of 0.62 on an FEP+ affinity benchmark, comparable to the physics-based OpenFE pipeline and over 1,000 times faster. source
Prerequisites
- First protein structure1958AI structure prediction learns from crystal structures collected since 1958
- Genome sequencing2003
- Deep Learning2012
- Cryo-EM Resolution Revolution2017Cryo-EM now supplies many of the new structures that protein-AI models learn from