AlphaFold is a family of deep learning models from Google DeepMind that predict the three-dimensional structure of proteins from their amino acid sequences, and in its latest version the structure of complexes with DNA, RNA and small molecules. Among named artificial intelligence models, it is the clearest example of a scientific AI system rather than a chat assistant or content generator.
AlphaFold 2 effectively addressed the decades-old protein structure prediction problem at the CASP14 assessment in 2020, reaching accuracy comparable to experimental methods for many targets. The work led to the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper, shared with David Baker for protein design.
How AlphaFold predicts protein structure
AlphaFold 2 starts from a multiple sequence alignment (MSA) of evolutionarily related proteins and optional structural templates. Residues that mutate together across evolution tend to be in physical contact, so the alignment carries indirect evidence about the fold. The model turns this evidence into two internal representations: an MSA representation and a pair representation describing every residue pair.
The core of the network is the Evoformer, a stack of 48 attention blocks that repeatedly exchanges information between the two representations and enforces geometric consistency through triangle updates. A structure module using invariant point attention then converts the pair information into atomic coordinates. The whole process is repeated through recycling, typically three extra passes, which refines the result.
The table below summarizes the main stages of the AlphaFold 2 pipeline and what each contributes.
| Stage | Input | Output |
|---|---|---|
| Database search | Query sequence | MSA and structural templates |
| Evoformer | MSA and pair features | Refined residue-pair geometry |
| Structure module | Single and pair representations | Backbone frames and side chains |
| Recycling | Previous prediction | Improved coordinates |
| Confidence heads | Final representations | pLDDT and PAE scores |
AlphaFold versions and how they differ
AlphaFold has evolved through three major generations, each widening the range of molecules it can model. The first version used convolutional networks to predict distance distributions, the second introduced the end-to-end attention architecture, and the third replaced the structure module with a diffusion-based generator that handles many molecule types.
| Version | Year | Key change | Scope |
|---|---|---|---|
| AlphaFold 1 | 2018 | Distance prediction plus optimization | Single protein chains |
| AlphaFold 2 | 2020 | Evoformer and structure module | Single chains with near-experimental accuracy |
| AlphaFold-Multimer | 2021 | Training on multi-chain complexes | Protein-protein complexes |
| AlphaFold 3 | 2024 | Pairformer and diffusion module | Proteins, DNA, RNA, ligands, ions |
Access also differs. AlphaFold 2 code was released under an open-source license in 2021, and the AlphaFold Protein Structure Database, run with EMBL-EBI, provides more than 200 million predicted structures. AlphaFold 3 is available through the web-based AlphaFold Server for non-commercial use, and its code and weights were later released for academic research under restrictive terms.
Reading AlphaFold confidence scores
Every AlphaFold prediction comes with per-residue and pairwise confidence measures, and correct interpretation depends on them. The predicted local distance difference test (pLDDT) scores local accuracy from 0 to 100, while the predicted aligned error (PAE) estimates, in angstroms, how reliably two regions are positioned relative to each other.
The bands below are the conventional guide for reading pLDDT values.
| pLDDT range | Confidence label | Practical meaning |
|---|---|---|
| Above 90 | Very high | Backbone and most side chains reliable |
| 70 to 90 | Confident | Backbone generally correct |
| 50 to 70 | Low | Treat topology with caution |
| Below 50 | Very low | Often disordered or unresolved |
A frequent error is to trust a model with high pLDDT but high PAE between domains. Each domain may be well predicted, yet their relative orientation can be arbitrary, which matters greatly for interfaces, allosteric sites and drug binding pockets.
AlphaFold compared with other AI models
Unlike general assistants such as Claude, Gemini or GPT-5, AlphaFold is a narrow, domain-specific model trained on the Protein Data Bank and sequence databases. It does not converse, follow instructions or generate text; it maps biological sequences to coordinates. Its closest relatives in the DeepMind line are AlphaGo and AlphaZero, which also combined deep learning with a precisely defined scientific or game objective.
| System | Primary task | Output |
|---|---|---|
| AlphaFold | Biomolecular structure prediction | 3D atomic coordinates with confidence |
| AlphaZero | Game playing by self-play | Moves and position evaluations |
| General LLM assistants | Language understanding and generation | Text, code and tool calls |
| Physics-based docking tools | Ligand placement by scoring functions | Ranked binding poses |
Competing structure predictors, including RoseTTAFold and open reproductions of AlphaFold 3, follow similar ideas. Their existence matters for commercial and reproducible work, because AlphaFold 3 licensing restricts some industrial uses.
Limitations and common misunderstandings
AlphaFold predicts a single static structure, usually the most populated or best-represented state in training data, not the full range of conformations a protein adopts in a cell. It therefore says little about dynamics, folding pathways or conformational changes triggered by binding or phosphorylation.
- Intrinsically disordered regions receive low pLDDT and may appear as unrealistic ribbons; low confidence there is often biologically correct.
- Point mutation effects are predicted unreliably, because the MSA dominates and a single substitution rarely changes the output.
- Orphan proteins with few homologs produce shallow alignments and weaker predictions.
- Ligand poses from AlphaFold 3 can be physically implausible and still require validation by docking, molecular dynamics or experiment.
The most important warning is that a prediction is a hypothesis, not a measurement. Crystallography, cryo-electron microscopy and NMR remain necessary when exact geometry decides a scientific or clinical conclusion. Predictions also inherit biases of the Protein Data Bank, which overrepresents soluble, crystallizable proteins and underrepresents membrane proteins and transient complexes.
Stripped to the essentials
Proteins are long chains that fold into precise shapes, and the shape decides what a protein does. Measuring that shape in a laboratory can take months or years.
AlphaFold reads the chain and predicts the folded shape in minutes to hours, together with a score that shows which parts it is sure about. It works much like an expert who has studied every known protein and can sketch a new one from its recipe, while still admitting where the sketch is uncertain.
Where it is applied
A university laboratory studying a parasite enzyme has no experimental structure for its target, a protein of about 420 residues. The team downloads the AlphaFold Protein Structure Database entry, which shows pLDDT above 90 for the catalytic domain and below 50 for a flexible 40-residue tail.
The researchers remove the low-confidence tail, use the core model for molecular replacement in X-ray crystallography, and solve the experimental structure in weeks rather than the months a de novo approach would have required. The final crystal structure differs from the prediction by under 1 angstrom over the core. The outcome is typical of current practice: AlphaFold accelerates experiments rather than replacing them.