What an AI protein structure can tell us

A protein image can look definite. Colored ribbons wind through space, perhaps beside a small molecule drawn as sticks. The image is a rendering of coordinates, with a view and color scale chosen by software or a person. To interpret it, ask what sequence and method produced the coordinates, what the confidence display measures, and which experimental condition the biological question concerns.

Proteins are built from one or more chains of amino-acid residues. The sequence records those residues in order. Each amino acid joins the chain through its backbone and has a side chain with its own size, charge, and chemical behavior. A structure assigns three-dimensional coordinates to these atoms. The sequence specifies which chemical groups are present; the coordinates show how they are arranged and which groups may be near one another. That geometry helps explain a fold or a possible binding site, though proximity alone does not prove a biological interaction. A protein drawing is a visual convention, not a photograph of atoms.

The chemical differences among side chains help explain why sequence matters. Some side chains are charged, some polar, and some mostly nonpolar. In a folded protein, their positions can make certain contacts possible and others less likely. Residues far apart in the chain can also come together in space, so the coordinates reveal neighborhoods that are not obvious from the linear sequence. This can suggest a contact to examine or test, while the sequence identifies the chemical groups that would make that contact plausible.

Protein diagrams often use ribbons to trace the backbone: helices appear as coils, beta strands as arrows, and loops as connectors. Stick views add side chains and small molecules. In a coordinate viewer, rotate the model and select residues or a ligand to inspect their atoms. Residue labels link those positions back to the sequence.

Experimental structures also come from measurements interpreted through models. X-ray crystallography uses diffraction from a crystal. Nuclear magnetic resonance, or NMR, uses measurements from atomic nuclei to constrain structures and motions. Cryo-electron microscopy uses images of frozen particles and computational reconstruction. Each method gives evidence about a sample under particular conditions. EMBL-EBI's introduction to proteins and structural methods explains this distinction.

The method and the material being studied shape what the coordinates represent. A crystallized protein was measured in a crystal environment; an NMR result can reflect measurements from molecules in solution; a cryo-EM reconstruction is inferred from many particle images. Construct boundaries, bound molecules, and solution conditions can matter when comparing structures. These coordinate sets are strong experimental evidence, but their ribbon renderings are not photographs of atoms moving through time.

From sequence to predicted coordinates

AlphaFold2, described by Jumper and colleagues, predicts coordinates using a protein sequence and evolutionary information from aligned related sequences. The pipeline may also use structural templates. Training used experimentally determined structures in the Protein Data Bank. At CASP14 in 2020, organizers gave participants blind targets whose structures were not publicly disclosed, then assessed predictions against the structures established for those targets. A blind evaluation limits copying of known answers, while target selection and benchmark coverage still define what the result tests. An independent CASP14 assessment adds context to the reported performance.

An alignment arranges related sequences so corresponding positions can be compared. Patterns of conservation and co-variation can provide clues about which residues may be structurally related. Those clues are statistical evidence from sequence families, not direct measurements of distances between atoms. AlphaFold2 combines such information with patterns learned from structural data to produce a coordinate model; optional templates provide another source of structural information when available.

Jumper and colleagues reported a marked improvement at CASP14. The prediction remains a computational model, even when it matches a withheld structure. It does not record a protein folding in a cell or provide a complete physical movie of the folding pathway. Their AlphaFold2 paper was published online on 15 July 2021, appeared as the version of record on 18 August, and was included in the 26 August issue. The independent CASP14 assessment appeared online on 7 October 2021 and in the December 2021 issue of Proteins.

Public prediction databases made many coordinates available for inspection. The AlphaFold Protein Structure Database provides predicted coordinates and confidence estimates. The 2021 human-proteome study distinguished residue-level confidence and coverage from the mere presence of coordinates for a full chain. A visible trace can include stretches with low local confidence, so coverage and confidence answer different questions. Database contents change; the original 2021 counts are historical, not a current inventory. The human-proteome study appeared online on 22 July 2021. The database paper appeared online on 17 November 2021 and in the 7 January 2022 issue of Nucleic Acids Research.

One sequence, two experimentally observed states

A documented example shows why a protein image needs experimental context. Nishimasu and colleagues studied an archaeal hexokinase and reported four X-ray structures, two of which illustrate the states discussed here. The apo form, recorded as PDB entry 2E2N, is in an open state. PDB entry 2E2Q contains xylose, magnesium ions, and ADP and is in a closed state. The authors reported a sugar-binding conformational change.

These are separate experimental snapshots, released on 16 January 2007, rather than frames from a timed movie. They show two arrangements under different ligand and cofactor conditions; they do not reveal the path or rate of movement between them. Here, apo identifies the form without the sugar, ADP, and magnesium of the complex. The entry also models other small molecules, so the label does not mean that its coordinate file is completely ligand-free. Comparing the structures shows why a sequence alone does not specify which condition or conformation matters to a question. “Open” and “closed” describe the arrangements reported in these entries. Nishimasu et al.'s paper appeared in the Journal of Biological Chemistry, volume 282, pages 9923-9931, in 2007. The original study describes the structures and reported change.

Local confidence is residue by residue

A common AlphaFold display colors residues by pLDDT, a per-residue estimate of local structural confidence from 0 to 100. It concerns the predicted shape around a residue. EMBL-EBI's guide describes scores above 90 as the highest accuracy category. Scores above 70 often indicate that the backbone is reliable, while side-chain details may be less reliable. The backbone is the connected chain that forms the structural frame; side chains branch from it and give residues much of their distinct chemical character. The pLDDT guide explains the scale.

The color on a ribbon is only a visual encoding of those per-residue values. A chain-wide average can conceal a low-confidence loop beside a high-confidence core, so read the scores along the region involved in a claim. A score is also not a physical property such as rigidity or temperature: it is a model's confidence estimate, whose interpretation stays tied to AlphaFold2 and the cited guide.

A low score may reflect limited information in the input, flexibility, or disorder; the score alone does not distinguish these explanations. It marks where a precise local interpretation needs care. A low-confidence loop may still appear as a smooth curve between well-supported segments, but the rendering does not turn that path into a measured trajectory. For a backbone claim, inspect residue-level confidence and coverage; a side-chain contact requires added care because its details may be less reliable.

Relative placement and the PAE score

AlphaFold2 also reports predicted aligned error, or PAE. In the matrix, find the cell where residue X on one axis meets residue Y on the other. X is the residue whose position is being assessed; Y is the reference residue. Aligning on Y means superposing the predicted and actual structures using Y as the anchor, then asking where X falls relative to it. Switching X and Y changes which residue anchors the comparison and asks a different placement question. The cell reports expected positional error in angstroms. Smaller values indicate higher confidence in that relative placement. PAE is an estimate, not a measured error or a probability. EMBL-EBI's PAE guide describes the measure.

A domain is a region of a protein chain with its own local fold. When each domain occupies a range of residue indices, pairings within that range form a block in the matrix; cells pairing residues from different ranges describe between-domain placement. These values are positional errors under the chosen reference alignment, not direct distances between atoms.

Consider a plainly hypothetical two-domain protein. Cells comparing residues within either domain might show expected errors of about 2 angstroms, while cells spanning the domains might show about 20 angstroms. These invented display values are neither AlphaFold output nor probabilities. The pattern indicates more confidence in each domain's internal geometry than in the placement of one domain relative to the other. An apparent pocket or contact spanning domains is therefore less secure than one inside a domain, even if each domain has high local confidence by pLDDT.

What a prediction can support

A predicted fold can help organize a hypothesis about which residues may be near each other or suggest a region to examine experimentally. Its usefulness depends on the question. A proposed contact within one domain calls for confidence in those local coordinates. A claim about domain orientation needs information about relative placement. A proposed interaction with another molecule needs a model that includes that partner and evidence suited to the binding question.

For AlphaFold2, the maintainers describe predictions as structural snapshots. The method does not normally predict full dynamics, folding pathways, or binding energies. Some conformations depend on ligands, partners, and the surrounding environment, while disordered regions may not have one fixed structure. These limits describe AlphaFold2 in the cited guidance and should not be generalized without qualification to later methods. The maintainers' AlphaFold2 guide discusses these cases.

AlphaFold 3, reported by Abramson and colleagues in 2024, extends prediction to complexes that may include proteins, DNA, RNA, small molecules, ions, and modified residues. Its coordinate generator uses a diffusion-based method. The paper reported improvements on its tested interaction benchmarks. That is a result about those evaluated tasks. The Nature paper and its accessible full text describe the method and benchmarks.

This broader scope matters when asking about a ligand or a partner: a single-chain fold prediction and a complex prediction answer different modeling questions. Check which molecules were included in the input and which benchmark supports the claim. A complex model provides coordinates for a proposed arrangement; experimental work remains the way to test that particular arrangement under relevant conditions.

Reading the image against the question

Identify the method and version, sequence range, chain, and color legend before making a claim. Then name the scale of the question: local fold, domain placement, ligand pose, or movement between states. For local shape, inspect pLDDT and coverage; for relative placement, inspect PAE; for a ligand pose, identify the ligand and conditions in the structure. If the question concerns the xylose-bound hexokinase state, 2E2Q is the relevant snapshot, while neither 2E2N nor 2E2Q gives a rate of movement. Match coordinates and evidence to the state or process being asked about.

Sources

• EMBL-EBI, “What are proteins and how do we know their structures?”, training guide checked 29 October 2026; publication date not stated.
• Jumper et al., “Highly accurate protein structure prediction with AlphaFold”, Nature, published online 15 July 2021; version of record 18 August 2021; issue dated 26 August 2021.
• Kryshtafovych et al., “Critical Assessment of Methods of Protein Structure Prediction (CASP) - Round XIV”, published online 7 October 2021; December 2021 issue of Proteins.
• Tunyasuvunakool et al., “Highly accurate protein structure prediction for the human proteome”, Nature, published online 22 July 2021.
• Varadi et al., “AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models”, published online 17 November 2021; issue dated 7 January 2022.
• Nishimasu et al., “Crystal structures of an ATP-dependent hexokinase with broad substrate specificity from the hyperthermophilic archaeon Sulfolobus tokodaii”, Journal of Biological Chemistry 282, 9923-9931 (2007). PDB 2E2N and PDB 2E2Q, released 16 January 2007.
• EMBL-EBI, “pLDDT: understanding local confidence”, training guide checked 29 October 2026; publication date not stated.
• EMBL-EBI, “PAE: A measure of global confidence in AlphaFold2 predictions”, training guide checked 29 October 2026; publication date not stated.
• EMBL-EBI, “Strengths and limitations of AlphaFold 2”, training guide checked 29 October 2026; publication date not stated.
• Abramson et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3”, Nature, published online 8 May 2024; accessible full text.