{
  "id": "protein-structure-prediction",
  "title": "Protein Structure Prediction",
  "lang": "en",
  "source_lang": "en",
  "summary": "Protein structure prediction uses computational methods to infer three-dimensional molecular structures from amino-acid sequences and other biological information.",
  "category": "science",
  "subcategory": "biology",
  "type": "concept",
  "aliases": [
    "Protein three-dimensional structure prediction",
    "computational protein structure prediction"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/protein-structure-prediction/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/protein-structure-prediction/",
      "json": "https://aiwiki.page/zh/protein-structure-prediction.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Protein structure prediction is the computational inference of a protein’s three-dimensional arrangement from its amino-acid sequence, sometimes supplemented by evolutionary relationships, experimental measurements, or information about molecular partners. Its outputs range from approximate folds to atomic models of individual chains and molecular complexes. Because molecular shape helps determine binding, catalysis, regulation, and assembly, structure prediction connects sequence information to biological mechanisms.\n\nThe field combines biophysics, evolutionary analysis, mathematical optimization, and machine learning. Earlier methods relied principally on experimentally determined structural templates, simplified physical models, or fragment assembly. From the late 2010s, neural networks increasingly learned structural constraints and coordinates directly from large biological datasets. AlphaFold2’s performance in the 2020 Critical Assessment of protein Structure Prediction marked a major advance in the accuracy attainable for many individual protein domains.\n\nA predicted structure is nevertheless a model, not an experimental observation or a complete account of a protein’s behavior. Proteins can adopt multiple conformations, contain disordered segments, and depend on membranes, chemical modifications, or binding partners. Understanding prediction therefore requires distinguishing structural accuracy from confidence, a representative conformation from an ensemble, and a plausible molecular arrangement from a demonstrated biological mechanism."
    },
    {
      "heading": "Physical foundations and the prediction problem",
      "text": "A protein chain consists of amino-acid residues joined by peptide bonds. Its primary structure is the residue sequence; secondary structure includes local arrangements such as the alpha helix and beta sheet; tertiary structure describes the overall fold of a chain; and quaternary structure describes associations among multiple chains. A prediction task must specify which of these levels is being modeled. Recovering an isolated domain’s backbone is different from positioning every side chain in a membrane-embedded, ligand-bound assembly.\n\nThe physical basis of structure formation includes backbone geometry, side-chain packing, electrostatic interactions, hydrogen bonding, and the hydrophobic effect. Hydrophobic groups tend to become buried away from water, while polar groups often remain solvent-exposed or form compensating interactions inside the protein. Disulfide bonds and coordinated metal ions can further constrain a fold. These contributions are coupled: burying a charged side chain, for example, can be unfavorable unless surrounding interactions compensate for its loss of hydration.\n\nChristian Anfinsen’s work on ribonuclease, recognized in the 1972 Nobel Prize in Chemistry, established an influential connection between sequence and native structure. Under suitable conditions, denatured ribonuclease could recover its active conformation, showing that sequence can contain the information needed for folding. This principle does not imply that every protein independently reaches one unique structure under all conditions. Cellular folding may involve chaperones, cotranslational events, membranes, partner molecules, and kinetic barriers.\n\nFor an equilibrium description, conformations can be represented schematically by a distribution\n\n\\[P(x\\mid s,c)=\\frac{1}{Z(s,c)} \\exp\\!\\left[-\\frac{G(x;s,c)}{k_{\\mathrm B}T}\\right],\\]\n\nwhere \\(x\\) denotes a conformation, \\(s\\) the sequence, \\(c\\) the environmental conditions, \\(G\\) an effective free energy, and \\(Z\\) a normalization factor. This expression emphasizes that a sequence may populate an ensemble rather than a single geometry. In practice, accurate free energies are difficult to calculate, and many prediction systems learn structural regularities without explicitly evaluating this distribution.\n\nThe distinction between structure prediction and protein folding is fundamental. Prediction seeks one or more likely structures; folding research also asks how the chain reaches them and on what timescales. Cyrus Levinthal’s 1969 argument highlighted why an exhaustive search through all possible conformations cannot explain rapid folding. Modern prediction systems likewise avoid exhaustive enumeration, using learned constraints, templates, restricted representations, or biased sampling to navigate the enormous conformational space."
    },
    {
      "heading": "Structural data, sequence information, and molecular representation",
      "text": "The central experimental archive for structural modeling is the Protein Data Bank (PDB), established in 1971. It contains structures determined primarily through X-ray crystallography, nuclear magnetic resonance spectroscopy, and cryo-electron microscopy. These methods provide different kinds of evidence. Crystallographic and cryo-electron microscopy models are fitted to density information; solution NMR structures are constrained by measurements such as interatomic distance information and dihedral-angle restraints.\n\nExperimental coordinates are not uniformly precise ground truth. Mobile loops may be unresolved, alternative side-chain conformations may be present, and sample preparation may stabilize particular states. A deposited construct can differ from the natural protein through truncations, substitutions, purification tags, or engineered disulfides. Crystal packing can also affect local conformations or intermolecular contacts. Training and evaluation therefore require attention to the experimental model’s quality, biological context, and correspondence with the sequence being predicted.\n\nThe choice of representation strongly affects computational difficulty. A coarse-grained model may retain only one point per residue, commonly the alpha carbon, whereas an all-atom model includes backbone and side-chain atoms. Internal-coordinate representations use bond lengths, bond angles, and torsion angles. The backbone angles \\(\\phi\\) and \\(\\psi\\) describe rotations around bonds adjacent to the alpha carbon; sterically allowed combinations are organized by the Ramachandran plot. Most peptide bonds are approximately planar and trans, although cis configurations occur, particularly before proline.\n\nOther representations describe relationships between residues rather than absolute coordinates. A contact map records which residue pairs are close under a chosen definition. A distance map gives more detailed separations, and orientation features describe how local residue frames face one another. These relational descriptions are naturally independent of the protein’s position and orientation in space. Converting them into coordinates requires a geometrically consistent embedding, because an arbitrary set of pairwise distances need not correspond to a realizable three-dimensional structure.\n\nSequence databases provide information complementary to structural archives. A multiple sequence alignment arranges evolutionarily related sequences so that corresponding positions can be compared. Large sequencing and metagenomic projects greatly expanded the available diversity, often supplying informative homologues for proteins with no close structural template. However, database abundance is uneven: some protein families have millions of relatives, while others have few detectable matches. Structural archives also disproportionately represent proteins that can be expressed, purified, stabilized, and experimentally resolved, creating biases that predictive systems can inherit."
    },
    {
      "heading": "Comparative modeling and fold recognition",
      "text": "Comparative modeling, also called homology modeling, predicts a target structure by using an experimentally characterized evolutionary relative as a template. Its basis is that three-dimensional structure generally changes more slowly than amino-acid sequence during evolution. Two proteins can therefore retain similar folds even after extensive sequence divergence. When an appropriate close template exists, much of the target backbone can be modeled reliably without reconstructing the fold from first principles.\n\nA typical workflow identifies candidate templates, aligns the target sequence to them, transfers compatible structural features, builds unmatched regions, and refines the result. Template selection considers not only sequence similarity but also coverage, experimental quality, ligand occupancy, oligomeric state, and biological conformation. A slightly more distant template in the relevant functional state may be more informative than a closer relative captured in a different state. Multiple templates can contribute complementary information, although combining them requires resolving inconsistent geometries.\n\nAlignment quality is often the decisive limitation. A one-residue shift in a buried beta strand can place hydrophobic and polar side chains on the wrong faces, disrupting packing and hydrogen-bond patterns. Insertions and deletions are especially difficult because no direct structural correspondence exists for the missing or additional residues. Loop modeling uses fragment libraries, conformational sampling, and energetic or statistical evaluation, but long flexible insertions remain challenging. Side chains are commonly modeled using rotamer libraries that describe favored combinations of side-chain torsion angles.\n\nWhen sequence similarity becomes too weak for straightforward alignment, fold recognition or threading evaluates whether a sequence is compatible with a known structural framework. A scoring system may favor hydrophobic residues in buried sites, suitable residue types in helices or sheets, and plausible interactions between aligned positions. Threading does not require that a close sequence relative has already been recognized. It instead asks whether a known fold provides a convincing structural explanation of the target sequence.\n\nPrograms such as MODELLER, introduced in the 1990s, formalized comparative modeling through the satisfaction of spatial restraints derived from templates. Automated resources made such methods widely accessible before end-to-end neural prediction became dominant. Comparative modeling remains useful because templates provide interpretable evidence, can identify biologically relevant conformations, and may include bound cofactors or experimentally established interfaces. Its characteristic errors remain important even within newer pipelines: an incorrect alignment, an unsuitable domain arrangement, or a template-specific loop can be carried into an otherwise plausible model. Template resemblance alone also does not establish identical biochemical function, because related folds can support different substrates, reactions, and interaction partners."
    },
    {
      "heading": "Conformational search, energy functions, and molecular simulation",
      "text": "Template-free prediction historically attempted to identify native-like structures by searching conformational space and ranking candidates with physical or knowledge-based scoring functions. The label ab initio has been used broadly, but many methods described this way incorporate information extracted from known structures. A strict distinction is therefore needed between methods based primarily on physical interactions and those using statistically learned structural preferences. Both face the same practical problem: generating sufficiently good candidate conformations and recognizing them when they appear.\n\nAn atomistic force field approximates molecular potential energy using terms for bond stretching, angle bending, torsional preferences, electrostatics, and nonbonded interactions. Molecular dynamics propagates atomic motion under such a model, usually with explicit or implicit solvent. It can describe fluctuations and transitions that a single static prediction cannot. However, a physically plausible simulation does not guarantee a correct structure: sampling may be insufficient, and inaccuracies in the force field or solvent treatment can favor the wrong states.\n\nConventional molecular dynamics is constrained by the short integration steps required to resolve rapid atomic motions, typically on the femtosecond scale. Biologically important folding events may occur over microseconds, milliseconds, or longer. Specialized computing and improved algorithms have enabled folding simulations for a number of small, relatively fast-folding proteins, but routine simulation of arbitrary proteins from an extended chain remains computationally demanding. Enhanced-sampling approaches alter the exploration of conformational space, for example through replica exchange or biasing along selected coordinates, while requiring careful interpretation of the resulting statistics.\n\nFragment assembly offers another way to reduce the search. In Rosetta, developed from the late 1990s, short sequence-compatible backbone fragments from known structures are assembled and rearranged, commonly through Monte Carlo sampling. Coarse-grained scoring identifies promising folds before more detailed side-chain packing and all-atom refinement. Related approaches combine templates, fragments, restraints, and iterative optimization. Their effectiveness depends on both the scoring landscape and the probability of sampling the correct topology.\n\nStatistical potentials derive preferences from observed structural distributions, such as the frequency of particular residue contacts or torsion-angle combinations. They are not identical to directly measured physical energies, because structural databases contain evolutionary and experimental selection biases. Nevertheless, such potentials can distinguish useful structural patterns at comparatively low computational cost. Refinement then attempts to improve local packing, remove clashes, and optimize geometry. It is often easier to repair a nearly correct model than to change an incorrect topology: a short minimization can resolve steric overlaps, but it rarely reorganizes a misfolded domain into its correct architecture."
    },
    {
      "heading": "Evolutionary constraints and learned sequence representations",
      "text": "Evolution supplies indirect information about three-dimensional contacts. If two residues interact, a substitution at one position may be compensated by a substitution at the other. For example, a change in side-chain size can be accommodated by an opposite change nearby, while a charge reversal may be accompanied by another substitution that preserves electrostatic compatibility. Across a sufficiently diverse family, these coupled changes can reveal spatial relationships that are not apparent from any single sequence.\n\nRaw sequence correlation is difficult to interpret. Two sites can covary because both interact with a third site, because related organisms share ancestry, or because the alignment mixes subfamilies with different functions. Direct coupling analysis and related approaches attempt to distinguish direct statistical dependencies from indirect correlations. A common formulation uses a Potts model in which each aligned site has amino-acid preferences and pairs of sites have coupling parameters. Strong inferred couplings can identify residue contacts that constrain a protein’s fold.\n\nAround 2011, studies demonstrated that sufficiently informative evolutionary couplings could support the construction of three-dimensional models for proteins lacking close structural templates. This helped bridge a major gap in earlier template-free methods: long-range contacts sharply restrict how secondary-structure elements can be arranged. A contact between distant sequence positions may distinguish among otherwise similar candidate topologies. As metagenomic sequence collections grew, more families acquired the alignment depth and diversity needed for useful coupling analysis.\n\nThe relationship between alignment size and useful information is not simply numerical. Thousands of nearly identical sequences may provide less independent evidence than hundreds of diverse homologues. Sequence reweighting reduces overrepresentation, while alignment filtering and domain-specific analysis can improve signal quality. Incorrectly grouped paralogues, poorly aligned insertions, or mixed domain architectures can create misleading correlations. Predicting contacts across two different proteins is harder still because corresponding interaction partners must often be paired correctly across organisms.\n\nProtein language models provide another route to extracting structural information from sequence collections. Trained through objectives such as predicting masked amino acids, these models learn contextual representations that can encode aspects of secondary structure, residue environments, and long-range organization. Systems including ESMFold showed that such representations could support structure prediction without constructing an explicit multiple sequence alignment for each target. This can accelerate large-scale prediction, although performance remains dependent on sequence family, architecture, training coverage, and model design. Evolutionary alignments and language-model representations are related but not interchangeable: one explicitly compares homologous sequences, while the other compresses patterns learned across a training corpus."
    },
    {
      "heading": "Deep-learning architectures and the AlphaFold transition",
      "text": "Early neural approaches often predicted intermediate features, including secondary structure, solvent exposure, contacts, or distance distributions. These predictions then guided a separate modeling program. The 2018 AlphaFold system exemplified the growing effectiveness of learned interresidue geometry. AlphaFold2, evaluated in 2020 and described in detail in 2021, instead integrated sequence relationships, pairwise representations, geometric reasoning, and coordinate generation into a more unified trainable system.\n\nAlphaFold2 maintains representations of aligned sequences and residue pairs. Its Evoformer modules exchange information between these representations, allowing evolutionary signals and possible spatial relationships to influence one another. Triangle-based operations update residue-pair information using relationships involving a third residue, helping the network reason about geometric compatibility. A structure module then generates residue frames and atomic coordinates using invariant point attention, an architecture designed to handle three-dimensional relationships without depending on an arbitrary global orientation.\n\nThe system repeatedly feeds predictions and internal representations back into itself through recycling. This permits successive updates of the proposed structure rather than a single irreversible coordinate assignment. Training combines several objectives, including frame-aligned point error, which evaluates geometry in local coordinate frames, and penalties or auxiliary losses related to structural plausibility and confidence. The trained model therefore learns not just typical folds but ways to reconcile sequence, evolutionary, template, and geometric information within one inference process.\n\nRoseTTAFold, introduced in 2021, pursued a related integration through a three-track network linking sequence-level, pairwise, and three-dimensional information. These developments demonstrated that substantial accuracy gains could come from exchanging information across representations rather than treating sequence analysis, contact inference, and coordinate construction as isolated stages. They also shifted computational expenditure: expensive training could produce models capable of generating useful structures far faster than exhaustive physical simulation.\n\nAlphaFold3, reported in 2024, extended joint modeling to a wider range of biomolecular interactions, including proteins, nucleic acids, small molecules, and certain chemical modifications. It replaced AlphaFold2’s coordinate-generation approach with a diffusion-based module that generates atomic arrangements through learned denoising. Such systems broaden what can be represented in one prediction, but their outputs still require evaluation appropriate to each molecular class and interaction type.\n\nThe 2024 Nobel Prize in Chemistry recognized this broader transformation: Demis Hassabis and John Jumper received half of the prize for protein structure prediction, while David Baker received half for computational protein design. The distinction matters scientifically. Predicting a structure from a sequence and designing a sequence for a desired structure are related inverse problems, but success in either does not automatically establish the other."
    },
    {
      "heading": "Benchmarking, structural accuracy, and confidence estimates",
      "text": "The Critical Assessment of protein Structure Prediction (CASP), initiated in 1994, evaluates methods using targets whose experimental structures are not yet publicly available to predictors. Its blind design limits the possibility of selecting favorable examples or unknowingly using the answer. CASP also separates different difficulties, such as template-based modeling, difficult single-chain targets, and assemblies. The 2020 CASP14 assessment established AlphaFold2’s major improvement over previous systems, although performance varied among targets and structural regions.\n\nNo single numerical score fully describes structural accuracy. The root-mean-square deviation after optimal rigid-body superposition is\n\n\\[\\mathrm{RMSD} = \\sqrt{\\frac{1}{N} \\sum_{i=1}^{N} \\left\\|R\\mathbf{x}_i+\\mathbf{t}-\\mathbf{y}_i\\right\\|^2},\\]\n\nwhere corresponding coordinates \\(\\mathbf{x}_i\\) and \\(\\mathbf{y}_i\\) are aligned using rotation \\(R\\) and translation \\(\\mathbf{t}\\). RMSD is intuitive but sensitive to outliers: one displaced domain or flexible tail can dominate the result even when much of the structure is accurate. It also depends on which atoms and residues are included.\n\nGlobal Distance Test scores assess the fraction of aligned residues falling within several distance thresholds; GDT_TS conventionally averages results at 1, 2, 4, and 8 angstroms. TM-score uses a length-dependent normalization that makes overall fold comparisons more interpretable across proteins of different sizes. The local Distance Difference Test, or lDDT, assesses preservation of local interatomic distances without requiring one global superposition. These measures emphasize different aspects of correctness and can rank the same models differently.\n\nConfidence estimates are distinct from retrospective accuracy measurements. AlphaFold’s predicted lDDT, or pLDDT, assigns a local confidence score on a 0–100 scale. Values above 90 commonly indicate highly reliable local geometry, while values below 50 indicate low confidence and may occur in disordered or incorrectly modeled regions. Predicted aligned error, or PAE, estimates uncertainty in the relative positions of residues or domains under a specified alignment. A model can therefore contain two confidently folded domains whose mutual orientation is uncertain.\n\nReliable benchmarking also requires controlling information leakage. A test sequence may resemble a training protein even when their identifiers differ, and structural similarity can persist after obvious sequence similarity is lost. Time-based splits, sequence clustering, and fold-aware analyses address different aspects of this problem. Results must additionally be interpreted by task: accurate backbone coordinates do not establish correct ligand placement, side-chain chemistry, oligomeric state, or biological function. Confidence calibration itself can deteriorate outside the conditions represented during model development, so high internal confidence is evidence about a model’s estimated reliability rather than independent experimental confirmation."
    },
    {
      "heading": "Complexes, conformational diversity, and environmental context",
      "text": "Many proteins function as parts of assemblies rather than isolated chains. Predicting a complex requires determining component identities, stoichiometry, individual conformations, and relative orientations. Even if each subunit is modeled accurately, the interface can be wrong. Symmetric assemblies introduce additional geometric constraints, while transient interactions may have small or flexible interfaces that provide weaker structural signals. A computationally plausible contact also does not prove that two proteins interact in a cell.\n\nProtein docking traditionally separates this problem into sampling possible relative arrangements and scoring their compatibility. Rigid-body docking treats subunits as approximately fixed; flexible docking allows interface side chains, loops, or larger backbone regions to adjust. AlphaFold-Multimer and later joint-modeling systems integrate more of these decisions into one prediction process. Interface-specific confidence estimates are useful, but incorrect partner pairing, missing components, or an inappropriate stoichiometry can still yield misleading arrangements.\n\nProteins also occupy multiple functional states. G-protein-coupled receptors, for example, can adopt different conformations associated with signaling partners and bound ligands. Transporters alternate accessibility between opposite sides of a membrane, and enzymes may close loops around substrates. A model that accurately resembles one state may not explain another. Standard prediction often favors a representative state supported by training data and input information; repeated predictions do not necessarily sample conformations according to their physical probabilities.\n\nIntrinsically disordered proteins and regions present a different challenge because a single stable structure may not exist under the conditions of interest. Their biological roles can depend on flexible ensembles, transient contacts, or folding upon binding. Low-confidence extended segments may be consistent with disorder, but low confidence can also reflect insufficient information or model failure. Conversely, a network can generate a compact and chemically plausible structure for a sequence whose experimentally relevant behavior is disordered.\n\nEnvironmental and chemical details further shape the prediction target. Membrane proteins experience a lipid environment unlike bulk water. Protonation states depend on local chemistry and pH, while metal coordination, cofactors, glycosylation, phosphorylation, and covalent ligands can alter stability or conformation. Predicting a protein sequence alone generally does not determine which of these components is present in a particular biological sample.\n\nExperimental restraints help resolve ambiguities that sequence-based models cannot settle. Cross-linking mass spectrometry can constrain proximity between sites; cryo-electron microscopy can locate subunits within assemblies; and NMR measurements can constrain local geometry or reveal motion. Integrative modeling combines such evidence while accounting for its uncertainty. The result may appropriately be an ensemble or a set of alternative arrangements rather than one apparently definitive structure."
    },
    {
      "heading": "Applications, structural databases, and remaining scientific limits",
      "text": "Predicted structures have expanded access to structural information far beyond proteins with experimentally determined models. The AlphaFold Protein Structure Database, launched in 2021, expanded in 2022 to more than 200 million predictions. This changed the scale of structural analysis: entire proteomes could be examined for folds, domain boundaries, candidate active sites, and similarities that are difficult to detect from sequence alone. Database-scale availability does not make every entry equally reliable, particularly for disordered segments and uncertain interdomain arrangements.\n\nOne major application is functional annotation. A predicted arrangement of catalytic residues or a recognizable ligand-binding pocket can help connect an uncharacterized protein to a biochemical family. Structural comparison can reveal remote relationships that sequence searches miss. Such evidence is strongest when combined with conserved residues, genomic context, cellular localization, and experiments. A shared fold alone is insufficient because many structural scaffolds support multiple functions, and small local changes can alter substrate specificity or reaction chemistry.\n\nPredictions also assist experimental structural biology. In crystallography, an accurate model can provide a search model for molecular replacement, supplying initial phase information needed to interpret diffraction measurements. In cryo-electron microscopy, predicted domains can help identify components and build atomic models into density maps. Independent agreement with the experimental data remains essential: fitting an attractive prediction into ambiguous density can otherwise reinforce an incorrect interpretation. Flexible regions and interfaces often require particular attention during model building.\n\nIn biotechnology and pharmaceutical research, structural models help characterize binding sites, interpret sequence variants, and formulate experimentally testable mechanisms. However, a predicted substitution environment does not by itself establish whether a human variant causes disease. Likewise, placing a small molecule in a plausible pocket does not establish binding affinity, selectivity, pharmacological activity, or clinical efficacy. Those properties depend on energetics, solvation, dynamics, concentrations, and biological processes not captured by a static coordinate model.\n\nPrediction is closely connected to protein design, which seeks sequences capable of adopting desired structures or functions. A structure predictor can serve as one filter in a design pipeline, but agreement between a design model and a prediction is not fully independent validation, especially when the systems share training data or learned assumptions. Expression, folding, stability, assembly, and function remain separate experimental questions.\n\nPersistent challenges include predicting conformational populations, determining how mutations shift those populations, modeling large heterogeneous assemblies, and resolving chemical states at catalytic sites. Computational scalability and reproducibility are also consequential: model weights, inference settings, sequence-database versions, template cutoffs, and stochastic seeds can affect results. Recording those details, together with confidence measures and the precise molecular composition being modeled, preserves the distinction between a useful structural hypothesis and an experimentally established molecular description."
    }
  ],
  "links": [
    "protein",
    "amino-acid",
    "mathematical-optimization",
    "machine-learning",
    "peptide-bond",
    "alpha-helix",
    "hydrogen-bond",
    "hydrophobic-effect",
    "water",
    "nobel-prize-in-chemistry",
    "protein-folding",
    "protein-data-bank",
    "x-ray-crystallography",
    "nuclear-magnetic-resonance",
    "cryo-electron-microscopy",
    "ramachandran-plot",
    "multiple-sequence-alignment",
    "evolution",
    "potential-energy",
    "molecular-dynamics",
    "alphafold",
    "critical-assessment-of-protein-structure-prediction",
    "ph",
    "mass-spectrometry",
    "alphafold-protein-structure-database"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}