Protein Threading

Recognize a plausible known fold for a weak-homology protein sequence by aligning it onto structural templates and optimizing a residue–environment compatibility score.

Version

v2 · 2026-09-06 · History

Domain-specific #

2558

Origin domain

computational biology

Subdomain

protein structure prediction

Aliases

Sequence Structure Threading, Protein Fold Recognition

Core Idea

Protein threading is the template-based fold-recognition method that asks an inverse-folding question: which already-known three-dimensional protein framework is most compatible with this target amino-acid sequence, even when ordinary sequence similarity is too weak to identify a reliable close homolog? It aligns the target sequence directly to positions in candidate structural templates, scores the fit of residues to the environments and contacts those positions create, and ranks the resulting sequence–structure alignments. The winning alignment supports a template-conditioned structural hypothesis; it is not itself experimental proof of the native structure.

The structural move is cross-representational. The target is one-dimensional sequence order. The template contributes a three-dimensional arrangement of backbone positions, solvent exposure, secondary structure, and residue-contact neighborhoods. A threading method proposes a monotone assignment of target residues to template positions, with insertions and deletions where allowed, then evaluates whether the assigned amino acids make sense in those structural environments. Bowie, Lüthy, and Eisenberg's foundational 3D-profile method characterized template positions by burial, polar environment, and local secondary structure, demonstrating detection of structural relationships not visible from sequence similarity alone.^[1] Jones, Taylor, and Thornton made the alignment more explicitly three-dimensional by fitting sequences to known backbone coordinates and evaluating the modeled contacts with empirical potentials.^[2]

The identity therefore contains four inseparable operations: construct or select a nonredundant library of known folds; generate feasible sequence-to-structure alignments; score residue–environment and often residue–residue compatibility; and compare or rank the scored hypotheses. When pairwise contacts and flexible gaps are included, finding a globally optimal threading becomes computationally hard: Lathrop proved the corresponding decision problem NP-complete and the optimization problem NP-hard under those conditions.^[3] Practical systems consequently use restricted score decompositions, dynamic programming, branch-and-bound, heuristics, learned confidence estimators, or ensembles rather than treating an exact global optimum as free.

Threading occupies a deliberately difficult template-based regime. Homology modeling starts from a detectable evolutionary relative of known structure and usually relies heavily on sequence alignment. Threading is invoked when close sequence homology does not settle the template but the target may still adopt a fold already represented in the structure library. This is a use-condition, not a metaphysical claim that the target has no homologs anywhere. Nor is the boundary perfectly sharp: contemporary fold-recognition pipelines mix sequence profiles, predicted secondary structure, contact information, learned representations, and structural potentials. The node is retained when the load-bearing inference still aligns a target sequence against candidate structural frameworks and judges sequence–structure compatibility.

Structural Signature

Sig role-phrases:

• the unresolved target sequence — an amino-acid sequence whose fold is not fixed by a sufficiently close known structural homolog
• the known-fold template library — experimentally determined or otherwise admitted structural frameworks against which the target can be tested
• the sequence-to-structure alignment — a monotone placement of target residues onto template backbone positions, with explicitly governed insertions, deletions, and unaligned regions
• the structural environment encoding — solvent exposure, secondary-structure state, contact neighborhood, geometry, or other features attached to template positions
• the compatibility objective — a score or empirical energy combining residue–environment fit, pair contacts, sequence terms, predicted features, and gap costs
• the constrained search procedure — an exact, approximate, heuristic, or learned procedure that explores template/alignment hypotheses under topology and ordering constraints
• the ranked fold hypotheses — templates and alignments ordered by score, statistical significance, consensus, or calibrated confidence rather than accepted merely because a best item exists
• the template-conditioned model — a target structural model or set of spatial restraints inherited from the selected alignment and exposed to refinement and validation
• the coverage boundary — the method can recognize only folds or useful structural fragments represented closely enough in its admitted template universe

The locked recognition test is conjunctive. A candidate qualifies as protein threading only if it begins from a protein sequence and candidate three-dimensional fold templates; places sequence positions into template structural roles rather than comparing sequence strings alone; evaluates the placements with a sequence–structure compatibility criterion; searches or optimizes across alignments or templates; and returns a fold/alignment hypothesis whose warrant remains conditional on template coverage and score calibration. A pipeline may add profiles, neural networks, predicted contacts, consensus, or assembly without losing the identity, provided this sequence-to-structure test remains load-bearing.

The main diagnostics follow the signature. If rankings collapse when near-duplicate templates are removed, apparent confidence may be library redundancy. If the correct fold is selected but residue correspondence is wrong, fold recognition succeeded while alignment failed. If low-scoring decoys outrank native-like templates, the compatibility function is mis-specified. If no candidate separates from the background distribution, the evidence supports “unresolved,” not the top template. If a domain-sized region scores well but a full-length model does not, the target may be multidomain or only partially covered. If alternative score terms reverse the winner, the inference is score-sensitive and should not be reported as a robust fold call.

Interventions target the failing role. Reduce template redundancy and calibrate against decoys for library bias. Add predicted secondary structure, sequence profiles, contact constraints, or improved environment potentials for score discrimination. Change gap penalties, domain boundaries, or alignment search for correspondence error. Use multiple independent threaders and consensus for method-specific instability. Escalate to free-modeling or hybrid assembly when no represented fold receives credible support. In every case, preserve the distinction between “best among tested templates” and “adequately supported native-like template.”

What It Is Not

• Not ordinary sequence alignment. Sequence alignment compares symbols or profiles along sequences. Threading makes template positions structurally typed and judges whether residues fit three-dimensional environments and contacts.
• Not homology modeling in the close-template regime. Comparative modeling normally begins from a detectable homolog and a sequence-based correspondence. Threading addresses the weaker-homology fold-recognition regime, although modern pipelines blur the technical boundary.
• Not de novo or free modeling. Free modeling searches conformations without requiring the native fold to be represented by a known template. Threading's decisive hypothesis space is constrained by a template library.
• Not structural alignment. Structural alignment compares two already-known three-dimensional structures. Threading aligns a sequence of unknown structure to a known structural framework.
• Not molecular simulation of folding. It does not reproduce the physical trajectory by which a polypeptide reaches its native state. It evaluates compatibility with candidate end-state frameworks.
• Not a statistical potential alone. A potential is one scoring component. Threading adds templates, feasible alignments, search, ranking, and a template-conditioned output.
• Not a database lookup alone. Retrieving a template by annotation or sequence hit is not threading unless the target is placed into structural roles and evaluated for sequence–structure fit.
• Not one named server. THREADER, GenTHREADER, LOMETS, and I-TASSER-related pipelines instantiate or incorporate the method; no implementation exhausts it.
• Not experimental structure determination. A high-scoring model is a computational hypothesis requiring confidence assessment and, where consequential, experimental or orthogonal validation.

Scope of Application

Protein threading lives in computational structural biology wherever structural information must be inferred for a sequence beyond the reliable reach of close-homolog comparative modeling but within plausible reach of known fold space.

• Remote fold recognition. A sequence may have diverged beyond straightforward pairwise similarity while retaining a known topology. Threading tests compatibility with structural environments that are more conserved than residue identity.
• Template selection for hard targets. When several weak sequence hits or unrelated-looking templates compete, sequence–structure scoring can distinguish which framework best accommodates the target.
• Sequence–structure alignment. Selecting the right fold is not enough. Threading supplies residue correspondences that determine where conserved core positions, insertions, loops, and gaps enter the model.
• Template-based model construction. Aligned backbone coordinates provide an initial target framework; loops, side chains, and unaligned regions can then be built or refined.
• Spatial-restraint generation. Threading alignments imply target contacts and distances. Meta-threading systems can aggregate such restraints to guide subsequent assembly or refinement.^[4]
• Proteome-scale structural annotation. Fast fold-recognition systems can scan many translated open reading frames, producing fold or domain hypotheses and confidence estimates. GenTHREADER combined sequence alignment, threading-derived evaluation, and a neural-network confidence measure for genomic-scale use.^[5]
• Domain decomposition. Local threading support may reveal that only one region maps credibly to a known fold, prompting separate modeling of domains rather than forcing one full-length template.
• Hybrid prediction pipelines. Multiple threading alignments can seed fragment assembly and structural refinement. I-TASSER is a canonical example: it identifies templates through threading and then assembles and refines fragments rather than treating the selected template as the finished model.^[6]

The method is weakest when the native topology is absent from, or too remote from, the template library; when disorder or large conformational changes violate the fixed-framework assumption; when oligomeric context determines the fold; or when the target contains extensive novel insertions. Those cases may still yield useful local restraints, but they do not license an unsupported full-chain model.

Clarity

Three distinctions keep the term precise. First, fold recognition and alignment accuracy are different outputs. A method can identify the correct broad fold yet shift residues, strands, or helices relative to their true correspondences. The first error changes the framework; the second corrupts the model built within a correct framework. Evaluation must report them separately.

Second, best score and significant support are different claims. Any finite library produces a highest-ranked template. The top score is evidential only relative to a decoy or background distribution, template redundancy, target length, score calibration, and agreement across signals. An honest method needs an abstention state. “No confident fold recognized” is a valid result.

Third, the homology boundary is operational. The method is often described as applying when no homolog of known structure exists, but complete non-homology is rarely demonstrable and remote evolutionary relationships can coexist with negligible sequence identity. The more exact statement is that no sufficiently close structural homolog is detected to make ordinary sequence-based comparative modeling reliable. Threading then exploits structural constraints and environment compatibility. It may discover remote homology, analogous fold reuse, or simply a physically plausible template; the score alone does not decide which historical explanation is correct.

The word “threading” can also be too broad. In its strict historical sense, the target sequence is placed through a full three-dimensional template and pairwise contacts contribute to the score. Some later systems called fold recognition or threading use profile–profile alignment plus predicted structural features without explicit all-atom contact evaluation. This entry uses a controlled genus: sequence-to-known-structure alignment with a load-bearing compatibility score. A purely sequence-profile classifier sits outside; a hybrid whose structural template roles govern alignment remains inside.

Manages Complexity

Protein folding presents an enormous conformational search space. Threading compresses that problem by replacing unrestricted coordinate generation with a finite repertoire of previously observed frameworks. The decision variables become template identity, domain boundary, residue correspondence, and gap placement; the objective measures compatibility. This restriction sacrifices the possibility of discovering a truly novel fold but makes many otherwise inaccessible predictions computationally tractable.

The template encodes many correlated constraints at once. A backbone framework determines which positions are buried, which are solvent-exposed, which residues approach one another, and which secondary-structure contexts occur. Scoring a target against those roles avoids enumerating all atomic conformations. Bowie and colleagues' 3D profiles are a particularly clear compression: each structural position becomes an environment class against which amino-acid preferences can be evaluated.^[1]

Decomposition controls the remaining search. Dynamic programming can optimize alignment when the score decomposes into local terms and manageable gap costs. Pairwise three-dimensional interactions couple distant sequence positions, making the full problem harder; exact methods may then restrict the model, while heuristics trade global guarantees for speed. Lathrop's complexity result explains why practical approximations are structural responses to the problem rather than mere implementation laziness.^[3]

Meta-threading manages model uncertainty by running diverse component methods and aggregating their outputs. LOMETS used multiple threading programs to produce consensus models and contact or distance restraints, then passed these to downstream assembly.^[4] Consensus does not make shared library omissions disappear, but it can reduce dependence on one score function or search heuristic.

The complexity reduction remains conditional. A larger template library improves potential coverage but increases multiple-testing burden, redundancy, and compute. A richer compatibility score distinguishes more subtle fits but creates harder optimization and more parameters to calibrate. Threading manages complexity by choosing an explicit, auditable compromise among hypothesis-space coverage, score realism, alignment flexibility, and tractability.

Abstract Reasoning

Inverse-fold formulation. Hold a candidate three-dimensional framework fixed and ask which sequence placements are compatible with it. This reverses the direct folding question and turns conformation generation into comparative hypothesis testing over known structures.

Role-typed alignment. Treat each template position not as another letter but as a bundle of structural obligations: burial, polarity, secondary structure, contact partners, and geometry. Aligning a residue means assigning it those obligations. This blocks a high sequence score from silently substituting for structural fit.

Constrained optimization. Define the candidate set, ordering and gap constraints, score orientation, and acceptable approximation. When pair interactions couple separated positions, record whether the solver proves optimality, bounds it, or returns only a heuristic candidate.

Decoy discrimination. A compatibility function is useful only if it ranks native-like or correct-fold models above plausible alternatives. Test on withheld targets and structurally realistic decoys rather than interpreting raw energy magnitude in isolation. Bryant and Lawrence evaluated a residue-contact potential by threading sequences through many alternative folding motifs, illustrating this discriminative criterion.^[7]

Uncertainty decomposition. Separate uncertainty about template coverage, fold identity, residue alignment, score calibration, domain boundaries, and final-coordinate refinement. A confident fold family with uncertain loops is not the same result as an uncertain fold.

Counterfactual library test. Ask what the pipeline would report if the correct template were removed or duplicated many times. This exposes whether confidence measures compatibility or simply reflects database composition.

Failure-directed intervention. If the fold is right and alignment wrong, adjust correspondence search or gap treatment. If both are wrong, improve the score or library. If all known templates are weak, stop threading rather than refining a false premise. Diagnosis chooses the repair layer.

Knowledge Transfer

Within structural bioinformatics, the threading package transfers across proteins, template databases, and implementations. A user can move from an enzyme domain to a membrane-associated domain and ask the same questions: Is an appropriate fold represented? Which target residues occupy which structural roles? Which score terms drive the match? Is the alignment stable? What evidence distinguishes the top template from decoys? The answers remain expressed in protein-specific vocabulary, but the diagnostic procedure is reusable.

It also transfers between method generations. Early environment profiles, pair-contact potentials, profile-enhanced threaders, meta-threading systems, and hybrid assembly pipelines differ technically. They remain comparable by mapping each implementation onto the locked roles: target, structural library, alignment space, structural encoding, objective, search, confidence, and output model. A new neural component does not by itself create a new abstraction; the question is whether it changes one role or replaces the sequence-to-structure inference entirely.

The node enables disciplined handoff. A fold-recognition stage should pass downstream not only coordinates but template identifiers, alternative alignments, per-region confidence, unsupported insertions, and the score or consensus basis. Refinement teams can then distinguish inherited template certainty from newly modeled regions. Experimentalists can choose validation targets at the weakest or most consequential parts of the hypothesis.

Outside protein structure prediction, the generic skeleton—align an unresolved object to typed templates and optimize compatibility—is already owned by broader abstractions such as optimization, representation, classification, and comparison. Calling document-template matching or shape registration “protein threading” would be metaphorical. The transferable reasoning is real, but the name and validity conditions remain biological.

Examples

Canonical: Thread one weak-homology sequence against two structural frameworks

Suppose a target protein sequence has no close structural homolog detectable by ordinary pairwise sequence comparison. A template library contains fold A, whose core positions alternate buried hydrophobic and exposed polar environments, and fold B, whose corresponding positions reverse several of those roles. The method aligns the target monotonically to each backbone, allows a loop insertion at a permitted gap, and computes

score = environment fit + pair-contact fit + secondary-structure agreement + gap cost.

For fold A, hydrophobic target residues occupy buried positions, charged residues remain exposed, predicted helices align with template helices, and long-range residue pairs fit the template contact network. Fold B has slightly better raw sequence similarity but buries several charged residues and violates predicted secondary structure. The structural compatibility score favors A. The method reports A plus the alignment, the score separation from B and decoys, and low confidence around the inserted loop. This is the inverse-folding logic of the foundational environment-profile and three-dimensional threading methods.^[1][2]

Mapped back: The target sequence is unresolved by close homology; A and B come from the known-fold library; residue placements are sequence-to-structure alignments; burial, contacts, and secondary structure form the environment encoding; the composite score is optimized across alignments; A becomes the ranked fold hypothesis; and the model remains template-conditioned with an explicitly weak loop.

Applied / In Practice: Meta-threading supplies restraints to an assembly pipeline

A target is submitted to a multi-threader pipeline. Several component methods search the same broad structure library using different profiles, structural features, scores, and alignment heuristics. Four methods independently support related templates over the target's N-terminal domain; two disagree on the C-terminal region; none supplies a credible full-length template. The system clusters the N-terminal alignments, derives consensus residue contacts and distance restraints, and passes those constraints to iterative assembly. The C-terminal region is treated with weaker restraints and broader conformational search rather than copied confidently from an arbitrary top template. LOMETS and I-TASSER exemplify this staged relationship between multiple threading alignments, spatial restraints, and subsequent assembly/refinement.^[4][6]

Mapped back: The component methods instantiate distinct constrained searches and objectives over a shared template universe; agreement stabilizes the ranked N-terminal fold hypothesis; disagreement exposes domain-specific uncertainty; the alignment-derived contacts become a template-conditioned output; and unsupported C-terminal structure is not promoted beyond the coverage boundary.

Structural Tensions

Library coverage ↔ novel-fold blindness. Restricting search to known frameworks produces tractability and leverage, but a missing native fold cannot win. Diagnostic: remove close structural neighbors and measure whether confidence degrades honestly or a wrong template remains falsely decisive.

Local environment fit ↔ nonlocal contact consistency. Positionwise scores optimize efficiently but can accept mutually incompatible placements; pair-contact terms add three-dimensional realism while coupling distant alignment choices. Diagnostic: compare winners under local-only and contact-aware scores and inspect which long-range contacts reverse the ranking.

Alignment flexibility ↔ opportunistic fit. Generous gaps accommodate real insertions and domain variation, yet enough freedom lets unrelated sequences exploit favorable template positions. Diagnostic: track score gain per added gap and validate gap patterns against secondary structure and independent constraints.

Score richness ↔ computational tractability. Detailed potentials and learned features may discriminate remote folds, but enlarge the search and calibration burden. Diagnostic: report marginal discrimination on held-out decoys relative to runtime, parameter count, and solver guarantees.

Sensitivity ↔ false-positive control. Lower thresholds recover remote relationships but make every library yield a seemingly plausible top hit. Diagnostic: calibrate significance by target length and library size, and require an explicit unresolved state.

Template fidelity ↔ refinement freedom. Copying the template stabilizes well-supported cores; allowing assembly to move away can correct divergence but also destroy the only reliable signal. Diagnostic: stratify restraints by alignment confidence and test whether refinement improves independent quality measures rather than only its own energy.

Method autonomy ↔ reduction to generic optimization. The pipeline can be described abstractly as constrained optimization and ranking, but that reduction erases the biological facts that make a score valid: amino-acid chemistry, three-dimensional residue environments, fold-library coverage, and protein-model validation. Diagnostic: replace residues and folds with arbitrary tokens and templates; if the same validity tests still suffice, the account is too generic, whereas if protein-specific compatibility must be restored, the domain node remains autonomous.

Structural–Framed Character

Protein Threading is structural-leaning with an aggregate framedness score of 0.2. Its procedure is formal and descriptive rather than evaluative or institution-dependent, but its identity does not literally travel outside computational structural biology.

1. Vocabulary travels (0.5). “Alignment,” “template,” “score,” and “optimization” travel. “Amino-acid sequence,” “fold,” “residue environment,” and “threading” in this sense do not.
2. Evaluative weight (0.0). The concept contains no built-in social or normative approval. Better and worse scores are defined operationally against predictive validity.
3. Institutional origin (0.0). Databases and assessment communities support the practice, but no institution's authority constitutes whether an algorithm aligns a sequence to a structural fold.
4. Human-practice bound (0.0). Automated systems instantiate the method without a human interpretive act at each alignment.
5. Import versus recognize (0.5). The method recognizes structural compatibility, yet its template library and empirical potentials impose a deliberately constructed hypothesis frame.

Its character: a substrate-neutral optimization skeleton whose scientific identity and validity remain bound to protein sequence–structure inference.

Structural Core vs. Domain Accent

Structural core. An unresolved target is aligned to typed candidates under constraints; a compatibility objective evaluates the assignments; search produces ranked hypotheses; confidence is judged against alternatives; and the output inherits both power and blind spots from the candidate library.

Domain accent. The target positions are amino acids; the candidate types are three-dimensional protein folds; role features include burial, secondary structure, solvent exposure, geometry, and residue contacts; gaps represent biological insertions or deletions; and success is judged by fold and alignment accuracy, not arbitrary fit.

Why this does not clear the prime bar. If protein vocabulary is stripped away, the remainder decomposes into live abstractions including optimization, representation, classification, comparison, and selection. Those primes travel literally. Protein threading does not: a legal brief fitted to a precedent or a task fitted to a workflow is at most analogous. Retaining sequence chemistry and fold-space validity preserves a reusable specialist method, while generalizing it would duplicate the existing prime layer.

Instantiates / Related Primes

• optimization. Threading specifies candidate template/alignment configurations, a compatibility objective, constraints, and an operative sense of best or acceptable. This is the strongest proposed direct parent.
• representation. A structural template represents selected geometric and environmental relations that the unknown target is hypothesized to preserve. The mapping is approximate and its faithfulness boundary must be stated.
• classification. Fold recognition can assign a target to a known fold or superfamily category, but some pipelines return ranked templates or local models without a definitive class assignment.
• comparison. Competing sequence–structure hypotheses are placed under a shared score and ranked; comparison is an operation inside the pipeline, not exact coverage of it.
• selection. A compatibility criterion retains one or several templates from a library for downstream modeling. Selection describes the retention step, not the alignment or biological score.
• parsing. Both recover latent organization from a linear sequence, but parsing uses a grammar and hierarchical derivation, whereas threading places residues into physical structural roles.

Relationships to Other Abstractions

Current abstraction Protein Threading Domain-specific

Parents (1) — more general patterns this builds on

• Protein Threading is a kind of Optimization Prime

  optimization. Threading specifies candidate template/alignment configurations, a compatibility objective, constraints, and an operative sense of best or acceptable.

Hierarchy path (1) — routes to 1 parentless root

• Protein Threading → Optimization

Neighborhood in Abstraction Space

Protein Threading sits in a sparse region of the domain-specific corpus (87^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Protein Structure & Antigen Recognition (7 abstractions)

Nearest neighbors

• Predicted Aligned Error — 0.81
• Levinthal's Paradox — 0.80
• Nitrogen Cycle — 0.80
• Contact order — 0.79
• THEMATICS — 0.79

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

• Homology modeling. Both are template-based, and modern methods overlap. Tell: if a close homolog plus sequence alignment supplies the decisive template and correspondence, the case is comparative modeling; if three-dimensional environment compatibility is needed to recognize a remote fold, it is threading.
• De novo / free modeling. Tell: remove every known structural template; if the method can still generate its main fold hypothesis, it is not template-dependent threading.
• Sequence-profile search. Tell: ask whether template positions carry structural environments or only sequence statistics. The latter alone is profile alignment.
• Structural alignment. Tell: structural alignment begins with coordinates on both sides; threading begins with coordinates only for the template.
• Protein folding simulation. Tell: a simulation models a conformational process or landscape; threading scores candidate frameworks without claiming to reproduce the folding path.
• Model-quality assessment. Tell: quality assessment scores already-built candidate models; threading jointly chooses a template and sequence-to-structure correspondence, though it may reuse similar potentials.
• Molecular docking. Tell: docking places separate molecules relative to one another; threading places one covalent sequence along an internal fold framework.
• parsing. Tell: parsing's acceptance conditions come from a grammar over symbols; threading's come from protein structure and residue compatibility.
• The everyday computing sense of threading. Tell: execution threads schedule instruction streams; protein threading aligns amino-acid positions to a molecular structure.

References

[1] Bowie, J. U., Lüthy, R., & Eisenberg, D. (1991). A method to identify protein sequences that fold into a known three-dimensional structure. Science, 253(5016), 164–170. https://doi.org/10.1126/science.1853201 registry ↩a ↩b ↩c

[2] Jones, D. T., Taylor, W. R., & Thornton, J. M. (1992). A new approach to protein fold recognition. Nature, 358, 86–89. https://doi.org/10.1038/358086a0 registry ↩a ↩b

[3] Lathrop, R. H. (1994). The protein threading problem with sequence amino acid interaction preferences is NP-complete. Protein Engineering, 7(9), 1059–1068. https://doi.org/10.1093/protein/7.9.1059 registry ↩a ↩b

[4] Wu, S., & Zhang, Y. (2007). LOMETS: A local meta-threading-server for protein structure prediction. Nucleic Acids Research, 35(10), 3375–3382. https://doi.org/10.1093/nar/gkm251 registry ↩a ↩b ↩c

[5] Jones, D. T. (1999). GenTHREADER: An efficient and reliable protein fold recognition method for genomic sequences. Journal of Molecular Biology, 287(4), 797–815. https://doi.org/10.1006/jmbi.1999.2583 registry ↩

[6] Zhang, Y. (2008). I-TASSER server for protein 3D structure prediction. BMC Bioinformatics, 9, 40. https://doi.org/10.1186/1471-2105-9-40 registry ↩a ↩b

[7] Bryant, S. H., & Lawrence, C. E. (1993). An empirical energy function for threading protein sequence through the folding motif. Proteins: Structure, Function, and Genetics, 16(1), 92–112. https://doi.org/10.1002/prot.340160110 registry ↩