How to Predict Quaternary Structure of Proteins: A Decision Guide for Homo-Oligomers and Hetero-Complexes

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-21

Quaternary structure prediction answers a specific question: how many copies of each subunit assemble into a functional protein complex and how are those subunits arranged in three-dimensional space. This decision guide covers the workflow from sequence input to validated complex models, with emphasis on choosing between homology-based assembly, docking, and deep learning approaches such as AlphaFold-Multimer. The practical outcome is a reproducible pipeline that produces stoichiometry predictions, ranked structural models, and quality assessments suitable for downstream experimental validation.

Key Takeaways

• Quaternary structure prediction determines how many copies of each subunit assemble into a protein complex and how those subunits are arranged in three-dimensional space.
• AlphaFold-Multimer is the state-of-the-art tool for predicting protein complex quaternary structure, with its accuracy depending heavily on the depth and diversity of the multiple sequence alignment.
• The MULTICOM system samples diverse MSAs and templates, ranks predictions through multiple metrics, and achieved average TM-scores around 0.72 for CASP15 assembly targets.
• DeepMSA2 improves prediction accuracy by iteratively searching genomic and metagenome sequence databases to construct alignments that capture coevolution between subunits.
• The ModFOLDdock2 server provides independent global and local quality scores for predicted quaternary structures, useful for evaluating models from any source.

Scope and Reader Context

This guide serves biology students, researchers, laboratory professionals, and life-science practitioners who need to determine the stoichiometry and arrangement of subunits in a protein complex from sequence or monomer structures. The workflow assumes you have a protein sequence or a set of sequences for the subunits of interest, and you need to decide which computational strategy will produce the most reliable quaternary structure model for your specific system.

The decision process differs substantially between homo-oligomers, where identical chains assemble symmetrically, and hetero-complexes, where different subunits occupy defined positions. For homo-oligomers, symmetry restraints can dramatically reduce the search space and improve model quality. For hetero-complexes, the stoichiometry itself is often unknown and must be predicted before any structural arrangement can be proposed.

The field has moved rapidly since 2021. AlphaFold-Multimer emerged as the state-of-the-art tool for predicting the quaternary structure of protein complexes, and subsequent systems have built on this foundation by improving the inputs fed to the predictor and by refining and ranking its outputs [<a href="#ref-1">1</a>]. Understanding what these tools can and cannot do, and how to combine them with classical methods, is the core skill this guide develops.

At a Glance: Method Selection for Quaternary Structure Prediction

The table below summarizes the primary approaches available, their data requirements, and the situations where each method performs best. Use this table as an initial decision aid before reading the detailed workflow sections.

Core Principles of Quaternary Structure Prediction

The Stoichiometry Problem

Before predicting how subunits arrange in space, you must determine how many copies of each chain are present in the biological assembly. This stoichiometry problem is distinct from the folding problem and requires different evidence. Experimental techniques such as size exclusion chromatography, native mass spectrometry, and crosslinking can provide stoichiometry information, but computational prediction must often proceed without this data.

MultiFOLD2 includes integrated stoichiometry prediction for quaternary structures, meaning the server attempts to determine both the number of subunits and their arrangement in a single workflow [<a href="#ref-2">2</a>]. This integration is valuable because stoichiometry errors propagate directly into structural errors. If you predict a trimer when the biological assembly is a tetramer, every interface prediction is wrong.

For homo-oligomers, symmetry provides a powerful constraint. Most homo-oligomeric assemblies use cyclic, dihedral, or cubic symmetries, and these symmetries can be imposed during model building or used to score candidate assemblies. The MULTICOM system samples diverse multiple sequence alignments and templates to generate structural predictions, and this diversity helps capture the correct stoichiometry when standard approaches fail [<a href="#ref-3">3</a>].

The Arrangement Problem

Once stoichiometry is established, the arrangement problem asks how the subunits pack together. This includes the relative orientation of each chain, the buried surface area at interfaces, and the overall shape of the complex. For hetero-complexes with defined stoichiometry, the arrangement problem resembles a docking problem with additional constraints from sequence coevolution and known interaction motifs.

Deep learning methods approach the arrangement problem directly. AlphaFold-Multimer takes sequences for all subunits as input and produces a full complex model, learning the arrangement from evolutionary couplings in the multiple sequence alignment [<a href="#ref-1">1</a>]. The quality of this arrangement prediction depends on the information content of the MSA, which is why MSA construction has become a central focus of method development.

The Role of Multiple Sequence Alignments

The multiple sequence alignment is the primary information source for deep learning structure prediction. DeepMSA2 demonstrated that iterative alignment search through genomic and metagenome sequence databases can remarkably increase the accuracy of protein tertiary and quaternary structure predictions compared with current state-of-the-art methods [<a href="#ref-4">4</a>]. This finding has practical implications: the quality of your quaternary structure prediction is limited by the quality of your MSA.

For quaternary structure prediction, the MSA must capture coevolution between subunits. This requires that the sequences in the alignment include homologs of all subunits from the same organism, so that correlated mutations between chains are visible. Standard sequence databases may lack sufficient coverage for poorly studied complexes, and metagenome databases can fill this gap.

The MULTICOM system samples diverse MSAs and templates for AlphaFold-Multimer, using both traditional sequence alignments and Foldseek-based structure alignments [<a href="#ref-3">3</a>]. This diversity is important because a single MSA may not capture all the evolutionary information needed for an accurate prediction. By generating multiple candidate MSAs and evaluating the resulting models, the system can select the most reliable prediction.

Practical Workflow for Quaternary Structure Prediction

Step 1: Define the Biological Question

Start by writing down exactly what you need to know about your complex. Are you trying to determine the stoichiometry of an uncharacterized complex? Do you need a model of a known assembly to interpret mutagenesis data? Are you testing whether two proteins physically interact?

The answer determines which tools you should use. If stoichiometry is unknown, you need a method with integrated stoichiometry prediction such as MultiFOLD2 [<a href="#ref-2">2</a>]. If you already know the stoichiometry and need an arrangement model, AlphaFold-Multimer or a docking approach may suffice. If you have experimental data on the complex, such as crosslinks or low-resolution electron microscopy maps, you may need to use these as restraints in model building.

Document your biological question, the sequences you will use, and any experimental evidence about the complex before starting computation. This documentation will guide your method choices and help you interpret results.

Step 2: Collect and Validate Sequences

Obtain the amino acid sequences for all subunits of your complex. The National Center for Biotechnology Information provides search systems and sequence resources that allow you to retrieve sequences by accession, gene name, or organism [<a href="#ref-5">5</a>]. Verify that your sequences are complete, that signal peptides or transmembrane regions are correctly annotated, and that you have the correct isoform for your experimental system.

For each sequence, check the following:

• Sequence length matches the expected protein length for your organism
• No internal stop codons or ambiguous residues
• The correct species and strain are represented
• Known domains and motifs are present

Record the accession numbers and sequence versions in your laboratory notebook or electronic records. This provenance is essential for reproducibility, as sequence databases are updated regularly and older accessions may be deprecated.

Step 3: Construct Multiple Sequence Alignments

The MSA is the foundation of deep learning structure prediction. For quaternary structure prediction, you need a concatenated or paired MSA that captures coevolution between subunits. The DeepMSA2 pipeline constructs uniform protein single-chain and multichain MSAs through iterative alignment search against genomic and metagenome sequence databases [<a href="#ref-4">4</a>].

If you are using a web server such as MultiFOLD2, the MSA construction is handled automatically [<a href="#ref-2">2</a>]. If you are running AlphaFold-Multimer locally or through a pipeline such as MULTICOM, you have control over MSA construction parameters.

Practical considerations for MSA construction:

• Use the full-length sequences when possible, as truncation removes coevolutionary signal
• Include sequences from diverse species to maximize evolutionary information
• Consider adding metagenome sequences when standard databases lack depth
• Generate multiple MSAs with different parameters to sample the alignment space

The MULTICOM system demonstrates the value of MSA diversity. By sampling diverse MSAs and templates, the system generates structural predictions that are evaluated and ranked through multiple complementary metrics [<a href="#ref-3">3</a>]. This approach improves average TM-scores compared with standard AlphaFold-Multimer, indicating that MSA sampling captures information that a single alignment misses.

Step 4: Generate Candidate Models

With MSAs prepared, you can generate candidate quaternary structure models. The choice of model generator depends on your computational resources and the complexity of your target.

AlphaFold-Multimer is the baseline approach. It accepts sequences for all subunits and produces a complex model with predicted confidence scores. The standard implementation uses sequence alignments to generate models, and its performance in CASP15 assembly targets produced average TM-scores around 0.72 for first predictions [<a href="#ref-3">3</a>].

The MULTICOM system enhances AlphaFold-Multimer by improving the input and refining the output. It samples diverse MSAs and templates, generates structural predictions using both traditional sequence alignments and Foldseek-based structure alignments, ranks predictions through multiple complementary metrics, and refines the structural predictions via a Foldseek structure alignment-based refinement method [<a href="#ref-3">3</a>]. In CASP15, the MULTICOM server ranked third among 26 server predictors, and the human predictor ranked seventh among 87 server and human predictors [<a href="#ref-3">3</a>].

The Foldseek Structure Alignment-based Multimer structure Generation method outperforms the widely used sequence alignment-based multimer structure generation [<a href="#ref-3">3</a>]. This finding suggests that structure-based alignment can capture evolutionary relationships that sequence alignment misses, particularly for distantly related complexes.

For researchers without local compute resources, the MultiFOLD2 server provides integrated stoichiometry prediction and improved sampling and scoring [<a href="#ref-2">2</a>]. This server is freely accessible through a web interface and has demonstrated high performance in continuous independent benchmarks such as CAMEO [<a href="#ref-2">2</a>].

Step 5: Evaluate and Rank Models

Model evaluation is a critical step that determines which candidate structure you will use for downstream analysis. The ModFOLDdock2 server uses a hybrid consensus approach to generate global and local quality scores for predicted quaternary structures [<a href="#ref-2">2</a>]. This server can evaluate models from any source, making it useful for independent assessment of AlphaFold-Multimer or docking results.

The MULTICOM system ranks structural predictions through multiple complementary metrics [<a href="#ref-3">3</a>]. This multi-metric ranking is important because no single score captures all aspects of model quality. A model with a good global score may have local errors at functionally important interfaces, and a model with moderate global quality may have accurate predictions at the binding site.

Quality assessment should include:

• Global model quality scores such as TM-score or predicted TM-score
• Local per-residue confidence scores
• Interface quality assessment, including buried surface area and residue conservation
• Comparison with experimental data when available

The average TM-score of the first predictions submitted by MULTICOM for CASP15 assembly targets was approximately 0.76, which is 5.3 percent higher than the approximately 0.72 achieved by standard AlphaFold-Multimer [<a href="#ref-3">3</a>]. The best of the top five predictions achieved an average TM-score of approximately 0.80, about 8 percent higher than the standard approach [<a href="#ref-3">3</a>]. These improvements demonstrate that model evaluation and selection contribute substantially to final model quality.

Step 6: Validate with Experimental Data

Computational predictions require experimental validation before they can be used to guide experiments or interpret biological data. The level of validation needed depends on the intended use of the model.

For hypothesis generation, a model with high confidence scores and consistent predictions across multiple methods may be sufficient. For structure-based mutagenesis or drug design, additional validation is essential.

Validation approaches include:

• Crosslinking mass spectrometry to identify residue pairs in close proximity
• Hydrogen-deuterium exchange to map solvent-exposed regions
• Site-directed mutagenesis of predicted interface residues followed by binding assays
• Small-angle X-ray scattering to compare predicted and experimental envelopes
• Cryo-electron microscopy for high-resolution validation when feasible

The NLRP3 inflammasome study provides an example of experimental validation of a predicted complex. The cryo-electron microscopy structure of inactive human NLRP3 in complex with NEK7 revealed an earring-shaped NLRP3 consisting of curved leucine-rich-repeat and globular NACHT domains, with the C-terminal lobe of NEK7 nestling against both NLRP3 domains [<a href="#ref-6">6</a>]. Structural recognition between NLRP3 and NEK7 was confirmed by mutagenesis both in vitro and in cells [<a href="#ref-6">6</a>]. This study demonstrates the value of combining structural prediction with experimental validation to establish biological mechanisms.

Options and Tradeoffs in Method Selection

Web Servers versus Local Pipelines

Web servers such as MultiFOLD2 and ModFOLDdock2 offer accessibility and ease of use. They require no local compute infrastructure and provide user-friendly interfaces [<a href="#ref-2">2</a>]. The tradeoff is limited control over parameters, potential queue times, and constraints on batch processing.

Local pipelines such as MULTICOM offer control and reproducibility. The source code is available, allowing researchers to modify parameters and integrate the pipeline with their own workflows [<a href="#ref-1">1</a>]. The tradeoff is the need for significant computational resources, particularly for MSA sampling and model refinement.

For most research groups, a hybrid approach works best. Use web servers for initial exploration and validation of results, and use local pipelines for production runs where parameter control and batch processing are needed.

Sequence Alignment versus Structure Alignment

Traditional sequence alignment methods search for homologous sequences using tools such as PSI-BLAST and HHsearch. Structure alignment methods such as Foldseek search for structural similarity in protein structure databases. The MULTICOM system uses both approaches to generate diverse MSAs and templates [<a href="#ref-3">3</a>].

The Foldseek Structure Alignment-based Multimer structure Generation method outperforms sequence alignment-based generation [<a href="#ref-3">3</a>]. This suggests that structure alignment captures evolutionary relationships that are invisible to sequence comparison, particularly for proteins with divergent sequences but conserved structures.

For practical use, consider running both sequence-based and structure-based alignment searches. If the two approaches identify different template structures, the diversity may improve model quality by providing alternative starting points for the predictor.

Single Model versus Ensemble Approaches

Standard AlphaFold-Multimer generates a small number of models and selects the best based on confidence scores. Ensemble approaches such as MULTICOM generate many models from diverse inputs and rank them using multiple metrics [<a href="#ref-3">3</a>].

The ensemble approach has a clear advantage in CASP15 benchmarks. The best of the top five models from MULTICOM achieved higher average TM-scores than the first predictions [<a href="#ref-3">3</a>]. This finding suggests that generating multiple models and selecting the best is more reliable than relying on a single prediction.

For practical use, generate at least five models for each complex and evaluate all of them. Do not discard models with lower global scores without examining their local quality, as a model with moderate global quality may have accurate predictions at the interface of interest.

Observations and Measurements in Quaternary Structure Prediction

Confidence Scores as Predictions of Accuracy

Deep learning structure predictors produce confidence scores that estimate the accuracy of each predicted residue position and the overall model. These scores are calibrated on training data and provide a useful guide for model selection, but they are not guarantees of accuracy.

The MULTICOM system ranks structural predictions through multiple complementary metrics [<a href="#ref-3">3</a>]. This multi-metric approach is more reliable than any single score because different metrics capture different aspects of model quality.

When reporting confidence scores, record the specific metric used and the version of the software. Different versions of AlphaFold-Multimer produce different score distributions, and comparisons across versions can be misleading.

TM-Score as a Global Quality Measure

The TM-score measures the global similarity between two protein structures, with values ranging from 0 to 1. A TM-score above 0.5 generally indicates the same fold, while scores above 0.8 indicate highly similar structures.

In CASP15, the average TM-score of the first predictions submitted by MULTICOM for assembly targets was approximately 0.76, compared with approximately 0.72 for standard AlphaFold-Multimer [<a href="#ref-3">3</a>]. The best of the top five predictions achieved approximately 0.80 [<a href="#ref-3">3</a>]. These values provide a benchmark for expected performance on assembly targets.

For your own predictions, compare TM-scores against these benchmarks. If your models consistently score below 0.5, the prediction is likely unreliable and experimental structure determination should be considered.

Interface Metrics

Interface quality is a critical aspect of quaternary structure prediction. A model with accurate monomer folds but incorrect interfaces is not useful for understanding complex function.

Interface metrics to record include:

• Buried surface area at each interface
• Number of interface residues
• Hydrogen bonds and salt bridges across the interface
• Conservation of interface residues across homologs
• Predicted local confidence at interface positions

The NLRP3-NEK7 structure illustrates the importance of interface analysis. The C-terminal lobe of NEK7 nestles against both NLRP3 domains, and modelling of an active NLRP3-NEK7 conformation predicted an additional contact between an NLRP3-bound NEK7 and a neighbouring NLRP3 [<a href="#ref-6">6</a>]. Mutations to this interface abolished the ability of NEK7 or NLRP3 to rescue NLRP3 activation in knockout cells [<a href="#ref-6">6</a>]. This example shows how interface predictions can be tested experimentally and how accurate interface models have biological consequences.

Records and Measurements for Reproducible Quaternary Structure Prediction

Documentation Requirements

Reproducible quaternary structure prediction requires detailed documentation of inputs, parameters, and outputs. The following records should be maintained for each prediction project:

• Sequence accessions and versions for all subunits
• MSA construction parameters, including databases searched and search iterations
• Software versions for all tools used
• Model generation parameters, including number of models and random seeds
• Quality assessment scores for all models
• Final model selection criteria and rationale

The Carpentries provides foundational computing and data lessons that emphasize reproducible research practices [<a href="#ref-7">7</a>]. These lessons cover shell, Git, and programming skills that support reproducible bioinformatics workflows.

Data Management

Store all input sequences, MSAs, models, and quality scores in a structured directory. Use version control for scripts and parameter files. Record the date and time of each computation, as database updates can change results.

Practitioners seeking reproducible bioinformatics education will find GTN's open-access tutorial library a valuable starting point for pipeline design.

For large-scale analyses, the nf-core documentation describes community pipeline standards for usage and configuration [<a href="#ref-8">8</a>]. These standards support reproducible workflow execution across different computing environments.

Benchmarking Against Known Structures

When predicting a complex with a known experimental structure, benchmark your pipeline by comparing predictions against the experimental structure. This comparison provides a measure of pipeline accuracy for your specific system and helps calibrate confidence scores.

The EMBL-EBI Training program provides bioinformatics learning pathways and data-resource training that include practical analysis education [<a href="#ref-9">9</a>]. These resources can help you develop benchmarking protocols and interpret quality metrics.

Common Failure Patterns in Quaternary Structure Prediction

Incorrect Stoichiometry Prediction

The most common failure is predicting the wrong number of subunits. This error is particularly common for homo-oligomers, where the predictor may propose a dimer when the biological assembly is a tetramer, or for hetero-complexes with weak interfaces.

MultiFOLD2 includes integrated stoichiometry prediction, which addresses this problem directly [<a href="#ref-2">2</a>]. If your method does not include stoichiometry prediction, you must determine stoichiometry independently using experimental methods or sequence-based predictions.

Signs of incorrect stoichiometry include:

• Low confidence scores at interfaces between predicted subunits
• Inconsistent stoichiometry across different prediction runs
• Buried surface areas that are unusually small or large for stable complexes
• Predicted assemblies that violate known symmetry preferences for the protein family

Poor MSA Quality

The MSA is the primary information source for deep learning predictors. Poor MSA quality leads to poor predictions regardless of the model generator used.

DeepMSA2 demonstrated that iterative alignment search through genomic and metagenome sequence databases can increase prediction accuracy [<a href="#ref-4">4</a>]. If your predictions are poor, improve the MSA by searching additional databases or using more sensitive search parameters.

Signs of poor MSA quality include:

• Very few homologous sequences identified
• Low sequence diversity in the alignment
• Gaps in the alignment at conserved positions
• Disagreement between predictions from different MSAs

Overconfidence in Model Scores

Deep learning predictors produce confidence scores that are calibrated on training data. These scores can be misleading for targets that differ substantially from the training distribution.

The MULTICOM system uses multiple complementary metrics to rank predictions [<a href="#ref-3">3</a>]. This approach reduces the risk of overconfidence in any single score.

To avoid overconfidence, compare predictions from multiple methods and examine the agreement between them. If AlphaFold-Multimer and docking approaches produce different arrangements, the prediction is uncertain regardless of confidence scores.

Ignoring Conformational Changes

Many protein complexes undergo conformational changes upon assembly. The monomer structures used for docking may not represent the conformation adopted in the complex.

The NLRP3 inflammasome study provides an example of conformational complexity. The inactive NLRP3 structure is earring-shaped with curved leucine-rich-repeat and globular NACHT domains [<a href="#ref-6">6</a>]. Modelling of an active conformation based on the NLRC4 inflammasome predicted additional contacts that are not present in the inactive structure [<a href="#ref-6">6</a>]. This conformational change would not be captured by docking the inactive monomer structures.

For complexes with known conformational changes, use deep learning methods that can model the complex directly from sequence instead of docking rigid monomer structures.

Limitations of Current Methods

Sequence Coverage Limits

Deep learning structure prediction relies on evolutionary information in the MSA. Complexes with few homologs, such as species-specific immune proteins or recently evolved interactions, have limited sequence coverage and produce less reliable predictions.

DeepMSA2 addresses this limitation by searching metagenome databases, which contain sequences from uncultured organisms [<a href="#ref-4">4</a>]. This approach can dramatically increase the number of homologous sequences available for alignment construction.

If your complex has poor sequence coverage, consider experimental structure determination or use of crosslinking data to constrain the assembly.

Symmetry Assumptions

Many prediction methods assume symmetric assemblies for homo-oligomers. While this assumption is often correct, some homo-oligomers form asymmetric assemblies or have symmetry that is broken by ligand binding or post-translational modification.

The NLRP3 inflammasome provides an example of symmetry considerations. The inactive NLRP3-NEK7 complex has a defined stoichiometry, and modelling of the active conformation predicts that NEK7 bridges adjacent NLRP3 subunits with bipartite interactions [<a href="#ref-6">6</a>]. This arrangement has specific symmetry properties that may not be captured by generic symmetry restraints.

For homo-oligomers with suspected asymmetry, generate models without symmetry restraints and compare the results with symmetric models.

Computational Cost

Deep learning structure prediction is computationally intensive. MSA construction, model generation, and refinement require significant CPU and GPU resources.

The MULTICOM system samples diverse MSAs and templates, which increases computational cost compared with standard AlphaFold-Multimer [<a href="#ref-3">3</a>]. The improved accuracy justifies this cost for challenging targets, but for simple complexes with deep sequence coverage, standard approaches may suffice.

For researchers with limited compute resources, web servers such as MultiFOLD2 and ModFOLDdock2 provide access to state-of-the-art methods without local infrastructure [<a href="#ref-2">2</a>].

Interpretation Limits

Quaternary structure models are predictions, not experimental structures. Even high-confidence models may have errors at functionally important positions, and models with moderate confidence scores require experimental validation.

The NLRP3 study demonstrates the value of combining prediction with experiment. The cryo-electron microscopy structure confirmed the overall arrangement predicted by modelling, but mutagenesis was required to establish the functional importance of specific interfaces [<a href="#ref-6">6</a>].

For any model used to guide experiments, validate the key interface predictions experimentally before drawing biological conclusions.

Quality Controls and Professional Escalation Criteria

Quality Control Checks

Before accepting a quaternary structure prediction, perform the following quality control checks:

• Verify that all subunits are present in the predicted assembly with the correct stoichiometry
• Check that the predicted interfaces have reasonable buried surface areas for stable complexes
• Compare predictions from multiple methods and examine the agreement
• Examine local confidence scores at functionally important positions
• Check for steric clashes between subunits
• Verify that the predicted assembly is consistent with known biochemical data

The ModFOLDdock2 server provides global and local quality scores for predicted quaternary structures [<a href="#ref-2">2</a>]. Use this server to independently evaluate models from any source.

When to Escalate to Experimental Structure Determination

Escalate to experimental structure determination when computational predictions are unreliable or when the biological conclusions depend critically on the structure.

Indications for experimental structure determination include:

• Low confidence scores across the entire model
• Disagreement between different prediction methods
• Poor sequence coverage in the MSA
• Conformational changes that cannot be modelled reliably
• Biological conclusions that depend on atomic details of the interface

Cryo-electron microscopy is the method of choice for large complexes, as demonstrated by the NLRP3-NEK7 structure determined at 3.8 angstrom resolution [<a href="#ref-6">6</a>]. X-ray crystallography and NMR remain valuable for smaller complexes and for systems that resist cryo-electron microscopy.

Documentation and Reporting Standards

Report quaternary structure predictions with full documentation of methods and limitations. Include the software versions, database versions, and parameters used. Report confidence scores and quality metrics. Describe the limitations of the prediction and the level of experimental validation required.

Each Bioconductor package submission undergoes technical review for documentation completeness, ensuring reproducible R-based genomic analysis workflows.

Safety and Regulatory Context

Responsible Use of Structural Models

Quaternary structure models are used in drug discovery, protein engineering, and biotechnology. Models that are used to guide experimental work should be validated experimentally before being used for high-stakes decisions.

The NLRP3 inflammasome is a drug target for inflammatory diseases. The structural model of the NLRP3-NEK7 complex provides a framework for understanding activation mechanisms and for designing experiments [<a href="#ref-6">6</a>]. However, the model alone is not sufficient for drug design without experimental validation.

Data Sharing and Reproducibility

Share the sequences, MSAs, models, and quality scores associated with published predictions. This sharing enables other researchers to reproduce the predictions and to build on the work.

Reproducibility-first pipeline training is a hallmark of the Galaxy Training Network's hands-on GTN lesson catalogue.

Professional Escalation

If you are uncertain about the reliability of a quaternary structure prediction, consult with colleagues who have structural biology expertise. If the prediction will be used for regulatory submissions or clinical decisions, involve structural biology experts and consider experimental structure determination.

The EMBL-EBI Training program provides bioinformatics learning pathways that can help you develop the skills needed to evaluate structural predictions critically [<a href="#ref-9">9</a>]. The NCBI provides official descriptions of databases and analysis services that support sequence-based analysis [<a href="#ref-5">5</a>].

Decision Framework for Selecting Prediction Strategies by Complex Type

The choice between homology modeling, docking, and deep learning approaches depends on the specific characteristics of your target complex. A structured decision framework helps you match the method to the biological problem instead of defaulting to the most popular tool. This section provides a practical framework organized around complex type, available data, and the questions you need to answer at each stage.

Framework for Homo-Oligomer Prediction

For homo-oligomers, the central question is symmetry. Most homo-oligomeric assemblies use cyclic symmetry, where subunits arrange in a ring, or dihedral symmetry, where subunits arrange in two stacked rings. The symmetry operation relates each subunit to the others, and this relationship dramatically reduces the number of possible arrangements.

Start by determining whether your protein has a known oligomeric state from experimental evidence. Size exclusion chromatography, native mass spectrometry, or analytical ultracentrifugation can establish the number of subunits in the biological assembly. If you have this information, record it before running any prediction. If you do not have experimental stoichiometry, use a method with integrated stoichiometry prediction such as MultiFOLD2, which includes stoichiometry prediction for quaternary structures [<a href="#ref-2">2</a>].

For homo-oligomers with known stoichiometry, apply the following decision path:

1. If you have a high-confidence monomer structure from experimental determination or deep learning prediction, test whether symmetry-constrained docking reproduces the known interfaces. This approach works best when the monomer structure is accurate and the conformational changes upon assembly are small.

1. If you have sequences but no reliable monomer structure, use AlphaFold-Multimer with a paired MSA that includes multiple copies of the same sequence. The predictor will attempt to model the symmetric assembly directly from sequence information.

1. If standard AlphaFold-Multimer produces low-confidence models or inconsistent stoichiometry across runs, apply the MULTICOM enhancement strategy. This system samples diverse MSAs and templates, generates structural predictions using both traditional sequence alignments and Foldseek-based structure alignments, and ranks predictions through multiple complementary metrics [<a href="#ref-3">3</a>].

The MULTICOM approach demonstrated measurable improvements in CASP15 assembly targets. The average TM-score of the first predictions submitted by MULTICOM was approximately 0.76, which is 5.3 percent higher than the approximately 0.72 achieved by standard AlphaFold-Multimer [<a href="#ref-3">3</a>]. The best of the top five predictions achieved an average TM-score of approximately 0.80, about 8 percent higher than the standard approach [<a href="#ref-3">3</a>]. These improvements came from MSA diversity and model ranking instead of from changes to the core predictor.

Framework for Hetero-Complex Prediction

Hetero-complexes present a different challenge because the stoichiometry is often unknown and the subunits are not related by symmetry. The decision framework for hetero-complexes must address both the stoichiometry question and the arrangement question separately.

For hetero-complexes with unknown stoichiometry, use MultiFOLD2 as the first step because it includes integrated stoichiometry prediction [<a href="#ref-2">2</a>]. The server attempts to determine both the number of subunits and their arrangement in a single workflow. This integration is valuable because stoichiometry errors propagate directly into structural errors.

For hetero-complexes with known stoichiometry, the decision path depends on the number of subunits and the availability of homologous sequences:

1. For complexes with two or three subunits and deep sequence coverage, AlphaFold-Multimer with a paired MSA is the appropriate starting point. The paired MSA must include sequences from organisms that contain all subunits, so that coevolution between chains is visible.

1. For complexes with many subunits or with poor sequence coverage, consider the DeepMSA2 approach. This pipeline constructs uniform protein single-chain and multichain MSAs through iterative alignment search against genomic and metagenome sequence databases [<a href="#ref-4">4</a>]. Large-scale benchmarks show that DeepMSA2 MSAs can increase the accuracy of protein tertiary and quaternary structure predictions compared with current state-of-the-art methods [<a href="#ref-4">4</a>].

1. For complexes where the individual subunit structures are known but the assembly is uncertain, use docking approaches with experimental restraints. Crosslinking mass spectrometry data can provide distance restraints between specific residues, and these restraints can guide the docking search.

The NLRP3 inflammasome provides a useful example of hetero-complex prediction challenges. The cryo-electron microscopy structure of inactive human NLRP3 in complex with NEK7 revealed an earring-shaped NLRP3 consisting of curved leucine-rich-repeat and globular NACHT domains, with the C-terminal lobe of NEK7 nestling against both NLRP3 domains [<a href="#ref-6">6</a>]. Modelling of an active NLRP3-NEK7 conformation based on the NLRC4 inflammasome predicted an additional contact between an NLRP3-bound NEK7 and a neighbouring NLRP3 [<a href="#ref-6">6</a>]. This example shows that hetero-complex prediction may require modelling multiple conformational states and that the active conformation can differ substantially from the inactive structure.

Framework for Complexes with Experimental Restraints

When you have experimental data about the complex, the decision framework changes. Experimental restraints can come from crosslinking mass spectrometry, hydrogen-deuterium exchange, small-angle X-ray scattering, or low-resolution electron microscopy maps. These data can be used to filter candidate models or to guide model generation.

The decision path for complexes with experimental restraints:

1. Generate a diverse set of candidate models using AlphaFold-Multimer or docking approaches. The MULTICOM system generates structural predictions using both traditional sequence alignments and Foldseek-based structure alignments, providing diversity in the candidate model set [<a href="#ref-3">3</a>].

1. Score each candidate model against the experimental restraints. Crosslinking distance restraints should be satisfied within the expected distance range. Small-angle X-ray scattering profiles should match the predicted envelope.

1. Rank the models by combined scores that include both the prediction confidence and the experimental restraint satisfaction. The MULTICOM system ranks structural predictions through multiple complementary metrics [<a href="#ref-3">3</a>], and you can add experimental restraint satisfaction as an additional metric.

1. If no model satisfies the experimental restraints, revisit the stoichiometry assumption or consider conformational changes that are not captured by the prediction methods.

Record System for Method Selection Decisions

Document the rationale for each method choice so that the decision process is reproducible and auditable. The following record fields should be maintained for each prediction project:

• Complex name and biological context
• Subunit sequences with accession numbers and versions
• Known experimental stoichiometry or the source of the stoichiometry assumption
• Symmetry assumption for homo-oligomers, if applicable
• Method selected and the reason for the selection
• Alternative methods considered and the reason they were not selected
• Software versions and database versions used
• Date of the analysis

The Galaxy Training Network (GTN) hosts peer-reviewed, browser-accessible tutorials that walk researchers through reproducible analysis pipelines.

Troubleshooting Method Selection Failures

When a selected method produces poor results, the failure often indicates a mismatch between the method and the biological problem. Use the following troubleshooting path:

If AlphaFold-Multimer produces low-confidence models, the MSA may lack depth or diversity. Apply the DeepMSA2 approach to construct MSAs through iterative alignment search against genomic and metagenome sequence databases [<a href="#ref-4">4</a>]. The major advantage of DeepMSA2 lies in its balanced alignment search and effective model selection, and in the power of integrating huge metagenomics databases [<a href="#ref-4">4</a>].

If docking approaches produce inconsistent arrangements, the monomer structures may not represent the conformation adopted in the complex. Consider whether conformational changes accompany assembly. The NLRP3 study demonstrated that the active conformation of the NLRP3-NEK7 complex differs from the inactive structure, with additional contacts predicted in the active state [<a href="#ref-6">6</a>].

If different methods produce conflicting arrangements, the conflict itself is informative. It indicates that the evolutionary signal is insufficient to determine the arrangement uniquely. In this case, experimental data are required to distinguish between the candidate models. Crosslinking mass spectrometry or site-directed mutagenesis of predicted interface residues can provide the discriminating evidence.

If all computational methods fail, escalate to experimental structure determination. Cryo-electron microscopy is the method of choice for large complexes, as demonstrated by the NLRP3-NEK7 structure determined at 3.8 angstrom resolution [<a href="#ref-6">6</a>]. The decision to escalate should be documented with the specific reasons why computational predictions were insufficient.

Comparison of Method Selection Outcomes

The table below summarizes the expected outcomes for different method selections across complex types. Use this comparison to calibrate your expectations and to identify when a method is underperforming.

The MULTICOM system demonstrated that the average TM-score of the first predictions for CASP15 assembly targets was approximately 0.76, compared with approximately 0.72 for standard AlphaFold-Multimer [<a href="#ref-3">3</a>]. The best of the top five predictions achieved approximately 0.80 [<a href="#ref-3">3</a>]. These benchmarks provide reference points for evaluating your own predictions on challenging targets.

Professional Escalation Criteria for Method Selection

Escalate to a more computationally intensive method or to experimental structure determination when the following conditions are met:

• The selected method produces models with TM-scores below 0.5, indicating that the predicted arrangement is likely incorrect
• Different methods produce conflicting arrangements and the conflict cannot be resolved with available experimental data
• The biological conclusions depend critically on atomic details of the interface, and the prediction confidence at interface positions is low
• The complex undergoes conformational changes that cannot be modelled reliably with any available method

The EMBL-EBI Training program provides bioinformatics learning pathways and data-resource training that include practical analysis education [<a href="#ref-9">9</a>]. These resources can help you develop the skills needed to evaluate structural predictions critically. The NCBI provides official descriptions of databases and analysis services that support sequence-based analysis [<a href="#ref-5">5</a>].

Frequently Asked Questions

What is the difference between tertiary and quaternary structure prediction?

Tertiary structure prediction determines the three-dimensional fold of a single protein chain. Quaternary structure prediction determines how multiple chains assemble into a complex, including the stoichiometry and arrangement of subunits. Quaternary structure prediction requires additional information about inter-chain interactions, which is captured in paired multiple sequence alignments that show coevolution between subunits.

When should I use AlphaFold-Multimer versus classical docking approaches?

AlphaFold-Multimer is the preferred approach when you have sequences for all subunits and want to predict the complex directly from sequence. It is particularly effective when the MSA has deep coverage. Classical docking approaches are useful when you have experimental monomer structures and want to explore alternative arrangements or when the complex undergoes conformational changes that deep learning methods cannot capture.

How do I determine the stoichiometry of an unknown complex?

MultiFOLD2 includes integrated stoichiometry prediction for quaternary structures [<a href="#ref-2">2</a>]. If you are using a method without stoichiometry prediction, determine stoichiometry experimentally using size exclusion chromatography, native mass spectrometry, or crosslinking. Sequence-based methods can also provide stoichiometry predictions based on symmetry preferences and interface conservation.

What is the role of multiple sequence alignments in quaternary structure prediction?

The MSA provides the evolutionary information that deep learning predictors use to infer contacts between residues, both within and between chains. DeepMSA2 demonstrated that iterative alignment search through genomic and metagenome databases can increase prediction accuracy [<a href="#ref-4">4</a>]. The quality of the MSA is the primary determinant of prediction quality.

How can I improve predictions for complexes with poor sequence coverage?

Search additional sequence databases, particularly metagenome databases that contain sequences from uncultured organisms. DeepMSA2 uses this approach to improve predictions [<a href="#ref-4">4</a>]. You can also use structure-based alignment methods such as Foldseek to identify distantly related templates [<a href="#ref-3">3</a>]. If sequence coverage remains poor, consider experimental structure determination.

What quality metrics should I report for quaternary structure predictions?

Report the global model quality score such as TM-score or predicted TM-score, local per-residue confidence scores, and interface quality metrics including buried surface area and interface residue conservation. The ModFOLDdock2 server provides global and local quality scores for predicted quaternary structures [<a href="#ref-2">2</a>]. Report the software versions and parameters used to generate the models.

How do I validate a predicted quaternary structure experimentally?

Use crosslinking mass spectrometry to identify residue pairs in close proximity, site-directed mutagenesis of predicted interface residues followed by binding assays, and small-angle X-ray scattering to compare predicted and experimental envelopes. The NLRP3 study used mutagenesis to confirm structural recognition between NLRP3 and NEK7 [<a href="#ref-6">6</a>]. For high-resolution validation, use cryo-electron microscopy or X-ray crystallography.

When should I escalate to experimental structure determination?

Escalate when computational predictions are unreliable, when different methods disagree, when the biological conclusions depend critically on atomic details, or when the complex undergoes conformational changes that cannot be modelled. Cryo-electron microscopy is the method of choice for large complexes, as demonstrated by the NLRP3-NEK7 structure determined at 3.8 angstrom resolution [<a href="#ref-6">6</a>].

Related Bioinformatics Guides

• How To Use Alphafold To Predict Structure: Structural Analysis and Computational Methodologies in Bioinformatics
• AlphaFold and Beyond: Predicting Viral Protein Structures for Antiviral Target Discovery
• Structural Prediction of Bat Coronavirus Spike Proteins: Insights from AlphaFold2 and Molecular Dynamics
• What Is the Monomer of a Protein? Structure & Synthesis
• Genomic Prediction in Livestock: A Decision Framework for Breeders

References and Further Reading

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.