# How to Use AlphaFold-Multimer for Protein-Protein Complex Prediction: A Practical Tutorial and Limitations

AlphaFold-Multimer is a deep learning system that predicts the three-dimensional structure of protein complexes from amino acid sequences. This tutorial covers the practical steps for preparing inputs, running predictions, and interpreting confidence scores, with specific attention to large complexes and common failure cases. The guidance is intended for biology students, researchers, laboratory professionals, and life-science practitioners who need reproducible protocols instead of theoretical background alone.

The workflow described here follows the standard AlphaFold-Multimer pipeline: sequence retrieval, multiple sequence alignment generation, model inference, and confidence-based structure interpretation. Each stage has specific decision points that affect output quality. Understanding these decision points helps you avoid wasted compute time and misinterpreted results.

## At a Glance

| Workflow Stage | Primary Decision | Common Output Problem | Practical Check |
| --- | --- | --- | --- |
| Sequence retrieval | Which database and identifier system to use | Wrong isoform or species variant | Verify sequence length and annotation against the source database |
| Multiple sequence alignment | Which MSA depth and diversity to accept | Shallow alignments for orphan or fast-evolving proteins | Inspect alignment coverage and sequence count before running inference |
| Model inference | Which model parameters and recycling settings to use | High memory usage or slow runtime for large complexes | Match model settings to complex size and available GPU memory |
| Confidence interpretation | Which score threshold indicates a reliable interface | Overinterpretation of low-confidence regions | Use pLDDT and PAE together, not either score alone |
| Large complex handling | Whether to predict full complex or subcomplexes | Memory exhaustion or poor interface prediction | Split into logical subcomplexes and assemble with docking tools |

## Understanding the Input Requirements

### Sequence Data Sources and Format

AlphaFold-Multimer requires amino acid sequences for every chain in the complex. The standard input format is FASTA, where each chain is represented by a header line starting with the greater-than symbol followed by the sequence on subsequent lines. For multimer prediction, all chains are placed in a single FASTA file, and the order of chains in the file determines the chain identifiers in the output.

The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide official access to sequence databases including GenBank, RefSeq, and the Protein Database. When retrieving sequences for a complex, you must verify that each sequence corresponds to the correct isoform and species. A common error is retrieving a sequence from a different splice variant or a homologous protein from a related species. The NCBI search systems allow you to filter by organism, sequence length, and annotation status, which reduces the risk of using the wrong sequence.

For researchers who need training in sequence retrieval and database navigation, the [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers structured learning pathways covering sequence databases, alignment tools, and structure resources. These materials are useful for laboratory professionals who want to standardize their sequence handling procedures.

### Chain Ordering and Stoichiometry

The order of chains in the FASTA file matters for two reasons. First, AlphaFold-Multimer assigns chain IDs based on file order, so the output structure will label chains A, B, C, and so on according to that order. Second, the model uses the sequence context to determine which chains interact, and the stoichiometry must be encoded by repeating identical sequences when the complex contains multiple copies of the same protein.

For a homodimer, the same sequence appears twice in the FASTA file. For a heterotrimer with two copies of one protein and one copy of another, the first protein sequence appears twice and the second appears once. The order of repeated sequences does not affect prediction quality, but consistent ordering helps you track which chain corresponds to which biological subunit.

### Sequence Length and Complex Size Considerations

The computational cost of AlphaFold-Multimer scales with the total number of residues in the complex. A complex with 1,000 total residues requires substantially more memory and time than a complex with 300 residues. The practical upper limit depends on your hardware, but complexes exceeding 2,000 to 3,000 total residues often require specialized handling.

For large complexes with six to twenty chains, the [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) demonstrates an alternative strategy. This method uses AlphaFold-Multimer to predict pairwise models and then assembles them using a reinforcement learning framework. The pairwise prediction strategy reduces the memory burden because each individual prediction involves only two chains, and the assembly step combines the pairwise models into the full complex. This approach was developed specifically for complexes where direct full-complex prediction is computationally prohibitive.

## Preparing Multiple Sequence Alignments

### Why the MSA Matters

AlphaFold-Multimer uses multiple sequence alignments as evolutionary context for structure prediction. The model learns from patterns of coevolution between residues, where correlated mutations across homologous sequences indicate spatial proximity in the folded structure. A deep and diverse MSA provides more evolutionary signal, which generally improves prediction accuracy.

The MSA is generated by searching sequence databases for homologs of each chain in the complex. The search is typically performed with tools such as MMseqs2 or JackHMMER, which iteratively search databases and build alignments. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible tutorials for running these search tools and understanding the output formats, which is valuable for researchers who are new to command-line bioinformatics.

### MSA Depth and Diversity Checks

Before running structure prediction, you should inspect the MSA for each chain. The key metrics are the number of aligned sequences and the sequence coverage across the full length of the target protein. A chain with fewer than 50 aligned homologous sequences may produce lower-confidence predictions, particularly for regions that lack evolutionary conservation.

For proteins with many homologs, such as conserved transcription factors, the MSA will be deep and the prediction confidence will generally be high. The [evolutionary analysis of transcription elongation factors](https://doi.org/10.1371/journal.pbio.3003855) demonstrates that conserved proteins across the tree of life retain similar domain architectures, which provides strong evolutionary signal for structure prediction. In contrast, proteins with few homologs, such as rapidly evolving viral proteins or species-specific immune receptors, will have shallow MSAs and may produce lower-confidence predictions.

### Handling Shallow MSAs

When the MSA is shallow, you have several options. First, you can relax the search parameters to include more distant homologs, accepting that the alignment may contain more noise. Second, you can use the structure prediction with the available MSA and interpret the confidence scores carefully, recognizing that low-confidence regions may reflect genuine uncertainty. Third, you can combine the shallow MSA with structural information from related proteins if such structures exist.

The [SymProFold pipeline](https://doi.org/10.1038/s41467-024-52138-3) illustrates an approach for proteins that form symmetrical assemblies where individual chains may have limited homologs. This pipeline uses AlphaFold-Multimer predictions to derive symmetrical assemblies from protein sequences, testing known symmetry operations and identifying the most likely symmetry for the assembly. The approach was validated for S-layer arrays and viral capsids, where the symmetry constraints compensate for weaker evolutionary signal in individual chains.

## Running AlphaFold-Multimer

### Model Selection and Parameters

AlphaFold-Multimer provides several model parameter sets that differ in architecture and training data. The choice of model affects prediction speed and accuracy, and the optimal choice depends on your specific complex. For most applications, the default model parameters are appropriate, but you should document which model version you used for reproducibility.

The recycling setting controls how many times the model refines its prediction. More recycling steps generally improve accuracy but increase runtime. The default of three recycling steps is a reasonable starting point, and you can increase recycling for complexes where initial predictions show low confidence.

### Hardware Requirements and Runtime

AlphaFold-Multimer requires a GPU for practical runtime. The memory requirement depends on the total number of residues in the complex, with larger complexes requiring more GPU memory. A complex with 500 total residues can typically run on a GPU with 8 GB of memory, while a complex with 1,500 residues may require 16 GB or more.

The [nf-core Documentation](https://nf-co.re/docs) provides guidance on configuring computational pipelines for reproducibility, including resource allocation and containerization. Using a standardized pipeline configuration ensures that your predictions can be reproduced by collaborators or in subsequent studies.

### Running the Prediction

The standard workflow involves running the AlphaFold-Multimer script with the FASTA file as input. The script performs the following steps automatically: MSA generation, model inference, and output writing. The output directory contains the predicted structure files in PDB format, along with JSON files containing confidence scores.

For researchers who prefer a graphical interface or who need to run predictions on shared infrastructure, the [Galaxy Training Network](https://training.galaxyproject.org/) offers workflows that wrap AlphaFold-Multimer and related tools. These workflows provide a reproducible environment and reduce the need for command-line expertise.

## Interpreting Confidence Scores

### pLDDT Scores

The predicted local distance difference test (pLDDT) score measures the confidence in the local structure prediction for each residue. Scores range from 0 to 100, with higher scores indicating higher confidence. Residues with pLDDT above 90 are typically predicted with high confidence, while residues below 50 are considered low confidence and may be disordered or incorrectly predicted.

For protein complexes, the pLDDT scores at the interface regions are particularly important. If the interface residues have high pLDDT scores, the predicted interaction is likely to be reliable. If the interface residues have low pLDDT scores, the interaction may be an artifact or may require experimental validation.

### PAE Scores

The predicted aligned error (PAE) score measures the confidence in the relative position of two residues. PAE scores are presented as a matrix, where each cell represents the predicted error in the position of residue i relative to residue j. Low PAE values indicate high confidence in the relative positioning, while high PAE values indicate uncertainty.

For protein complexes, the PAE matrix is essential for interpreting the relative orientation of chains. If the PAE values between two chains are low, the relative orientation is well-defined. If the PAE values are high, the orientation is uncertain, and the predicted interface may not be biologically relevant.

### Using pLDDT and PAE Together

The two confidence scores provide complementary information. pLDDT indicates the confidence in the local structure, while PAE indicates the confidence in the relative positioning of residues. A region with high pLDDT but high PAE may be well-structured locally but poorly positioned relative to other chains. A region with low pLDDT but low PAE may be flexible but consistently positioned.

The [SymProFold study](https://doi.org/10.1038/s41467-024-52138-3) used AlphaFold-Multimer confidence scores to validate symmetrical assemblies, confirming that high-confidence interfaces correspond to experimentally determined structures. This validation approach demonstrates the importance of using confidence scores to filter predictions before investing in experimental follow-up.

## Handling Large Complexes

### Direct Prediction Limits

Direct prediction of large complexes with more than six chains often fails due to memory constraints or produces low-confidence interfaces. The memory requirement grows with the total number of residues, and the attention mechanism in the model scales quadratically with sequence length. For complexes exceeding 2,000 total residues, direct prediction may be impractical on standard hardware.

### Subcomplex Prediction and Assembly

The [AFM-RL method](https://doi.org/10.1101/2024.01.20.576386) provides a practical alternative for large complexes. The approach predicts pairwise models for all chain combinations and then assembles the pairwise models using reinforcement learning. This strategy reduces the memory burden because each prediction involves only two chains, and the assembly step combines the pairwise information.

The AFM-RL approach was demonstrated for complexes with six to twenty chains, showing improved modeling performance compared to existing methods. The pairwise prediction strategy also provides a diverse set of models, which aids the assembly step by offering multiple possible orientations for each chain pair.

### Symmetrical Complexes

For complexes with symmetry, the [SymProFold pipeline](https://doi.org/10.1038/s41467-024-52138-3) offers a specialized approach. This pipeline tests known symmetry operations (p1, p2, p3, p4, and p6) and identifies the most likely symmetry for the assembly. The approach was validated for S-layer arrays and viral capsids, where the symmetry constraints improve the reliability of the assembly.

The SymProFold approach is particularly useful for complexes where the individual chains have limited evolutionary signal. The symmetry constraints compensate for the weaker MSA by enforcing consistent interfaces across the assembly.

## Common Failure Patterns

### Chain Misassignment

A frequent failure occurs when the wrong isoform or species variant is used for one or more chains. This error is often silent because the prediction completes successfully, but the resulting structure does not represent the biological complex of interest. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide tools for verifying sequence identity, including BLAST searches and annotation reviews.

To prevent chain misassignment, verify each sequence against the source database before running the prediction. Check the sequence length, the annotation, and the species. If the complex involves proteins from multiple species, confirm that each sequence is from the correct organism.

### Shallow MSA Leading to Low Confidence

Proteins with few homologs produce shallow MSAs, which often lead to low-confidence predictions. The low confidence may appear as low pLDDT scores across the entire chain or as high PAE values between chains. This failure pattern is common for rapidly evolving proteins, such as viral proteins or immune receptors.

When you encounter low confidence due to shallow MSAs, consider whether the prediction is still useful. If the overall fold is conserved but the interface is uncertain, the prediction may be useful for hypothesis generation but not for detailed mechanistic interpretation.

### Interface Overprediction

AlphaFold-Multimer sometimes predicts interfaces that are not biologically relevant. This overprediction can occur when the model identifies a stable packing arrangement that does not correspond to a real interaction. The confidence scores can help identify overpredicted interfaces: interfaces with low pLDDT or high PAE are more likely to be artifacts.

The [ProEnd database study](https://doi.org/10.1186/s12864-024-10864-4) illustrates the importance of experimental validation for predicted interactions. The study identified HbYX motif-containing proteins using a bioinformatics pipeline and then validated two novel proteins experimentally by pulldown assays, confirming that they directly interact with the proteasome. This validation step is essential for distinguishing genuine interactions from prediction artifacts.

### Memory Exhaustion

Large complexes can exhaust GPU memory, causing the prediction to fail. The failure may occur during MSA generation or during model inference. If you encounter memory exhaustion, reduce the complex size by splitting it into subcomplexes or use the pairwise prediction strategy described in the AFM-RL approach.

## Practical Workflow for a Typical Complex

### Step 1: Retrieve and Verify Sequences

Retrieve the amino acid sequences for all chains in the complex from [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). Verify that each sequence matches the expected isoform and species. Record the accession numbers and sequence lengths in your laboratory notebook or electronic lab notebook.

### Step 2: Prepare the FASTA File

Create a single FASTA file containing all chains in the order you want them labeled. Repeat sequences for homomeric complexes according to the stoichiometry. Verify that the file contains the correct number of sequences and that each sequence is complete.

### Step 3: Generate and Inspect the MSA

Run the MSA generation step and inspect the output. Check the number of aligned sequences and the coverage for each chain. If any chain has a shallow MSA, consider adjusting the search parameters or documenting the limitation.

### Step 4: Run the Prediction

Run AlphaFold-Multimer with the prepared FASTA file. Monitor the runtime and memory usage. If the prediction fails due to memory exhaustion, split the complex into subcomplexes or use the pairwise prediction strategy.

### Step 5: Interpret the Confidence Scores

Examine the pLDDT and PAE scores for the predicted structure. Identify high-confidence regions and low-confidence regions. Pay particular attention to the interface residues and the relative orientation of chains.

### Step 6: Document the Prediction

Record the model version, parameters, MSA statistics, and confidence scores. This documentation is essential for reproducibility and for interpreting the prediction in the context of subsequent experiments.

## Records and Measurements

### What to Record

For each AlphaFold-Multimer prediction, record the following information: the FASTA file used, the source database and accession numbers for each sequence, the MSA generation parameters, the model version and parameters, the recycling setting, the runtime and hardware used, and the confidence scores for the final prediction.

The [nf-core Documentation](https://nf-co.re/docs) emphasizes the importance of standardized pipeline configuration for reproducibility. Using a versioned pipeline configuration ensures that your predictions can be reproduced by collaborators or in subsequent studies.

### Quality Metrics

The primary quality metrics are the pLDDT scores and the PAE matrix. For each chain, record the mean pLDDT and the fraction of residues with pLDDT above 70. For each interface, record the mean PAE between the two chains and the fraction of interface residues with low PAE.

These metrics provide a quantitative basis for comparing predictions across different complexes or different parameter settings. They also provide a basis for deciding whether a prediction is reliable enough for experimental follow-up.

## Limitations and Interpretation Boundaries

### Prediction Accuracy Limits

AlphaFold-Multimer predictions are computational models, not experimental structures. The accuracy varies depending on the complex, the MSA depth, and the conformational state. Predictions for well-studied complexes with deep MSAs are generally reliable, while predictions for novel complexes with shallow MSAs may contain substantial errors.

The [SymProFold study](https://doi.org/10.1038/s41467-024-52138-3) validated predictions against experimental data, confirming that high-confidence predictions can match experimentally determined structures. However, the study also noted that experimental structural characterization of S-layers is challenging due to their self-assembly properties and high sequence variability, highlighting the need for experimental validation even when predictions are confident.

### Conformational Diversity

AlphaFold-Multimer predicts a single conformation for each complex. Proteins that undergo conformational changes, such as allosteric transitions or induced-fit binding, may not be accurately represented by a single prediction. The predicted structure may represent one stable state, but it may not capture the full range of biologically relevant conformations.

### Post-Translational Modifications

AlphaFold-Multimer does not account for post-translational modifications such as phosphorylation, glycosylation, or ubiquitination. These modifications can affect protein structure and interactions, and their absence in the prediction may lead to incorrect conclusions about the complex.

### Experimental Validation Requirements

Predictions from AlphaFold-Multimer should be validated experimentally before drawing strong conclusions. The [ProEnd database study](https://doi.org/10.1186/s12864-024-10864-4) demonstrated this principle by validating two novel HbYX proteins experimentally, confirming that they directly interact with the proteasome. The experimental validation confirmed the computational predictions and provided additional functional information.

## Professional Escalation Criteria

### When to Seek Expert Assistance

If you encounter persistent failures in AlphaFold-Multimer predictions, consider seeking assistance from a structural bioinformatics specialist. Specific situations that warrant escalation include: repeated memory exhaustion for complexes that should be feasible, consistently low confidence scores across multiple parameter settings, and predictions that contradict strong experimental evidence.

### When to Use Alternative Methods

If AlphaFold-Multimer predictions are consistently unreliable for your complex, consider alternative approaches. Molecular docking methods can provide complementary information, particularly for complexes where the individual structures are known. The [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) combines AlphaFold-Multimer with reinforcement learning for large complexes, and other docking methods may be appropriate for specific cases.

### When to Consult the Literature

Before investing substantial time in computational prediction, consult the literature for existing structural information about your complex. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides guidance on finding and using structural data resources, which can help you identify existing structures or experimental data that inform your prediction strategy.

## Safety and Reproducibility Context

### Computational Reproducibility

Reproducibility in computational biology requires careful documentation of all parameters and inputs. The [nf-core Documentation](https://nf-co.re/docs) provides standards for reproducible pipelines, including version control, containerization, and parameter documentation. Following these standards ensures that your predictions can be reproduced and verified.

### Data Management

Protein sequence data and structure predictions should be managed according to your institution's data management policies. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide guidance on data submission and access, and the [Bioconductor Project](https://bioconductor.org/) offers tools for reproducible genomic analysis that can be applied to structure prediction workflows.

### Training and Skill Development

Researchers who are new to structural bioinformatics should invest in foundational training. The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational computing, data, shell, Git, and programming training that is essential for running computational pipelines effectively. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers more advanced training in bioinformatics data resources and analysis methods.

## Decision Framework for Choosing Between Full-Complex Prediction and Subcomplex Assembly

Selecting the correct prediction strategy is the most consequential decision in an AlphaFold-Multimer workflow. Running the full complex when subcomplex assembly is warranted wastes computational resources and often produces misleading confidence scores. Splitting into subcomplexes when full prediction is feasible introduces unnecessary assembly errors. This section provides a structured decision framework based on measurable criteria that you can apply before launching any prediction run.

### Primary Decision Criteria

The first criterion is total residue count across all chains in the complex. Count every residue in every chain, including tags and linkers that may be present in your construct. Complexes up to 1,000 total residues are generally suitable for direct full-complex prediction on hardware with 16 GB or more of GPU memory. Complexes between 1,000 and 2,000 residues require careful assessment of available memory and may need reduced recycling settings. Complexes above 2,000 residues should default to subcomplex prediction unless you have access to high-memory hardware and have verified that the MSA depth supports full prediction.

The second criterion is MSA depth for each chain. Examine the number of aligned sequences and the coverage for every chain before deciding on the prediction strategy. A complex where all chains have more than 1,000 aligned homologs with coverage above 90 percent is a strong candidate for full prediction even at moderate size. A complex where any chain has fewer than 100 aligned homologs should trigger subcomplex prediction regardless of total residue count, because the low-confidence regions in the shallow-MSA chain will propagate uncertainty across the entire complex.

The third criterion is the availability of prior structural information. If experimental structures exist for individual chains or subcomplexes, you can use those structures to guide the assembly strategy. The [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) demonstrates that pairwise models generated by AlphaFold-Multimer can be assembled into larger complexes using reinforcement learning, and this strategy becomes more reliable when individual chain structures are known.

### Decision Matrix for Common Scenarios

| Complex Scenario | Total Residues | MSA Depth | Recommended Strategy | Expected Outcome |
| --- | --- | --- | --- | --- |
| Homodimer of well-conserved enzyme | 600 | Deep for both chains | Full complex prediction | High-confidence interface with pLDDT above 80 at interface |
| Heterotrimer with one orphan protein | 900 | Deep for two chains, shallow for one | Full prediction with careful confidence interpretation | Moderate confidence overall, low confidence in orphan chain regions |
| Six-chain complex of conserved proteins | 1,800 | Deep for all chains | Full prediction on high-memory GPU or subcomplex assembly | Full prediction may succeed but verify memory usage |
| Ten-chain complex with mixed MSA depth | 3,000 | Variable across chains | Subcomplex prediction with pairwise assembly | Improved memory usage, assembly errors possible |
| Symmetrical assembly with 12 identical chains | 2,400 | Deep for the single unique chain | SymProFold pipeline for symmetry detection | Symmetry constraints improve assembly reliability |
| Large complex with 20 chains | 5,000 | Variable | AFM-RL pairwise prediction and assembly | Memory feasible, assembly quality depends on pairwise model diversity |

### Step-by-Step Decision Procedure

Follow this procedure for every new complex before running any prediction.

Step 1: Retrieve all sequences and verify their identity against the source database. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide official access to sequence databases and search systems that allow you to confirm the correct isoform and species for each chain. Record the accession numbers and sequence lengths.

Step 2: Calculate the total residue count by summing the length of every unique chain multiplied by its copy number in the complex stoichiometry. Write this number in your records before proceeding.

Step 3: Generate the MSA for each chain and record the number of aligned sequences and the coverage percentage. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible tutorials for running MSA generation tools and understanding the output formats.

Step 4: Apply the decision matrix above. If the total residue count is below 1,000 and all chains have adequate MSA depth, proceed with full-complex prediction. If the total residue count is above 2,000 or any chain has shallow MSA, proceed with subcomplex prediction.

Step 5: For subcomplex prediction, define the subcomplex boundaries based on biological knowledge of the interaction network. Predict each subcomplex separately, then assemble the subcomplex models using the confidence scores to guide the relative orientation.

Step 6: Document the decision and the rationale in your records. This documentation is essential for interpreting the final prediction and for reproducing the workflow.

### Subcomplex Boundary Selection

When you decide to split a complex into subcomplexes, the choice of boundaries affects the quality of the final assembly. The optimal boundaries separate the complex at regions of low predicted interface confidence, which you can estimate from prior knowledge or from preliminary pairwise predictions.

For complexes with known interaction domains, split at the boundaries between functional modules. For example, if the complex contains a catalytic domain and a regulatory domain that are connected by a flexible linker, predict the catalytic core and the regulatory module separately. The flexible linker will produce low-confidence predictions in full-complex runs, and splitting at this region avoids propagating that uncertainty.

For complexes without prior structural information, run pairwise predictions for all chain combinations first. The [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) uses this strategy to generate a diverse set of pairwise models that inform the assembly. Examine the PAE matrices from the pairwise predictions to identify which chain pairs have well-defined relative orientations. Group chains that have low PAE values between them into the same subcomplex, and place boundaries between chain pairs that show high PAE values.

### Handling Symmetrical Complexes

Symmetry introduces a special case in the decision framework. For complexes where identical chains assemble into symmetrical structures, the [SymProFold pipeline](https://doi.org/10.1038/s41467-024-52138-3) provides a dedicated approach that tests known symmetry operations including p1, p2, p3, p4, and p6. This pipeline uses the high accuracy of AlphaFold-Multimer predictions to derive symmetrical assemblies from protein sequences, and it was validated for S-layer arrays and spherical viral capsids.

When you suspect symmetry in your complex, apply the SymProFold approach instead of attempting full-complex prediction. The symmetry constraints reduce the effective number of independent interfaces that need to be predicted, which compensates for weaker evolutionary signal in individual chains. The pipeline identifies the most likely symmetry operation and produces an assembly that satisfies the symmetry constraints.

The decision to use SymProFold should be made when the complex shows evidence of symmetry, such as identical chains arranged in a repeating pattern or experimental data suggesting a symmetrical assembly. The [SymProFold study](https://doi.org/10.1038/s41467-024-52138-3) confirmed that the predicted models matched available experimental data at the cellular level, and additional crystal structures validated the symmetry and interfaces of several assemblies.

### Resource Allocation Based on Decision

The decision framework also guides resource allocation. Full-complex prediction for a 1,500-residue complex may require 16 GB of GPU memory and several hours of runtime. Subcomplex prediction with pairwise assembly may require less memory per run but more total runs, and the assembly step adds computational overhead.

The [nf-core Documentation](https://nf-co.re/docs) provides guidance on configuring computational pipelines for reproducibility, including resource allocation and containerization. Use this guidance to estimate the memory and runtime requirements for each strategy before committing to a run. If your available hardware cannot support the full-complex prediction, switch to subcomplex prediction instead of attempting to reduce the model quality by lowering recycling settings.

### Common Decision Errors

The most common error in strategy selection is attempting full-complex prediction for a complex that exceeds the memory capacity of the available hardware. This error manifests as a runtime failure or as a prediction that completes but produces low-confidence interfaces due to reduced model quality. The second most common error is splitting a complex into subcomplexes when full prediction would have succeeded, which introduces unnecessary assembly uncertainty.

A third error is ignoring MSA depth when making the decision. A complex with 800 total residues but one chain with fewer than 50 aligned homologs will produce unreliable predictions regardless of the favorable size. The shallow-MSA chain will have low pLDDT scores, and the interfaces involving that chain will have high PAE values. This pattern indicates that the prediction should be interpreted with caution or that additional sequence search should be performed to improve the MSA.

### Recording the Decision

Record the following information for every prediction strategy decision: the total residue count, the MSA depth for each chain, the available hardware specifications, the chosen strategy, and the rationale for the choice. This record allows you to compare prediction quality across different complexes and to identify patterns in which strategies succeed or fail.

The [Bioconductor Project](https://bioconductor.org/) offers tools for reproducible genomic analysis that can be applied to structure prediction workflows, including data management and documentation practices. The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational computing and data management training that supports consistent record keeping.

### Escalation Criteria for Strategy Decisions

If you are uncertain about the appropriate strategy for a specific complex, consult a structural bioinformatics specialist before running the prediction. Specific situations that warrant escalation include complexes with more than 20 chains, complexes where the stoichiometry is uncertain, and complexes where prior experimental data contradicts the predicted assembly.

The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers structured learning pathways covering structural bioinformatics resources and analysis methods. These materials can help you build the expertise needed to make strategy decisions independently, and they provide a reference for consulting the literature on specific complex types.

### Validation of the Decision Framework

The decision framework described here is based on the practical constraints of AlphaFold-Multimer and the demonstrated capabilities of alternative approaches. The [AFM-RL study](https://doi.org/10.1101/2024.01.20.576386) showed that pairwise prediction followed by reinforcement learning assembly improves modeling performance for large complexes with six to twenty chains compared to existing methods. The [SymProFold study](https://doi.org/10.1038/s41467-024-52138-3) validated symmetry-based assembly against experimental data, confirming that the approach produces reliable models for symmetrical assemblies.

These studies provide evidence that the strategy decision matters for prediction quality. Applying the decision framework consistently will help you avoid the common failure patterns of memory exhaustion, low-confidence interfaces, and assembly errors that arise from mismatched strategy selection.

## Frequently Asked Questions

### What is the minimum MSA depth needed for reliable AlphaFold-Multimer predictions?

There is no universal minimum MSA depth that guarantees reliable predictions. The required depth depends on the protein family and the complexity of the structure. Proteins with deep MSAs, such as conserved transcription factors, generally produce more reliable predictions than proteins with shallow MSAs. The [evolutionary analysis of transcription elongation factors](https://doi.org/10.1371/journal.pbio.3003855) demonstrates that conserved proteins across the tree of life retain similar domain architectures, providing strong evolutionary signal for structure prediction. For proteins with shallow MSAs, you should interpret confidence scores carefully and consider experimental validation.

### How do I choose between predicting the full complex and predicting subcomplexes?

The choice depends on the total number of residues and the available hardware. Direct prediction of the full complex is appropriate when the total residue count is within your hardware limits and the MSA is sufficiently deep. For large complexes with six to twenty chains, the [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) uses pairwise prediction followed by assembly, which reduces memory requirements and improves modeling performance. For complexes with symmetry, the [SymProFold pipeline](https://doi.org/10.1038/s41467-024-52138-3) offers a specialized approach that tests symmetry operations and identifies the most likely assembly.

### What does a low pLDDT score at the interface mean?

A low pLDDT score at the interface indicates low confidence in the local structure prediction for those residues. This may mean that the interface is flexible, disordered, or incorrectly predicted. Low pLDDT scores at the interface should be interpreted as a warning that the predicted interaction may not be reliable. You should examine the PAE scores between the two chains to determine whether the relative orientation is also uncertain.

### How do I interpret the PAE matrix for a protein complex?

The PAE matrix shows the predicted error in the relative position of each residue pair. Low PAE values between two chains indicate high confidence in their relative orientation, while high PAE values indicate uncertainty. For a reliable complex prediction, the PAE values between interface residues should be low. If the PAE values are high across the interface, the relative orientation of the chains is not well-defined.

### Can AlphaFold-Multimer predict complexes with post-translational modifications?

No, AlphaFold-Multimer does not account for post-translational modifications. The prediction is based solely on the amino acid sequence, and modifications such as phosphorylation, glycosylation, or ubiquitination are not included. If post-translational modifications are important for your complex, you should interpret the prediction with caution and consider experimental validation.

### What should I do if my prediction has low confidence scores?

If your prediction has low confidence scores, first check the MSA depth and diversity. A shallow MSA may be the cause of low confidence. Consider adjusting the MSA search parameters to include more distant homologs. If the MSA is adequate but the confidence remains low, the complex may be flexible or the interaction may be weak. In this case, consider alternative methods such as molecular docking or experimental validation.

### How do I validate an AlphaFold-Multimer prediction experimentally?

Experimental validation methods include co-immunoprecipitation, pulldown assays, cross-linking mass spectrometry, and structural methods such as cryo-electron microscopy or X-ray crystallography. The [ProEnd database study](https://doi.org/10.1186/s12864-024-10864-4) used pulldown assays to validate two novel HbYX proteins, confirming that they directly interact with the proteasome. The choice of validation method depends on the specific complex and the available experimental resources.

### What are the hardware requirements for running AlphaFold-Multimer on a large complex?

The hardware requirements depend on the total number of residues in the complex. Larger complexes require more GPU memory and longer runtime. For complexes exceeding 2,000 total residues, direct prediction may be impractical on standard hardware. The [AFM-RL approach](https://doi.org/10.1101/2024.01.20.576386) reduces the memory burden by predicting pairwise models and assembling them, which allows prediction of complexes with six to twenty chains on more modest hardware.

## Related Bioinformatics Guides

- [How To Use Alphafold To Predict Structure: Structural Analysis and Computational Methodologies in Bioinformatics](/knowledge/bioinformatics/how-to-use-alphafold-to-predict-structure)
- [Gene Set Enrichment Analysis in R: A Practical Tutorial for Interpreting Omics Data](/knowledge/bioinformatics/gene-set-enrichment-analysis-in-r-a-practical-tutorial-for-interpreting-omics-data)
- [Protein Language Models in Bioinformatics: A Practical Guide to Selection and Application](/knowledge/bioinformatics/protein-language-models-in-bioinformatics-a-practical-guide-to-selection-and-application)
- [Metagenomic Assembly and Binning: A Practical Workflow for Recovering Genomes from Complex Microbial Communities](/knowledge/bioinformatics/metagenomic-assembly-and-binning-a-practical-workflow-for-recovering-genomes-from-complex-microb)
- [AlphaFold and Beyond: Deep Learning for Protein Structure Prediction in Veterinary Virology](/knowledge/bioinformatics/alphafold-deep-learning-protein-structure-prediction-veterinary-virology)

## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [Alishewanella Phage LSH1 from the Sea Surface Microlayer Provides a Novel Minimalistic View of the Siphoviral Hub Structure](https://europepmc.org/article/PMC/PMC13243795)
- [Evolutionary analysis of transcription elongation factors reveals conserved and lineage-specific regulatory domains.](https://doi.org/10.1371/journal.pbio.3003855). 2026.
- [SymProFold: Structural prediction of symmetrical biological assemblies.](https://doi.org/10.1038/s41467-024-52138-3). 2024.
- [ProEnd: a comprehensive database for identifying HbYX motif-containing proteins across the tree of life.](https://doi.org/10.1186/s12864-024-10864-4). 2024.
- [AFM-RL: Large Protein Complex Docking Using AlphaFold-Multimer and Reinforcement Learning](https://doi.org/10.1101/2024.01.20.576386). bioRxiv, 2024.

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