# A Beginner's Guide to Protein-Protein Docking: How to Generate and Rank Complex Models

Protein-protein docking predicts the three-dimensional structure of a protein complex from the free, unbound structures of its component proteins. This article provides a working pipeline for newcomers: preparing input structures, performing global sampling with established docking servers, refining candidate models, and ranking them with scoring functions and experimental restraints. The workflow emphasizes reproducible records, common failure patterns, and clear criteria for when to escalate to experimental validation.

## Scope and Reader Context

This article addresses biology students, researchers, laboratory professionals, and life-science practitioners who need a practical protocol for protein-protein docking. The focus is on computational prediction of complex structures from unbound components, a problem distinct from small-molecule docking or de novo structure prediction. The intended outcome is a ranked list of candidate complex models with documented confidence levels and limitations.

The pipeline described here uses publicly available tools and follows the logic of community-wide assessment experiments such as CAPRI, where successful algorithms combine physical principles with experimental or bioinformatics-derived information. Information-driven approaches such as HADDOCK exemplify this strategy, using known or predicted interface residues to guide sampling. For antibody-antigen complexes, a two-step protocol using ClusPro for global docking followed by SnugDock for local refinement has been demonstrated as an effective computational strategy.

## At a Glance: Docking Pipeline Overview

| Pipeline Stage | Primary Tools | Key Inputs | Main Output | Common Pitfall |
| --- | --- | --- | --- | --- |
| Structure preparation | NCBI databases, structural repositories | Unbound component structures, sequence identifiers | Cleaned PDB files with correct protonation and residue numbering | Missing residues or alternate conformations left unaddressed |
| Global sampling | ClusPro, HADDOCK web server | Prepared component structures, optional restraints | 10 to 1000 candidate complex poses | Insufficient sampling for highly flexible interfaces |
| Refinement | SnugDock, Rosetta refinement protocols | Top-ranked global poses, interface information | Locally optimized complex models | Over-refinement that moves models away from experimental evidence |
| Scoring and ranking | Rosetta scoring functions, HADDOCK scores, cluster analysis | Refined models, experimental restraints | Ranked model list with confidence metrics | Over-reliance on a single scoring function without cross-validation |

## Core Principles of Protein-Protein Docking

### The Docking Problem Defined

Protein-protein docking aims to predict the three-dimensional structure of a protein complex starting from the free forms of the individual partners. This definition distinguishes docking from complex structure prediction methods that use homologous complex templates or deep learning approaches trained on known complex structures. The docking problem is computationally challenging because the search space includes both translational and rotational degrees of freedom for each partner, plus conformational changes that occur upon binding.

Community-wide experiments such as CAPRI have established that the most successful docking algorithms combine physical principles with information from experimental or bioinformatics sources. Pure physics-based sampling alone rarely suffices for accurate predictions, particularly for complexes involving flexible loop regions or large conformational changes upon binding.

### Information-Driven Approaches

Information-driven docking methods such as HADDOCK use experimental data to guide the sampling process. Usable information includes chemical shift perturbations from NMR titrations, mutagenesis data identifying interface residues, cross-linking mass spectrometry distance restraints, and bioinformatics predictions of interface residues. The key principle is that ambiguous interaction restraints derived from such data bias the sampling toward regions of the search space consistent with experimental observations.

For antibody-antigen complexes, a practical protocol involves two-step docking: first, global docking with ClusPro to identify the binding pose when binding information is unknown, then local docking with SnugDock to refine the predicted pose. This approach acknowledges that global sampling methods may identify approximately correct binding regions, but local refinement is needed to optimize side-chain conformations and backbone adjustments at the interface.

### Scoring and Ranking Fundamentals

Scoring functions evaluate candidate complex models based on physical and statistical terms. Common terms include van der Waals interactions, electrostatics, hydrogen bonding, solvation energy, and statistical potentials derived from known complex structures. The challenge is that scoring functions are approximate, and the true complex structure is not always the lowest-energy model according to a given function.

Cluster analysis provides a practical solution: models that cluster together in structural space are more likely to represent a genuine binding mode than isolated low-energy outliers. Most docking servers implement clustering as a standard post-processing step, grouping similar poses and ranking clusters by the average score of their members.

## Preparing Input Structures

### Obtaining Component Structures

The first step is to obtain high-quality structures of the unbound components. The National Center for Biotechnology Information provides access to sequence databases and structure resources that support this process. When an experimentally determined structure exists for your protein of interest, download it from the appropriate structural database using its accession identifier.

When no experimental structure is available, computational structure prediction becomes necessary. Deep learning methods such as AlphaFold2 have become standard tools for generating component structures, and tutorials exist for integrating such predictions with docking workflows. The quality of the input structures directly limits the quality of the docking prediction, so document the source and confidence of each component structure.

### Structure Cleaning and Validation

Raw structure files often contain issues that interfere with docking: missing side-chain atoms, alternate conformations, water molecules, ligands, or multiple chains in the same file. Prepare each component as a separate file containing only the relevant chain or chains. Remove water molecules unless they are known to mediate the interface. Decide whether to retain ligands, cofactors, or metal ions based on whether they are expected to participate in the interaction.

Check the protonation state of titratable residues at your working pH. Many docking servers assume standard protonation states, which may be incorrect for catalytic residues or residues in unusual electrostatic environments. If your system has known pH sensitivity, consider whether the default protonation is appropriate or whether you need to adjust specific residues.

### Handling Missing Residues and Flexibility

Missing residues in crystal structures often correspond to flexible loops or termini. For docking, you have three options: leave the missing region absent, model it with loop-building tools, or use a structure prediction method to generate a complete model. The choice depends on whether the missing region is expected to participate in the interface.

If the missing region is at the interface, leaving it absent will produce incorrect models. If it is distant from the interface, its absence may be acceptable for initial docking, but document this limitation in your records. For highly flexible proteins, consider whether ensemble docking approaches are appropriate, where multiple conformations of each partner are used as input.

## Global Search and Sampling Strategies

### Rigid-Body Docking with ClusPro

ClusPro performs global rigid-body docking using fast Fourier transform correlation methods. It samples billions of possible orientations and translations, then clusters the top-scoring models and ranks clusters by their average score. ClusPro is appropriate as a first-pass global search tool, particularly when no experimental information about the interface is available.

For antibody-antigen complexes, ClusPro serves as the first step in a two-step protocol, generating a set of candidate binding poses that are then refined locally. The server requires only the two component structures as input and returns a ranked list of clusters with downloadable models.

### Information-Driven Sampling with HADDOCK

HADDOCK uses ambiguous interaction restraints derived from experimental or predicted interface information to drive docking. The protocol involves three stages: rigid-body docking with restraint-guided sampling, semi-flexible refinement of the interface, and explicit solvent refinement. The user must supply active and passive residues, which define the expected interface region.

The quality of HADDOCK results depends heavily on the quality of the restraint information. Active residues are those known to be at the interface, while passive residues are surface neighbors that may contact the partner. When experimental data are unavailable, bioinformatics predictions of interface residues can substitute, but with lower confidence.

### Sampling Limitations and Conformational Change

Rigid-body docking assumes that the component structures do not change upon binding. This assumption fails for complexes with induced-fit binding, where conformational changes occur at the interface. HADDOCK addresses this through its semi-flexible refinement stage, which allows backbone and side-chain movements in the interface region. Rosetta-based docking protocols similarly include refinement stages that permit conformational adjustment.

For large conformational changes, such as domain rearrangements or disorder-to-order transitions, standard docking protocols may fail entirely. In such cases, consider whether the docking problem is well-posed or whether you need to model the conformational change separately before docking.

## Refinement of Candidate Models

### Local Refinement with SnugDock

SnugDock performs local docking refinement, optimizing the orientation of the components and the conformations of interface side chains. It is designed for antibody-antigen complexes but can be applied to other protein-protein interactions. The input is a starting complex model from global docking, and the output is a refined model with optimized interface geometry.

The IsAb protocol demonstrates the utility of this approach: ClusPro generates the global binding pose, then SnugDock refines it locally. This two-step strategy balances the need for broad sampling with the need for atomic-level accuracy at the interface.

### Rosetta Refinement Protocols

The Rosetta software suite provides multiple refinement protocols applicable to docking. The RosettaDock protocol performs Monte Carlo sampling of rigid-body orientations and side-chain conformations, starting from an approximate complex model. The Rosetta scoring function evaluates the refined models using physical and statistical terms.

Rosetta has been used to predict the structure of protein-protein complexes and to design altered specificity protein-protein interactions. The suite also supports integration of sparse experimental data, such as NMR or EPR restraints, into the modeling process. For beginners, the Rosetta web server provides a more accessible interface than the full command-line suite.

### Solvent and Energy Minimization

Final refinement typically includes energy minimization in the presence of explicit or implicit solvent. This step relieves steric clashes introduced during sampling and optimizes electrostatic and hydrogen-bonding interactions at the interface. HADDOCK includes an explicit solvent refinement stage as the final step of its protocol.

Energy minimization alone cannot fix fundamentally incorrect docking poses. If the global sampling stage placed the components in the wrong relative orientation, refinement will only optimize that incorrect pose. The purpose of refinement is to improve the geometry of approximately correct models, not to correct sampling failures.

## Scoring Functions and Model Ranking

### Physics-Based and Statistical Scoring

Rosetta scoring functions combine physics-based terms, such as Lennard-Jones interactions and electrostatics, with statistical terms derived from known protein structures. The total score of a model reflects its estimated free energy, with lower scores indicating more favorable interactions. However, absolute scores are not directly comparable across different protein systems.

HADDOCK scores combine the restraint energy with interaction energies, including van der Waals and electrostatic terms. The HADDOCK score is used to rank clusters of models, with the cluster having the lowest average score ranked first.

### Cluster Analysis for Ranking

Cluster analysis groups similar models based on pairwise root-mean-square deviation of interface residues. The underlying assumption is that the correct binding mode will be sampled multiple times and will therefore form a large cluster. Small clusters or isolated models are less reliable, even if their individual scores are favorable.

When evaluating docking results, examine both the cluster size and the average score. A large cluster with a good average score is more reliable than a small cluster with an excellent individual score. Most docking servers provide cluster information as part of their output.

### Cross-Validation with Experimental Data

The most reliable ranking strategy is cross-validation against independent experimental data. If you have cross-linking mass spectrometry data, compare the predicted complex models against the observed distance restraints. If you have mutagenesis data, check whether residues known to be important for binding are located at the predicted interface.

The integration of structural mass spectrometry data with computational protein structure prediction has been demonstrated in tutorial form, including strategies for incorporating covalent labeling, ion mobility, and surface-induced dissociation data into Rosetta-based workflows. These approaches use Rosetta scoring terms and specialized applications to evaluate whether predicted models are consistent with experimental measurements.

## Practical Implementation Steps

### Step 1: Define the Biological Question

Write a clear statement of what you are trying to predict. Are you asking whether two proteins interact, where the interface is located, or how a mutation affects binding? The docking protocol differs depending on the question. For interaction prediction, docking alone is insufficient because docking will always produce some complex model, even for non-interacting proteins.

Document the biological context: the species, the cellular compartment, the known or suspected function of the complex, and any experimental evidence for the interaction. This context guides your choice of input structures and restraints.

### Step 2: Collect and Prepare Input Structures

Obtain the unbound structures of both components. Record the accession identifiers, resolution, and any known issues with each structure. If you are using predicted structures, record the prediction method and confidence metrics.

Prepare separate structure files for each component. Remove water molecules, ligands, and other chains that are not part of the docking problem. Check for missing residues and decide how to handle them. Document all preparation steps in your records.

### Step 3: Identify Available Experimental Information

Search the literature and public databases for experimental information about the interaction. Useful data include mutagenesis results, chemical shift perturbations, cross-linking restraints, and electron microscopy maps. Even qualitative information, such as which domain of one protein binds to which domain of the other, can guide the docking.

If no experimental information is available, use bioinformatics tools to predict interface residues. These predictions are less reliable than experimental data but can still improve docking performance compared to blind sampling.

### Step 4: Perform Global Docking

Run global docking with ClusPro or HADDOCK. For ClusPro, submit the two prepared structures and collect the ranked clusters. For HADDOCK, define active and passive residues based on your experimental or predicted interface information.

Record the parameters used for each run, including any restraint definitions, the number of models generated, and the version of the server or software. This information is essential for reproducibility.

### Step 5: Refine Top Candidates

Select the top-ranked clusters from global docking for refinement. The number of clusters to refine depends on the diversity of the top-ranked solutions. If the top clusters are structurally similar, fewer need refinement. If they are diverse, refine more.

Run local refinement with SnugDock or RosettaDock for each selected cluster. Record the starting model, the refinement protocol, and the final refined model for each candidate.

### Step 6: Score and Rank Refined Models

Apply scoring functions to the refined models. If you have experimental restraints, evaluate each model against those restraints. Combine the scoring results with cluster information to produce a final ranking.

Compare the top-ranked models to identify conserved interface residues and interactions. If multiple top-ranked models share a common interface, that interface is more likely to be correct than one that appears in only a single model.

### Step 7: Document and Report

Record the complete pipeline: input structures, preparation steps, docking parameters, refinement protocols, scoring results, and final ranking. Include the limitations of each stage and the confidence level of the final prediction.

When reporting results, distinguish between predictions that are strongly supported by multiple lines of evidence and those that are speculative. Provide the ranked list of models with their scores and the criteria used for ranking.

## Records and Measurements

### Essential Records for Reproducibility

Maintain a laboratory notebook or electronic record that captures every step of the docking pipeline. For each docking run, record the software version, the input structures, all parameters, and the date. This record allows you or others to reproduce the prediction exactly.

For HADDOCK runs, record the active and passive residue lists, the restraint definitions, and the number of models generated at each stage. For ClusPro runs, record the server version and the clustering parameters. For Rosetta runs, record the score function, the number of Monte Carlo steps, and the random seed.

### Quality Metrics to Track

Track the following metrics for each docking prediction:

| Metric | What It Measures | How to Record | Interpretation |
| --- | --- | --- | --- |
| Interface root-mean-square deviation | Structural similarity between predicted and known complex models | Calculate after superposition of interface residues | Lower values indicate better predictions when a reference structure exists |
| Cluster size | Number of models sharing a common binding mode | Record from server output or cluster analysis | Larger clusters are more reliable |
| HADDOCK score or Rosetta score | Estimated interaction energy | Record from scoring function output | Lower scores indicate more favorable predicted interactions |
| Restraint satisfaction | Fraction of experimental restraints satisfied by the model | Count satisfied restraints divided by total restraints | Higher values indicate better agreement with experimental data |
| Buried surface area | Surface area buried at the predicted interface | Calculate with a surface area tool | Larger buried areas are typical of stable interfaces but are not definitive |

### Benchmarking Against Known Complexes

If you are docking proteins for which a known complex structure exists, use that structure to benchmark your pipeline. Calculate the interface root-mean-square deviation between your top-ranked model and the known complex. This benchmark tells you how well your pipeline performs for your specific protein system.

For novel complexes, no such benchmark is possible. In this case, rely on cross-validation with experimental data and on the consistency of the top-ranked models. If multiple independent docking runs with different parameters produce similar top-ranked models, confidence increases.

## Common Failure Patterns

### Failure Pattern 1: Incorrect Input Structures

The most common cause of docking failure is poor input structures. Structures with missing interface residues, incorrect protonation states, or conformational states incompatible with binding will produce incorrect complex models regardless of the quality of the docking protocol.

Prevention: Validate input structures before docking. Check for missing residues, especially in surface loops. Verify that the structure represents a biologically relevant conformation. If using predicted structures, check the confidence metrics and consider whether the prediction is suitable for docking.

### Failure Pattern 2: Insufficient Sampling

Rigid-body docking may miss the correct binding mode if the interface involves conformational changes. The correct pose may be sampled but with high energy due to steric clashes that would be relieved by conformational adjustment.

Prevention: Use refinement protocols that allow backbone and side-chain flexibility. Consider ensemble docking approaches that use multiple conformations of each partner. If the complex is known to involve large conformational changes, consider whether docking is the appropriate method.

### Failure Pattern 3: Over-Reliance on a Single Scoring Function

Different scoring functions make different errors. A model that ranks first according to one function may rank poorly according to another. Over-reliance on a single function can lead to confident but incorrect predictions.

Prevention: Use multiple scoring functions and compare their rankings. Cross-validate with experimental restraints. Pay attention to cluster size in addition to individual scores. If different scoring functions rank different models first, treat the prediction as low confidence.

### Failure Pattern 4: Misinterpretation of Docking Results

Docking always produces a complex model, even for proteins that do not interact. A low-scoring model does not prove that the proteins bind. Docking predicts the structure of a complex assuming that the complex forms, it does not predict whether the complex forms.

Prevention: Use docking only when there is independent evidence for the interaction. For interaction prediction, use experimental methods or specialized prediction tools. Report docking results as structural predictions conditional on the interaction occurring.

### Failure Pattern 5: Ignoring Experimental Data

Docking without experimental restraints is substantially less accurate than docking with restraints. Ignoring available experimental data, such as mutagenesis results or cross-linking restraints, reduces the reliability of the prediction.

Prevention: Search the literature thoroughly before docking. Use all available experimental information to guide sampling and to validate results. If experimental data conflict with the top-ranked model, investigate the source of the conflict.

## Limitations and Interpretation Boundaries

### Accuracy Limits of Docking Methods

Docking predictions are approximate. Even the best methods produce models with errors at the interface, particularly for side-chain conformations. The accuracy of a docking prediction depends on the quality of the input structures, the availability of experimental restraints, and the conformational flexibility of the system.

Community-wide assessments have shown that docking methods can produce near-native models for many complexes, but the success rate is far from perfect. For any given complex, the top-ranked model may be incorrect. The confidence in a prediction should reflect the convergence of multiple models, the agreement with experimental data, and the consistency of the scoring results.

### When Docking Is Not Appropriate

Docking is not appropriate for predicting whether two proteins interact. It is not appropriate for complexes involving large conformational changes that cannot be modeled by refinement protocols. It is not appropriate when the input structures are of poor quality or when the binding interface is unknown and no experimental data are available.

For intrinsically disordered proteins that fold upon binding, standard docking protocols are inadequate. For membrane protein complexes, the hydrophobic environment and the constraints of the membrane must be considered. For very large complexes with many subunits, multi-body docking is required, which is substantially more complex than pairwise docking.

### The Role of Experimental Validation

Computational docking predictions require experimental validation before they can be used for downstream applications. Validation methods include site-directed mutagenesis of predicted interface residues, cross-linking mass spectrometry, and biophysical binding assays. The choice of validation method depends on the biological question and the available experimental systems.

The integration of computational modeling with structural mass spectrometry provides a particularly powerful validation strategy. By comparing predicted models against experimental measurements such as solvent accessibility, collision cross sections, and energy-resolved mass spectrometry data, researchers can identify models that are consistent with experimental observations.

## Professional Escalation Criteria

### When to Seek Expert Assistance

Consult a structural bioinformatics expert or a computational biology core facility when any of the following conditions apply:

- The docking problem involves more than two components or multi-body assembly
- The components undergo large conformational changes upon binding
- The input structures are of poor quality or contain large disordered regions
- The top-ranked models are inconsistent across different scoring functions
- The prediction will be used for critical applications such as drug design or clinical interpretation

### When to Move to Experimental Methods

Move to experimental structure determination when the computational prediction is not sufficiently reliable for your application. If the docking prediction is ambiguous, with multiple diverse top-ranked models, experimental methods such as cross-linking mass spectrometry, hydrogen-deuterium exchange, or cryo-electron microscopy may be necessary to resolve the ambiguity.

For antibody engineering applications, the IsAb protocol includes experimental validation of computational predictions through bioassays. The computational design of antibody variants is followed by experimental testing to confirm that the designed variants have the predicted binding properties.

### Documentation for Publication

When publishing docking results, report the complete pipeline with sufficient detail for reproduction. Include the software versions, the input structures, the parameters, and the scoring results. Deposit the top-ranked models in a public repository. Report the limitations of the prediction and the confidence level.

The reproducibility standards promoted by community initiatives such as the Galaxy Training Network and nf-core provide useful models for documenting computational workflows. These standards emphasize version control, parameter documentation, and the use of containerized or virtualized environments to ensure that workflows can be reproduced exactly.

## Decision Framework for Selecting a Docking Strategy

### Matching the Docking Method to Your Biological Question

The choice between rigid-body docking, information-driven docking, and refinement-focused protocols depends on what you already know about your system. A structured decision framework prevents wasted compute time and misinterpretation of results. The framework below organizes the decision process around three questions: what information you have, what conformational changes you expect, and what accuracy you need.

### Decision Point 1: Available Experimental Information

Start by inventorying what you know about the interaction. This single factor most strongly determines which docking strategy will succeed. The community-wide CAPRI assessments have consistently shown that methods combining physical principles with experimental or bioinformatics-derived information outperform blind sampling approaches.

**Tier 1: No interface information available.** When you have no experimental data about the binding interface, use ClusPro for global rigid-body docking. ClusPro performs exhaustive sampling using fast Fourier transform correlation methods and clusters the top-scoring models. The server requires only the two component structures as input. This approach is appropriate for initial exploration but produces lower-confidence predictions than information-driven methods.

**Tier 2: Qualitative interface information available.** If you know which domain or region of one protein contacts the other, or if you have mutagenesis data identifying key residues, use HADDOCK with ambiguous interaction restraints. The IsAb protocol for antibody design demonstrates this tier: when binding information is unknown, ClusPro performs global docking first, then SnugDock refines the predicted pose locally. This two-step approach acknowledges that global sampling identifies approximately correct binding regions while local refinement optimizes the interface geometry.

**Tier 3: Quantitative distance restraints available.** When you have cross-linking mass spectrometry data, NMR chemical shift perturbations, or other quantitative distance information, use these restraints to drive the docking directly. HADDOCK accepts ambiguous interaction restraints derived from such data and biases sampling toward regions consistent with experimental observations. The integration of structural mass spectrometry data with computational protein structure prediction has been demonstrated in tutorial form, including strategies for incorporating covalent labeling, ion mobility, and surface-induced dissociation data into Rosetta-based workflows through scoring terms and specialized applications.

### Decision Point 2: Expected Conformational Change

The magnitude of conformational change upon binding determines whether rigid-body docking will suffice or whether you need flexible refinement protocols.

**Small conformational changes, side-chain rearrangements only.** Most protein-protein interactions involve side-chain adjustments at the interface with minimal backbone movement. Standard rigid-body docking followed by side-chain refinement handles this case well. HADDOCK includes a semi-flexible refinement stage that allows side-chain movements in the interface region. RosettaDock performs Monte Carlo sampling of rigid-body orientations and side-chain conformations starting from an approximate complex model.

**Moderate backbone movements in loop regions.** When interface loops are expected to move upon binding, use protocols that permit backbone flexibility in the interface region. HADDOCK's semi-flexible refinement allows backbone movements in addition to side-chain adjustments. Rosetta refinement protocols similarly include stages that permit conformational adjustment. The key limitation is that these methods only sample local conformational space around the starting model, so the global docking stage must place the components in approximately correct relative orientations.

**Large domain rearrangements or disorder-to-order transitions.** Standard docking protocols fail for these systems. If the components undergo large conformational changes upon binding, consider whether the docking problem is well-posed. You may need to model the conformational change separately before docking, using methods such as normal mode analysis to generate alternative conformations of each partner. NRGSuite-Qt provides normal mode analysis functionality through the ENCoM method, which can generate alternative conformations for ensemble docking approaches. If the conformational change is too large, experimental structure determination may be necessary.

### Decision Point 3: Required Output Accuracy

The accuracy you need determines how much refinement and cross-validation to perform.

**Screening-level accuracy, identifying approximate binding regions.** For initial hypothesis generation, the top-ranked clusters from ClusPro or HADDOCK global docking may suffice. The goal is to identify which regions of each protein participate in the interface, not to obtain atomic-level detail. Cluster analysis provides the ranking criterion: large clusters of similar models are more reliable than isolated low-energy outliers.

**Interface-level accuracy, identifying specific contact residues.** For mutagenesis design or understanding binding determinants, you need accurate interface residue identification. This requires refinement of top-ranked global poses and cross-validation with experimental data. The IsAb protocol demonstrates this level: after ClusPro global docking and SnugDock local refinement, in silico alanine scanning predicts potential hotspot residues, which are then validated experimentally.

**Atomic-level accuracy for drug design or engineering.** When the prediction will guide therapeutic antibody engineering or drug design, atomic-level accuracy at the interface is required. This demands extensive refinement, multiple scoring functions, and cross-validation with quantitative experimental data. The Rosetta software suite has been used to accurately design novel protein structures and predict protein-protein complex structures with atomic-level accuracy in benchmark cases. However, such accuracy is not guaranteed for every system, and experimental validation remains essential.

### Decision Matrix for Common Scenarios

| Scenario | Recommended Strategy | Expected Output | Confidence Level |
| --- | --- | --- | --- |
| Two globular proteins, no interface data | ClusPro global docking | Ranked clusters of rigid-body poses | Low to moderate |
| Antibody-antigen, binding site unknown | ClusPro global docking, then SnugDock refinement | Refined antibody-antigen complex models | Moderate |
| Known interface residues from mutagenesis | HADDOCK with ambiguous interaction restraints | Restraint-guided complex models | Moderate to high |
| Cross-linking MS data available | HADDOCK with distance restraints or Rosetta with MS scoring terms | Models consistent with experimental restraints | High |
| Highly flexible proteins, large conformational change | Ensemble docking or separate conformational modeling before docking | Multiple complex models reflecting conformational diversity | Low, may require experimental methods |

### Implementing the Decision Framework

**Step 1: Document your information inventory.** Before running any docking software, write down what you know about the interaction. Include experimental evidence for the interaction itself, any interface information from mutagenesis or biophysical studies, and any structural data about the individual components. This inventory determines which tier of the decision framework applies.

**Step 2: Assess conformational flexibility.** Examine your input structures for flexible regions, disordered loops, and known conformational states. Check whether the unbound structures differ from what you expect in the bound state. If you have multiple structures of the same protein in different conformations, consider whether ensemble docking is appropriate.

**Step 3: Select the docking strategy.** Apply the decision framework to choose between ClusPro, HADDOCK, or a multi-stage protocol. Record the rationale for your choice in your laboratory notebook. This documentation is essential for interpreting results and for reproducing the pipeline.

**Step 4: Define stopping criteria.** Decide in advance how you will judge the docking results. Will you accept the top-ranked cluster, or do you require consistency across multiple scoring functions? Will you cross-validate against experimental restraints? Defining stopping criteria prevents the common failure pattern of over-interpreting a single low-energy model.

**Step 5: Escalate when the framework indicates.** If your system falls into a category where docking is unlikely to succeed, such as large conformational changes or intrinsically disordered regions, escalate to experimental methods or consult a structural bioinformatics expert before investing substantial compute time.

### Common Mistakes in Strategy Selection

**Mistake 1: Using HADDOCK without restraint information.** HADDOCK requires active and passive residue definitions. Running HADDOCK without meaningful restraints is equivalent to blind docking but with additional complexity. If you have no interface information, use ClusPro instead.

**Mistake 2: Expecting rigid-body docking to handle induced fit.** Rigid-body docking assumes the unbound structures do not change upon binding. If your system is known to undergo conformational changes, rigid-body results will be incorrect regardless of the scoring function. Use refinement protocols that permit flexibility or model the conformational change separately.

**Mistake 3: Refining incorrect global poses.** Refinement optimizes the geometry of approximately correct models. If the global docking stage placed the components in the wrong relative orientation, refinement will only optimize that incorrect pose. The decision framework emphasizes getting the global sampling stage right before investing in refinement.

**Mistake 4: Ignoring the difference between interaction prediction and structure prediction.** Docking predicts the structure of a complex assuming the complex forms. It does not predict whether the proteins interact. If your question is whether two proteins bind, docking is the wrong tool. Use experimental binding assays or specialized interaction prediction methods instead.

**Mistake 5: Selecting a strategy based on software familiarity instead of the biological question.** The choice between ClusPro, HADDOCK, RosettaDock, and SnugDock should follow from your information inventory and accuracy requirements, not from which software you have used before. The decision framework provides a systematic basis for this choice.

### Recording Strategy Decisions

Record the following in your laboratory notebook for each docking project:

| Record Item | Example Entry | Purpose |
| --- | --- | --- |
| Information inventory | Mutagenesis data identifies residues 45-52 of protein A as critical for binding | Determines docking tier and restraint definition |
| Conformational assessment | Protein B has a flexible loop (residues 120-135) near the predicted interface | Determines whether flexible refinement is needed |
| Strategy selection rationale | Chose HADDOCK because interface residues are known from mutagenesis | Documents the decision process |
| Stopping criteria | Accept top cluster if it satisfies 80 percent of cross-linking restraints | Prevents over-interpretation |
| Escalation triggers | If top-ranked models are inconsistent across scoring functions, consult expert | Defines when to seek assistance |

The decision framework provides a systematic method for selecting the docking strategy that matches your biological question, available information, and accuracy requirements. Applying this framework before running any docking software reduces wasted compute time, prevents common failure patterns, and produces results that are interpretable and defensible. The framework also supports reproducibility by documenting the rationale for each methodological choice, which is essential for publication and for downstream experimental validation.

## Frequently Asked Questions

### What is the difference between protein-protein docking and small-molecule docking?

Protein-protein docking predicts the structure of a complex formed by two or more proteins, treating each protein as a rigid or semi-flexible body with hundreds or thousands of atoms. Small-molecule docking predicts the binding pose of a small molecule, typically a drug candidate, within a protein binding site. The search spaces and scoring functions differ substantially. Protein-protein docking involves six degrees of freedom for the relative orientation of the partners plus conformational flexibility, while small-molecule docking involves the conformational degrees of freedom of the ligand within a defined binding pocket. Tools such as NRGSuite-Qt provide separate functionality for protein-protein interaction analysis and small-molecule docking.

### Which docking server should I use as a beginner?

For a first docking experiment, start with ClusPro, which requires only the two component structures as input and provides a ranked list of clusters without requiring restraint definitions. After gaining familiarity with the output format and the interpretation of clusters, move to HADDOCK, which requires active and passive residue definitions but provides more accurate results when experimental information is available. The choice between servers depends on the availability of experimental restraints and the complexity of the system.

### How do I choose active and passive residues for HADDOCK?

Active residues are those experimentally shown to be at the interface, such as residues with significant chemical shift perturbations upon binding or residues that abolish binding when mutated. Passive residues are surface neighbors of active residues that may contact the partner. When experimental data are unavailable, use bioinformatics predictions of interface residues, but recognize that predicted restraints are less reliable than experimental ones. The quality of the restraints directly affects the quality of the docking prediction.

### Can I dock proteins using AlphaFold2-predicted structures as input?

Yes, predicted structures can serve as input for docking, but the quality of the prediction limits the docking accuracy. AlphaFold2 predictions are highly accurate for many proteins, but they may be less reliable for flexible regions, multi-domain proteins, or proteins that undergo large conformational changes. Check the per-residue confidence scores and consider whether the predicted structure is suitable for docking. Tutorials exist that demonstrate the integration of AlphaFold2 predictions with docking workflows.

### How many models should I generate in a docking run?

The number of models depends on the sampling strategy and the complexity of the system. ClusPro generates and clusters a large number of models internally, and the user receives the ranked clusters. HADDOCK generates a user-specified number of models at each stage, with typical values ranging from 1000 rigid-body models to 200 refined models. The goal is to sample the search space sufficiently to include the correct binding mode, which requires more models for flexible or large systems.

### How do I know if my docking prediction is correct?

Without experimental validation, you cannot know with certainty that a docking prediction is correct. Indicators of a reliable prediction include a large cluster of similar models, agreement with experimental restraints, consistency across different scoring functions, and a buried interface with favorable interactions. The strongest validation comes from experimental methods such as mutagenesis, cross-linking mass spectrometry, or structure determination. The integration of structural mass spectrometry data with computational modeling provides a systematic approach to validating predicted models against experimental measurements.

### What is the role of clustering in docking result analysis?

Clustering groups similar models based on structural similarity, typically measured by interface root-mean-square deviation. The assumption is that the correct binding mode will be sampled multiple times and will form a large cluster, while incorrect modes will be sampled less frequently. Clustering provides a more reliable ranking criterion than individual model scores because it identifies binding modes that are consistently sampled. Most docking servers implement clustering as a standard post-processing step.

### When should I escalate to experimental structure determination?

Escalate to experimental methods when the docking prediction is not sufficiently reliable for your application, when multiple diverse top-ranked models are produced, or when the prediction will be used for critical decisions such as drug design or clinical interpretation. Experimental methods such as cross-linking mass spectrometry, cryo-electron microscopy, or X-ray crystallography can provide definitive structural information. The choice of method depends on the size of the complex, the availability of purified protein, and the resolution required for your application.

## Related Bioinformatics Guides

- [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)
- [AutoDock Vina Receptor-Ligand Docking: Practical Protocols for Protein-Small Molecule Docking](/knowledge/bioinformatics/autodock-vina-receptor-ligand-docking)
- [How To Use Alphafold To Predict Structure: Structural Analysis and Computational Methodologies in Bioinformatics](/knowledge/bioinformatics/how-to-use-alphafold-to-predict-structure)
- [Structure-Guided Antiviral Design: In Silico Docking and Molecular Dynamics of SARS-CoV-2 Spike Protein Variants](/knowledge/bioinformatics/structure-guided-antiviral-design-sars-cov-2-spike-protein-docking-molecular-dynamics)
- [What Is the Monomer of a Protein? Structure & Synthesis](/knowledge/bioinformatics/protein-monomers-amino-acids-peptide-synthesis)

## 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.
- [Comprehensive Tutorial for Computational Methods of Protein Structure Prediction Incorporating Mass Spectrometry Data.](https://pubmed.ncbi.nlm.nih.gov/42415383). Mass spectrometry reviews, 2026.
- [IsAb: a computational protocol for antibody design.](https://pubmed.ncbi.nlm.nih.gov/33876197). Briefings in bioinformatics, 2021.
- [Practically useful: what the Rosetta protein modeling suite can do for you.](https://pubmed.ncbi.nlm.nih.gov/20235548). Biochemistry, 2010.
- [Information-driven structural modelling of protein-protein interactions.](https://pubmed.ncbi.nlm.nih.gov/25330973). Methods in molecular biology (Clifton, N.J.), 2015.
- [NRGSuite-Qt: a PyMOL plugin for high-throughput virtual screening, molecular docking, normal-mode analysis, the study of molecular interactions, and the detection of binding-site similarities.](https://pubmed.ncbi.nlm.nih.gov/40538444). Bioinformatics advances, 2025.

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