# Inferring Protein Complexes from Proteomics Interaction Networks: A Practical Guide to Computational Methods Proteomics interaction networks derived from mass spectrometry experiments contain rich information about protein complexes, but extracting stable biological assemblies from noisy interaction data requires computational methods that researchers often do not know how to select or apply. This article explains the principles behind protein complex prediction algorithms such as MCL and ClusterONE, describes the data preparation steps needed before network analysis, and provides a practical workflow for applying these methods to proteomics datasets. The intended reader is a biology student, researcher, or laboratory professional who has generated or obtained protein-protein interaction data and needs concrete decisions about which algorithm to use, how to validate results, and how to interpret predictions within experimental limitations. ## At a Glance Protein complex prediction from interaction networks is a computational task that transforms pairwise interaction evidence into multi-protein assembly models. The choice of algorithm depends on network properties, data quality, and the biological question being asked. | Algorithm | Input Requirements | Output Format | Best Use Case | Key Limitation | |-----------|-------------------|---------------|---------------|----------------| | MCL (Markov Clustering) | Weighted or unweighted edge list, inflation parameter | Disjoint clusters | Large networks with clear density differences | Produces only non-overlapping clusters, may split complexes sharing subunits | | ClusterONE | Weighted edge list, minimum cluster size, density threshold | Overlapping clusters with cohesion scores | Networks where proteins participate in multiple complexes | Requires tuning of density parameters, sensitive to noise | | Spectral clustering | Affinity matrix or edge list | Partitioned clusters | Networks with well-separated dense regions | Number of clusters must be specified in advance | | Hierarchical clustering | Distance or similarity matrix | Dendrogram with cut levels | Exploratory analysis of network structure | Cut level selection is subjective, scales poorly to large networks | The workflow described in this article follows a standard sequence: quality control of proteomics identifications, construction of the interaction network, selection of a clustering algorithm, parameter optimization, validation against known complexes, and biological interpretation. Each step involves concrete decisions that affect the final output. ## Understanding Protein Complexes in Proteomics Data Protein complexes are stable assemblies of multiple polypeptide chains that perform cellular functions as a unit. Examples include the proteasome, RNA polymerase, and the spindle assembly checkpoint complex. These assemblies appear in proteomics interaction networks as dense regions where member proteins show many mutual interactions, while proteins outside the complex show fewer connections to the complex members. Mass spectrometry-based proteomics generates interaction data through several experimental approaches. Affinity purification followed by mass spectrometry (AP-MS) identifies proteins that copurify with a bait protein. Cross-linking mass spectrometry provides distance constraints between protein residues. Co-fractionation mass spectrometry separates protein complexes by biochemical properties before identification. Each approach produces different types of evidence with different noise characteristics. The relationship between mRNA transcript abundance and protein abundance is not straightforward. Studies of colon and rectal tumors have shown that messenger RNA transcript abundance does not reliably predict protein abundance differences between tumors, which means interaction networks built from transcriptomic data alone may not reflect the actual protein interaction landscape [8]. This finding supports the use of direct proteomic measurements when constructing interaction networks for complex prediction. Single-cell proteomics technologies are expanding the ability to measure proteins in specific cell types and spatial contexts. Spatial single-cell mass spectrometry has resolved the context-dependent proteome of murine hepatocytes at a depth of approximately 1,700 proteins from a single cell slice, with half of the proteome differentially regulated in a spatial manner [10]. This spatial information matters for complex prediction because a protein complex may exist in one cell type or tissue region but not another. ## Data Inputs for Interaction Network Construction ### Mass Spectrometry Identification Outputs The starting point for interaction network construction is a list of identified proteins from mass spectrometry experiments. These identifications typically come from database search engines that match tandem mass spectra to peptide sequences. The output includes protein identifiers, peptide-spectrum matches, and quantitative values such as spectral counts or intensity measurements. Before network construction, researchers must apply quality filters to the protein identifications. Common filters include minimum peptide count, false discovery rate thresholds, and minimum number of observed spectra. The specific thresholds depend on the experimental design and the stringency required for downstream analysis. NCBI provides access to sequence databases and search systems that support protein identification workflows [1]. ### Interaction Evidence Types Interaction networks can be constructed from different types of evidence: **Copurification data** from AP-MS experiments provides direct evidence that two proteins exist in the same complex. The bait protein pulls down interaction partners, and the presence of both proteins in the same purification indicates a potential interaction. **Co-fractionation data** from size exclusion chromatography or other separation methods shows that proteins elute in the same fractions, suggesting they migrate together as part of a larger assembly. **Cross-linking data** provides spatial proximity information, showing that specific residues on two proteins are close enough to be covalently linked by the cross-linking reagent. **Literature-curated interactions** from databases provide previously reported interactions that can supplement experimental data. Each evidence type has different reliability characteristics. Copurification data can include contaminants that bind non-specifically to the affinity matrix. Co-fractionation data can group proteins that happen to have similar biochemical properties without being in the same complex. Cross-linking data is highly specific but covers only a fraction of the interactions in a complex. ### Quantitative Filtering and Scoring Raw interaction evidence requires scoring to distinguish genuine interactions from noise. Common scoring approaches include: **Spectral count-based scoring** uses the number of observed spectra for each protein as a proxy for abundance. Proteins with higher spectral counts in the purification are more likely to be genuine interaction partners. **Reproducibility filtering** requires that an interaction be observed in multiple biological replicates before it is included in the network. This approach reduces false positives from stochastic sampling. **Negative control subtraction** uses control purifications to identify proteins that bind non-specifically. Proteins appearing in both experimental and control purifications are removed or down-weighted. **Quantitative comparison** uses intensity-based quantification to compare protein abundance between bait purifications and controls. Proteins significantly enriched in the bait purification are considered interaction candidates. The choice of scoring method affects the final network topology and therefore the complex prediction results. Researchers should document the scoring approach and justify the thresholds used. ## Core Principles of Complex Prediction Algorithms ### Network Representation Protein interaction networks are represented as graphs where nodes are proteins and edges are interactions. Each edge can carry a weight representing the confidence or strength of the interaction evidence. The network can be directed if the interaction evidence has an orientation, such as bait-prey relationships in AP-MS data, or undirected if the evidence is symmetric. Network properties that matter for complex prediction include: **Density** is the fraction of possible edges that are actually present. Protein complexes appear as regions of high density relative to the rest of the network. **Connectivity** describes how easily one node can reach another through edges. Complex members are typically well-connected to each other. **Modularity** measures the strength of division of a network into modules. Complexes contribute to high modularity when they form distinct dense regions. **Degree distribution** describes how many edges each node has. Hub proteins that participate in many complexes have high degree and may connect multiple dense regions. ### Clustering Approaches Complex prediction algorithms generally fall into several categories: **Density-based clustering** identifies regions of the network where the edge density exceeds a threshold. These methods work well when complexes are dense and well-separated from the rest of the network. **Flow-based clustering** simulates information flow through the network and identifies regions where flow becomes trapped. The Markov Clustering algorithm (MCL) uses this approach by simulating random walks and strengthening edges within dense regions while weakening edges between regions. **Modularity optimization** seeks partitions of the network that maximize the difference between observed edges and expected edges in a random network. These methods are widely used for community detection but may not capture the specific properties of protein complexes. **Seed-based expansion** starts from seed proteins or seed complexes and expands outward by adding proteins that increase the density or cohesion of the cluster. ClusterONE uses this approach with a cohesion measure that accounts for both internal density and external connections. ### Overlapping versus Non-overlapping Clusters A key decision in complex prediction is whether proteins can belong to multiple complexes. Many proteins participate in more than one complex, either through shared subunits or through dynamic assembly and disassembly. Non-overlapping clustering methods such as MCL assign each protein to exactly one cluster, which can split complexes that share subunits. Overlapping methods such as ClusterONE allow proteins to appear in multiple clusters, which better reflects the biological reality of shared complex components. The choice between overlapping and non-overlapping methods depends on the biological question. If the goal is to identify stable core complexes with minimal shared components, non-overlapping methods may suffice. If the goal is to capture the full repertoire of complexes including those with shared subunits, overlapping methods are more appropriate. ## The Markov Clustering Algorithm in Practice ### How MCL Works MCL simulates random walks on the interaction network to identify densely connected regions. The algorithm alternates between two operations: expansion, which allows flow to spread through the network, and inflation, which strengthens strong edges and weakens weak edges. After repeated iterations, the network separates into distinct clusters where flow cannot easily escape. The inflation parameter controls the granularity of the clustering. Higher inflation values produce smaller, more numerous clusters. Lower inflation values produce larger, fewer clusters. The choice of inflation parameter is the main user decision in MCL and requires optimization for each dataset. ### Parameter Selection for MCL The inflation parameter typically ranges from 1.5 to 5.0 in practice. Values around 2.0 are commonly used as a starting point. Researchers should test multiple inflation values and evaluate the resulting clusters against known complexes or functional annotations. Selection criteria for the inflation parameter include: **Cluster size distribution** should match expectations for the organism and cellular context. Very small clusters of two or three proteins may represent incomplete complexes, while very large clusters may represent merged complexes. **Functional coherence** of clusters can be assessed using gene ontology enrichment analysis. Clusters with high functional coherence are more likely to represent genuine complexes. **Recovery of known complexes** provides a direct benchmark if the organism has well-annotated complexes available. ### Running MCL on Proteomics Data The input to MCL is an edge list with weights. The weights should reflect the confidence in each interaction. Unweighted networks can be used, but weighted networks generally produce better results because they incorporate quantitative information. The MCL algorithm requires the network to be represented as a matrix. For large networks, sparse matrix representations are necessary for computational efficiency. The output is a cluster assignment for each protein, which can be mapped back to protein identifiers for downstream analysis. ## ClusterONE for Overlapping Complex Detection ### The Cohesion Measure ClusterONE identifies protein complexes by searching for groups of proteins that are internally dense and have few connections to the rest of the network. The algorithm uses a cohesion measure that combines the number of internal edges with penalties for missing internal edges and external connections. The cohesion score for a group of proteins is calculated from the total weight of internal edges, the total weight of external edges, and a penalty term for missing internal edges. Groups with high cohesion are considered candidate complexes. ### Growth and Merging Steps ClusterONE grows candidate complexes from seed proteins by iteratively adding proteins that increase cohesion and removing proteins that decrease cohesion. After the growth phase, the algorithm merges highly overlapping candidate complexes to produce the final set of predicted complexes. The minimum cluster size parameter determines the smallest complex that will be reported. The density threshold determines how dense a cluster must be to be considered a complex. Both parameters require tuning based on the network properties and the expected complex characteristics. ### Comparison with MCL ClusterONE and MCL differ in several important ways: **Overlap handling** is the most significant difference. MCL produces disjoint clusters, while ClusterONE allows overlap. For networks where proteins participate in multiple complexes, ClusterONE may produce more biologically accurate results. **Input requirements** are similar, with both algorithms accepting weighted edge lists. ClusterONE requires additional parameters for minimum cluster size and density. **Computational complexity** differs between the algorithms. MCL is generally faster on large networks because it operates through matrix multiplication. ClusterONE requires iterative growth from many seeds, which can be computationally intensive. **Sensitivity to noise** differs based on the algorithm design. MCL is relatively robust to noise because the inflation step suppresses weak edges. ClusterONE can be sensitive to spurious edges that connect otherwise distinct complexes. ## Preparing Proteomics Data for Network Analysis ### Quality Control of Protein Identifications The quality of complex prediction depends directly on the quality of the input protein identifications. Poor quality identifications introduce spurious nodes and edges that confuse clustering algorithms. Quality control steps include: **Filtering by peptide count** removes proteins identified by a single peptide, which are more likely to be false positives. The minimum peptide count should be based on the experimental design and the expected protein abundance range. **Filtering by false discovery rate** ensures that the protein identifications meet a defined statistical threshold. Standard thresholds are often set at 1% or 5% FDR depending on the stringency required. **Removing contaminants** such as keratins and other common laboratory contaminants prevents these proteins from appearing as false interaction partners. **Checking replicate consistency** identifies proteins that appear in only one replicate, which may represent stochastic identifications instead of genuine biological signals. ### Constructing the Edge List The edge list is the core input for complex prediction algorithms. Each row represents an interaction between two proteins with a weight indicating confidence. Edge list construction involves: **Pairing proteins** from copurification data. In AP-MS experiments, the bait protein is paired with each prey protein identified in the purification. For co-fractionation data, proteins are paired if they show correlated elution profiles. **Assigning weights** to each edge. Weights can be based on spectral counts, intensity values, or statistical scores from interaction prediction tools. **Applying thresholds** to remove low-confidence edges. The threshold determines the tradeoff between sensitivity and precision in the final network. **Handling missing data** is important because proteomics experiments have incomplete coverage. Proteins that were not detected in a particular experiment are not necessarily absent from the complex. ### Network Visualization and Inspection Before running complex prediction algorithms, researchers should visualize the network to identify potential issues. Network visualization tools can reveal: **Disconnected components** that may represent separate biological processes or experimental artifacts. **Highly connected hubs** that may represent contaminants or proteins with many genuine interaction partners. **Dense regions** that may represent candidate complexes. **Isolated nodes** with few or no edges that will not participate in complex prediction. Visual inspection provides a qualitative check on the data quality before quantitative analysis. The Galaxy Training Network provides accessible workflow training that includes network analysis approaches suitable for researchers who want to develop reproducible analysis pipelines [4]. ## Step-by-Step Workflow for Complex Prediction ### Step 1: Define the Biological Question The first decision is what type of complexes to look for. Are you interested in stable complexes that exist under the experimental conditions, or dynamic assemblies that form transiently? Are you looking for complexes in a specific cell type or condition, or across multiple conditions? The biological question determines the experimental data needed and the computational approach. Stable complexes are better captured by methods that identify dense, persistent clusters. Dynamic assemblies may require time-resolved data or integration of multiple experimental conditions. ### Step 2: Collect and Prepare Interaction Data Assemble all available interaction evidence for the proteins of interest. This may include: **Experimental data** from your own mass spectrometry experiments or from public repositories. **Literature-curated interactions** from databases that collect published interaction data. **Predicted interactions** from computational methods that infer interactions from genomic context, co-evolution, or structural information. Each data source should be documented with its provenance and reliability characteristics. The integration of multiple data sources can improve coverage but also introduces heterogeneity in data quality. ### Step 3: Build the Interaction Network Construct the network from the prepared edge list. The network should be checked for: **Self-loops** where a protein is paired with itself, which should be removed. **Duplicate edges** that may arise from multiple experimental observations of the same interaction. **Edge weight normalization** to ensure that weights from different data sources are comparable. The network representation should be saved in a standard format that can be used by multiple analysis tools. ### Step 4: Select and Run the Clustering Algorithm Choose the clustering algorithm based on the network properties and the biological question. Run the algorithm with initial parameter settings and examine the output. For MCL, start with an inflation parameter of 2.0 and examine the cluster size distribution. For ClusterONE, start with a minimum cluster size of 3 and a density threshold appropriate for the network density. ### Step 5: Optimize Parameters Test multiple parameter settings and evaluate the results using validation metrics. The optimization process should be documented so that the final parameter choice is reproducible. Parameter optimization criteria include: **Recovery of known complexes** provides a direct measure of performance if known complexes are available for the organism. **Functional enrichment** of predicted complexes indicates whether the clusters correspond to coherent biological processes. **Cluster stability** across parameter values indicates whether the predictions are robust or depend heavily on specific parameter choices. ### Step 6: Validate Predictions Validation of predicted complexes can use several approaches: **Comparison with known complexes** from curated databases provides a direct benchmark. The overlap between predicted and known complexes can be quantified using measures such as the Jaccard index or the fraction of predicted complexes that match known complexes. **Functional annotation analysis** checks whether the proteins in each predicted complex share functional annotations. Complexes with coherent functions are more likely to be genuine. **Cross-validation** splits the interaction data into training and test sets to check whether the predicted complexes are reproducible across data subsets. **Literature validation** checks whether the predicted complexes have been reported in previous studies. ### Step 7: Interpret and Report Results The final output should include the predicted complexes with their member proteins, the confidence scores, and the parameter settings used. The interpretation should acknowledge the limitations of the computational approach and the experimental data. Reporting should include: **The complete list of predicted complexes** with member proteins and scores. **The parameter settings** used for each algorithm. **The validation results** showing how well the predictions match known complexes or functional annotations. **The limitations** of the analysis, including data coverage gaps and potential false predictions. ## Reproducibility and Workflow Management ### Version Control for Analysis Code Reproducible analysis requires version control for all analysis code and parameter settings. The Carpentries provides foundational lessons in shell, Git, and programming that support reproducible research practices [6]. Researchers should track: **The exact version of each software tool** used in the analysis. **The parameter settings** for each algorithm run. **The input data files** with their version identifiers. **The analysis scripts** that transform inputs to outputs. ### Containerization and Pipeline Management Containerization packages the analysis environment so that the same software versions and dependencies are used across different computing systems. Workflow management systems track the execution of analysis steps and ensure that each step uses the correct inputs and produces the expected outputs. The nf-core community provides documentation for building and using reproducible bioinformatics pipelines that follow community standards [5]. These pipelines can be adapted for protein complex prediction workflows, ensuring that the analysis is reproducible across different computing environments. ### Documentation Standards Documentation should record: **The experimental design** that generated the proteomics data. **The data processing steps** from raw mass spectrometry data to protein identifications. **The network construction parameters** including thresholds and scoring methods. **The clustering algorithm parameters** and the rationale for their selection. **The validation approach** and the results obtained. This documentation enables other researchers to reproduce the analysis and to understand the decisions that shaped the final predictions. ## Common Failure Patterns in Complex Prediction ### Over-clustering and Under-clustering Over-clustering occurs when a single complex is split into multiple predicted clusters. This failure pattern arises when the inflation parameter is too high in MCL or when the density threshold is too strict in ClusterONE. Over-clustering produces many small clusters that may represent fragments of larger complexes. Under-clustering occurs when multiple complexes are merged into a single predicted cluster. This failure pattern arises when the inflation parameter is too low or when the density threshold is too permissive. Under-clustering produces large clusters that contain multiple distinct complexes. Diagnosis of clustering granularity problems involves comparing the cluster size distribution with expectations for the organism and examining whether known complexes are recovered as single clusters or split across multiple clusters. ### Noise Sensitivity Interaction networks from proteomics data contain false positive edges from contaminants, non-specific binding, and experimental artifacts. These spurious edges can connect distinct complexes, causing them to merge, or can create spurious dense regions that are not genuine complexes. Noise reduction strategies include: **Stricter interaction thresholds** that remove low-confidence edges before clustering. **Contaminant filtering** that removes known contaminants from the network. **Edge weighting** that down-weights interactions with low reproducibility. **Network pruning** that removes edges connecting otherwise distinct dense regions. ### Missing Data Problems Proteomics experiments have incomplete coverage, meaning that some complex members may not be detected. Missing proteins create gaps in the network that can prevent the algorithm from identifying the complete complex. Missing data can arise from: **Detection limits** where low-abundance proteins are not identified by mass spectrometry. **Dynamic range problems** where high-abundance proteins mask low-abundance proteins. **Sample preparation losses** where certain proteins are lost during purification or digestion. **Technical variability** where stochastic sampling misses proteins in some replicates. The impact of missing data can be reduced by integrating multiple experimental replicates and by combining data from different experimental approaches. ### Parameter Overfitting Parameter optimization against known complexes can lead to overfitting, where the parameters are tuned to perform well on the benchmark set but generalize poorly to novel complexes. Overfitting is detected when the performance on the benchmark set is much higher than the performance on held-out data or on novel predictions. Strategies to reduce overfitting include: **Cross-validation** that evaluates performance on data not used for parameter optimization. **Limited parameter search** that tests a small number of parameter combinations instead of exhaustive search. **Biological plausibility checks** that evaluate whether the predicted complexes make biological sense beyond the benchmark set. ## Validation Approaches for Predicted Complexes ### Benchmarking against Known Complexes Known complexes from curated databases provide a benchmark for evaluating prediction performance. The evaluation measures how well the predicted complexes match the known complexes. Common evaluation metrics include: **Precision** measures the fraction of predicted complexes that match known complexes. **Recall** measures the fraction of known complexes that are recovered by the predictions. **F-measure** combines precision and recall into a single score. **Matching criteria** define what constitutes a match between a predicted and known complex. Common criteria require a minimum fraction of overlap between the predicted and known complex members. ### Functional Enrichment Analysis Gene ontology enrichment analysis tests whether the proteins in each predicted complex share functional annotations. Complexes with significant enrichment for specific biological processes, molecular functions, or cellular components are more likely to represent genuine biological assemblies. The enrichment analysis should account for the background distribution of annotations in the network. Proteins that are annotated with common functions may show enrichment by chance. ### Cross-species Conservation Protein complexes that are conserved across species are more likely to represent genuine biological assemblies. Conservation can be assessed by checking whether the predicted complex members have orthologs in other species that also interact. Cross-species conservation analysis requires orthology mapping and interaction data from multiple species. The absence of conservation does not rule out a genuine complex, as some complexes are species-specific. ### Experimental Validation The most direct validation of predicted complexes is experimental confirmation. Approaches include: **Co-immunoprecipitation** to confirm that predicted complex members physically associate. **Size exclusion chromatography** to confirm that predicted complex members co-migrate at the expected molecular weight. **Cross-linking mass spectrometry** to confirm spatial proximity between predicted complex members. **Fluorescence microscopy** to confirm co-localization of predicted complex members. Experimental validation is resource-intensive and is typically applied to a subset of high-confidence predictions instead of the full set. ## Limitations of Computational Complex Prediction ### Incomplete Interaction Coverage Interaction networks from proteomics data are incomplete. Not all interactions are detected in a given experiment, and the coverage depends on the experimental approach, the detection limits, and the dynamic range of protein abundances. The relationship between mRNA and protein abundance is complex, and transcriptomic data cannot reliably substitute for direct protein measurements [8]. This limitation means that interaction networks built from indirect evidence may miss interactions that are present at the protein level. ### Dynamic Complex Behavior Protein complexes are not static entities. Many complexes assemble and disassemble in response to cellular signals, and the composition of a complex can vary across conditions, cell types, and time points. Computational prediction from a single experimental condition captures only a snapshot of the complex repertoire. Integrating data from multiple conditions can reveal condition-specific complexes but also introduces complexity in interpreting the results. ### Spatial and Temporal Context Protein complexes exist in specific cellular locations and at specific times. Spatial proteomics studies have shown that protein levels can vary dramatically across tissue regions, with half of the hepatocyte proteome differentially regulated in a spatial manner [10]. This spatial variation means that a complex present in one cell type or tissue region may be absent in another. Single-cell and spatial proteomics technologies are beginning to provide the resolution needed to study complexes in their native context, but these technologies are still limited in depth and throughput. ### Algorithm Assumptions Each clustering algorithm makes assumptions about the structure of protein complexes. MCL assumes that complexes are dense regions that trap random walks. ClusterONE assumes that complexes are internally dense with few external connections. These assumptions may not hold for all complexes. Some complexes are not dense in the interaction network because their members interact primarily through a few key contacts instead of through many pairwise interactions. Other complexes have many external connections that make them difficult to separate from the rest of the network. ### Interpretation Challenges The output of complex prediction algorithms requires biological interpretation. A predicted cluster may represent a genuine complex, a subcomplex, a dynamic assembly, or an artifact of the clustering process. Interpretation requires knowledge of the biological system under study, including the functions of the proteins involved and the cellular processes in which they participate. Computational predictions should be treated as hypotheses to be tested instead of as definitive answers. ## Integrating Complex Prediction with Other Proteomics Analyses ### Proteogenomic Context Proteogenomic analysis integrates proteomics data with genomic data to understand how genetic variation affects protein abundance and function. Studies of colon and rectal tumors have shown that somatic variants display reduced protein abundance compared to germline variants and that copy number alterations have relatively few effects at the protein level [8]. This proteogenomic context matters for complex prediction because genetic alterations can affect complex composition and stability. A mutation that reduces the abundance of one complex member may destabilize the entire complex, leading to its absence from the interaction network. ### Protein Quantitative Trait Loci Large-scale proteogenomic studies have identified thousands of protein quantitative trait loci that affect circulating protein levels [11]. These genetic effects can influence complex formation and stability by altering the abundance of complex members. Understanding the genetic regulation of protein levels can help interpret complex prediction results. A complex that is absent from the interaction network in a particular individual or condition may be absent because of genetic variation that reduces the abundance of one or more members. ### Chemical Proteomics and Complex Disruption Chemical proteomics approaches can identify compounds that disrupt protein complexes. Studies using activity-based protein profiling have identified compounds that stereoselectively and site-specifically disrupt MAD2L1BP interactions with the spindle assembly checkpoint complex, leading to delayed mitotic exit in cancer cells [9]. These chemical tools provide a way to validate predicted complexes and to study the functional consequences of complex disruption. A predicted complex that can be disrupted by a specific compound provides evidence that the complex is genuine and functionally relevant. ### Single-Cell and Spatial Context Single-cell proteomics technologies are expanding the ability to measure proteins in individual cells and in spatial context. Spatial single-cell mass spectrometry has defined zonation of the hepatocyte proteome, showing that protein levels change dramatically in proximity to the central vein [10]. These technologies provide the resolution needed to study complexes in specific cell types and tissue regions. Complex prediction from single-cell or spatial data can reveal cell-type-specific complexes that are masked in bulk measurements. ## Records and Documentation for Complex Prediction Analyses ### Analysis Records Each complex prediction analysis should produce a record that includes: **The input data files** with their versions and provenance. **The software tools** with their versions and the computing environment. **The parameter settings** for each algorithm run. **The output files** with the predicted complexes and their scores. **The validation results** with the metrics used and the values obtained. These records enable the analysis to be reproduced and audited. ### Parameter Optimization Records The parameter optimization process should be documented, including: **The parameter values tested** for each algorithm. **The evaluation metrics** used to compare parameter settings. **The rationale** for the final parameter choice. **The sensitivity of the results** to parameter variation. This documentation helps other researchers understand the decisions that shaped the final predictions. ### Validation Records Validation records should include: **The validation approach** used for each predicted complex. **The validation results** showing which complexes were confirmed and which were not. **The confidence levels** assigned to each prediction. **The limitations** of the validation approach. These records support the interpretation of the predicted complexes and guide decisions about which predictions to pursue experimentally. ## Professional Escalation Criteria ### When to Seek Computational Expertise Researchers should consider consulting a computational biologist or bioinformatics specialist when: **The network is very large** with more than several thousand proteins and tens of thousands of interactions, requiring specialized computational resources. **The clustering results are inconsistent** across parameter settings, suggesting that the network structure does not support clear complex boundaries. **The validation results are poor** with low recovery of known complexes, suggesting that the data or the approach needs revision. **The analysis requires custom methods** that are not available in standard software packages. ### When to Seek Experimental Collaboration Computational predictions should be validated experimentally before being reported as definitive findings. Researchers should seek experimental collaboration when: **The predicted complex is novel** and has not been reported in the literature. **The predicted complex has potential clinical relevance** and may inform disease mechanisms or treatment decisions. **The predicted complex is central to the research question** and the conclusions depend on its validity. **The computational evidence is ambiguous** and experimental validation is needed to resolve uncertainty. ### When to Question the Input Data Researchers should revisit the input data when: **The network contains many isolated nodes** suggesting that the interaction coverage is too sparse for reliable complex prediction. **The network contains very large connected components** suggesting that the interaction thresholds are too permissive. **The clustering results are dominated by a few large clusters** suggesting that the network structure does not support the identification of distinct complexes. **The validation results are inconsistent across data subsets** suggesting that the data quality varies across experiments. ## Frequently Asked Questions ### What is the difference between MCL and ClusterONE for protein complex prediction? MCL simulates random walks on the interaction network and uses an inflation step to separate the network into disjoint clusters. ClusterONE grows candidate complexes from seed proteins using a cohesion measure and allows clusters to overlap. The main practical difference is that MCL assigns each protein to exactly one complex, while ClusterONE allows proteins to belong to multiple complexes. For networks where proteins participate in several complexes through shared subunits, ClusterONE may produce more biologically accurate results. For large networks where computational efficiency is important, MCL is generally faster. ### How do I choose the inflation parameter for MCL? The inflation parameter controls the granularity of MCL clustering. Higher values produce smaller and more numerous clusters, while lower values produce larger and fewer clusters. A common starting point is 2.0, with testing of values in the range of 1.5 to 5.0. The optimal value depends on the network properties and the expected complex characteristics. Evaluation criteria include the recovery of known complexes, the functional coherence of the predicted clusters, and the cluster size distribution. The parameter optimization process should be documented for reproducibility. ### What input data do I need for protein complex prediction? The minimum input is an edge list representing pairwise protein interactions with confidence weights. The edge list can be derived from affinity purification mass spectrometry data, co-fractionation experiments, cross-linking studies, literature-curated interactions, or a combination of these sources. Before network construction, the protein identifications should pass quality filters including minimum peptide counts and false discovery rate thresholds. The edge weights should reflect the confidence in each interaction, and low-confidence edges should be removed before clustering. ### How do I validate predicted protein complexes? Validation can use several complementary approaches. Comparison with known complexes from curated databases provides a direct benchmark using metrics such as precision, recall, and F-measure. Gene ontology enrichment analysis tests whether the proteins in each predicted complex share functional annotations. Cross-species conservation checks whether the complex members have interacting orthologs in other species. Experimental validation through co-immunoprecipitation, size exclusion chromatography, or cross-linking mass spectrometry provides the most direct confirmation but is resource-intensive and typically applied to a subset of high-confidence predictions. ### Can I use transcriptomic data to build interaction networks for complex prediction? Transcriptomic data can provide indirect evidence for interactions, but messenger RNA abundance does not reliably predict protein abundance differences [8]. Interaction networks built from transcriptomic data alone may miss interactions that are present at the protein level or include interactions that do not occur because the proteins are not co-expressed. Direct proteomic measurements provide more reliable evidence for interaction network construction. If transcriptomic data must be used, the limitations should be acknowledged and the predictions should be treated with appropriate caution. ### What should I do if my clustering results are unstable across parameter settings? Unstable clustering results suggest that the network structure does not support clear complex boundaries. This can arise from noisy interaction data, incomplete coverage, or complexes that are not well-separated in the network. Strategies to address instability include improving the input data quality through stricter interaction thresholds and contaminant filtering, testing a wider range of parameter values to identify stable regions, and using multiple algorithms to check whether the predictions are consistent across methods. If the results remain unstable, the limitations should be reported and the predictions should be treated as low confidence. ### How do I handle proteins that appear in multiple predicted complexes? Proteins that appear in multiple predicted complexes may be genuine shared subunits that participate in several complexes, or they may be artifacts of the clustering process. Overlapping clustering methods such as ClusterONE are designed to handle shared components. The biological interpretation should consider whether the protein has known functions in multiple complexes and whether the predicted complexes are supported by independent evidence. Experimental validation can help distinguish genuine shared components from clustering artifacts. ### What are the main limitations of computational complex prediction? The main limitations include incomplete interaction coverage from proteomics experiments, the dynamic nature of protein complexes that assemble and disassemble in response to cellular signals, the spatial and temporal context of complex formation, and the assumptions made by clustering algorithms about complex structure. Computational predictions should be treated as hypotheses to be tested experimentally instead of as definitive answers. The limitations of the input data and the computational approach should be reported alongside the predictions. ## Related Bioinformatics Guides - [Modeling Host-Pathogen Protein-Protein Interaction Networks: Computational Approaches for Veterinary Virology](/knowledge/bioinformatics/modeling-host-pathogen-protein-protein-interaction-networks) - [STRING Database and Protein-Protein Interaction Networks](/knowledge/bioinformatics/string-database-and-protein-protein-interaction-networks) - [Olink Proteomics: A Practical Guide to Panel Selection and Data Interpretation](/knowledge/bioinformatics/olink-proteomics-a-practical-guide-to-panel-selection-and-data-interpretation) - [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) - [Proteomics Data Analysis in R: A Practical Workflow for Differential Expression and Visualization](/knowledge/bioinformatics/proteomics-data-analysis-in-r-a-practical-workflow-for-differential-expression-and-visualization) ## 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. - [A multi-omic single-cell landscape of human gynecologic malignancies.](https://pubmed.ncbi.nlm.nih.gov/34739872). Molecular cell, 2021. - [Proteogenomic characterization of human colon and rectal cancer.](https://pubmed.ncbi.nlm.nih.gov/25043054). Nature, 2014. - [Multi-tiered chemical proteomic maps of tryptoline acrylamide-protein interactions in cancer cells.](https://pubmed.ncbi.nlm.nih.gov/39138346). Nature chemistry, 2024. - [Spatial single-cell mass spectrometry defines zonation of the hepatocyte proteome.](https://pubmed.ncbi.nlm.nih.gov/37783884). Nature methods, 2023. - [Multi-cohort proteogenomic analyses reveal genetic effects across the proteome and diseasome.](https://pubmed.ncbi.nlm.nih.gov/42097137). Cell, 2026. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.