# Differential Abundance Analysis in Single-Cell Proteomics: Statistical Methods and Pitfalls Single-cell proteomics measures protein abundances in individual cells, and differential abundance analysis seeks to identify which proteins or cell populations change significantly between experimental conditions. The core challenge is that single-cell protein measurements carry high technical variability, missing values, and compositional constraints that violate assumptions of conventional statistical tests. This article provides a practical framework for selecting statistical methods, avoiding false discoveries, and interpreting results in single-cell proteomics experiments. ## Scope and Reader Context This article addresses researchers, laboratory professionals, and bioinformatics practitioners who generate single-cell proteomics data and need to compare protein abundances or cell population frequencies between conditions. The methods discussed apply to mass spectrometry-based single-cell proteomics and antibody-based approaches such as cytometry by time of flight and flow cytometry. The statistical principles also extend to spatial proteomics and single-cell multi-omics studies that integrate protein measurements with transcriptomic data. The primary problem is identifying statistically significant changes in cell population abundances between conditions when data are noisy, sparse, and inherently compositional. A secondary problem is avoiding false discoveries that arise from inappropriate statistical models, inadequate sample sizes, or failure to account for batch effects and technical variation. ## At a Glance: Method Selection for Differential Abundance Analysis | Method | Best Use Case | Key Assumptions | Primary Limitation | |--------|---------------|-----------------|-------------------| | t-test or Wilcoxon test | Comparing protein abundance between two conditions with adequate replication | Normal distribution or symmetric distribution, independent samples | Does not account for cell-level correlation or batch structure | | Mixed-effects models | Experiments with multiple batches, donors, or technical replicates | Random effects are normally distributed, residuals are homoscedastic | Requires sufficient number of batches or donors for stable variance estimates | | Compositionality-aware methods | Cell population frequencies or relative abundance data | Total measured signal carries no biological information | Cannot be applied to absolute abundance measurements without transformation | | Zero-inflated models | Data with excessive missing values or dropout events | Missingness mechanism is ignorable or modeled explicitly | Complex model fitting, risk of overparameterization with small sample sizes | ## Core Principles of Single-Cell Proteomics Data ### Data Structure and Measurement Characteristics Single-cell proteomics data differ fundamentally from bulk proteomics data. Each cell represents one observation, and the number of cells measured per sample can range from dozens to thousands depending on the platform. Mass spectrometry-based approaches such as tandem mass tag labeling and label-free quantification produce protein intensity values for individual cells, while antibody-based approaches measure fluorescence or mass signals from targeted panels of proteins [9]. The cellular abundance of proteins varies even between isogenic single cells, and this variability can have regulatory roles in processes such as apoptosis induction and proliferation decisions [10]. This biological variability is a genuine signal, also noise, and statistical methods must distinguish between meaningful biological heterogeneity and technical measurement error. ### Missing Values and Dropout Single-cell proteomics data contain a high proportion of missing values. A protein may be undetectable in one cell because it is genuinely absent, because its abundance falls below the detection limit, or because of stochastic sampling during the measurement process. The distinction between these mechanisms matters for statistical analysis. Missing values that depend on protein abundance require different handling than missing values that occur completely at random. The pulsed stable isotope labeling approach demonstrates that approximately 4,000 proteins can be detected in single HeLa cells using state-of-the-art workflows [8]. This coverage represents a fraction of the total proteome, and the undetected proteins are not randomly distributed across cells or conditions. Ignoring the structure of missingness can bias differential abundance results. ### Compositional Nature of Measurements Many single-cell proteomics measurements are relative instead of absolute. The measured intensity for a protein reflects its proportion of the total detectable signal in that cell, not an absolute copy number. This compositional structure means that an increase in one protein necessarily corresponds to a decrease in others when the total signal is fixed. Compositional data require specialized statistical treatment. Standard approaches that analyze each protein independently ignore the constraint that abundances sum to a constant, which can produce spurious correlations and false differential abundance calls. Methods that explicitly model compositionality, such as centered log-ratio transformations, address this issue but require careful interpretation. ## Statistical Methods for Differential Abundance Analysis ### Conventional Approaches and Their Limitations The simplest approach to differential abundance analysis applies a t-test or Wilcoxon rank-sum test to each protein, comparing mean or median abundance between two conditions. These tests are appropriate when cells are independent observations and the data meet distributional assumptions. However, single-cell proteomics data frequently violate these assumptions. Cells from the same biological sample are not independent. They share the same donor, batch, and processing history. Treating each cell as an independent observation inflates the effective sample size and produces artificially small p-values. This pseudoreplication problem is a common source of false discoveries in single-cell analyses. The t-test also assumes normally distributed data. Protein intensity values are typically skewed and heteroscedastic, with variance increasing with mean abundance. Log transformation addresses skewness but does not solve the correlation problem. ### Mixed-Effects Models Mixed-effects models provide a framework for accounting for the hierarchical structure of single-cell proteomics data. A typical model includes fixed effects for the experimental condition and random effects for biological replicates, batches, or other sources of technical variation. The key advantage of mixed-effects models is that they correctly attribute variance to the appropriate level. Cell-to-cell variability within a sample is treated as a random effect, while the condition effect is estimated at the sample level. This approach prevents pseudoreplication and produces valid inference when the number of biological replicates is adequate. Mixed-effects models require sufficient numbers of random effect levels. A model with three biological replicates per condition may produce unstable variance estimates, and the model may fail to converge. Researchers should assess the number of replicates needed for their specific experimental design before committing to this approach. ### Compositionality-Aware Methods When the research question concerns cell population abundances instead of individual protein levels, compositionality-aware methods are appropriate. These methods recognize that the frequency of one cell population is constrained by the frequencies of all other populations. The centered log-ratio transformation is a common tool for compositional data analysis. It transforms relative abundances into a coordinate system where standard multivariate statistical methods can be applied. The interpretation of results changes, however, because the transformed values represent log ratios relative to the geometric mean of all components. Compositionality-aware methods are particularly relevant for studies that identify changes in immune cell populations or other cell types within a tissue. For example, single-cell analysis of the bone marrow microenvironment in B cell acute lymphoblastic leukemia demonstrated remodeling of immune cell populations upon disease initiation and re-emergence during chemotherapy [7]. Such population-level changes require methods that respect the compositional structure of cell frequency data. ### Zero-Inflated and Mixture Models Zero-inflated models address the excess of missing values in single-cell proteomics data. These models combine a logistic component that models the probability of detection with a count or continuous component that models abundance when the protein is detected. Mixture models extend this approach by allowing multiple latent states. For example, a protein might exist in a low-abundance state and a high-abundance state, with cells transitioning between states. Differential abundance analysis then tests whether the proportion of cells in each state differs between conditions. These models are more flexible than conventional approaches but require larger sample sizes and more computational resources. Model fitting can be unstable, and the choice of distributional assumptions can substantially affect results. ## Practical Workflow for Differential Abundance Analysis ### Step 1: Define the Biological Question and Experimental Design Before collecting data, specify the exact comparison to be made. Is the question about differences in individual protein abundance between conditions, or about differences in cell population frequencies? The answer determines the appropriate statistical framework. The experimental design must include adequate biological replication. Technical replication, where the same sample is measured multiple times, does not substitute for biological replication, where independent samples are collected from multiple donors or animals. The number of biological replicates needed depends on the expected effect size and the variability of the measurement platform. ### Step 2: Preprocess and Quality Control the Data Data preprocessing includes normalization, batch correction, and quality filtering. Normalization adjusts for differences in total signal between cells or batches. Batch correction removes systematic technical variation that is not related to the biological condition. Quality control should identify and remove low-quality cells, such as those with very low total protein signal, high levels of contamination, or abnormal measurement profiles. The criteria for excluding cells should be established before the analysis to avoid bias. The Galaxy Training Network provides accessible tutorials for quality control and preprocessing of single-cell data [4]. These resources offer practical guidance for implementing standard workflows. ### Step 3: Select the Statistical Method Based on Data Characteristics The choice of statistical method depends on the data structure, the experimental design, and the biological question. The following decision criteria guide method selection: - If the analysis compares individual protein abundances between two conditions with adequate biological replication, mixed-effects models are preferred over t-tests because they account for cell-level correlation. - If the analysis compares cell population frequencies, compositionality-aware methods should be used. - If the data contain a high proportion of missing values, zero-inflated models may be necessary. - If the experimental design includes multiple batches, batch correction and mixed-effects models with batch as a random effect are required. ### Step 4: Apply Multiple Testing Correction Differential abundance analysis typically tests thousands of proteins simultaneously. Multiple testing correction is essential to control the false discovery rate. The Benjamini-Hochberg procedure is a common choice, but more stringent methods may be appropriate when the cost of false positives is high. The choice of significance threshold should be justified in the methods section of any report. A false discovery rate of 5 percent is a common default, but the appropriate threshold depends on the downstream validation strategy. ### Step 5: Validate Findings with Independent Methods Statistical findings should be validated with independent approaches. Validation can include targeted protein measurements, orthogonal assays, or replication in an independent cohort. The single-cell spatial omics and bulk proteomics approaches used in the APOE Christchurch study demonstrate how multiple measurement modalities can corroborate findings [11]. Validation is particularly important for unexpected or novel findings. The reproducibility of single-cell proteomics results across platforms and laboratories is still being established, and independent confirmation strengthens confidence in biological conclusions. ## Options and Tradeoffs in Method Selection ### Parametric versus Nonparametric Tests Parametric tests such as the t-test assume a specific distributional form and are more powerful when assumptions are met. Nonparametric tests such as the Wilcoxon test make fewer assumptions and are more robust to outliers and skewed distributions. For single-cell proteomics data, nonparametric tests are often safer choices because the data frequently deviate from normality. However, nonparametric tests have lower power, and they do not address the correlation structure within samples. ### Cell-Level versus Sample-Level Analysis A fundamental decision is whether to analyze data at the cell level or aggregate to the sample level. Cell-level analysis preserves information about heterogeneity but risks pseudoreplication. Sample-level analysis, where each biological replicate contributes one summary statistic, avoids pseudoreplication but discards information. A compromise approach uses two-stage analysis. In the first stage, cell-level data are summarized within each sample. In the second stage, sample-level summaries are compared between conditions using conventional statistical tests. This approach is simple and robust, though it may have lower power than a well-specified mixed-effects model. ### Fixed versus Random Effects for Batches Batch effects are a major source of technical variation in single-cell proteomics. The choice between treating batch as a fixed or random effect depends on the number of batches and the goal of the analysis. Fixed effects for batch are appropriate when the analysis includes a small number of batches and the goal is to adjust for their influence. Random effects for batch are appropriate when batches are sampled from a larger population of possible batches and the goal is to generalize beyond the specific batches in the study. ### Model Complexity versus Interpretability More complex models can capture more features of the data but are harder to fit and interpret. Zero-inflated models, mixture models, and hierarchical Bayesian models offer flexibility but require expertise and computational resources. The principle of parsimony suggests choosing the simplest model that adequately addresses the research question. A model that is too simple will produce biased results, but a model that is too complex may overfit the data and produce unstable estimates. ## Observations and Measurements in Practice ### Variability Sources in Single-Cell Proteomics Understanding the sources of variability in single-cell proteomics data is essential for designing experiments and interpreting results. The main sources are: - Biological variability between cells of the same type - Technical variability from sample preparation and measurement - Batch effects from processing samples at different times or with different reagents - Missing value variability from stochastic detection limits The relative contribution of these sources varies by platform and protocol. Mass spectrometry-based approaches have different variability profiles than antibody-based approaches [10]. Researchers should characterize the variability in their specific system before designing differential abundance experiments. ### Protein Turnover as a Confounding Factor Protein abundance reflects the balance between synthesis and degradation. Two cells with identical protein abundance may have very different turnover rates, and these differences can affect biological interpretation. The SC-pSILAC approach demonstrates that protein turnover dynamics can be measured simultaneously with abundance in single cells [8]. Turnover dynamics can confound differential abundance analysis when conditions affect protein stability or synthesis rates. A protein that appears unchanged in abundance may actually have altered turnover, and this distinction can be biologically important. Researchers studying dynamic processes should consider whether turnover measurements are needed to interpret abundance changes. ### Cell Size and Abundance Scaling Protein abundance does not always scale with cell size. The SC-pSILAC study found that histones and some cell-cycle proteins do not scale with cell diameter [8]. This observation has implications for differential abundance analysis because cell size differences between conditions can create apparent abundance changes that reflect size effects instead of regulatory changes. Normalization strategies that adjust for cell size or total protein content can address this issue, but the choice of normalization method affects results. Researchers should examine whether their normalization approach is appropriate for the biological question. ## Records and Documentation for Reproducibility ### Essential Records for Differential Abundance Analysis Reproducible analysis requires comprehensive documentation of data processing and statistical decisions. The following records should be maintained: - Raw data files and instrument output - Preprocessing scripts and software versions - Quality control metrics and exclusion criteria - Normalization and batch correction parameters - Statistical model specifications - Multiple testing correction methods - Analysis scripts and computational environment details The nf-core documentation provides standards for reproducible bioinformatics pipelines that can be adapted for single-cell proteomics analysis [5]. These standards emphasize version control, containerization, and automated workflow management. ### Version Control and Computational Environment Analysis code should be maintained under version control, and the computational environment should be documented. The Carpentries lessons provide foundational training in version control with Git and reproducible computing practices [6]. These skills are essential for maintaining analysis reproducibility over time. Software versions matter for single-cell proteomics analysis. Different versions of analysis packages can produce different results, and the specific versions used should be recorded. Containerization tools can preserve the computational environment and ensure that analyses can be rerun identically. ### Reporting Standards for Publications Publications reporting differential abundance results should include sufficient detail for readers to evaluate the analysis. The methods section should describe: - The statistical method and rationale for its selection - The software and version used - The normalization and batch correction approach - The multiple testing correction method - The significance threshold and its justification - The number of biological and technical replicates The EMBL-EBI Training resources provide guidance on bioinformatics data analysis and reporting standards [2]. These resources support researchers in producing analyses that meet community expectations for rigor and transparency. ## Quality Controls and Their Implementation ### Pre-Analysis Quality Checks Before running differential abundance analysis, verify that the data meet quality standards. Key checks include: - Total protein signal per cell should follow a consistent distribution across samples - The number of detected proteins per cell should be similar across conditions - Batch effects should be visualized and assessed before correction - Technical replicates should cluster together in dimensionality reduction plots The Bioconductor project provides packages for quality control and visualization of single-cell data [3]. These tools support systematic assessment of data quality before statistical analysis. ### Post-Analysis Quality Checks After running differential abundance analysis, verify that results are robust to analysis choices. Sensitivity analyses can assess whether findings change with different normalization methods, statistical tests, or significance thresholds. A common post-analysis check is to examine the distribution of p-values. A well-behaved analysis should produce a uniform distribution of p-values under the null hypothesis, with a peak near zero indicating true signals. An excess of very small p-values may indicate model misspecification or unaccounted technical variation. ### Benchmarking Against Known Controls Including positive and negative controls in the experimental design provides a reference for evaluating analysis performance. Positive controls are proteins or cell populations known to change between conditions. Negative controls are known to be unchanged. Benchmarking against controls can reveal systematic biases in the analysis. If negative controls show significant differences, the analysis may have inadequate batch correction or inappropriate statistical modeling. ## Common Failure Patterns in Differential Abundance Analysis ### Pseudoreplication and Inflated Significance The most common failure pattern is treating cells as independent observations when they are clustered within samples. This practice inflates the effective sample size and produces p-values that are too small. The result is a high rate of false discoveries that do not replicate in validation studies. The solution is to account for the hierarchical structure of the data using mixed-effects models or sample-level analysis. Researchers should report the number of biological replicates, beyond the number of cells, in their methods. ### Inadequate Handling of Missing Values Missing values in single-cell proteomics data are often handled by simple imputation or by excluding proteins with high missingness. Both approaches can introduce bias. Imputation methods that assume missing values are random will underestimate the uncertainty in the data. Excluding proteins with high missingness can remove biologically important proteins that are difficult to detect. The appropriate handling of missing values depends on the missingness mechanism. Researchers should examine whether missingness is related to protein abundance or experimental condition and choose methods accordingly. ### Ignoring Compositional Structure Analyzing relative abundance data as if it were absolute can produce spurious results. When the total signal is constrained, changes in one protein force compensatory changes in others. Standard statistical tests do not account for this constraint. Compositionality-aware methods address this issue but require careful interpretation. The results of compositional analysis describe relative changes, not absolute changes, and this distinction should be clear in reporting. ### Overcorrection for Batch Effects Batch correction is essential for combining data across batches, but overcorrection can remove biological signal. Some batch correction methods assume that batch effects are orthogonal to biological effects, which is not always true. If batch is correlated with condition, correction can eliminate the very differences the study aims to detect. Researchers should assess the degree of batch confounding before applying correction methods. If batch is strongly confounded with condition, the experimental design should be reconsidered instead of relying on computational correction. ### Multiple Testing Correction Omission Failing to apply multiple testing correction is a straightforward but common error. Testing thousands of proteins without correction produces a large number of false positives by chance alone. The false discovery rate should be controlled and reported. ## Limitations of Current Methods ### Statistical Power with Limited Replicates Single-cell proteomics experiments are expensive, and many studies include only a small number of biological replicates. Limited replication reduces statistical power and increases the risk of both false positives and false negatives. Power analysis should be conducted before the experiment to determine the number of replicates needed to detect the expected effect size. If the required number of replicates is not feasible, the study may need to focus on larger effect sizes or use a more targeted measurement approach. ### Platform-Specific Variability Different single-cell proteomics platforms have different coverage, sensitivity, and variability profiles. Mass spectrometry-based approaches offer broad coverage but may have lower throughput. Antibody-based approaches offer higher throughput but are limited to targeted panels [9]. Results obtained on one platform may not replicate on another. Researchers should be cautious about generalizing findings across platforms and should validate key results with an independent method. ### Interpretation Challenges with Compositional Data Compositional data analysis produces results that are relative, not absolute. A finding that a protein is increased in one condition means that its proportion of the total signal is higher, not necessarily that its absolute abundance is higher. This distinction is important for biological interpretation. Researchers should consider whether absolute or relative abundance is the biologically relevant quantity for their question. If absolute abundance matters, methods that provide absolute quantification should be used. ### Reproducibility Across Laboratories The reproducibility of single-cell proteomics across laboratories is still being established. Differences in protocols, reagents, and instrumentation can produce systematic differences in results. The NCBI provides data resources that support sharing and comparison of proteomics data across studies [1]. Researchers should be aware that findings from a single laboratory may not replicate elsewhere and should design studies with this limitation in mind. ## Safety and Regulatory Context ### Data Management and Privacy Considerations Single-cell proteomics data from human samples may contain sensitive information. Researchers must comply with applicable regulations for data protection and privacy. De-identification of samples and secure data storage are essential practices. The NCBI provides guidance on data submission and access for human data [1]. Researchers should review the requirements for their specific data type and follow approved procedures for data sharing. ### Ethical Use of Animal Models Studies using animal models must follow institutional and regulatory requirements for animal care and use. The APOE Christchurch study used mouse models of Alzheimer's disease, and such studies require ethical approval and adherence to established guidelines [11]. Researchers should document animal welfare considerations in their protocols and ensure that experimental designs minimize animal numbers while maintaining statistical power. ### Reporting Standards for Clinical Translation When single-cell proteomics findings have potential clinical implications, reporting standards should be followed. The B cell acute lymphoblastic leukemia study identified monocyte abundance as a predictor of patient survival, and such findings require careful validation before clinical application [7]. Researchers should avoid overstating the clinical significance of findings from single-cell studies and should clearly describe the limitations of their data. ## Professional Escalation Criteria ### When to Seek Statistical Consultation Researchers should seek statistical consultation when: - The experimental design is complex, with multiple factors or nested structures - The data show unusual patterns that are difficult to model - Different statistical methods produce conflicting results - The analysis requires methods beyond the researcher's expertise Statistical consultants can help with model specification, power analysis, and interpretation of results. Early consultation is preferable to seeking help after data collection is complete. ### When to Reconsider the Experimental Design The experimental design should be reconsidered when: - Batch effects are strongly confounded with the condition of interest - The number of biological replicates is too small for the expected effect size - The measurement platform cannot detect the proteins of interest with adequate reliability - The variability between replicates is too large to detect meaningful differences Redesigning the experiment may be more efficient than attempting to salvage an inadequate design with complex statistical methods. ### When to Escalate Technical Issues Technical issues that require escalation include: - Persistent batch effects that cannot be corrected - Unstable instrument performance across runs - High rates of missing values that suggest protocol problems - Inconsistent results between technical replicates These issues should be addressed before proceeding with differential abundance analysis. Analyzing data with unresolved technical problems produces unreliable results. ## Decision Framework for Method Selection Based on Experimental Constraints Selecting a statistical method for differential abundance analysis requires a structured evaluation of experimental constraints before data collection begins. The choice between t-tests, mixed-effects models, and compositionality-aware approaches depends on the number of biological replicates, the degree of batch confounding, the expected missingness rate, and whether the research question concerns absolute protein abundance or relative cell population frequencies. A decision framework that maps these constraints to specific methods reduces the risk of committing to an inappropriate analysis plan after data collection is complete. ### Constraint Assessment Before Data Collection The first step in the decision framework is to document the experimental constraints that will determine which statistical methods are viable. Four constraints require explicit assessment: **Biological replicate count.** The number of independent biological samples per condition determines whether mixed-effects models can produce stable variance estimates. Mixed-effects models require sufficient levels of the random effect to estimate between-sample variance reliably. With fewer than five biological replicates per condition, variance estimates become unstable and model convergence becomes uncertain. In such cases, sample-level analysis using aggregated summaries with conventional tests may be more reliable than a mixed-effects model that cannot estimate variance components accurately. **Batch structure and confounding.** The relationship between batch and condition determines whether batch correction is feasible and which correction approach is appropriate. If batch is completely confounded with condition, meaning all samples from one condition are processed in one batch and all samples from the other condition in a different batch, no computational correction can separate batch effects from biological effects. The experiment must be redesigned to interleave conditions across batches. If batch is partially confounded, meaning each condition appears in multiple batches but unevenly, mixed-effects models with batch as a random effect can adjust for batch variation while estimating the condition effect. **Expected missingness rate.** The proportion of missing values anticipated for the proteins or cell populations of interest influences whether zero-inflated models are necessary. Pilot data or published studies using the same platform can provide estimates of typical missingness rates. When more than 20 percent of measurements are expected to be missing for key analytes, standard tests that exclude missing values or use simple imputation will produce biased results. Zero-inflated models or missingness-aware approaches should be planned from the outset. **Measurement scale and compositionality.** The decision between analyzing individual protein abundances and cell population frequencies determines whether compositionality-aware methods are required. If the research question concerns the relative frequency of cell populations within a tissue or sample, the data are inherently compositional because the frequencies of all populations sum to one. If the research question concerns the absolute abundance of individual proteins, compositionality may still apply when the measurement platform reports relative intensities instead of absolute copy numbers. ### Method Selection Matrix Based on Constraint Profiles The following matrix maps constraint profiles to recommended analytical approaches. This matrix serves as a practical decision aid for planning differential abundance analysis. | Constraint Profile | Recommended Approach | Rationale | |--------------------|----------------------|-----------| | 5 or more replicates per condition, low missingness, absolute abundance question | Mixed-effects model with condition as fixed effect and biological sample as random effect | Accounts for cell-level correlation while preserving power from cell-level data | | 3 to 4 replicates per condition, low missingness, absolute abundance question | Sample-level analysis with aggregated summaries and t-test or Wilcoxon test | Avoids unstable variance estimates from mixed-effects models with few replicates | | Any replicate count, high missingness, absolute abundance question | Zero-inflated model or missingness-aware approach | Models the detection process explicitly instead of ignoring missing values | | Any replicate count, cell population frequency question | Compositionality-aware method with centered log-ratio transformation | Respects the constraint that population frequencies sum to one | | Batch partially confounded with condition | Mixed-effects model with batch as random effect | Adjusts for batch variation while estimating condition effects | | Batch completely confounded with condition | Redesign experiment before analysis | No computational method can separate confounded batch and condition effects | ### Implementation Steps for the Decision Framework Implementing the decision framework requires a systematic process that begins before data collection and continues through analysis. The following steps provide a practical sequence for applying the framework. **Step 1: Document the constraint profile.** Create a written record of the expected number of biological replicates per condition, the batch structure, the anticipated missingness rate, and the measurement scale. This record should be completed before data collection begins and updated if the experimental plan changes. **Step 2: Select the primary analytical approach.** Use the method selection matrix to identify the recommended approach for the documented constraint profile. Record the selection and the rationale in the analysis plan. **Step 3: Define the analysis workflow.** Specify the preprocessing steps, normalization method, statistical test, and multiple testing correction procedure that will be applied. The workflow should be written down before any analysis is performed to prevent post hoc decisions that could bias results. **Step 4: Conduct the analysis according to the plan.** Execute the predefined workflow without deviation. If unexpected data characteristics emerge, such as higher than expected missingness or unanticipated batch effects, document the deviation and consult the escalation criteria described below. **Step 5: Perform sensitivity analysis.** After completing the primary analysis, repeat the analysis using alternative methods to assess whether conclusions are robust to analytical choices. For example, if the primary analysis used a mixed-effects model, repeat the analysis using sample-level aggregation to determine whether the same proteins or populations are identified as significantly different. ### Record System for Analytical Decisions A structured record system supports reproducibility and provides a basis for troubleshooting when results are unexpected. The following records should be maintained for each differential abundance analysis: **Analysis plan document.** This document records the constraint profile, the selected method, the rationale for the selection, and the predefined workflow. The analysis plan should be dated and versioned, with changes documented as they occur. **Method selection log.** This log records the specific statistical tests, software packages, and parameter settings used in the analysis. The log should include the version numbers of all software packages because different versions can produce different results. **Constraint assessment worksheet.** This worksheet documents the actual values of the constraints after data collection. The worksheet should record the number of biological replicates per condition, the batch assignment of each sample, the observed missingness rate, and the measurement scale. Comparing the actual constraint values to the expected values reveals whether the analysis plan remains appropriate. **Sensitivity analysis record.** This record documents the alternative methods tested and whether they produced consistent conclusions. The record should note any discrepancies between methods and the potential explanations for those discrepancies. ### Troubleshooting Unexpected Results When differential abundance analysis produces unexpected results, the record system provides the information needed to identify the cause. The following troubleshooting sequence addresses common sources of unexpected findings. **Check the constraint assessment worksheet.** Verify that the actual number of biological replicates matches the planned number. If replicates were lost during the experiment, the analysis may have insufficient power, and the results should be interpreted with caution. **Examine the batch assignment.** Determine whether batch is confounded with condition in the actual data. If samples from one condition were processed predominantly in one batch, the observed differences may reflect batch effects instead of biological effects. **Review the missingness pattern.** Assess whether missingness differs between conditions. If one condition has systematically higher missingness for a protein, the differential abundance result may reflect detection differences instead of true abundance differences. **Compare sensitivity analysis results.** If the primary analysis and sensitivity analysis produce different conclusions, the results are not robust to analytical choices. The discrepancy should be investigated before drawing biological conclusions. ### Escalation Criteria for Method Selection Issues Certain situations require escalation beyond the standard decision framework. These situations include: **Persistent model convergence failures.** If a mixed-effects model fails to converge despite adequate replicate numbers, the model specification may be inappropriate for the data structure. Statistical consultation should be sought to identify alternative model specifications or analytical approaches. **Conflicting results across sensitivity analyses.** When different reasonable analytical methods produce substantially different conclusions, the biological interpretation is uncertain. The analysis should be reviewed with statistical consultation to determine whether the conflict reflects a genuine biological phenomenon or an analytical artifact. **Constraint violations discovered after data collection.** If the actual data reveal constraints that differ substantially from the planned constraints, such as much higher missingness than expected or unexpected batch structure, the analysis plan should be revised with statistical consultation before proceeding. **Results that contradict established biological knowledge.** Unexpected findings that contradict known biology should be treated with caution. The analysis should be reviewed for technical artifacts before considering the finding as a genuine discovery. The single-cell analysis of the bone marrow microenvironment in B cell acute lymphoblastic leukemia identified monocyte abundance as a predictor of patient survival, and such findings required careful validation before clinical interpretation [7]. Unexpected results warrant similar scrutiny. ### Integration with Existing Analytical Workflows The decision framework integrates with the preprocessing and quality control steps described in the main workflow. The constraint assessment should be completed before preprocessing begins because the choice of normalization and missing value handling depends on the measurement scale and missingness rate. The method selection should be completed before statistical analysis begins because the choice of test determines the required data format and preprocessing decisions. The framework also integrates with the validation step. Sensitivity analyses conducted as part of the framework provide evidence of robustness that strengthens validation efforts. When findings are robust across multiple analytical approaches, confidence in the biological conclusion increases. ### Practical Considerations for Implementation Implementing the decision framework requires discipline in documenting decisions and following the predefined workflow. The following practical considerations support successful implementation: **Involve statistical expertise early.** Statistical consultation should occur during experimental design, not after data collection. The constraint assessment and method selection are most effective when informed by statistical expertise. **Use available training resources.** The Galaxy Training Network provides accessible tutorials for single-cell data analysis workflows [4]. The Carpentries lessons provide foundational training in reproducible computing practices that support the record system [6]. These resources can help researchers implement the framework effectively. **Document decisions in real time.** Analytical decisions should be recorded as they are made, not reconstructed after the analysis is complete. Real-time documentation is more accurate and supports troubleshooting when unexpected results arise. **Review the framework periodically.** The decision framework should be reviewed and updated as new statistical methods become available and as the researcher gains experience with different data types. The Bioconductor project provides ongoing documentation of new packages and methods for single-cell analysis [3]. ## Frequently Asked Questions ### What is the difference between differential abundance and differential expression in single-cell proteomics? Differential abundance refers to changes in the quantity of proteins or cell populations between conditions. Differential expression is a term more commonly used in transcriptomics to describe changes in mRNA levels. In single-cell proteomics, differential abundance analysis tests whether the measured protein quantity differs between conditions, accounting for the technical and biological variability in the measurements. ### Why can't I treat each cell as an independent observation in a t-test? Cells from the same biological sample share the same donor, batch, and processing history. Treating them as independent observations inflates the effective sample size and produces p-values that are too small. This pseudoreplication leads to false discoveries that do not replicate in validation studies. Mixed-effects models or sample-level analysis account for the correlation between cells from the same sample. ### When should I use a compositionality-aware method instead of a standard test? Compositionality-aware methods should be used when the data are relative abundances that sum to a constant. This situation arises when analyzing cell population frequencies or when the measurement platform reports relative instead of absolute protein quantities. Standard tests that ignore the compositional constraint can produce spurious correlations and false differential abundance calls. ### How do I handle missing values in single-cell proteomics data? The appropriate handling of missing values depends on the mechanism that produces them. If missingness is related to protein abundance, simple imputation methods will introduce bias. Options include zero-inflated models that explicitly model the detection process, or sensitivity analyses that assess how results change under different missingness assumptions. The missingness mechanism should be examined before choosing a method. ### What is the minimum number of biological replicates needed for differential abundance analysis? The required number of biological replicates depends on the expected effect size, the variability of the measurement platform, and the desired statistical power. There is no universal minimum, but studies with fewer than three biological replicates per condition generally have very limited power and unstable variance estimates. Power analysis should be conducted before the experiment to determine the required sample size. ### How do batch effects affect differential abundance results? Batch effects are systematic technical differences between groups of samples processed at different times or with different reagents. If batch is correlated with the condition of interest, batch effects can create false differences or mask true differences. Batch correction methods can address this issue, but overcorrection can remove biological signal. The degree of batch confounding should be assessed before applying correction. ### Should I use parametric or nonparametric tests for single-cell proteomics data? Nonparametric tests are often safer for single-cell proteomics data because the data frequently deviate from normality and contain outliers. Nonparametric tests make fewer assumptions and are more robust, but they have lower power than parametric tests when assumptions are met. The choice should be based on the distribution of the data and the sample size. ### How do I validate differential abundance findings from single-cell proteomics? Validation can include targeted protein measurements with an orthogonal method, replication in an independent cohort, or functional experiments that test the biological significance of the finding. The single-cell spatial omics and bulk proteomics approaches used in the APOE Christchurch study demonstrate how multiple measurement modalities can corroborate findings [11]. Validation is particularly important for unexpected or novel findings. ## Related Bioinformatics Guides - [Single-Cell Sequencing Methods: A Comparative Overview](/knowledge/bioinformatics/single-cell-sequencing-methods-a-comparative-overview) - [Single-Cell Multi-Omics Integration: Methods and Applications](/knowledge/bioinformatics/single-cell-multi-omics-integration-methods-and-applications) - [Spatial Proteomics vs. Single-Cell Proteomics: Choosing the Right Approach](/knowledge/bioinformatics/spatial-proteomics-vs-single-cell-proteomics-choosing-the-right-approach) - [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) - [Single-Cell Sequencing Workflow: From Sample Preparation to Data Analysis](/knowledge/bioinformatics/single-cell-sequencing-workflow-from-sample-preparation-to-data-analysis) ## 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. - [Extensive Remodeling of the Immune Microenvironment in B Cell Acute Lymphoblastic Leukemia.](https://pubmed.ncbi.nlm.nih.gov/32470390). Cancer cell, 2020. - [Global analysis of protein turnover dynamics in single cells.](https://pubmed.ncbi.nlm.nih.gov/40168994). Cell, 2025. - [The Intriguing Landscape of Single-Cell Protein Analysis.](https://pubmed.ncbi.nlm.nih.gov/35199955). Advanced science (Weinheim, Baden-Wurttemberg, Germany), 2022. - [Single cell protein analysis for systems biology.](https://pubmed.ncbi.nlm.nih.gov/30072488). Essays in biochemistry, 2018. - [APOE Christchurch enhances a disease-associated microglial response to plaque but suppresses response to tau pathology.](https://pubmed.ncbi.nlm.nih.gov/39844286). Molecular neurodegeneration, 2025. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.