Proteomics data analysis follows a consistent pipeline: raw mass spectrometry files go through quality control, database searching, quantification, normalization, statistical testing, and finally biological interpretation. Each stage has specific tools and decision points that affect your results. The choices you make early, particularly around acquisition mode and quantification strategy, shape every downstream step.
How Mass Spectrometry Data Gets Acquired
Before analysis begins, you need to understand how your data was collected, because the acquisition mode determines which software and algorithms you’ll use. The two main approaches are data-dependent acquisition (DDA) and data-independent acquisition (DIA).
DDA works by automatically selecting the most abundant peptide signals in real time and fragmenting them one at a time to generate identification spectra. This is the more established approach and has well-developed analysis tools, but it has a fundamental limitation: the instrument picks what to fragment based on intensity, which means low-abundance peptides can be missed entirely. This stochastic sampling also means that if you run the same sample twice, you won’t get identical results.
DIA takes a different approach. Instead of picking individual peptides, the instrument divides the entire mass range into small windows (typically 5 to 25 daltons wide) and fragments everything within each window simultaneously. This produces more complex spectra since multiple peptides are fragmented together, but it captures a much more complete picture of the sample. DIA offers better reproducibility and quantification accuracy than DDA because it doesn’t rely on biased sampling. The tradeoff is that analyzing DIA data is more computationally demanding and historically required a spectral library built from DDA experiments, though newer tools are reducing that dependency.
Quality Control Before You Start
Quality control is the first real analysis step, and skipping it is one of the most common mistakes. You’re checking whether the instrument performed consistently and whether sample preparation worked as expected. Key metrics fall into two categories.
Identification-based metrics require that peptides have already been matched to sequences. These include the peptide identification rate (what fraction of spectra led to a confident match), tryptic miscleavage rates (which tell you how well the protein digestion step worked), and mass measurement accuracy (how close observed masses are to theoretical values). High miscleavage rates, for example, suggest the digestion was incomplete, which reduces the number of usable peptides and can bias quantification.
Chromatographic metrics track how well your liquid chromatography separation performed. Retention time standards, either spiked-in reference peptides or endogenous peptides modeled from your own data, let you monitor whether peptides eluted at the expected times. Drift in retention times across runs signals instrument instability that will affect quantification if not corrected.
Choosing a Quantification Strategy
Proteomics quantification falls into two broad categories: label-free and label-based. Your choice affects experimental design, cost, throughput, and the type of bias you’ll encounter.
Label-free quantification (LFQ) measures protein abundance by comparing peptide signal intensities or spectral counts across separate runs. Its biggest advantage is proteome coverage. Comparative studies have shown that label-free approaches can identify up to three times more proteins than label-based methods in replicate measurements. The downside is lower accuracy in measuring fold changes between conditions, since each sample is analyzed in a separate run and subject to run-to-run variability.
Label-based methods like TMT (tandem mass tags) attach chemical labels to peptides or proteins so that multiple samples can be combined and analyzed in a single run. This eliminates much of the run-to-run variation and improves quantification accuracy. However, TMT has a well-known ratio compression problem: true differences between conditions get underestimated because co-isolated peptides contribute interference to the reporter ion signals. Applying TMT labels at the protein level rather than the peptide level can reduce this compression effect. Both label-free and label-based approaches show comparable reproducibility in protein quantification, so the choice often comes down to whether you prioritize depth of coverage (label-free) or measurement accuracy (TMT).
Database Searching and Protein Identification
Raw spectra need to be matched against a protein sequence database to identify which proteins are present. The instrument gives you fragment ion patterns; search engines compare those patterns against theoretical spectra generated from known protein sequences. Several widely used tools handle this step.
MaxQuant is one of the most popular platforms for processing DDA data, particularly for label-free and SILAC experiments. It handles the full pipeline from raw files through protein quantification and outputs results as text files you can open in spreadsheet software, though meaningful interpretation typically requires dedicated downstream tools. FragPipe is a newer alternative that offers fast database searching with the MSFragger engine and has become increasingly popular for large-scale datasets. Skyline was originally designed for targeted proteomics but has expanded to support label-free quantification and DIA analysis, with particular strength in visualizing and validating peptide-level data.
Regardless of which tool you use, one critical parameter is the false discovery rate (FDR). Every database search produces some incorrect matches, and the FDR estimates what fraction of your identifications are false positives. The standard threshold in the field is 1% FDR at the peptide-spectrum match level. At this cutoff, roughly 1 in 100 of your identified spectra is expected to be wrong. Some studies use a stricter 0.1% threshold for higher confidence, while 5% is occasionally used for exploratory work but generally considered too lenient. Importantly, a low FDR at the spectrum level doesn’t guarantee a low FDR at the protein level, so many pipelines apply separate filtering at both levels. Research on very large datasets has shown that the maximum number of true protein identifications is typically reached at a spectrum-level FDR around 0.5%, meaning that loosening the threshold beyond that point mostly adds false positives rather than real discoveries.
Normalization: Removing Technical Bias
Raw protein intensities contain systematic biases from differences in sample loading, instrument sensitivity, and chromatographic conditions. Normalization corrects for these technical variations so that the remaining differences reflect actual biology.
Median centering is the simplest approach: you shift all values in each sample so that the medians align. This assumes that most proteins don’t change between conditions, which is usually reasonable. Quantile normalization goes further by forcing the entire distribution of intensities to be identical across samples. It’s effective but aggressive, and can distort real biological differences if the assumption of similar distributions doesn’t hold.
LOESS (locally estimated scatterplot smoothing) normalization fits a flexible curve to the relationship between two samples and corrects for intensity-dependent bias, where the systematic error is larger or smaller depending on the signal level. Variance-stabilizing normalization (VSN) transforms the data to make variance independent of the mean, which is particularly useful when you plan to use statistical methods that assume equal variance. There is no single best normalization method for all datasets. The right choice depends on your experimental design, the amount of technical variation, and how many proteins you expect to genuinely change between conditions.
Statistical Testing for Differential Expression
Once your data is normalized, you want to know which proteins are significantly more or less abundant between your experimental groups. This is differential expression analysis, and several approaches are available.
The simplest option is a standard t-test or its variants, comparing protein intensities between groups. This works for straightforward two-group comparisons but doesn’t handle the complexity of proteomics data well, particularly the problem of missing values (proteins detected in some samples but not others).
LIMMA, originally developed for gene expression microarrays, has become widely used in proteomics. It fits a linear model to each protein and uses an empirical Bayes method to borrow information across proteins, which improves statistical power when sample sizes are small. This borrowing of information is especially valuable in proteomics, where you often have only three to five replicates per group.
More proteomics-specific tools like MSstats and QPROT address challenges unique to mass spectrometry data. MSstats works directly with peptide-level intensities and handles the rollup from peptides to proteins within its statistical framework, rather than requiring you to summarize to protein level first. QPROT includes its own normalization procedure and explicitly accounts for missing data in its model, using a standardized score based on the posterior distribution of fold changes and controlling false discoveries through empirical Bayes estimation. Perseus, developed by the same group behind MaxQuant, provides a point-and-click interface for statistical analysis, annotation, and visualization without requiring programming skills.
Results are typically visualized as volcano plots, which display fold change on the horizontal axis and statistical significance on the vertical axis. Proteins that are both substantially changed and statistically significant appear in the upper corners of the plot, making it easy to identify candidates of interest.
Biological Interpretation
A list of differentially expressed proteins is a starting point, not an endpoint. The goal is to understand what biological processes or pathways are affected. Two main approaches help with this.
Gene Ontology (GO) enrichment analysis asks whether your list of significant proteins is enriched for particular biological processes, molecular functions, or cellular locations compared to what you’d expect by chance. If 15 of your 200 significant proteins are involved in oxidative stress response, and you’d only expect 3 by random chance, that’s a meaningful signal.
Pathway analysis maps your proteins onto known signaling or metabolic pathways. KEGG is one of the most widely used pathway databases, organizing proteins into diagrams of interconnected biological processes. DAVID, maintained by the National Institutes of Health, integrates multiple annotation sources and can perform both GO enrichment and KEGG pathway mapping in a single platform. It also groups functionally related proteins into clusters, which helps when your protein list is large and touches many overlapping categories.
String-DB is another commonly used resource that visualizes known and predicted protein-protein interactions, helping you see whether your differentially expressed proteins form connected networks rather than being isolated hits. Proteins that cluster together in an interaction network are more likely to represent a coherent biological response than scattered individual changes.
Handling Missing Values
Missing values are one of the most persistent challenges in proteomics. A protein might go undetected in some samples because it’s truly absent, because its abundance fell below the detection limit, or because of the stochastic nature of DDA sampling. How you handle these gaps matters.
The first decision is whether to filter. A common approach is to require that a protein be detected in at least 50% to 70% of samples in at least one experimental group. Proteins with too many missing values can’t be reliably quantified, and keeping them adds noise.
For the remaining missing values, imputation fills in estimates. Simple approaches replace missing values with a small constant (often the minimum detected value divided by two), assuming that missing means low abundance. More sophisticated methods like k-nearest neighbors or random draws from a downshifted normal distribution attempt to model the likely true values. The choice of imputation method can change which proteins appear statistically significant, so it’s worth testing more than one approach and noting which results are robust across methods.
Reporting and Reproducibility
Transparent reporting of every analysis decision is essential for reproducibility. This means documenting the software versions, database versions, FDR thresholds, normalization methods, imputation strategies, and statistical tests you used. The proteomics community has increasingly emphasized this, recognizing that the number of decision points in the pipeline makes it easy for two analysts to reach different conclusions from the same raw data. Depositing your raw files in public repositories like PRIDE (ProteomeXchange) allows others to reanalyze your data independently, which strengthens confidence in the findings.