XCMS is a free, open-source software package designed to process raw mass spectrometry data from metabolomics experiments, turning messy instrument output into a structured table of molecular features that researchers can analyze statistically. First released in 2006, it has become one of the most widely used tools for liquid chromatography–mass spectrometry (LC-MS) data preprocessing, available both as an R package for programmers and as a web-based platform called XCMS Online for those who prefer a point-and-click interface.1PubMed. Metabolomics Data Processing Using XCMS The software handles the core grunt work of metabolomics: finding peaks in raw data, lining up chromatographic runs from different samples so they can be compared, and filling in gaps where a compound was present in one sample but the instrument missed it in another. What makes it worth understanding, even if you never plan to run it yourself, is how central these preprocessing steps are to whether downstream biological conclusions hold up.
Why Raw LC-MS Data Needs Processing in the First Place
A mass spectrometer coupled to a liquid chromatography system produces an enormous amount of raw data per sample. Molecules elute off the chromatographic column at slightly different times across runs, ionize into multiple species (different charge states, sodium or potassium adducts, in-source fragments), and sit on top of background chemical noise. A single biological sample might generate hundreds of thousands of data points. The goal of preprocessing software like XCMS is to reduce that torrent to a clean table where each row is a sample, each column is a molecular feature (defined by a mass-to-charge ratio and a retention time), and each cell contains an intensity value representing how much of that feature was present. Without automated processing, comparing metabolite levels between a disease group and a healthy control group across dozens or hundreds of samples would be practically impossible.
How XCMS Detects Peaks
The first and arguably most consequential step in any XCMS workflow is peak detection, also called feature detection. XCMS offers several algorithms for this, but the one most commonly used for modern high-resolution instruments is called centWave. The centWave algorithm works in two stages. It first scans the raw data to find regions of interest, which are partial mass traces where a consistent mass-to-charge signal persists across consecutive scans. It then applies a mathematical technique called continuous wavelet transformation to identify genuine chromatographic peaks within those regions, separating real signal from noise and resolving overlapping peaks that elute close together.2PubMed Central. Highly sensitive feature detection for high resolution LC/MS
The sensitivity of this step matters enormously. One post-processing study found that roughly 35% of peaks detected by XCMS in their dataset were low-quality, mostly exhibiting poor signal-to-noise ratios, and needed to be filtered out by additional quality checks.3PubMed Central. Comprehensive Peak Characterization (CPC) in Untargeted LC-MS Analysis That does not mean XCMS is doing something wrong. Casting a wide net at the detection stage and then filtering afterward is a deliberate strategy in untargeted metabolomics, where missing a real compound is often worse than including some false positives that can be cleaned up later. But it does mean that the raw feature list XCMS produces is not the final answer. Researchers need to apply their own quality filters before drawing biological conclusions.
Retention Time Correction and Sample Alignment
Even in a well-controlled experiment, the exact moment a compound comes off the chromatographic column drifts slightly from one injection to the next. Column aging, small temperature fluctuations, mobile phase preparation differences, and instrument downtime all contribute. If you simply matched features across samples by exact retention time, you would split what is really the same compound into multiple features and lose statistical power. XCMS addresses this with a retention time correction step that estimates and removes systematic drift across the full set of samples.4PubMed Central. Data processing, multi-omic pathway mapping, and metabolite activity analysis using XCMS Online
After correction, features detected in individual samples are grouped across samples based on their corrected retention time and mass-to-charge ratio. The grouping step produces the final feature table. Where a feature was detected in some samples but not others, XCMS goes back to the raw data and tries to integrate signal at the expected location, a step called “fillPeaks” or gap-filling. This prevents a zero value from appearing simply because the peak-picking algorithm missed a low-abundance compound in a particular run. Whether that rescued signal is real or just noise is something downstream statistical analysis has to sort out.
The Parameter Problem
One of the most common frustrations new users encounter is that XCMS results are highly sensitive to the parameter values chosen for each processing step. The peak detection algorithm, for instance, requires the user to specify expected peak width ranges, a signal-to-noise threshold, and mass accuracy tolerances. The alignment and grouping steps have their own parameters controlling how much retention-time drift to tolerate and how close in mass two features must be to be considered the same compound. Getting these wrong can dramatically change how many features are reported and how accurate their quantification is.
A study that systematically optimized both instrument and XCMS parameters showed just how large the effect can be. For a set of dilute standard samples, optimization improved the median coefficient of variation from 24% down to 7%, retention time fluctuation dropped from about 9 seconds to half a second, and mass accuracy improved by more than threefold.5PubMed Central. Automated optimization of XCMS parameters for improved peak picking of liquid chromatography-mass spectrometry data using the coefficient of variation and parameter sweeping for untargeted metabolomics The researchers achieved this through a brute-force sweep of parameter combinations, which works but is computationally expensive.
A more elegant solution is the IPO package (Isotopologue Parameter Optimization), which automates the tuning of XCMS settings. IPO uses naturally occurring carbon-13 isotope peaks as internal quality markers. If peak detection is working well, the software should consistently find isotope pairs at the expected mass difference and intensity ratio. IPO iteratively adjusts XCMS parameters to maximize a score based on this isotope information, freeing the user from having to manually test dozens of parameter sets.6PubMed Central. IPO: a tool for automated optimization of XCMS parameters It also optimizes retention time correction and grouping parameters using pooled quality-control samples, where the same biological material is injected repeatedly and should ideally produce identical results.
XCMS Online and the Accessibility Question
XCMS was originally built as an R package, which means it lives in a programming environment. Users write scripts, call functions, and debug code. For analytical chemists and bioinformaticians, this is fine and arguably preferred because scripted workflows are reproducible and customizable. But metabolomics is increasingly used by biologists, clinicians, and environmental scientists who do not write code. XCMS Online was created specifically to bridge that gap. It provides a web-based graphical interface where users upload their raw data files, select parameters from dropdown menus, and launch processing jobs on a remote server.7PubMed Central. XCMS Online: a web-based platform to process untargeted metabolomic data
The cloud-based platform also bundles statistical analysis and visualization tools that would otherwise require separate software. Users can perform pairwise or multi-group comparisons, view interactive plots of their results, and even run metabolic pathway mapping to connect their features to known biochemical routes.4PubMed Central. Data processing, multi-omic pathway mapping, and metabolite activity analysis using XCMS Online An interactive version further extended this with features for multivariate statistical analysis and meta-analysis across multiple datasets.8PubMed Central. Interactive XCMS Online: simplifying advanced metabolomic data processing and subsequent statistical analyses
The trade-off is flexibility. Power users running the R package can insert custom code at any step, swap in alternative algorithms, or connect XCMS to the broader Bioconductor ecosystem of analysis tools. XCMS Online gives you a curated set of options with less room for customization. For many routine analyses, that is perfectly adequate. For methods development or unusual experimental designs, the R package remains the standard.
How XCMS Compares to Other Tools
XCMS is far from the only game in town. MZmine, MS-DIAL, and Compound Discoverer (a commercial tool from Thermo Fisher) are the most commonly compared alternatives. An evaluation that tested all four on the same spiked bovine saliva samples found that only about 8% of detected features were shared across all four software packages, which is a striking illustration of how much preprocessing choices affect the final dataset.9PubMed. Modular comparison of untargeted metabolomics processing steps Among the features that did overlap, MS-DIAL showed the closest match to manual integration, while XCMS and MZmine also performed well. Compound Discoverer struggled with peaks sitting on high baselines.
An earlier benchmark comparing five tools found that all showed similar ability to detect genuine features from known spiked compounds, but they diverged sharply on quantification accuracy. MZmine outperformed the others on the accuracy of relative quantification and on identifying true discriminating markers with fewer false positives.10PubMed. Comprehensive evaluation of untargeted metabolomics data processing software in feature detection, quantification and discriminating marker selection These comparisons do not mean one tool is categorically “best.” Performance depends on the data type, the instrument, the chromatographic method, and how well each tool’s parameters have been tuned. XCMS remains the most-cited option, in part because of its long track record and its integration with the R ecosystem, but researchers running critical analyses increasingly process data through more than one pipeline and look for agreement.
Companion Tools and the Broader Ecosystem
A raw XCMS feature table is a list of mass-to-charge ratios, retention times, and intensities. It does not tell you which features come from the same compound. A single metabolite can produce a dozen or more signals in LC-MS: the protonated molecule, a sodium adduct, a dehydration product, in-source fragments, and various isotope peaks. Treating each of these as a separate feature inflates the feature count and complicates statistical interpretation. The CAMERA package, built to work directly with XCMS output, addresses this by grouping co-eluting features and annotating which ones are likely isotopes, adducts, or fragments of the same parent compound.11PubMed Central. CAMERA: an integrated strategy for compound spectra extraction and annotation of liquid chromatography/mass spectrometry data sets The result is a shorter, more biologically meaningful list of putative compounds rather than a sprawling list of redundant signals.
The open-source, community-driven design of XCMS has encouraged a wide ecosystem of compatible packages.12PubMed Central. xcms in Peak Form: Now Anchoring a Complete Metabolomics Data Preprocessing and Analysis Software Ecosystem Real-time processing wrappers like SimExTargId, for example, automate the entire pipeline from vendor format conversion through XCMS peak-picking, CAMERA annotation, and statistical analysis, running autonomously while an experiment is still being acquired on the instrument.13Bioinformatics. SimExTargId: a comprehensive package for real-time LC-MS data acquisition and analysis This kind of automation is valuable in clinical or high-throughput settings where fast turnaround matters.
Beyond Liquid Chromatography
Although XCMS is most commonly associated with LC-MS, the software is not limited to liquid chromatography. It can handle gas chromatography–mass spectrometry (GC-MS) data as well, and it works with both centroid and profile mode acquisitions at any resolution level, from unit-resolution quadrupole instruments to ultrahigh-resolution Orbitrap and FT-ICR systems. Community training resources, such as those provided through the Galaxy bioinformatics platform, walk users through GC-MS-specific workflows that pair XCMS peak detection with downstream tools for retention index assignment and spectral matching against libraries. The core algorithms do not care what separated the molecules before they reached the mass spectrometer; they care about finding consistent signals in the data and aligning them across runs.
Batch Effects and Quality Control After XCMS
Running XCMS produces a feature table, but that table typically contains systematic artifacts that must be corrected before analysis. The most common is batch effects: gradual signal drift within an analytical run and jumps in intensity between runs performed on different days. Instruments lose sensitivity over the course of a long sequence as the ion source gets dirty, and recalibration between batches can shift absolute intensities. Batch correction methods applied after XCMS processing try to remove these trends, often using pooled quality-control samples injected at regular intervals as reference points.
How missing values are handled in this correction matters more than many researchers realize. One study of batch correction approaches found that replacing non-detected values with the limit of detection produced the best results, while substituting zero or half the detection limit led to clearly worse performance.14PubMed Central. Improved batch correction in untargeted MS-based metabolomics This kind of detail rarely appears in published methods sections, but it can influence which metabolites end up being called statistically significant.
From Features to Biological Meaning
A processed and corrected feature table is still just a list of numbers. Figuring out what those numbers mean biologically requires annotation (assigning putative identities to features) and, in many workflows, pathway analysis to see whether the changes cluster in particular metabolic routes. XCMS Online integrates a systems-biology workflow that maps features onto known metabolic pathways and supports integration with genomic and proteomic datasets to build a broader picture of what is going on in a biological system.4PubMed Central. Data processing, multi-omic pathway mapping, and metabolite activity analysis using XCMS Online
It is worth keeping expectations realistic about this step. In untargeted metabolomics, even the best preprocessing pipelines typically assign a confident identity to only a fraction of detected features. Many features remain annotated only at the level of “something with this mass exists,” without knowing whether it is a lipid, an amino acid derivative, or an artifact. Pathway enrichment analyses built on incomplete annotations can produce misleading results if the unidentified features happen to belong to the same pathway being tested. XCMS and its companion tools handle the data processing competently; the biological interpretation is where the hard thinking still falls to the researcher.
Data Standards and Reproducibility
Metabolomics as a field has been pushing toward standardized, reproducible data handling, and the open formats that XCMS reads are part of that effort. The software accepts vendor-neutral formats like mzML and mzXML, which are XML-based representations of raw mass spectrometry data. Converting proprietary instrument files into these formats (usually with the free ProteoWizard toolkit) is typically the first step in any XCMS workflow. The existence of these shared formats is what allows a single software package to process data from instruments made by different manufacturers.
On the reporting side, community initiatives like MetaboLights provide public repositories where researchers can deposit raw data and processing metadata in accordance with FAIR principles, meaning data should be findable, accessible, interoperable, and reusable.15Nucleic Acids Research. MetaboLights: open data repository for metabolomics When researchers deposit their XCMS parameter settings alongside raw data, anyone can reprocess the experiment and verify the findings. This kind of transparency is still aspirational in much of the field, but the infrastructure to support it exists and is slowly becoming the expectation rather than the exception.
When XCMS Might Not Be the Right Choice
For all its dominance, XCMS is not always the ideal starting point. Targeted metabolomics experiments, where researchers already know exactly which compounds they want to measure, typically use vendor software or specialized tools optimized for extracting known transitions with maximum precision. XCMS was designed for untargeted or semi-targeted work, where the goal is discovery rather than precise quantification of a predefined panel.
Studies involving very large cohorts (thousands of samples) can also strain the standard XCMS workflow. Memory requirements grow with sample count, and retention time correction algorithms that work well for 50 samples may struggle or slow to a crawl with 5,000. Workarounds exist, including processing samples in batches and then merging feature tables, but they introduce their own alignment challenges. Some commercial platforms and newer open-source tools have been engineered from the ground up for large-scale studies, which can make them a better practical choice even if XCMS remains the more cited option in the literature.
Researchers working with imaging mass spectrometry or direct-infusion approaches (where there is no chromatographic separation) will also find that XCMS is not designed for their data. The software’s algorithms assume a chromatographic time axis, and without one, the peak-detection and alignment steps do not apply. Other tools handle those specialized data types more naturally.