The work, published in BMC Bioinformatics by Anna Glodek and Joanna Polańska of the Silesian University of Technology together with Marta Gawin and Monika Pietrowska of the Maria Sklodowska-Curie National Research Institute of Oncology in Gliwice, introduces a two-stage computational pipeline that combines fuzzy logic with machine learning. The key insight is deceptively simple. True members of the same isotopic envelope belong to the same physical molecule, so they must be distributed across the tissue in essentially the same way. Peaks that merely happen to lie at the right mass spacing but originate from different molecules will, in general, paint different pictures on the tissue section. By quantifying how similar two spatial molecular maps are, DeisoLAB can decide whether two peaks are siblings from one molecule or strangers that simply share a mass neighborhood.

To understand why this matters, it helps to appreciate the scale of the problem. In a matrix-assisted laser desorption/ionization mass spectrometry imaging, or MALDI-MSI, experiment, the instrument raster-scans a laser across a tissue section mounted on a conductive slide, acquiring a full mass spectrum at every pixel. A single experiment on a tumor section can generate tens of thousands of spectra, each containing hundreds of peaks. Every peptide generates a characteristic envelope whose peaks are separated by roughly 1.003 daltons, the mass difference between carbon-12 and carbon-13, with intensities that follow a predictable pattern determined by the peptide’s elemental composition. The trouble is that in complex biological samples, peaks from entirely different peptides, or from the matrix compounds used to prepare the sample, can land in exactly those positions, masquerading as envelope members and corrupting downstream identification.

Existing deisotoping tools, most of them developed for liquid chromatography-mass spectrometry, rely almost exclusively on spectral information: the mass spacing between peaks and the expected intensity ratios within an envelope. That approach works reasonably well when the mass resolution is high enough to resolve every peak cleanly, but MALDI-MSI data often come from time-of-flight instruments operating at relatively low mass resolution, where peaks broaden and overlap and the spectral rules alone become ambiguous. The Polish-led team recognized that imaging data contain an extra dimension of information that chromatography-based methods simply do not have, and they built their entire algorithm around it.

The first stage of DeisoLAB is a Mamdani-Assilan fuzzy-inference system, a type of fuzzy logic controller that translates human-expert reasoning into mathematical rules expressed in natural language terms such as high, medium, and low. Fuzzy systems are well suited to this kind of problem because the criteria for judging whether two peaks might belong to the same envelope are inherently graded rather than binary. The system evaluates candidate peak pairs against the expected mass spacing and intensity relationships of isotopic patterns, and instead of issuing a hard yes-or-no verdict, it assigns each pair a degree of membership in the category of plausible envelope candidates. This soft filtering step dramatically reduces the number of peak pairs that need to be examined in the computationally heavier second stage, pruning away obvious non-candidates while keeping borderline cases under consideration.

The second stage is where the spatial magic happens. For every peak that survives the fuzzy filter, the software constructs a spatial distribution map, essentially an image in which the brightness of each pixel corresponds to the measured intensity of that mass-to-charge ratio at that location on the tissue. If two peaks truly belong to the same isotopic envelope, their maps should look nearly identical, because they are two views of the same underlying molecule. To turn that visual intuition into a number, the researchers applied image texture metrics, including measures derived from the gray-level co-occurrence matrix, a classical computer-vision tool that characterizes how pixel intensities are arranged relative to one another in an image. These metrics capture subtle differences in spatial patterning that simple correlation might miss, providing a rich feature vector describing the relationship between any pair of candidate peaks.

Those features then feed into a Naïve Bayes classifier, a probabilistic machine-learning model that, despite its simplicity, performs remarkably well when the underlying features are informative and roughly independent. The classifier was trained to distinguish genuine envelope pairs from coincidental ones, producing a final verdict for each candidate pair. The choice of a lightweight model is deliberate: with strong spatial features doing the heavy lifting, an elaborate deep-learning architecture is unnecessary, and the resulting tool remains fast and interpretable enough for routine laboratory use on large imaging datasets.

The performance figures reported in the paper are impressive. The method was evaluated on eight MALDI-MSI datasets spanning both fresh-frozen and formalin-fixed paraffin-embedded tissues, the two dominant preparation formats in clinical and biomedical imaging studies, including data from head and neck cancer research. Across these datasets, the spatial-distribution-based classification stage achieved recall values between 88.12 and 94.12 percent, meaning it correctly identified the vast majority of true envelope members, and precision values between 72.95 and 85.71 percent, indicating that most of its positive calls were correct. Perhaps most striking is the specificity, which exceeded 99 percent, meaning the tool almost never wrongly labels an unrelated peak as an envelope member. The overall deisotoping accuracy reached 96.98 percent, a figure that translates directly into cleaner peak lists and more confident peptide identifications downstream.

The implications extend well beyond a single laboratory’s convenience. Deisotoping quality is a bottleneck that quietly shapes every peptide-level conclusion drawn from imaging experiments, from biomarker discovery in oncology to mapping of molecular changes in neurodegenerative disease. False envelope assignments inflate the apparent complexity of a sample and can lead researchers to chase phantom peptides, while missed envelope members suppress the signals of real ones. Because DeisoLAB’s spatial criterion works even when mass resolution is modest, it could open peptide-level imaging studies to laboratories using less expensive time-of-flight instrumentation, democratizing a technique that has often demanded top-tier hardware to produce reliable proteomic results.

There is also a conceptual lesson here for the broader field of computational mass spectrometry. Most analytical software inherited from liquid chromatography treats each spectrum as an isolated object, discarding the spatial context that makes imaging experiments unique. DeisoLAB demonstrates that the tissue itself is a source of analytical information, a kind of built-in validation layer in which the biology of the sample corroborates or refutes the instrument’s spectral evidence. As mass spectrometry imaging continues its march toward clinical applications, including tumor margin assessment and personalized treatment planning, tools that fuse spectral and spatial evidence, as this fuzzy-logic-and-machine-learning pipeline does, are likely to become standard components of the imaging workflow rather than optional refinements.

The software arrives at a moment when the imaging community is actively wrestling with reproducibility and standardization, and an open, spatially aware deisotoping method addresses one of the noisiest steps in the pipeline. The authors acknowledge colleagues at the Silesian University of Technology who verified the fuzzy-inference design and provided pre-processed data, and the work was supported by institutional research grants from the same university. For researchers drowning in the thousands of peaks that every tissue section produces, the message of this study is refreshingly concrete: stop treating your imaging data as a pile of spectra, and start reading the maps, because the spatial pattern of a molecule is a fingerprint that coincidental peaks cannot easily forge.

Subject of Research: Computational identification of isotopic envelopes in MALDI mass spectrometry imaging data using fuzzy inference and spatial distribution analysis

Article Title: DeisoLAB – isotopic envelope identification by analysis of the spatial distribution of peptides in MALDI-MSI data

Article References: Glodek, A., Gawin, M., Pietrowska, M., & Polańska, J. (2026). DeisoLAB – isotopic envelope identification by analysis of the spatial distribution of peptides in MALDI-MSI data. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06642-6

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06642-6

Keywords: mass spectrometry imaging, MALDI-MSI, isotopic envelope, deisotoping, fuzzy-inference system, machine learning, Naïve Bayes classifier, peptide imaging, image texture metrics, proteomics, bioinformatics, cancer research