The framework integrates established open-source tools, including MCMICRO and SOPA for image processing, QuPath for pathology annotation, napari for visualization, and AnnData and SpatialData for spatial data management. It also incorporates laser microdissection, the physical isolation of cells or regions selected from microscope images, and downstream liquid chromatography–mass spectrometry. Two principal routes are offered. The first, called flashDVP, lets researchers manually annotate visible structures before collecting them, avoiding the need for expertise in computational image segmentation. The second is an extended workflow that supports cell segmentation, marker-based phenotyping, cellular-neighborhood analysis and integration of imaging and proteomic data.

Technical performance was demonstrated on 11 whole-slide images containing 6.34 million cells, which the pipelines segmented in less than two hours. The processing system is designed around standard, nonproprietary file formats so that individual modules can be replaced or expanded as methods improve. A napari plug-in allows users to inspect multichannel images, overlay segmentation masks and adjust marker thresholds with immediate feedback. The plug-in also generates summary tables containing the selected cutoff values for each marker-sample pair, making the filtering strategy visible and reproducible. A separate QuPath-to-LMD web application transfers annotations into laser-microdissection-ready contours and supports validation before collection. The developers tested alignment of single-cell and region contours with a Leica LMD7 system at 20-fold and 40-fold magnifications.

Using flashDVP, the team profiled first-trimester human placenta, where some cell types are difficult to isolate because they lack clear boundaries or resist dissociation into single-cell suspensions. Four-color fluorescence imaging was combined with manual annotation of syncytiotrophoblasts, cytotrophoblasts and Hofbauer cells, followed by laser microdissection and label-free dia-PASEF proteomics. Across three placental tissues, the analysis identified 3,994 unique proteins. Principal-component analysis separated the cell phenotypes regardless of tissue origin, while differential-abundance analysis found markers that broadly agreed with published placental transcriptomic data. Pathway analysis associated syncytiotrophoblast profiles with steroid-hormone metabolism, vesicular transport and energy demands, whereas Hofbauer-cell profiles included immune-regulation and messenger-RNA-splicing pathways.

The same framework was used to connect a spatial-transcriptomic map of early-stage lung adenocarcinoma with protein measurements from a serial tissue section. More than 340,000 cells and 960 cancer-related genes had previously been profiled by CosMx spatial transcriptomics, revealing 18 cell types and several tumor neighborhoods. Researchers selected a region containing tumor cells with an epithelial-to-mesenchymal signature, then collected tumor and stromal regions containing 50–100 cells. The proteomic measurements detected 4,138 proteins in stromal samples and 4,976 in tumor samples, for 5,956 in total. Proteins associated with the transcriptomically defined niche included NDRG1 and LGALS1, while pathway analysis showed elevated glycolysis, hypoxia, angiogenesis and epithelial-to-mesenchymal signaling. SLC2A1, encoding the glucose transporter GLUT1, was the most upregulated protein in the selected niche.

Public datasets provided additional context for the lung-cancer observation, rather than a new clinical trial. The researchers found higher SLC2A1 RNA expression in 483 lung adenocarcinoma tumors than in 347 control samples and examined the protein in Human Protein Atlas immunohistochemistry images. In a survival analysis of 239 patients per group divided at the median, high SLC2A1 expression was associated with lower overall survival, with a reported P value of 0.0003. These analyses support SLC2A1 as a niche-associated candidate for further investigation, but the results do not establish that targeting it would treat lung cancer or improve patient outcomes. The lung-cancer application was based on one archival specimen, with external transcriptomic and protein data used for corroboration.

The researchers also evaluated how tissue architecture changes between a primary triple-negative breast-cancer resection and a locally relapsed tumor collected from the same patient two years after adjuvant chemotherapy. Nine-marker multiplex imaging of 5-micrometer formalin-fixed paraffin-embedded sections produced more than 150 gigabytes of data per whole-slide image stack. The pipelines segmented 610,182 cells in the primary specimen and 1,005,051 cells in the relapse, identifying tumor, T-cell, B-cell, macrophage and stromal populations. Compared with the primary tumor, the relapse had markedly fewer B and CD8-positive T cells, a 16-fold increase in CD68-positive macrophages and a fourfold expansion of vimentin-positive stromal cells. Spatial neighborhood analysis identified seven recurring cellular neighborhoods, with macrophages closer to tumor cells in the relapse and lymphocytes preferentially surrounding ductal carcinoma in situ in the primary specimen.

To test whether these neighborhoods could be profiled efficiently, the team compared three deep-visual-proteomics strategies on adjacent, matched regions. Conventional DVP excised 100 individual cell shapes per sample, whereas cellular-neighborhood DVP, or cnDVP, excised one continuous shape containing 100 phenotypically matched cells. This reduced tissue-collection complexity by approximately 100-fold. Single-cell DVP was tested with individual tumor and immune cells; among 50 single immune-cell samples, only three passed a filter requiring more than 400 identified proteins, while individual tumor cells yielded as many as 1,500 proteins. Neighborhood-guided collection produced the highest average coverage in tumor and immune niches and increased protein identifications by an average of 26% relative to conventional DVP. Most proteins overlapped across methods, and tissue- and niche-specific quantitative signatures were highly correlated, indicating that the faster approach largely preserved the measured biology.

In a larger exploratory analysis, the researchers collected 210 regions from three neighborhood types in the breast-cancer tissues and processed them in a workflow reaching 80 samples per day. After quality control, 203 proteomes remained, each containing about 4,500–5,000 proteins in tumor-enriched neighborhoods and 3,500–4,000 in immune-enriched regions. Unsupervised clustering distinguished primary from relapsed tissue and revealed different protein programs. Immune-enriched regions in the relapse showed increased interferon-gamma signaling, complement and heme metabolism, while tumor-enriched regions showed greater glycolysis and activation of epithelial-to-mesenchymal, MYC and interferon signaling. The primary tumor also displayed greater proteomic variability, lower global correlation and stronger spatial autocorrelation than the relapse, a pattern that also appeared in DVP and scDVP samples collected as individual tumor shapes.

openDVP thus offers three complementary sampling scales: direct annotation of visible compartments, pooled phenotypically coherent cell populations, and larger spatially defined cellular neighborhoods. Each carries a trade-off between resolution, proteome depth, analytical effort and throughput. The neighborhood method may be especially useful for coherent, homotypic tissue structures, whereas rare cell phenotypes and cell-to-cell variation still require finer sampling. Results also depend on tissue quality, staining, imaging performance, annotation or segmentation accuracy, and the efficiency of laser microdissection. The demonstrations are not population-scale clinical studies: the lung and breast applications each involved one patient, while the placental analysis used tissue from three tissues. OpenDVP is therefore best understood as an extensible research platform intended to broaden participation in spatial proteomics, not as a clinically validated diagnostic system.

Subject of Research: Open-source image-guided spatial proteomics of human tissues

Article Title: OpenDVP: an experimental and computational framework for community-empowered deep visual proteomics

Article References: Nimo, J., Fritzsche, S., Valdes, D. S., Trinh, M., Pentimalli, T., Klingeberg, M., Schallenberg, S., Klauschen, F., Herse, F., Florian, S., Rajewsky, N., & Coscia, F. (2026). OpenDVP: an experimental and computational framework for community-empowered deep visual proteomics. Nature Methods. https://doi.org/10.1038/s41592-026-03262-z

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03262-z

Keywords: openDVP, spatial proteomics, deep visual proteomics, open-source software, mass spectrometry, cellular neighborhoods, laser microdissection, spatial transcriptomics, placenta, breast cancer, experimental, computational