Every vendor claims the leading system, so comparing spatial transcriptomics platforms on marketing material gets you nowhere. Two peer-reviewed benchmarks published within the last year changed that. Both ran competing platforms on matched serial sections from the same tissue blocks, which is the only way to separate platform performance from sample variation.
Key Takeaways
|
Platform-by-Platform Overview
Five systems dominate procurement conversations, and they do not all do the same job. Two are sequencing-based, capturing probes onto a barcoded slide for readout on a sequencer. Two are imaging-based, detecting transcripts in place on the instrument. One profiles user-defined regions rather than resolving individual cells. The underlying distinction between the sequencing and imaging families is covered in Imaging-Based vs. Sequencing-Based Spatial Transcriptomics: Which Fits Your Lab?.
Platform | Vendor | Family | Coverage | Notable Characteristic |
Visium / Visium HD | 10x Genomics | Sequencing-based | Whole transcriptome | HD uses a 2 micron barcoded grid, approx. 11 million features per 6.5 x 6.5 mm capture area |
Xenium | 10x Genomics | Imaging-based | Targeted panels, 377 and 5,000-gene versions benchmarked | Continuous scan up to 1.1 x 2.4 cm; padlock probes with rolling circle amplification |
CosMx | Bruker | Imaging-based | Standard 1,000-gene panel with add-ons; whole-transcriptome panel added since benchmarking | Discrete user-defined 0.5 x 0.5 mm fields of view; branch chain hybridization amplification |
MERSCOPE | Vizgen | Imaging-based | Fully customizable or standard panels with add-ons | Direct probe hybridization amplified by tiling each transcript with many probes |
GeoMx | Bruker | Region-of-interest profiling | Targeted, region-level rather than single-cell | Profiles defined regions; not directly comparable to the others on single-cell metrics |
Table 1. Platform overview. Coverage figures describe the configurations used in the benchmarks cited below, not full product specifications. Panel options change frequently; confirm current specifications with vendors before procurement.
The amplification chemistry differences in the last column are worth understanding, because they explain the sensitivity results that follow. As the Nature Communications team describes it, Xenium uses a small number of padlock probes with rolling circle amplification, CosMx uses a low number of probes amplified by branch chain hybridization, and MERSCOPE uses direct probe hybridization amplified by tiling each transcript with many probes. Three different routes to the same goal, with measurably different outcomes.
Visium vs. Xenium: How Do the Two Most-Compared Platforms Differ?
These two draw more comparison than any other pairing, partly because they come from the same vendor and are often evaluated together. They are not really competitors. Visium tells you what is there, and Xenium tells you exactly where it is.
Visium captures probes onto a spatially barcoded slide and reads them out by sequencing, giving whole-transcriptome coverage with no need to choose targets in advance. Xenium images transcripts in place on the instrument, resolving them to individual cells but only for a pre-selected panel. The technical comparison across six cancer types published in Genome Biology in January 2026 ran both on matched sections and found gene-level correlation above 0.85 between them, so they broadly agree about what is present. Where they diverge is density and resolution: Visium HD showed UMI density roughly 15 times lower than Xenium, while offering the whole transcriptome that Xenium’s panel cannot.
For Visium HD specifically, the practical resolution is lower than the specification suggests. The native 2 micron grid carries what the authors call extraordinary computational cost, so features are usually aggregated into 8 or 16 micron bins. At 8 microns, bins frequently contained transcripts from 2 to 4 cells, and at 16 microns from 5 to 10. Xenium avoids that problem but introduces a different one, since transcripts must be assigned to cells by image segmentation, and segmentation errors misassign them.
Two further findings matter for interpretation. Visium HD over-represented epithelial and stromal cells relative to immune cells, because larger cells span more bins and get counted more than once. And whole-transcriptome coverage let the same study score pathway activity robustly and identify macrophage subpopulations that the targeted platform missed entirely. If your question is pathway-level or you do not yet know which genes matter, that is the argument for sequencing-based capture regardless of resolution.
Lab Management Certificate
The Lab Management certificate is more than training—it’s a professional advantage.
Gain critical skills and IACET-approved CEUs that make a measurable difference.
Resolution, Plex, and Sensitivity Compared
The most useful cross-vendor evidence comes from the systematic benchmarking of imaging platforms in FFPE tissues published in Nature Communications in November 2025, which ran Xenium, MERSCOPE, and CosMx on matched serial sections from tissue microarrays spanning 17 tumor and 16 normal tissue types, imaging 263 cores in total. Its headline result is consistent with the Genome Biology work, which is what makes it persuasive.
Finding | Source and Operational Meaning |
On matched genes, Xenium generated higher transcript counts per gene without sacrificing specificity | Nature Communications, across all three imaging platforms. Two independent studies now agree on this, which is stronger than either alone |
Xenium showed far lower background than CosMx: negative probe rate 0.06% against 10.31% | Genome Biology. Signal-to-noise was roughly 450 versus roughly 3, and CosMx low-abundance transcripts were sometimes indistinguishable from background |
CosMx detected 2 to 3 times more genes and transcripts per cell using full panels | Genome Biology. Consistent with its larger 1,000-gene panel. On the 125 genes shared with Xenium, the advantage disappeared |
CosMx segmentation was more anatomically plausible, using nuclear plus membrane staining | Genome Biology. Xenium in the version tested expanded computationally from DAPI nuclei, inflating cell size in sparse regions |
CosMx detected MMP7, absent from the smaller Xenium panel, enabling spatial subtyping that would otherwise have been missed | Genome Biology. A concrete illustration that panel content, not just platform, determines what you can conclude |
All three imaging platforms performed spatially resolved cell typing, with Xenium and CosMx finding slightly more clusters | Nature Communications. Coarse cell typing is fairly robust to platform choice; fine distinctions are not |
A 5,000-gene panel produced roughly 3 times fewer transcripts per gene than a 377-gene panel on the same platform | Genome Biology. Optical crowding. MYC-positive cancer cell calls moved from 70% to 30% between panels on one sample |
Table 2. Findings from two independent peer-reviewed benchmarks. Genome Biology data reflect mid-2025 capabilities; the authors note CosMx has since added a whole-transcriptome panel. Neither study is intended as a platform recommendation.
Read that table as a set of trade-offs rather than a scoreboard. Xenium leads on per-gene sensitivity and background. CosMx leads on segmentation plausibility and standard panel breadth, and its larger panel caught biology a smaller panel could not. Both studies state explicitly that they are technical references rather than buying recommendations, and both note that coarse cell type composition estimates were broadly similar across platforms, which means platform choice matters most when you need fine distinctions.
Interested in life sciences?
How Do Sample Format and Throughput Differ?
This is where procurement decisions are actually made, and where the most useful finding in either paper sits. The platforms do not ask the same thing of your samples.
Platform | Sample QC Guidance | Area Covered | Throughput Consideration |
Xenium | H&E pre-screening suggested | Continuous scan, up to 1.1 x 2.4 cm | Approx. 50 hours for a 377-gene run, approx. 7 days for a 5,000-gene run |
CosMx | H&E pre-screening suggested | Discrete 0.5 x 0.5 mm fields of view, user-selected | Field of view count drives run length and total area covered |
MERSCOPE | DV200 above 60% recommended | Defined imaging region | Panel design flexibility may add lead time before the first run |
Visium / Visium HD | DV200 assessment before library prep | 6.5 x 6.5 mm capture area, larger areas available | Slide prep on site, then sequencer queue and heavy compute |
GeoMx | Region selection from stained section | User-defined regions of interest | Suits many samples with few regions each, rather than whole-section depth |
Table 3. Sample and throughput requirements. QC guidance as reported in the Nature Communications benchmark and vendor documentation. Run times are what the Genome Biology study observed, not vendor specifications, and depend on scanned area and configuration.
The Sample QC Difference Is a Real Procurement Filter The Nature Communications team notes that CosMx and Xenium suggest pre-screening samples on H&E, while MERSCOPE recommends a DV200 above 60%. DV200 is the proportion of RNA fragments longer than 200 nucleotides, and it falls as blocks age and as fixation conditions vary. If your intended cohort is a decade-old biobank, that single line may decide your platform before any performance figure does. For comparison, the Genome Biology group screened blocks above 34% DV200 for their sequencing-based work. Pull DV200 on a representative sample of your actual blocks before you shortlist, not after. Both studies also found detection correlated strongly with RNA quality on every platform tested, so a poor archive limits all of them, just not equally. |
One further throughput consideration favours continuous scanning over field-of-view selection. In the Genome Biology work, a single 0.5 by 0.5 mm field of view produced cancer cell proportion estimates ranging from 5% to 55%, against 22% for the whole section. If your study design depends on composition estimates, scanned area is not a convenience feature.
Consumable Economics
Vendors do not publish list prices, but academic core facilities often must publish their rates, which makes some of the economics visible. The Yale Keck Microarray Shared Resource rate card listed sequencing-based library preparation at $2,141 per 6.5 mm section and $4,151 per 11 mm section for internal users as of July 2025, with the high-definition chemistry at $3,478 and GeoMx library preparation at $768 per sample. Those are service rates including reagents, not instrument prices, and external users pay considerably more.
Two observations follow. First, capture area is a direct cost multiplier on sequencing-based workflows: the 11 mm area cost nearly twice the 6.5 mm area on the same rate card. Second, GeoMx sits at a materially different price point per sample, which is consistent with its design intent of profiling more samples at lower depth rather than resolving one section exhaustively. Per-section imaging platform rates are not published in comparable form, so those need to come from your own quotes. Full cost category modeling, including the reagent share of per-section cost and an embedded estimator, is in How Much Does Spatial Biology Cost? Budgeting for Instruments, Consumables, and Data.
One economic point the benchmarks make indirectly: a failed run costs the same as a successful one. Given that detection tracks RNA quality on every platform, the cheapest available saving is screening blocks before committing reagents to them.
Which Platform Fits Which Use Case?
Match the platform to the question and the constraint. These are starting points, not conclusions.
If Your Situation Is | Start With |
Targets unknown, or pathway-level analysis needed | Visium or Visium HD, for whole-transcriptome coverage |
Defined targets, maximum per-gene sensitivity needed | Xenium, on the smallest panel covering your targets |
Cell segmentation accuracy is critical to the biology | CosMx, whose membrane staining produced more plausible boundaries |
You need a fully custom panel | MERSCOPE or Xenium, both of which offer custom panel design |
Old or degraded FFPE archive | Check DV200 first. MERSCOPE’s stricter guidance may rule it out |
Many samples, few regions of interest each | GeoMx, which is designed for exactly that shape of study |
No storage, compute, or analyst capacity yet | None of them. Outsource this project and close the gap first |
Table 4. Best-fit starting points. Sample type, tissue area, budget, and existing infrastructure all constrain the answer further.
The last row is not a formality. Both benchmarks generated datasets far beyond what a benchtop workstation handles, and every platform here pushes work downstream onto storage, compute, and analyst time. Work through the procurement process in Choosing Between Spatial Biology Platforms: A Lab Manager's Buyer's Guide before committing, and see the operational picture across the whole workflow in Spatial Biology in the Lab: A Manager's Guide to Evaluating, Implementing, and Scaling Spatial Technologies.
This article was produced under Lab Manager's AI Editorial Guidelines.
