Cell typing and malignant calls for single-cell, spatial and H&E data

h5ad in, 425 cell types out.

Fast, simple, accurate, consistent across 10x Chromium, CosMx, Xenium, Visium HD, Singular & H&E. It’s really that easy.

Try MiraTyper free
425
Cell Types
~250M
Cells in Source Collection
<1 day
New Panel Onboarding

Drag to compare. Malignancy and lymphocyte calls from Atera transcriptomics, alongside the matched H&E section.

The Challenge

Cell Annotation Is a Cornerstone of Biology

Understanding a tissue starts with knowing which cells are present. Bad annotations compromise your downstream analysis. Consistent annotation enables researchers to compare cell types and states confidently across datasets, instruments, and studies.

Manual

Slow and Subjective

Hand-labeling can consume weeks of scientist time, and annotations can vary between analysts.

Reference

Reference-Dependent

Label-transfer tools inherit the gaps and biases of the available reference, and need a tissue-specific model each time.

Batch

Batch-Sensitive

Accuracy drifts across platforms, donors, and runs, so results stop being comparable when conditions change.

Supported Platforms

Start with the Data You Already Have

Available models and validation depend on the assay and panel.

Technology Supported data and workflows
Xenium Spatial RNA panels: 5K, 500, and Immuno-Oncology 380.
CosMx Spatial RNA panels: 1K and 6K.
Atera Whole-transcriptome spatial RNA (Xenium WTA).
Singular G4X Spatial RNA and protein panels; lung-panel benchmarks available.
Visium Spatial gene expression for cell typing.
Visium HD Spatial gene expression with H&E-based cell segmentation. GPU required for image-based segmentation.
10x Chromium Single-cell RNA, 3′ and 5′, supplied as AnnData.
H&E Whole-slide tissue measurements and image-based cell segmentation. GPU required.
Takara Seeker Prototype Spatial RNA cell typing demonstrated in mouse brain tissue.
See panel and benchmark technical details
The MiraTyper Advantage

One Tool, Deterministic by Design

MiraTyper returns the same labels on every run, across tissues, platforms, and species.

Repeatable

Deterministic

The same input yields the same cell types and states on every run. MiraTyper is not LLM-based, so there are no tokens and no run-to-run variance.

Efficient

Laptop-Class Hardware

Transcriptomics inference runs on CPU: roughly 300–480 cells per second on a laptop in whole-slide runs. H&E models, including image-based segmentation for Visium HD, require a GPU.

Universal

No Model Selection

The same model runs on every tissue. It reads the full panel rather than a few markers, so it is robust to dropout, low sensitivity, and segmentation errors.

See what MiraTyper finds in your data

Add your details, pick a time and we will run two datasets for free. Quick and easy.

  1. 1. Your details
  2. 2. Choose a time

Differentiation

Build a Biological Coordinate System

Cell types and states are the join key across every supported modality. When each platform is typed the same way, results from different instruments align instead of requiring reconciliation.

Measurement Layer

Integration happens where measurements are made, not after each platform has produced its own report, so biology and context survive the join.

One Ontology

The same 425-type Cell Ontology spans single-cell, spatial, protein, and H&E data, so an instrument change does not reset the analysis.

Re-queryable

Existing data answers new questions. Adding a platform or a primitive does not invalidate earlier measurements.

Agent-Ready

Skills and tools are deterministic and token-efficient, and plug into an existing agentic system or one MiraOmics builds.

Benchmarks

Measured, Per Panel, in Public

Every supported panel is benchmarked separately, and the scoring rules are published with the numbers. Nothing here is averaged into a single headline figure.

Cell type

0.894–0.909 Hierarchical F1

Across supported gene panels on a held-out test set of 110,952 cells and 425 classes. On real vendor-labeled tissue the figure is lower, and that number is published too.

Malignancy · historical benchmarks

0.965–0.996 AUROC

Earlier v1.3-era panel evaluations, using different test and validation sets. Released v1.9 models include colorectal training data; performance and thresholds should be validated for the intended tissue and assay.

Against other tools

Highest F1 at Every Stringency

Compared with four published annotation tools across seven datasets, six tissues, and about a million cells, at three levels of label-matching strictness.

See the per-panel tables and scoring rules
How It Thinks

Labels on a Tree, Not a Flat List

MiraTyper predicts into the Cell Ontology, so a near miss is scored as a near miss rather than a total error.

cell› leukocyte› lymphocyte› T cell› CD8-positive T cell

425 cell types arranged across the standardized Cell Ontology, plus a dedicated malignant-versus-normal classifier. Every label arrives pre-harmonized, so results merge cleanly across studies.

Data

~250M Cells

Our source collection contains about 250 million cells. Training uses a carefully curated subset with harmonized Cell Ontology labels. Recreating this training resource requires the selection, curation, and label harmonization across source studies, as well as the data itself.

Labels

Ontology and Confidence

Pre-harmonized Cell Ontology labels, each with a confidence score.

Reference-Free

No Reference Set

Nothing to curate or transfer from. Custom cell types can be fine-tuned in under an hour.

Multi-Modal

RNA, Protein, H&E

Transfer learning carries the same label space into protein panels and H&E.

Primitives and Skills

Build Each Detector Once, Reuse It Everywhere

Primitives detect one thing reliably. Skills combine them into analyses. Both share one data structure, so they work together across instruments and modalities.

Primitives

Validated detectors, with typical build effort.

  • LymphocytesH&E · any H&E scanner · ~1 week to build Available
  • Malignant cellsH&E · any H&E scanner · ~1 week Available
  • 425 normal cell typesFull transcriptome · 10x, Visium, Atera · ~1 month Available
  • Cancer cellsFull transcriptome · 10x, Visium, Atera · ~1 week Available
  • Cancer vs. normalTranscriptome panels · Xenium, CosMx, Singular · ~1 week Available
  • H&E segmenterImage · H&E, Visium, Xenium, Singular · ~1 month Available
  • Normal cell detectorProtein + RNA panel · Singular, Xenium · ~1 week Available
  • PSMA expressionImage · predicted from H&E · ~1 week Early

Skills

Analyses composed from primitives, light on compute and tokens.

  • Triad detectorCD4 + CD8 + dendritic-cell communities · ~1 day Available
  • TLS detectorTertiary lymphoid structures · ~1 week Available
  • Niche detectorRecurring cellular neighborhoods · ~1 week Available
  • Report generatorPer-patient and cohort reports · ~1 day Available
  • Data manager~1 week to build Available

In Development

Planned primitives and the applications they open up.

  • HER2 protein and RNA from H&E Planned
  • MSI / MMR colorectal structure from H&E Planned
  • Tumor-infiltrating lymphocytes from H&E Planned
  • Tumor–stroma ratio from H&E Planned
  • Hot vs. cold tumors from H&E Planned
  • EGFR, KRAS, TP53 mutation status from H&E and RNA Planned
  • HRD / BRCA-ness: ovarian, breast, prostate, pancreatic Planned
  • CAF content, desmoplasia, macrophage polarization Planned
  • Normal and cancer cells from protein Planned
  • Mouse normal and cancer cells from RNA Planned
  • Glioblastoma classifier for CosMx Planned
  • MASH fibrosis and steatohepatitis scoring Planned
  • IBD, kidney, and lung fibrosis activity Planned
Across Platforms

The Same Engine on Different Instruments

The label space stays fixed while the data source changes, which is what makes niche, ligand–receptor, and trajectory analyses comparable across studies.

Spatial map of a cervical tumor with cells colored by MiraTyper type and circles marking CD4–CD8–dendritic-cell triads
Atera Figure 1. Immune triads (CD4 + CD8 + dendritic cells) in a cervical tumor, a structure linked to checkpoint response. The analysis ran one hour after download, with no retraining.
Cell boundaries drawn over a Visium HD H&E image, separating individual nuclei
Visium HD Figure 2. MiraTyper uses the H&E image to identify cell boundaries, combines the Visium HD expression measurements within each cell, and assigns a cell type. Destriping improves the map for visual review; it did not significantly improve cell-typing accuracy in this evaluation.
Spatial map of a mouse brain section with cells colored by fine-grained predicted cell type
Takara Seeker (Prototype) Figure 3. Cell types predicted in a mouse brain section from Takara Seeker data using the prototype mouse model. The map distinguishes cortical layers, white matter, and cerebellum from the cell-type predictions, without anatomical region labels supplied to the model.

Put MiraTyper on Your Own Data

Available as a Docker package for your own environment: CPU for transcriptomics, GPU for H&E, including on an air-gapped network.

Try MiraTyper free