VivOME
Two omes. One living space.
A versioned atlas that places RNA and protein measurements of single cells in one shared space, and says so when it can't.
Scroll to see how
Every cell is written twice.
Its RNA shows which genes are switched on. Its proteins show what the cell is actually doing.
Single-cell RNA sequencing has large, labelled references to read the first against. Single-cell proteomics is younger: its measurements are sparser, its panels differ from lab to lab, and it has few labelled references.
So a protein measurement of a single cell often arrives without a name.
One space for both.
VivOME is a joint latent atlas: one coordinate space where RNA and protein measurements of single cells sit side by side, so a cell measured either way is read against the same reference.
The reference is built from RNA cells across immune and blood cell types, over one shared gene feature space:
- RNA cells in the reference
- immune and blood cell types
- genes in the feature space
- dimensions in the latent space
It's built for labs that measure single cells and need to know which cells they have, from a reference that says what it doesn't know.
Trained on RNA. Read on protein.
The reference encoder learns from RNA alone, with labels.
Protein measurements are projected into the same space when they arrive, zero-shot. No protein labels and no paired RNA and protein cells are used in training.
The reference has never seen the kind of data it's asked to place. That is what makes its placements worth trusting, and worth checking.
Bring your own matrix.
-
Upload
An RNA or protein matrix. DIA-NN reports and FragPipe output are read as they come.
-
Align
Gene symbols, Ensembl and UniProt identifiers are mapped onto the reference's genes. Raw intensities are recognised and moved to a log scale. Cells with too few observed genes are refused, not guessed.
-
Place
Each cell is encoded into the shared space and lands beside the reference cells it most resembles.
-
Decide
Each cell comes back with coordinates, a set of plausible labels and a confidence, or an abstention and its reason. The same file always returns the same answer.
The projection service runs locally and is in preview. It is not yet hosted. See the projection service
When unsure, it says so.
Most classifiers always return a label.
VivOME abstains when a cell falls outside the space it has evidence for, and tells you why.
How the service decidesScored the same way.
Integration methods are easy to flatter. VivOME's benchmark:
- gives every method the same RNA reference and the same labels;
- touches protein labels only at the final scoring;
- reports a modality probe beside every accuracy. The probe checks whether RNA and protein truly share one space, or only point the same way.
Results are published once they're signed off. Until then, the page says Pending.
See the benchmarkA living atlas, versioned.
-
The first model, trained on RNA and protein together. It's kept as the documented baseline, not erased.
-
, atlas
A frozen, RNA-only reference, with a model card and its known failure modes written down. Still selectable.
-
, atlas
The current release. Five encoders, trained only on labelled blood scRNA-seq, answer together; no protein labels are used. Each answer is given at the level the evidence supports: one of the reference's blood cell types, its group, or its lineage. Where the evidence runs out, it abstains and says why.
-
Next, in testing
A sharper test for cells unlike any in the reference, and a measured false-abstention rate. It ships only if it passes.
Every release follows one protocol: frozen inputs, a regenerated manifest, a new version number.
Read the model cardNo cell placed without evidence.
Every number on this site is computed from the data behind it. When a number doesn't exist yet, the page says Pending.