{02/09} · Omics
All workVivOME Atlas: Single Cell RNA and Protein Atlas
Mentor Dr Michael P Snyder, Stanford W. Ascherman Professor of Genetics
Can a model trained only on RNA label cell types in protein data it has never seen?
A versioned atlas and web service that labels cell types in single cell mass spectrometry proteomics data, using a human blood RNA reference (Tabula Sapiens). It reads DIA-NN, Spectronaut, MaxQuant and FragPipe outputs, and runs on Google Cloud Run behind an invite only site.
- 5model PyTorch ensemble, trained on RNA only
- 7methods benchmarked under one shared protocol
- ~970automated tests
What the record supports
Claim
A five model ensemble trained on RNA only can be applied to protein zero shot, giving each cell a conformal answer set.
Evidence
Benchmarked against seven methods, including Harmony, scANVI, Seurat and MaxFuse, under one shared protocol.
Limits, stated plainly
The service declines to answer, with a stated reason, when unsure. Every model change was preregistered, and every held out test was scored once.