SCimilarity: a search engine for cells

Query a cell, and SCimilarity returns its nearest neighbors across a reference of 23.4 million cells.

Foundation models are large machine learning models trained on huge amounts of data and then reused for many different tasks, the same way one language model can draft an email or debug code. SCimilarity applies that idea to cells: trained on tens of millions of single cells, then reused as a search engine. Think of it as reverse image search for cells: hand the model a cell, and it returns every similar cell we've ever seen, along with the tissue, disease, and context that came with them. The question behind this is deceptively simple: have we ever seen a cell like this before, and if so, in what tissue, what disease, what context? Single-cell sequencing has produced enormous public and internal datasets, but almost nobody was asking that question at scale. Information retrieval, the thing that makes web search and protein search possible, barely exists yet for single-cell genomics. That's still true today; it remains one of the most underdeveloped areas of the field, with very few tools or benchmarks built around it.

Answering it meant solving three problems at once: collecting data at a scale nobody had assembled for this purpose, curating consistent metadata across studies that were never designed to be compared, and designing the model around metric learning, training it to place similar cells near each other in its internal representation, rather than the classification or generative objectives most single-cell models use. The result, published in Nature (2025), is a foundation model that turns "what is this cell" into a search problem: embed a cell (convert it into a list of numbers describing its biological state), find its nearest neighbors across a reference of millions of cells, and read off the biological context that comes with them.

SCimilarity method diagram: precomputed cell database compared against a query cell through the same foundation model

The mechanism itself: a query cell is embedded into the same space as the 23.4-million-cell reference, and its nearest neighbors are tabulated by study, tissue, and disease.

As one worked example from the paper: take fibrosis-associated macrophages (FMs), immune cells that build up in scarred tissue and are of interest as a target for treating diseases like pulmonary fibrosis, and ask SCimilarity which cells across the entire reference actually look like them.

Scoring every monocyte and macrophage (one dot per cell) from 1,041 tissue samples across 143 studies against the FM profile, ordered by mean score: most samples score low, but a clear tail of high-scoring samples emerges.

SCimilarity is now built into Genentech's standard compute environments, has an internal community of roughly 130 people, and its outputs (similarity searches, embeddings) show up in governance decisions and target nomination, the process of deciding which biological molecule to pursue as a drug target, not just as a research artifact.

Each circle's size is the number of cells (tissue, then disease, then individual study), colored by the percentage of monocytes and macrophages scoring in the top 1% for similarity to the FM profile. Zooming into two of those studies, an interstitial lung disease cohort and a pancreatic cancer cohort, shows the FM signature concentrated in one distinct cluster in both.

Heimberg, G., et al. A cell atlas foundation model for scalable search of similar human cells. Nature, 2025.