LatentCellBio × AI

Wiki · Bio × AI Methods · concept

Single-cell foundation model

A transformer pretrained on millions of single cells that turns a cell's gene expression into a reusable representation.

A single-cell foundation model is a transformer pretrained on millions of single-cell gene-expression profiles. It treats each cell like a sentence — the set of genes a cell expresses, and how strongly — and learns to read that sentence the way a language model reads text. The payoff is a reusable representation: hand it a cell and it returns a vector capturing what kind of cell it is, which you can then use for many downstream tasks without training from scratch.

How it works

Each cell is turned into tokens: its expressed genes, tagged with expression level. The model is then pretrained on huge public atlases — tens of millions of cells — using a masked objective: hide some of a cell’s genes and make the model predict them from the rest. To do that well it has to learn which genes travel together and what defines a cell state. Out come two kinds of embeddings: a vector per cell and a vector per gene. You then fine-tune the pretrained model for specific jobs — labelling cell types, predicting how a cell responds to a perturbation, or integrating datasets collected on different machines so batch effects wash out.

Why it matters (for bio × AI)

This is an early step toward a virtual cell: one model that has seen enough biology to reason about a cell it was never explicitly shown. Those cell representations feed perturbation-prediction — guessing what a drug or gene knockout will do before running the experiment. The best-known exemplars are scGPT (Bo Wang’s lab) and Geneformer (from the Ellinor/Theodoris groups). One honest caveat: whether these models genuinely beat much simpler baselines is actively debated, and several careful comparisons have found the gains thinner than the headlines suggest.

See foundation model for the pretrain-then-adapt recipe, embeddings for the cell and gene vectors these models produce, and transformer for the architecture underneath.