Executive Summary

Single-cell RNA sequencing (scRNA-seq) reads out, for each individual cell, which genes it is currently expressing and how strongly — as tens of thousands of numbers. Most attempts over the past few years to build a foundation model from this data borrowed the natural-language playbook wholesale. A gene becomes a word (a token), a cell becomes a sentence of those words, and the whole thing goes into a transformer exactly as a language model is trained. Geneformer (2023) and scGPT (2024) are the flagship examples, and scGPT — trained on 33 million cells — became the field's baseline.

The trouble lies in the act of translating into a sentence. Expression levels are originally continuous magnitudes, but to handle them as tokens you have to bin them into a few categories or convert them into ranks. This binning erases much of the magnitude information, and genes end up treated largely as an unordered set, blurring the structure of which gene co-fires with which. The preprocessing step meant to make data "AI-Ready" becomes the very place where the signal gets shaved off.

scVision's idea starts from skipping tokenization. It draws a cell's entire transcriptome as one continuous two-dimensional image. Every gene is assigned a single fixed coordinate on a 104×104 grid, and all cells share that same layout. The only thing that changes from cell to cell is the brightness at each coordinate. Highly expressed genes are drawn bright, lowly expressed ones dark, so expression magnitude is carried directly in pixel intensity without any binning.

How the coordinates are chosen is the heart of the method. scVision optimizes the gene layout with a technique called Gromov-Wasserstein optimal transport, placing co-regulated genes next to one another on the image. As a result, a gene program that works as one unit shows up as a local texture in the image — a bright patch in a particular region. The cell's state gets translated into the texture of the picture.

The training procedure follows the computer-vision standard. An 86-million-parameter vision transformer (ViT) is pretrained with masked image modeling — an image-domain fill-in-the-blank that hides part of the image and reconstructs the original from the rest. The training data is 72 million human cells, among the largest single-cell pretraining runs to date. The habit of reading cells as if handling language shifts here into seeing cells as if handling photographs.

Before you read the results, look at how they were evaluated. scVision froze its pretrained encoder — no fine-tuning to the downstream task — and evaluated directly in a zero-shot setting. The design is meant to test the claim that it won not by seeing more training data but by changing only the representation.

The results split four ways. First, on all six independent tissue atlases not used in training, it ranked first in cell-type annotation accuracy. Balanced accuracy varied by tissue between 0.47 and 0.83, but it never lost the top spot in any tissue. Second, it was especially strong when labels were extremely scarce. scVision given a single label matched the performance a competing model reached after seeing fifty. Third, on the integration task of combining multiple studies, it performed on par with the best token-based model while preserving more biological structure — without ever seeing a batch label. Fourth, on the task of recovering the gene programs a cell is running with no labels at all, it also outperformed existing foundation models and classical methods together.

Good performance alone can't tell you whether the cause is the image representation or just a good architecture in the vision transformer. The authors designed a controlled experiment that answers this directly: randomly shuffling the coordinates assigned to genes. Leaving the network structure intact and breaking only the image layout, the drop in performance tells you how much the representation contributes.

When the coordinates were shuffled, accuracy fell sharply. More importantly, that drop was larger than removing the vision transformer itself. Scattering the spatial positions of genes broke performance more than changing the architecture did. The authors confirmed for themselves that the source of the performance is not the transformer's excellence but the biologically meaningful spatial structure embedded in the gene layout.

Switching the representation to an image brought a bonus, too. Because regions of the image correspond directly to groups of co-expressed genes, the model's attention maps can be read like activity maps of gene programs. Spatial masking experiments — occluding an entire neighborhood of related genes to observe the perturbation effect — become possible. Token-based models have no clean operation that corresponds to this.

Overhype should be kept in check. As a critique in a specialist bio newsletter points out, cell-type annotation is a relatively easy subtask in single-cell analysis. This result alone is too early to get excited about. Whether this representation actually helps on harder downstream tasks such as perturbation prediction or drug response remains an open question. So does the scalability of the optimal-transport computation, whether one grid layout holds across all tissues, and performance on rare cell types.

Pull this story out of genomics and a question familiar to anyone who works with data remains. When we say we're preparing AI-Ready Data, we usually picture filling in missing values, filtering outliers, aligning formats — treating it as a problem of making data clean. But what scVision showed is a question one step earlier: what do we represent the same data as?

The scRNA-seq data is the same either way. What changed is only whether you feed it as a sequence of gene tokens or as a gene-expression image. Fed as tokens, the relationships between genes and the magnitude of expression thin out; fed as an image, those two signals come back to life as the position and brightness of pixels. It was not the quality of the data but its representation that separated performance. However well you clean it, if the representation format is leaking information, that loss cannot be undone at the model stage.

Put as a practical question, it goes like this. When we prepare data, is the thing we should actually be touching the missing values, or the representation? How you design the schema, how you encode continuous values, in what structure you carry relationships — these are decided before cleaning. Asking first what changes if you represent something as a graph or an image when a table seems obvious, and what disappears when you cut it into categories, belongs at the starting point of data preparation.

• 1.Yesiloglu, R., Mostafa, S., Zou, J., Alizadeh, A., Wu, J., Xing, L., Adeli, E., & Islam, M. T. (2026). "A vision foundation model for single-cell biology via spatial gene cartography." arXiv:2607.14163.
• 2.Cui, H., Wang, C., Maan, H., Pang, K., Luo, F., Duan, N., & Wang, B. (2024). "scGPT: toward building a foundation model for single-cell multi-omics using generative AI." Nature Methods, 21, 1470–1480.
• 3.OmicsML. (2026). "Awesome Foundation Model for Single-Cell Papers." GitHub.
• 4.kiin.bio. (2026). "Stanford's scVision, Ohio State's conDitar, and the AllTheBacteria consortium." kiin.bio newsletter.