virtual spatial transcriptomics use case

A new chapter for your collection.

An archive holds years of careful work, in slides kept long after the studies that needed them ended. Read for predicted spatial gene expression, each of those slides can become the starting point for research nobody had in mind when the collection was built.

For example

The slide below came out of an open archive: a melanoma that had spread to a lymph node. Its virtual molecular view locates the node's B-cell follicles and its collagen, and nobody outlined either.

Out-of-the-box prediction

Worked through for Pharma & Biotech, Academia & Research and Biobanks.

The research decision

What could the archive tell you if every slide carried the same predicted genes?

Give the cohort one comparable reading of virtual spatial transcriptomics: predicted, spatially resolved gene expression from H&E images across 19,338 genes, at every retained tile, on each slide's own coordinates. The reading consumes no extra section.

Where the work gets stuck

Cohorts outrun the assay that could read them

H&E slides scale with every case, while spatial assays scale with budget. Because a stain reports morphology and nothing else, the image that exists for practically every patient cannot record expression. A pathologist still diagnoses cancer on tissue.

The laboratory cuts a block, stains a section and scans it, so an image exists for practically every case the service sees. What that image records, it records completely: how the disease arranged itself, which cells sit where, where the boundary runs.

Spatial transcriptomics measures those programmes in place and keeps the architecture around them. Each sample it reaches costs a section, a kit and a slot in a queue, so a study measures the number its budget allows and leaves the rest of the cohort alone.

Meanwhile the tissue that could have answered moves on: a block runs out, a section stops yielding material an assay can use, or the consent it came under will not stretch to a new experiment.

Coverage, not method, sets the limit

Reviewing artificial intelligence in drug discovery for Nature Reviews Drug Discovery, Bender and colleagues identify the availability of suitable datasets as a persistent constraint. That framing matches cohort work. You often need coverage across cases before you can justify measured follow-up. Because the assay budget sets how many samples a study characterises molecularly, every statistic downstream inherits the budget. A cohort of four hundred cases reaches a spatial assay a dozen samples at a time, and the association anyone draws afterwards rests on the dozen.

Virtual molecular views give the archive a second reading

A question that used to wait for an assay can go to a picture. DeepSpot-M predicts spatial gene expression from the stained image alone: 19,338 genes, at every retained tile, on the slide's own coordinates. Slides from different archive periods enter the same workflow, while preservation and protocol differences remain sources of uncertainty. Reading a cohort makes the data available. Deciding what to validate stays your job.

Across a research cohort, DeepSpot-M lets you examine candidate expression axes in every sample, in place, with the architecture of the tissue intact. Because nothing selected them out in advance, genes the study did not order arrive in the same file. However, an axis found this way remains a hypothesis until a measured experiment tests it.

What becomes possible

Explore the whole archive on the same predicted genes, and keep measured follow-up for the hypotheses that survive.

Every sample in the cohort
A spatial assay prices itself per sample, so a study measures what it can afford. DeepSpot-M runs on every slide in the cohort, and the number of samples you can look at stops tracking the number you can pay to measure.
Nothing is consumed
Reading a slide uses none of it up, so the cohort survives the study and the same section can still answer a second question a year later.
Genes the study did not order
The panel belongs to DeepSpot-M and your question does not choose it, so genes you would never have selected arrive in the same file.

Proof on real tissue

The DeepSpot-M preprint demonstrates transfer to held-out cancers and a 28,664-slide virtual atlas across 32 cancer types. This page makes the cohort-scale workflow inspectable on one open-access whole slide.

One diagnostic slide, read for expression

This section was not staged for a figure. Nobody chose a region, outlined a structure or marked tissue. It came off a diagnostic service, went into an open archive, and the maps under it show what DeepSpot-M returned from the image.

Every photograph below comes from one open-access diagnostic slide: a melanoma that had spread to a lymph node.