virtual spatial transcriptomics use case
Let discovery move beyond the original question.
A targeted panel is fixed before the experiment, and the result that matters is often the one that points past it. With our team, the panel you measured can anchor predictions for the genes it left out, on the spots your assay already reported, so an emerging idea can go where it leads.
For example
A panel of a few hundred genes is a sliver of the 19,338 that DeepSpot-M predicts. After adapting on that sliver, the rest of the panel arrives as predictions on the same spots, beside the genes you measured.
Worked through for Academia & Research and Sequencing & Core Facilities.
The research decision
Which off-panel genes should move into validation after a targeted spatial experiment?
Extend a measured panel into virtual spatial transcriptomics on the same tissue coordinates. Predicted genes arrive beside, never over, measured columns so the experiment can support a second round of questions.
Where the work gets stuck
Panels fix the question a year before the answer
You lock the gene list before you see the biology. A targeted spatial assay measures a few hundred chosen genes in place, at high resolution, and it measures them well. Somebody chooses which few hundred before the experiment, out of the understanding of the disease that existed then and the panel the vendor offered.
That choice ages faster than anything else in the design. A result earns attention precisely when it points somewhere the panel does not reach: a population that separates on a gene the panel never included, or a target whose neighbours never got measured.
Following a result off-panel costs a second experiment
New sections, new reagents, a second budget, and a block that may not survive the cut. Chasing an off-panel result means going back to the tissue. You order new sections, new reagents and a second run, and the sample you already cut once may not survive being cut again.
Because that cost falls due before anyone knows whether the lead holds, most leads never get followed at all.
Until now, extending a panel usually meant another experiment: new sections, new training, another wait. A model learned one cohort's panel as a fixed set of outputs, so it answered for that panel and stopped.
Ask the off-panel question on spots you already measured
Unchosen genes arrive on the spots your assay already reported. Adapting on the panel you measured puts the second question to the first experiment. DeepSpot-M reads your panel alongside the image, and the rest of the genes come back at the positions your assay reported, in the coordinate system your analysis already speaks.
For target validation this changes what the experiment can answer. Whether a target gets expressed is what the panel measured. But whether the tissue around it behaves the way the hypothesis requires is almost always a question about genes the panel never carried. Asking it without consuming another section moves the question much earlier in a programme.
It also changes what heterogeneity you can see. A targeted panel resolves the axes it was designed around and averages the rest into silence, a reasonable trade when the axes are known and a costly one when they are not. Off-panel genes come back to the same spatial frame, where they cluster and compare beside the measured ones.
What becomes possible
Predict the genes your panel left out, on the spots it measured, then choose which leads to validate.
The measured genes stay measured
What your assay reported stands. Predicted columns arrive beside it and overwrite none of them, so the two stay distinguishable in everything that happens afterwards.
One tissue, one coordinate system
The predicted genes land on the positions your assay reported, so they join the analysis you already wrote. Nothing needs registering back to a second set of coordinates.
No second section
Nothing gets cut, stained or shipped. The sample you consumed once stays consumed once, which for a rare or irreplaceable block usually amounts to the whole of the constraint.
Proof on real tissue
The DeepSpot-M preprint demonstrates prediction for genes excluded from training. This page shows panel coverage against the 19,338-gene output and where predicted values return on the measured spatial frame.
What a panel covers, and where the rest arrives
A targeted panel spans a few hundred genes out of the 19,338 DeepSpot-M carries. The panel is large; your assay measured a slice of it. Predicted genes fill in on the spots you already have.
Every photograph below comes from one open-access diagnostic slide: a melanoma that had spread to a lymph node.
Your input and output
- Input
- An H&E image and the panel you measured, with a position per spot.
- Output
- Virtual spatial transcriptomics for off-panel genes at the measured positions.
Take the next step
Start with an example slide, or talk to us about your cohort.