virtual spatial transcriptomics use case
Bring AI closer to the biology you study.
Every laboratory's slides carry its own fixation, stain and scanner, and the questions that matter most are asked of that material. With our team, a few slides you have measured adapt DeepSpot-M to your protocol before it predicts the rest of your cohort.
For example
Adaptation can begin from very little: the DeepSpot-M preprint demonstrates it on a new cohort from one measured slide.
Worked through for Sequencing & Core Facilities.
The research decision
Should a few measured slides adapt the model before it reads the rest of this cohort?
Anchor virtual spatial transcriptomics to measured examples from your own protocol before scaling across the remaining slides. In the paper's restoration demonstration, the same adaptation approach returned selected failed-QC genes as explicit predictions.
Where the work gets stuck
Your protocol changes what a section looks like
Measured data from your own protocol can anchor predictions to your material. Fixation time, stain lot, section thickness and scanner all change what a section looks like.
A model that has never read your laboratory reads your sections through somebody else's conventions, and you pay for that in predictions that look plausible and sit slightly off in ways that are easy to trust and hard to catch.
Material from other protocols cannot fully characterise that gap. Its size depends on the distance between your material and the model's prior experience, so measured slides from your own cohort provide the most relevant anchor.
Closing the gap used to cost a cohort of its own
Matching a model to your protocol meant paying for the assay you were trying to avoid. Until recently, adapting a model to a laboratory meant building a paired dataset at that laboratory: measure enough of your own slides, alongside their images, to train on. But that is the expense the study was trying to avoid in the first place, and it landed before anyone could tell whether adaptation would help.
DeepSpot-M adapts from as little as one slide. Because the price of matching the model to your material now falls to a handful of samples you have already run, the question changes from whether you can afford to adapt to which slides to adapt on.
Measure a handful, read the rest
Measure a few slides once, then read the rest from H&E images on the same panel. A few of your own measured samples are those data. DeepSpot-M adapts on them and then reads the slides you never measured. The expensive step stays where it was and stops having to be repeated for every sample behind it.
After a small measured anchor, the rest of the cohort reads from H&E slides on the same gene panel. Biology can size a study where the cost of a kit used to. The archive page starts from slides nobody measured. This one starts from a measurement you would otherwise repeat.
The same step rescues a run that partly failed. Tell us which genes did not survive quality control, and DeepSpot-M holds them out and predicts them from the image, so nobody writes off a run over a handful of columns and nobody cuts another section to recover them.
What becomes possible
Adapt the model to your protocol from a few measured slides, or restore genes that failed quality control as labelled predictions.
Start from measured examples
Adaptation reads the samples you have already measured. The paper demonstrates a benefit from one cohort-matched slide, while the right number for a new cohort remains an empirical question.
Your protocol, your scanner
Adaptation sees your staining and your scanner, so the predictions DeepSpot-M makes afterwards run over material that resembles the material it adapted on.
Genes back without re-sequencing
Tell us the genes a run lost. DeepSpot-M holds them out and predicts them from the image, which is what makes a restored column a rescue.
Proof on real tissue
The DeepSpot-M preprint demonstrates test-time adaptation from one slide and transcriptome restoration. Two diagrams make the distinct workflows explicit without presenting an illustration as measured evidence.
Two modes, one starting point
Both modes start from the same thing: slides you have measured, paired with their images. They differ in what stays back from the adaptation, and why.
Your input and output
- Input
- A few slides you have measured, each with its image.
- Output
- Cohort-adapted virtual spatial transcriptomics and labelled predictions for restored genes.
Take the next step
Start with an example slide, or talk to us about your cohort.