Announcing OCTO-VirtualCell (vc) a multi-scale, multimodal transformer trained to predict gene expression for a virtual cell in cellular contexts within patient tissue samples. Complete wth the Celleporter demo app to explore the data! 1/

OCTO-vc simulations place virtual cells in local patches across patient tumors and predict context dependent gene expression changes. For example, we can ask how the biology of a T cell would be impacted at different locations within or outside of a tumor, or across different patient tumors. Check out this cool demo of querying a virtual cell across a whole patient sample: 2/

We trained OCTO-vc on 40 million cells of spatial transcriptomics data from more than 1,000 patient samples generated in-house at NOETIK. This is one of the largest unified datasets of its kind. We've found that spatial, single cell data provide a much richer training set of biological context than single cell RNAseq. With this depth of data, the model learns patient-specific biology. 3/

To explore OCTO-vc, we are releasing a demo of CELLEPORTER, an interactive tool for viewing virtual cell simulations with different prompts. We've loaded it up with a selection of real tumor samples and a handful of genes to explore across different virtual cell types. Even with a curated set of samples, virtual cells, and genes you can appreciate the scale of potential questions to ask! 4/

There are plenty of examples in the technical report. Here we simulate a virtual B cell in a core with a TLS and you can see the model provides a much richer representation of B cell biology for this gene than the raw data. 5/

How do we use OCTO-vc to power drug discovery? Here we place virtual T cells in tumor samples from a cohort of patients that are sensitive to immune checkpoint inhibitors and a cohort that are resistant. We compare simulated gene expression in the two cohorts and find that T cell transcripts associated with cytotoxic activity are enriched in the sensitive population. 6/

Next we counterfactually simulate knock out of genes in our resistant cohort, and ask which genes impact these T cell cytotoxicity transcripts- to reveal potential drug targets that could lead to therapeutics for these patients. These are not in vitro experiments, they are simulations on real patient tissues! 7/

The power of this model is now we can learn patient biology, simulate perturbations, and perform drug discovery research directly on patient tissues. And its not limited to cancer, we can apply these models to any human healthy or disease tissue! Read more here:

Link to Technical Report:

You can make virtual tumor environments - can you simulate response if gene knockout across many cells in parallel to see if target knockout treats tumor vs cell level? For T cell example you knockout periphery cells around T Cells and then can predict T Cell response?

Yes exactly. In the example we simulate gene KO (for all genes) in the tumor environment, and predict expression in virtual T cells to screen for genes that revert cytotoxic gene expression.