We NOETIK built a new foundation model for spatial biology: OCTO-VirtualCell. This model is prompted to simulate single cell gene expression in "virtual cells," which are placed within real, intact patient tissue. Read about & explore these simulations below:

You can get an intuitive feel for how virtual cell simulations work through Celleporter, an interface for viewing OCTO-vc predictions in patient samples: Or read how we're using this model for drug discovery in a technical report:

OCTO-vc was trained on a @NOETIK_ai-generated spatial transcriptomics dataset of 40M cells across more than 1000 tissue samples from patients with cancer. More about how we generate massive, high-quality datasets for self-supervised learning:

OCTO-vc simulates based on a prompt, which consists of 1. a few genes set as on or off in the virtual cell 2. the local and patient-level biological context, such as genes expressed by nearby cells. The model then predicts expression of the remaining genes in the virtual cell.

Virtual cell simulations reveal a huge amount of biology. One cool example: we can use virtual B cells to see the spatial and functional structure of tertiary lymphoid structures, sites where B cells mature and produce antibodies in response to infection or cancer.

In another case, OCTO-vc predicts that virtual killer (CD8) T cells will be in different functional states in different regions of a patient's tumor. This heterogeneity is central to @NOETIK_ai 's mission of finding cancer immunotherapies that work in each patient.

Counterfactual simulations with virtual cells can reveal even more differences between patients. Most patients' killer T cells are predicted to be more effective the more cancer cells are presenting antigens on their surface -- the expected biology. But some patients don't show this response. This could be a clue as to why many patients don't respond to existing immunotherapies.

We can scale up this approach to run a "virtual screen" on an entire patient cohort. In this one, we first found that virtual killer T cells are predicted to be in a less functional state in a specific set of patients known to be unresponsive to immunotherapy.

Then, we can simulate the knockdown of different genes, one by one, in each patient's tissue. Some of these simulated interventions partially "rescue" the virtual T cells, as OCTO-vc predicts they'll become more effective at killing cancer cells. These are ingredients for finding the right drugs for each patient.

This is a transformative moment for understanding the world of biology and disease. So much of biology is spatial: how cells come together to form functional tissues and how they communicate with each other. But our brains didn't evolve to make sense of thousands of genes and proteins expressed across labyrinthine structures -- and there's no army of data labelers coming to tell us what it all means. Self-supervised learning on real patient data is the way. We've now crossed a threshold where data-generating technology and machine learning methods, combined, are powerful enough to reveal biology deeper than we've ever been able to see before. We're beyond excited to explore this new world and we're looking for collaborators!

It's awesome that just as we've started building these models, people are getting excited about using ML-based simulation to discover new biology -- with the idea of a "virtual cell" as a unifying framework: We're tackling a particular problem here -- focused especially on learning spatial and functional relationships between *different* cells -- where the concept of a "virtual cell" as a unit of simulation makes complete sense. We're excited about bringing in other data and data modalities and simulating other aspects of cell + patient biology, and looking forward to seeing what problems others are using the virtual cell concept to tackle.