Biology is spatial.
Cells communicate via direct contact and secretion of chemical mediators. Tissues organize themselves into functional multicellular structures to mediate complex biological programs. Therapeutic interventions unfold within this spatial milieu - and as such designing interventions that disrupt disease progression is a spatial problem.
That’s why we’re building foundation models for spatial biology - models that understand how cells and tissues behave in the context of health and disease. Models that can redefine the taxonomy of disease at the system level.
Our initial focus is cancer immunotherapy – where many of the critical differences between patients are spatial. In some patients, the immune system is able to infiltrate and attack a tumor; in others, the tumor evolves to disrupt and drive away immune cells, preventing effective treatment. The biological mechanisms that cause these differences are not well understood. Our core thesis is that self-supervised machine learning, given diverse data from real patients, can reveal them. This is why we built a multimodal, continuously growing dataset from thousands of human cancer samples and used it to train OCTO, our first foundation model for patient-level spatial biology.
OCTO-VirtualCell (OCTO-vc) is our next generation foundation model. OCTO-vc simulates spatially resolved single-cell gene expression in intact or virtual patient tissues, enabling high-fidelity prediction of the behavior of thousands of genes in response to changes in biological context. The idea of a virtual cell was proposed recently as framework for modeling many cellular processes at multiple scales. In the present work, a "virtual cell" is a concrete case of this broader concept: a prediction of which genes a cell would express in a certain environment, even if no such cell is really present. OCTO-vc was trained on nearly 40M cells across many cancer indications, which to our knowledge is the largest extant spatial transcriptomics dataset.
We place virtual cells in different regions of real tissue samples to probe specific biology in specific patients. For example, we can drop a virtual T cell into various parts of a tumor to infer whether the T cell will be activated – capable of killing cancer cells – or not. By running virtual cell simulations at many locations, we create a holistic picture of the underlying biology.
OCTO-vc is a transformer trained via masked token modeling to predict spatial patterns of gene expression. This is loosely analogous to large language models, except that tokens here encode gene expression – including user-provided “prompts” to specify genes expressed in the virtual cell itself. Self-supervised training gives rise to a remarkably rich representation of tissue and patient biology when orchestrated at scale. The generative capacity of OCTO-vc allows us to make predictions about precise cellular phenotypes, identify tissue-scale differences between patients, and simulate new therapeutic interventions. Below, we combine these abilities in a case study: virtual cell simulations reveal new biomarkers and potential therapeutic targets in a treatment-resistant population of lung cancer patients.
Finally, we introduce Celleporter: an interactive tool for viewing virtual cell simulations with different prompts, in a variety of patient samples, across millions of spatial locations. Visualizing virtual cells is a fundamentally new way to see the hidden biological structure in tissue from real patients. We have just begun to explore this world of latent spatial biology, and we are eager for others to join us.
We think of a virtual cell almost like a high-dimensional thermometer: by moving it around real tissue and seeing what happens, we can measure properties of the local biology. A prompt sets several genes to be on or off in the virtual cell, tuning the instrument to detect different properties – like genes that should be expressed in one cell type but not another. We test whether these virtual cell simulations are accurate by comparing predicted to true gene expression in data held out of training.
Model predictions are highly consistent with these data and known biology. For example, large aggregates of immune cells called tertiary lymphoid structures (TLS) contain many B cells (the cells that produce antibodies in their mature state). Immature B cells express the gene TCL1A and are highly enriched within TLSs. In a tissue sample with a TLS, virtual B cells – that is, OCTO-vc simulations prompted to express several (non-TCL1A) B cell genes – express much higher levels of TCL1A inside the TLS versus outside. This strongly aligns with TCL1A expression in the real data.
In this example, changing the virtual cell prompt to simulate T cells – a different immune cell type that does not express TCL1A – results in background-level predictions. This shows that OCTO-vc learns to integrate both local context and the prompt itself into a spatial representation of the tissue. Different prompts result in different predicted patterns of gene expression. In this tissue sample, for instance, virtual (CD4) T cells express the gene CXCL13 mainly in the outer region of the TLS; this is where T cells are known to congregate and recruit B cells via secretion of the CXCL13 protein, which acts as an attractive intercellular signal.
Virtual cells often reveal richer structure than what can be seen in the raw data. A good illustration is the immune cell gene PD1, the key target of leading cancer immunotherapies. PD1 is normally expressed at low levels and is detected only sparsely in the real data. Virtual T cells, though, reveal a clear pattern of expression, consistent with but far less noisy than the observations in this sample. Differences in predictions between two virtual T cell subtypes (CD4 versus CD8 T cells) further show how OCTO-vc integrates local context with complex inter- and intra-cellular gene expression patterns, learned across the whole training dataset.
%402x.avif)
%402x.avif)
%402x.avif)
The TLS in this sample is a good test for OCTO-vc because of its spatially stereotyped structure. The center of a TLS typically contains immature B cells, whereas the perimeter contains “naïve” B cells that are functional but have not yet encountered an antigen; mature, antigen-exposed B cells mainly reside outside of a TLS. OCTO-vc reflects this developmental trajectory, with virtual B cells expressing typical immature (TCL1A), naïve (IGHM), and mature (IGHA1) B cell genes in the center, perimeter, and outside of the TLS, respectively.
Cases like these illustrate the power of learning a spatial representation of single cell data: each virtual cell prompt is a lens for observing particular aspects of the representation and biology – often at significantly higher resolution and fidelity than the underlying assays.
At Noetik we think of this technology as a new kind of microscope - a new way to probe biological systems. The first place we’re pointing it is immuno-oncology because we believe the complexity of tumor-immune interaction is well abstracted by the language of machine learning. Why, for example, are immune cells able to infiltrate and kill a tumor in some patients but not in others? How do cells affect each other to create a suppressive or inflamed tumor microenvironment, and what should we change to shift that balance?
Virtual CD8 T cells are a good place to start. These cells are the backbone of cancer immunotherapy because they are the main cells capable of killing tumor cells. However, they can only do so after recognizing antigens on tumor cells; before this, they are incapable of their cytotoxic (cell-killing) function.
Simulations with virtual CD8 T cells (v-CD8s) reflect this biology. In a tissue sample with both tumor and non-tumor regions, OCTO-vc predicts that a gene highly expressed in the "naïve", non-cytotoxic state (TCF7) will be present in v-CD8s outside the tumor, whereas a gene marking the actively cytotoxic state (GZMB) will be expressed by v-CD8s inside the tumor – where they would recognize antigens on cancer cells. Since virtual cells are counterfactuals, we see these differences in cell state regardless of whether there are CD8 cells in the real tissue – they stem from the model’s latent representation of spatial biology, not just the data in this sample.
These v-CD8 experiments also reveal more subtle differences in the tumor-immune microenvironment. Some tumors respond to the immune system’s attack by expressing immune checkpoint genes like PD-L1, which interacts with its receptor PD1 on activated CD8 cells to inhibit them. The PD1 gene in v-CD8s may therefore mark tissue regions where CD8 cells would enter an inhibited state. In this sample, simulated PD1 is elevated only in some parts of the tumor – implying latent heterogeneity in the tissue, which OCTO-vc simulations expose. This is even starker when viewing expression patterns of all virtual cell genes, such as through principal components analysis of model predictions.
This hidden heterogeneity is at the heart of Noetik’s mission.
In most cases, we don’t know why one patient responds to immunotherapy and another doesn’t; there are countless ways tumors may evade the immune system. The promise of a model like OCTO-vc is in helping us see and address the relevant ones in each patient.
Virtual CD8 T cells are a promising lens because immunotherapy hinges on the CD8-tumor interaction. CD8 cells detect antigens (in this case, mutated proteins) on the surface of cancer cells. This activates CD8 cells to become cytotoxic, producing and secreting protein-degrading enzymes and inducing programmatic cell death in their targets. This interaction depends on MHC I, the protein that presents antigens on the surface of target cells.
Counterfactual simulations reveal that OCTO-vc has learned this fundamental cell-cell relationship. Since CD8 activation depends on MHC I, v-CD8 expression of cytotoxic genes should increase when we counterfactually edit real, nearby tumor cells to express higher levels of the MHC I gene. This is exactly what we see. OCTO-vc has learned this biological relationship from spatial data alone, without human help.
However, simulations expose patient-to-patient differences even in this canonical interaction. While v-CD8s in most tissue samples are sensitive to counterfactual changes in tumor MHC I, a subset of samples do not show this effect: elevating MHC I in tumor cells in these samples produces little or no increase in v-CD8 cytotoxic genes.
Work is ongoing to dissect the mechanisms behind this patient diversity – for instance, testing which cell types and genes are responsible for MHC I insensitivity in some tumors. More generally, this example shows how counterfactual virtual cell simulations can help us stratify patient populations and understand why some patients do not respond to specific therapies – pointing towards better ways of positioning new drugs.
If OCTO-vc learns established biology, such as the relationship between tumor MHC I and CD8 cells, it may also learn new biology underpinning the immune response to cancer. This is what we want to surface to discover better drugs.
Tumors can acquire genetic mutations that induce resistance to immunotherapy. These mutations change which genes tumor cells express and can have multiple downstream effects on the immune response, such as directly blocking CD8 cells from attacking the tumor. Patients with tumor mutations in the gene STK11 rarely respond to existing immunotherapies, but the reason is not fully understood.
We used virtual cell simulations to look for differences in CD8 cell state between patients with and without STK11 mutations. To do this we prompted OCTO-vc with v-CD8 cells at each position in each tumor from hundreds of patient samples and compared predicted gene expression across the aggregated STK11 normal and STK11-mutated cohorts. This revealed a specific subset of cytotoxic effector genes, including Granzyme A (GZMA) and Granzyme K (GZMK), predicted at lower levels in STK11-mutated patients – suggesting that CD8 cells in these patients may be less effective at killing tumor cells in a way not addressed by current immunotherapies. This raises the question of what factors in the tumor microenvironment are responsible for the v-CD8 phenotype.
Virtual cell simulations can speak to this, too. We ran a large-scale “virtual screen” for potential drug targets that could reverse the v-CD8 phenotype by knocking out each gene expressed in each tumor cell of each STK11-mutant patient, one by one. These counterfactuals simulate idealized therapeutics, which would reduce or block the function of a particular tumor molecule. We considered genes that showed a large effect in this digital assay – such as returning v-CD8 GZMA and GZMK levels closer to those of the STK11-normal patients – to be promising candidates for further validation in this population with unmet therapeutic needs.
Although better cancer therapies are our first focus, we think virtual cell simulations have far broader scope. Given the right data and the right scale, models like OCTO-vc can become a general interface for spatial biology, a foundational tool for basic research, and a discovery engine for drug development. The ability to control generative models allows large-scale in silico experimentation - a new way to conduct virtual screens. For the hardest problems in drug discovery, places where the complexity of biological systems is preventing us from getting treatments to patients, a new era is starting where we have the tools to grapple with that complexity. We are advancing fast into this new unknown - let us know if we can work together.
