Teaching a computer to read cancer the way a pathologist reads a slide - one virtual cell at a time.
Above: the Noetik wordmark. The "k" does a lot of quiet work for a company trying to make biology legible.
Somewhere on a screen in South San Francisco, a cell that does not exist is changing its mind. Switch off a gene, and the model predicts how the tumor next door reacts - which immune cells move in, which signals go quiet, which patient might finally respond. No mouse. No microscope. Not yet, anyway.
This is Noetik on an ordinary Tuesday. It calls itself an AI-native biotechnology company, which is a tidy way of saying the model came first and the lab grew up around it. Its job is unfashionably specific: find better cancer therapies, and find the patients they will actually help. Its method is the part that raises eyebrows - it built a foundation model of human cancer biology and is now renting it to pharma.
"An AI-native biotechnology company with a mission to leverage advanced machine learning methods to discover and develop cancer therapies."
- Noetik, in its own wordsMost biotechs sell a molecule. Noetik is trying to sell the thing that knows where the molecules should go. That is either the future of drug discovery or a very expensive way to make a spreadsheet. The company is betting it is the former, and in January 2026 one of the largest drug makers on earth agreed to find out.
A tumor is not a lump. It is a neighborhood - cancer cells, immune cells, blood vessels, signals crossing in every direction. The reason immunotherapy works miracles for some patients and nothing for others lives in that crowd, in how the cells are arranged and how they talk. For decades, the tools forced a brutal trade: you could study many cells shallowly, or a few cells in deep spatial detail. Rarely both.
So the field guessed. It ran expensive trials, watched most of them miss, and called the misses "biology we don't understand yet." Polite. Costly. The tumor microenvironment stayed a black box, and the black box kept winning.
"The tumor microenvironment is where immunotherapy is decided - and it was the one place we couldn't read at scale."
- The problem Noetik set out to solveThe honest version of the problem: oncology had plenty of data and almost no way to make it predictive. What it needed was less a new instrument than a new pair of glasses.
In 2022, Ron Alfa and Jacob Rinaldi left Recursion Pharmaceuticals - one of the original AI-drug-discovery houses - to start Noetik. Alfa, a physician-scientist, took the CEO chair; Rinaldi, the science. They pulled in people from Genentech and the Parker Institute for Cancer Immunotherapy. Doctors, ML researchers, computational biologists, all in one building, which is rarer than it sounds.
Their bet had two halves, and the second is the clever one. First: generate spatial data on human tumors at industrial scale - proteins, RNA, pathology, genotype, clinical outcomes, all mapped to where each cell physically sits. Second, and this is where most atlases stop and Noetik keeps going: run real perturbations. Their Perturb-Map platform fires CRISPR edits in living systems and watches, spatially, what breaks. That gives the model cause and effect, not just a very pretty correlation.
"We are doctors, scientists, engineers, machine learning researchers, computational biologists, with a shared vision to discover precision cancer therapies."
- The Noetik team pageInvestors found the pitch persuasive. A $14M seed in 2022 became an oversubscribed $40M Series A in August 2024, led by Polaris Partners, with Khosla Ventures, Wittington, Breakout, DCVC and Zetta along for the ride. Roughly $62M to teach a computer something pathologists spend a career learning. Ambitious. Possibly cheap.
// Four years, one virtual cell, two big partners
A biotech timeline with no failed Phase II on it - mostly because the clinic is still ahead, not behind.
The headline product is OCTO-VirtualCell, a foundation model of roughly 1.5 billion parameters. It was trained on one of the largest proprietary multimodal spatial datasets anywhere - spatial proteomics, spatial transcriptomics, H&E pathology, DNA genotyping and clinical metadata from nearly 200 million tumor and immune cells across thousands of patients. The model learns to simulate gene expression, cell states, and the crosstalk between tumor and immune cells.
In plain terms: you can ask it questions. What happens to this neighborhood if we knock out that gene? Which patients carry the signature that responds to this drug? It is a virtual cell you can poke without paying for a single mouse - and a companion tool, Celleporter, lets scientists actually see the answers.
The 1.5B-parameter virtual cell foundation model that simulates gene expression, cell states and tumor-immune interactions.
High-throughput in vivo CRISPR perturbation that generates spatially resolved cause-and-effect training data.
A visualization environment for exploring and interrogating the virtual cell models and spatial atlas.
"Among the first and largest transactions monetizing a biological foundation model as a scalable enterprise asset."
- Noetik's Chief Business Officer, on the GSK dealIdeas are cheap in biotech; the test is whether anyone pays for them. Noetik's came in January 2026, when GSK licensed OCTO-VC in a five-year anchor partnership covering non-small cell lung cancer and colorectal cancer. The structure is the tell: roughly $50M in upfront and near-term milestones, plus annual subscription-style fees for access. Pharma paying a license fee for a biology model, like enterprise software. That is new.
// Selected Noetik capital events, $ millions (approx.)
Before GSK there was Agenus, which in 2025 enlisted Noetik's models to predict response to its botensilimab/balstilimab combination - the exact "which patient" problem the company was built for. Two partners, one validating the science, one validating the business model. Skeptics will note the therapeutics still have to reach a clinic. Fair. But the model is already earning its keep.
"Noetik's partnership with GSK puts dollar signs on the AI licensing business model in techbio."
- DCVC, an early investorStrip away the model architecture and the mission is stubbornly human: fewer patients getting drugs that were never going to work for them. Noetik's name nods to "noetic" - to do with knowledge and the intellect - which is a touch grand for a company that ultimately wants oncologists to guess less. But the ambition tracks. If you can read the tumor microenvironment at scale, you can match therapy to biology instead of to hope.
The company frames itself, plainly, as doctors and engineers under one roof. The therapeutics pipeline is the long game; the model licensing is what funds the wait. It is a patient strategy from people who clearly do not enjoy waiting.
Return to the virtual cell that started this, the one quietly changing its mind in South San Francisco. A year ago it was a research demo. Now a global pharma pays to ask it questions, and a clinical-stage immunotherapy leans on its predictions. The cell still does not exist. It is just getting harder to argue it does not matter.
Here is the wager, in one line: that the next breakthrough cancer drug will be found in silico before it is ever made in glass. Noetik does not have the clinical wins to prove it - not yet. What it has is a foundation model that pharma now treats as an asset worth licensing, a lab built to keep feeding it, and a very specific definition of success: a patient who gets the right drug because a virtual cell saw it first. That is not a slogan. It is a deliverable, and the clock is running.