Noetik
Noetik builds AI models of how tumors behave and licenses them to drug companies for cancer research. The models help researchers find targets, run virtual experiments and predict which patients will respond to a therapy.
The models are trained on Noetik's own data: tumor samples from more than 4,500 patients, with tissue images, protein and gene activity maps and DNA sequencing lined up together. OCTO-VC simulates how cells behave inside a tumor, including how they interact with the immune system. TARIO-2 predicts how a patient will respond to treatment.
GSK licensed OCTO-VC in lung and colorectal cancer in January 2026, paying $50 million upfront for a five year subscription. Noetik later reported that the models met GSK's testing criteria and were deployed. With Agenus, it used TARIO-2 to predict response to an immunotherapy combination.
Noetik is based in South San Francisco. It calls itself a therapeutics company, but names no drug of its own. No price list is published.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
The models are the product. What Noetik sells is access to its foundation models. GSK licenses OCTO-VC virtual cell models by subscription, and Noetik has deployed them to GSK for running and fine tuning. Agenus works with Noetik on TARIO-2 predictions of patient response.
The spatial dataset Noetik generates exists to train those models. No laboratory service, data product or drug candidate is offered alongside them.
Oversight is left undescribed. Drug company scientists use the models for target discovery, virtual experiments and grouping patients, so people decide what to do with the outputs. But Noetik publishes nothing on how predictions are reviewed, when a simulated result should not be relied on, or how a prediction used to select patients for a trial is checked.
TARIO-2 is presented as predicting individual patient response, which raises the stakes of that gap.
Noetik names its models and describes their data, without versions or update practice. It names OCTO, OCTO-VC and TARIO-2. The models learn by predicting missing biological information from the surrounding tissue, without hand labeled data.
The training data is described in detail: tumor samples from more than 4,500 patients, 196,000 tissue cores, 450 million cells and 2.9 petabytes. H&E images, a custom spatial protein panel, whole spatial transcriptomes and DNA sequencing are lined up across the same tissue. Technical reports describe OCTO and the virtual cell models.
No parameter counts, version history or notice of model changes for licensees is published.
Noetik builds its own models. It generates its own spatial dataset and trains its own foundation models, and no outside model provider is named or implied.
Everything around the models is unpublished: the sources of the tumor samples, any laboratory or data partners, the cloud provider, and the parties that handle licensee data during fine tuning.
Partners and research are named, but results are not set out. GSK is a named licensee. Noetik reports that OCTO-VC met GSK's testing criteria in non small cell lung cancer and colorectal cancer before the models were deployed, without publishing the criteria or results.
Its research page lists technical reports on OCTO and on virtual cells, a white paper on predicting response to newer checkpoint inhibitors, and presentations at SITC 2024 and 2025, AACR 2025 and ASCO 2026. The ASCO 2026 work, with Agenus, predicted response to botensilimab with balstilimab. None of these is peer reviewed, and no figures, patient groups or comparisons are set out on Noetik's pages.
How patient data moves through Noetik's systems is not described anywhere. The company describes the size and types of its tumor dataset, but not where the samples came from, what consent covers, how the data was deidentified, or how long it is kept. It also does not say whether a licensee's data used for fine tuning goes back into Noetik's training sets. No safety engineering for the models is published.
There is no HIPAA position and no privacy document that covers the product. The company builds its models from tumor samples, imaging, sequencing and spatial data from more than 4,500 patients, and deploys fine tuning to licensees who may bring their own patient data.
Its site publishes terms of use that cover the website only and disclaim responsibility for the privacy of information sent over the internet. No privacy policy, business associate position or data processing agreement is published.
No security certification, attestation, trust center or penetration testing statement is published, and the website terms disclaim responsibility for the privacy of information sent over the internet. The company holds a large dataset drawn from patient tumors and deploys models into licensees' research.
Noetik makes no device claim, and the models are sold for research and development. Licensed OCTO-VC models support discovery and patient grouping research inside drug companies, and nothing presents any model as a diagnostic. The FDA's databases list no Noetik device.
Predicting which patients respond to a therapy becomes a companion diagnostic use if it moves into selecting patients for trials. Noetik publishes no regulatory position for TARIO-2.
Testing is described only in outline, with no results by group and no governance process. The company reports that its models met a licensee's testing criteria, and lists technical reports and conference presentations. No results compared across cancer types, tissue sources, laboratories or patient groups are set out. Nothing describes how model behavior is evaluated before release or monitored afterward.
The dataset's patient sources and demographics are not described. That leaves open whether predictions hold for groups under represented in it.
What happens when a model is wrong is not covered in any published material. The only published terms cover the website, which is provided as is. No error rate, warranty, remedy or correction route for model outputs is published, and the GSK license terms are not public beyond its commercial structure.
No clinical record integration exists, and delivery into a licensee's systems is reported without detail. Noetik reports deploying OCTO-VC to GSK for running and fine tuning, which shows the models reach a buyer's research environment.
No platform, cloud environment, data format or interface is named. Nothing describes how a licensee's own data is brought to the models.
Where the models run and where data rests both go unstated. The company reports deploying the models to GSK for running and fine tuning, without saying whether that happens in GSK's environment or on infrastructure Noetik hosts. Nothing describes where the tumor dataset is stored, which cloud provider is used, or whether data crosses borders.
No price is published, but one contract's shape is public. GSK's license is non exclusive, runs for five years and is a subscription, with annual fees after $50 million upfront, covering models in two cancer types. That shows how Noetik sells: subscriptions to model access for named cancer types.
A contract value is not a price another buyer can apply, and no rate or tier is published.
The models' scope is named, and a partner's acceptance stands behind part of it. The licensed OCTO-VC models cover non small cell lung cancer and colorectal cancer, which GSK accepted against its testing criteria. TARIO-2 is applied to response to checkpoint inhibitor combinations with Agenus. The underlying data comes from tumor samples of more than 4,500 patients, without the cancer types listed on Noetik's pages.
Buyers are oncology research and development teams. The response prediction work with Agenus also ties the company to clinical trials. No published figures show how the models perform in either licensed cancer type.
Alternatives to Noetik
Other products that do the same kind of work as Noetik, and what each one does.
Biomy makes DeepPathFinder, cloud software that uses AI to map immune and cancer cells on tissue slides for drug research.
BioMap builds large AI models of protein biology and uses them to find drug targets and design biologic drugs, working with pharmaceutical partners.
Drug discovery company built around super resolution live cell fluorescence microscopy, applying advanced optics descended from Nobel recognized single molecule imaging work together with machine learning to track how individual proteins behave and move inside living cells, rather than…
South San Francisco company founded in 2023 and launched publicly in April 2024 with more than $1 billion in committed funding, co led by ARCH Venture Partners and Foresite Labs.
Earendil Labs uses AI to design antibody drugs and develops them through its own pipeline, licensing individual programs to large pharmaceutical companies.
Immunai occupies a different position from most of this category. It is not developing its own drugs.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Not published
|
Subscription licenses to foundation models by indication, with annual fees; partnered programs. No rate published. | Not published. No privacy policy or data processing terms are published. | Not published. | Vendor Published |
The only commercial terms in public are GSK's: a non exclusive, five year license to OCTO-VC models in non small cell lung cancer and colorectal cancer, structured as a subscription with annual licensing fees after $50 million upfront. That is a contract value, not a price list, and the Agenus work is a partnered program on undisclosed terms. The general subscription basis, and the cost of fine tuning on a licensee's own data, are not published.