Noetik

Noetik builds AI foundation models of human tumor biology to discover cancer therapies and identify patients likely to respond to treatment. Its work spans virtual-cell simulations and TARIO-2, which predicts spatial gene-expression maps from routine pathology images. Alongside its therapeutic-development ambitions, Noetik has licensed foundation models to GSK in a collaboration that includes $50 million in upfront capital and near-term milestones.[6][7][5][8]

These models grew from an earlier effort to collect multimodal data from human tumors: Noetik began building that dataset after its founding in mid-2022.[3][6]

Ron Alfa and Jacob Rinaldi co-founded Noetik in mid-2022. Alfa had been head of research at Recursion; Rinaldi had led its oncology work and previously worked in cancer research at Genentech. At Noetik’s September 7, 2023 launch from stealth, Alfa was CEO, Rinaldi was chief scientific officer, and Lacey Padron was chief technical officer, having previously served as vice president of informatics at the Parker Institute for Cancer Immunotherapy. The South San Francisco company launched with a $14 million seed round led by DCVC.[3][4]

Alfa’s concern was the gap between promising preclinical immunotherapy results and disappointing human trials, particularly the difficulty of knowing which patients should respond. Noetik used patient tumor samples and self-supervised learning, allowing models to learn patterns directly from data rather than requiring people to specify what to find. At launch, Alfa described the goal as seeing underlying tumor biology and identifying new cancer-drug targets.[3]

In September 2023, DCVC reported that Noetik had analyzed over 1,000 human lung cancer samples and reached 18 employees in less than 12 months of operation. Noetik also reported generating hundreds of terabytes of human data in less than a year. An April 2026 Latent Space article described almost two years devoted to data collection and identified four modalities: spatial transcriptomics, spatial proteomics, H&E imaging, and whole exome sequencing.[4][3][1]

On August 29, 2024, Noetik announced that it had closed an oversubscribed $40 million Series A, led by Polaris Partners, with new investors including Khosla Ventures, Wittington Ventures, and Breakout Ventures. Polaris managing partner Amy Schulman was announced as joining the board. Planned uses included expanding the human spatial-omics atlas, scaling foundation-model training, and advancing therapeutic candidates toward the clinic. Its discovery platform paired human spatial-omics data with Perturb-Map, a massively multiplexed in vivo CRISPR perturbation platform, to train self-supervised models of tissue and tumor biology.[5]

OCTO was Noetik’s first foundation model for patient-level spatial biology. In its December 18, 2024 technical report, the company introduced the next generation, OCTO-VirtualCell (OCTO-vc), and reported training it on nearly 40 million cells across multiple cancer indications. OCTO-vc is a transformer trained through masked token modeling, with tokens encoding gene expression. A virtual cell predicts which genes a cell would express in a particular environment, even when no such cell is physically present. Neighboring-cell gene expression supplies local context; prompts specify genes that are expressed or unexpressed in the virtual cell.[6]

This lets researchers place a virtual T cell at different locations in a tumor and predict whether its gene-expression state would indicate activation or inhibition. Noetik introduced Celleporter, an interactive viewer for exploring these simulations across patient tissues and changing virtual-cell prompts. In one reported counterfactual experiment, increasing MHC I expression in tumor cells increased predicted cytotoxic gene expression in nearby virtual CD8 T cells for most patients, while some samples showed little or no response.[6]

Noetik also used OCTO-vc to compare virtual CD8 cells across hundreds of samples from patients with and without STK11 tumor mutations. The company reported lower predicted expression of cytotoxic effector genes GZMA and GZMK in the mutated cohort. It then simulated tumor-cell gene knockouts one by one, looking for changes that brought those virtual-cell expression levels closer to the comparison cohort. The resulting targets were candidates for further validation; the computational screen did not establish therapeutic efficacy.[6]

The next deployment challenge was access to data. Noetik’s rich spatial-transcriptomics measurements were research assays, while H&E pathology images were inexpensive and routinely collected. In April 2026, Noetik described TARIO-2, which extended its earlier TARIO architecture to sequences containing both H&E and spatial-transcriptomics tokens. Unlike OCTO-vc’s masked-token objective, TARIO-2 uses autoregressive next-token prediction. At inference, it needs only an H&E image to generate predicted spatial gene-expression maps, making the model usable with existing pathology images rather than requiring a new spatial-transcriptomics assay.[7]

The April report described a 200-million-parameter model, trained on 2,545 patients and validated on 213 held-out patients with no patient overlap. Noetik reported spatial correlations around 0.3–0.5 for the highest-signal genes, declining as noisier genes were included. Correlations for spatially aggregated gene abundance were around 0.6 across most cohorts. These measured gene-expression prediction, rather than treatment-response accuracy. The article credited contributions from Eshed Margalit, Daniel Bear, Lacey Padron, Dexter Antonio, Michela Meister, Ron Alfa, Dan Millman, and Dulce Ovando Morales.[7]

On June 15, 2026, Noetik announced a retrospective TARIO-2 study in a white paper authored with Agenus. The company reported that routine pretreatment H&E images identified patients with improved response and survival outcomes after treatment with botensilimab and balstilimab, the BOT+BAL immune checkpoint inhibitor combination. This supplied a treatment-response research example beyond transcriptome reconstruction, while remaining a retrospective finding rather than evidence of prospective patient benefit.[9]