Noetik’s partnership with GSK puts dollar signs on the AI licensing business model in TechBio

Of all the TechBio companies in DCVC’s portfolio, Noetik may have the business model that leans into AI the hardest. It doesn’t yet have its own pipeline of drugs in development. Its initial focus has been entirely on building world models for biology—large generative models that can simulate patients’ cells and make predictions that inform patient treatment and support the development of new cancer therapies. That makes it a relatively new beast in the biopharma universe: a company built mainly for the purpose of creating frontier AI models for biology, with broad partnering potential across the biopharma industry.

Industry watchers have been predicting the rise of generative-AI licensing companies as a new asset class in biotech­nology for a while now. But the big questions have always been: can they make money, and how? Don’t all successful life science companies need in-house drug assets? Now some answers are emerging. 

Today, Noetik announced a $50+ million collab­o­ra­tion and licensing agreement with pharma giant GSK. Under the terms of the five-year deal, GSK will pay upfront capital fees, as well as annual licensing or subscrip­tion fees, to access Noetik’s OCTO-VC virtual cell foundation models to help accelerate the development of drugs for non-small cell lung cancer and colorectal cancer. 

As far as we know, it’s the first example of a substantial pure-play AI licensing deal in TechBio — making it a kind of existence proof for this new business model. ​“We are moving the industry from AI services collab­o­ra­tions to licensing AI infra­struc­ture,” Noetik chief business officer Shafique Veroni said in the company’s announce­ment. ​“To our knowledge, this is among the first and largest trans­ac­tions monetizing a biological foundation model as a scalable enterprise asset.”

What makes Noetik’s technology so valuable to drug developers? In short, it uses the power of vision trans­formers, trained on vast amounts of proprietary multimodal spatial data from human tumor tissue, to predict how patients will respond to new therapies. That helps pharma companies short-circuit the trial-and-error approach that makes traditional drug development so slow, expensive, and risky.

Noetik begins by collecting tumor slices from thousands of cancer patients and analyzing them in its wet lab to highlight spatial gene and protein expression patterns, DNA sequences, and structural markers. These images become OCTO-VC’s training data — and through a self-supervised process of tokeniza­tion and masking, the model learns how each cell’s local environment affects its gene expression. Once it’s pre-trained on billions of tokens derived from patches of tissue slice images, the model can then be used to simulate unknown biology and answer questions such as: Which patients are most likely to respond to cancer immunotherapy, based on the types of cancer cells and fibroblasts found in their tumors? Which changes in gene expression might make tumor cells more susceptible to attack by the immune system?

As part of its non-exclusive licensing deal, GSK will also pay Noetik to generate bespoke human spatial datasets to train versions of OCTO-VC that answer questions of strategic interest to GSK as it hunts for treatments for non-small cell lung cancer and colorectal cancer. The datasets will allow GSK to ​“query tumor biology with a level of resolution that was previously impossible,” in the words of Noetik CEO and co-founder Ron Alfa.

The GSK deal leaves Noetik free to license OCTO-VC to other drug developers, and/or create other tailored models, trained on whatever type of tissue or data is most relevant to the diseases Noetik’s future partners want to treat. The quandary, up to now, has been knowing how to price such deals. We’re excited because we think the agreement with GSK creates a model for future part­ner­ships — and we can’t wait to see what kinds of new therapies and regimens Noetik and its drug development partners will unlock together.