Biotech & drug discovery · AI foundation models
CellType
A two person Yale spinout that translates single cell biology into a language large models can learn, and is using that model to steer a real cancer drug's next act. Founded in 2025, no priced round on the public record yet.
A cell becomes a sentence
Take a single cell's RNA sequencing profile, a list of tens of thousands of gene expression measurements, and rank the genes from most to least active. The resulting string of gene names is what David van Dijk and Ivan Vrkic call a cell sentence. Feed enough of them to a large language model and, the founders argue, it starts to learn the grammar of biology the same way GPT style models learn the grammar of English.
That idea, called Cell2Sentence, was published at ICML 2024 while van Dijk and Vrkic were still at Yale. It is the technical foundation the company was built on. “We're building an agentic drug company: foundation models that learn human biology + AI agents that run the full discovery pipeline,” the founders wrote in their October 2025 launch post. “Instead of mice and cell lines, we simulate what will happen in patients.”
The problem it is aimed at
About 90 percent of drugs that enter clinical trials fail, and a typical approved drug takes more than ten years and over $2 billion to develop, according to the company's own launch materials. A large share of that failure is attributed industry wide to preclinical models, mice and cultured cell lines, that do not reliably predict what a drug will do in a human patient. CellType's pitch is that a foundation model trained directly on human single cell data can stand in for some of that testing before it happens in a lab or a trial.
That pitch faces real competition. A tier list write up of the Y Combinator Winter 2026 batch called the space “brutally competitive with well-capitalized incumbents like Recursion” and said the open question is whether “working with Top 10 pharma” converts into signed, disclosed contracts.
Two founders, one from Yale, one from CERN
David van Dijk is a Yale professor with more than 11,000 citations and papers in Cell, Nature, ICML and NeurIPS. According to the company, he turned down a job at Google to build CellType. Ivan Vrkic co developed Cell2Sentence with van Dijk at Yale, studied at EPFL, previously led large scale foundation model training at a biotechnology company, and built software used to help control the Large Hadron Collider at CERN.
“We're David and Ivan,” the pair wrote at launch, adding that “we built this technology at Yale and are now turning it into a company.” The Y Combinator company page lists a team size of five, though every independent article reviewed for this profile, including a Y Combinator batch tier list, describes CellType as a two person company. Who the other listed employees are is not stated anywhere found.
A hypothesis that held up in living cells
The company's headline result did not start as a company result. On October 15, 2025, Google published a blog post describing C2S Scale 27B, a 27 billion parameter model built on its open Gemma architecture, as the product of “our research collaboration with Yale University.” The post does not mention CellType. Researchers used the model to run a dual context virtual screen of more than 4,000 drugs, looking for a compound that would boost antigen presentation, a signal that helps the immune system see a tumor, only in a patient relevant immune context. The model flagged silmitasertib (CX-4945), a CK2 inhibitor, as a conditional amplifier. Yale researchers then tested it in human neuroendocrine cell models the system had never seen during training. Silmitasertib alone did nothing. Low dose interferon alone had a modest effect. Combined, the two produced roughly a 50 percent increase in antigen presentation.
Google CEO Sundar Pichai posted about the result on X, calling it “an exciting milestone” for AI in science; independent coverage put the post's views at more than 6.8 million. CellType's own materials, published the same week the company launched, describe the finding as something the company “has already discovered and validated.”
From academic result to a pharma MOU
Five months after the Google post, on March 3, 2026, Senhwa Biosciences, a Taiwan listed clinical stage company (TPEx: 6492), announced a strategic memorandum of understanding with CellType to fold its AI platform into the clinical development of Senhwa's lead asset: the same drug, silmitasertib, that the model had flagged. The release describes a six month pilot covering indication expansion, biomarker discovery, and combination therapy evaluation, and says the deal “establishes Senhwa as CellType's founding strategic partner.” It is a memorandum of understanding, not a signed licensing or product deal, and it leaves room for “joint ventures, co-development, or licensing arrangements” later rather than committing to any of them now.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Cell2Sentence method | Published at ICML 2024, preprint on bioRxiv, code and a GitHub repository public since 2023 to 2024. | Public record |
| C2S Scale 27B result | Google's own October 15, 2025 blog post attributes the model and the silmitasertib finding to “our research collaboration with Yale University.” CellType is not named in Google's post or in independent coverage of it (Interesting Engineering, October 16, 2025). | Differs from company framing |
| Silmitasertib plus interferon result | Roughly 50 percent increase in antigen presentation in human neuroendocrine cell models, per Google's release, confirmed “multiple times in vitro.” | Public record |
| Senhwa Biosciences MOU | Signed March 3, 2026, announced by Senhwa (TPEx: 6492) via press release. A memorandum of understanding and six month pilot, not a definitive licensing or revenue agreement. | Partner-stated |
| “Working with Top 10 pharma” | Stated by the company in its launch post and repeated in coverage. No pharma partner besides Senhwa is named anywhere found. | Company-stated |
| FDA clearances, clinical trials, NIH or NSF awards | None found under CellType's name. | Not found |
| Team size | Y Combinator's company page lists 5. Independent coverage (BiopharmaTrend, StartupHub.ai, The YC Tier List, Hiretop) consistently describes CellType as a two founder company. | Differs from record |
Read plainly: the science behind CellType is real and independently confirmed by Google's own account of the work, but that work was framed by Google as a Yale University collaboration, not a CellType achievement, and it predates the company's commercial existence by only days. The Senhwa relationship is an early stage MOU built around that same molecule, not yet a signed commercial contract. Claims about unnamed Top 10 pharma engagements and the size of the team cannot be checked against any public record.
What to watch
- Whether the Senhwa pilot converts into a signed licensing or co-development agreement, and on what terms.
- Whether any of the “Top 10 pharma” relationships are named or produce disclosed deal terms.
- A first Form D or other financing disclosure filed directly under CellType, Inc.'s own name, rather than through investor side vehicles.
- Published, peer reviewed results beyond the single silmitasertib finding.
In their words
“We're building an agentic drug company: foundation models that learn human biology + AI agents that run the full discovery pipeline.”
David van Dijk and Ivan Vrkic, launch post, Oct 14, 2025 · Company
“Instead of mice and cell lines, we simulate what will happen in patients.”
David van Dijk and Ivan Vrkic, launch post, Oct 14, 2025 · Company
“It's not too often that a tenured professor steps away to build a company around their own work.”
Dave Messina, General Partner, Pioneer Fund's Future of Health fund, quoted by The Silicon Valley Post, Mar 24, 2026 · Investor
“With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer.”
Sundar Pichai, CEO, Google, post on X, Oct 2025 · Independent
“we are redefining the strategic positioning of CX-4945.”
Benny T. Hu, Chairman, Senhwa Biosciences, press release, Mar 3, 2026 · Partner
“Two people cannot fight that battle forever.”
Daniel Singer, StartupHub.ai, May 12, 2026 · Independent
“If they can close even one enterprise pharma deal with disclosed terms, this flips to S-tier.”
The YC Tier List, 2026 · Independent
Related companies
Sources
- CompanyCellType: The agentic drug company (Y Combinator company page and launch post)
- IndependentY Combinator Backs Yale Spinout Using Agentic AI to De-risk Clinical Trials
- IndependentCellType Uses AI to Simulate Human Biology for Drug Discovery
- IndependentCellType YC W2026: Teaching LLMs to Speak Biology (Claude's Corner)
- IndependentCellType tier list verdict, W26 batch
- IndependentCellType: The Agentic Future of Drug Discovery
- IndependentHow a Gemma model helped discover a new potential cancer therapy pathway
- Independent'Milestone': Google AI reveals new method to make cancer treatable
- PartnerSenhwa Biosciences Signs Major Strategic MOU with Y Combinator-Backed AI Biotech Company CellType
- Public recordCell2Sentence: Teaching Large Language Models the Language of Biology
- Public recordPioneer Fund Future of Health, portfolio page
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
