Daniel Levine and David van Dijk want to know what happens when you give a drug to a human patient. Before you give the drug to the human patient.
It's the kind of pitch that sounds almost too good—the computational biology equivalent of asking a crystal ball what side effects might emerge in Phase II. But on March 3, their New York-based startup CellType landed its first outside bet: Senhwa Biosciences, a Taiwanese biotech firm, committed $100,000 in pilot collaboration funding under a strategic MOU to test whether CellType's AI platform could say something useful about CX-4945, Senhwa's cancer compound. The deal also gives Senhwa a right of first discussion when CellType raises its next equity round.
A hundred thousand dollars is pocket change in the world of pharmaceutical development, where Phase III trials can run into the hundreds of millions. But for a two-person team that just graduated from Y Combinator's Winter 2026 batch—one of roughly 190 startups in that cohort—the Senhwa arrangement carries weight beyond the dollar figure. It's validation. And in drug discovery, where AI has made big promises and delivered mixed results, validation matters.
Perhaps more than the founders expected, their underlying technology attracted attention from one of the industry's most formidable players. In October 2025, Google DeepMind announced it had used a 27-billion-parameter version of CellType's foundation model, Cell2Sentence, to predict a novel cancer therapy combination. The prediction—pairing silmitasertib with interferon to amplify antigen presentation—was subsequently validated in lab experiments. The drug in question? Silmitasertib. The same compound Senhwa is now developing as CX-4945.
That's not coincidence. It's a case study in how AI-driven drug discovery might actually work.
Teaching Machines to Read Cells
At the core of CellType's approach is Cell2Sentence, a biological foundation model that treats individual cells as if they were sentences in a language. Multi-omics data—gene expression levels, protein markers, epigenetic signals—gets translated into text. Large language models then train on that text, learning patterns the way they learn grammar.
Cells become sentences. Biology becomes syntax.
The model emerged from van Dijk's lab at Yale, where he's an assistant professor of medicine and computer science with a research footprint spanning machine learning and single-cell biology. His work has been cited over 11,000 times. Levine, his co-founder, studied machine learning at Yale and EPFL, built control software for CERN's Large Hadron Collider, and spent time leading foundation model training at a biotech startup before launching CellType in 2025.
Cell2Sentence was published in the ICML 2024 proceedings—a collaboration between the van Dijk Lab and other researchers—and has since been released in multiple versions. Google's 27-billion-parameter variant, built on the Gemma architecture, is the largest publicly documented iteration. When Google used it to run a virtual screen last fall, analyzing how silmitasertib might interact with interferon signaling, the model predicted an amplification effect under certain immune conditions. Lab tests confirmed it. Sundar Pichai amplified the announcement. For a brief moment, the discovery rippled through research circles.
For CellType, it was both credential and proof-of-concept.
The Field Gets Crowded
CellType isn't alone in trying to apply large language models to single-cell biology. The field has accelerated sharply over the past year. Benchmarks like CellVerse arrived in May 2025; surveys like LLM4Cell followed in October. OpenAI released GPT-Rosalind, a biology-tuned LLM, in April 2026. The terrain is getting competitive.
What CellType claims as its distinguishing feature is the "agentic" layer—AI agents orchestrating discovery workflows on top of the foundation models, automating the iterative work of hypothesis generation and testing. Target discovery, translational prediction, virtual trials. All framed as solvable through better biological language models and smarter orchestration.
In a February 2026 launch post on Y Combinator's platform, the company described a pipeline that had already screened over 4,000 drugs and surfaced what it called a "new cancer treatment signal." The post also mentioned collaboration "with Top 10 pharma," though no names have been disclosed publicly. Perhaps they're waiting for the data to speak louder than the partnerships.
A Modest Deal With Larger Implications
The Senhwa arrangement is structured as a services agreement, not equity. CellType will apply its platform to Senhwa's CX-4945 program; Senhwa, a publicly traded biotech on Taiwan's TPEx, gets early access to whatever CellType's AI uncovers. According to Senhwa's public disclosure, the partnership will feature in CellType's Demo Day materials—logo, case study, the works.
For Senhwa, it's a low-risk experiment. For CellType, it's real-world traction at a moment when investors are increasingly skeptical of AI hype in life sciences. Drug discovery has seen waves of computational tools that promised to compress timelines and reduce failure rates. Some delivered incremental improvements. Many didn't. Foundation models may prove different, or they may run headlong into the same stubborn biological complexity that has humbled earlier technologies.
The next six months will offer data points.
Building a Team, Betting on Scale
CellType is hiring. As of early May 2026, three roles were listed on Y Combinator's job board: founding engineers for data systems and model training, and a head of strategy and partnerships. The company describes itself as "the agentic drug company," positioning AI agents as the orchestration layer that turns foundation models into discovery engines.
Y Combinator's standard investment is $500,000, typically structured as two SAFEs: $125,000 post-money for 7% equity and $375,000 on an MFN SAFE. Some databases list only the first tranche, which can create confusion about the company's actual funding. Pioneer Fund, a Y Combinator alumni-led fund, has been reported as an investor, though that hasn't been independently confirmed on Pioneer's portfolio page.
The company's stated vision is ambitious—virtual human simulations that can answer questions about toxicity, efficacy, and patient stratification before clinical trials begin. Whether that vision holds up is another question entirely. Drug development is an industry that punishes overconfidence and rewards skepticism.
But Levine and van Dijk have something many AI startups don't: academic credibility, external validation from Google, and now a pharma partner willing to put money on the table. Modest money, yes. But money nonetheless.
If the agentic drug company thesis holds, if Cell2Sentence can genuinely predict what happens in human biology before the first patient enrollment, then $100,000 from Senhwa might look, in hindsight, like the beginning of something. If not, it will be another footnote in the long history of technologies that promised to revolutionize drug discovery.
The Senhwa collaboration will run through early September. By then, we'll know a little more.
