The Wrong Company, the Right Question About AI Drug Discovery

The recent headline about an AI agent partnership in drug discovery names the wrong company. The deal was announced July 27, 2026, between Phylo and Ono Pharmaceutical - maker of Opdivo, the pioneering cancer immunotherapy, founded in Osaka in 1717. It was not Chugai, a separate Japanese pharma that is majority-owned by Roche. The mix-up is easy to make. Both are Japanese oncology players. Both invest in AI.

But the error points to something worth looking at. If the headline gets the name wrong, the reader probably doesn't know what story is actually being told.

Phylo is a startup founded by Kexin Huang, a Stanford computer scientist who originally released Biomni as an open-source AI agent for biomedical research in June 2024. Tens of thousands of scientists across 7,000+ research labs adopted it. In February 2026, Phylo launched commercially with a $13.5 million seed round co-led by Andreessen Horowitz and Menlo Ventures' Anthology Fund, with support from Anthropic. Biomni Lab integrates over 300 databases, software systems, and analytical tools into what they call an Integrated Biology Environment. Scientists use it to synthesize experimental history, design experiments, reason over internal data, and run computational biology workflows.

That's not the same thing as the companies everyone has been paying attention to.

The dominant story in AI drug discovery has been about platforms that promise to design molecules. Insilico Medicine, now valued above $4 billion, is pushing Rentosertib toward Phase III as the first clearly AI-designed drug to reach late-stage trials. Recursion has built automated laboratories and accumulated more than 50 petabytes of multimodal data, pursuing closed-loop experimentation at industrial scale. These companies are selling the idea that AI will replace the human intuition that has guided molecule design for decades.

Phylo is selling something different. It's not trying to find the next blockbuster compound. It's trying to make every biologist 10 times faster at the work they already do.

Most people think the most valuable company in AI drug discovery will be the one that finds the first approved AI-designed drug. That assumes the prize is a single molecule. But a drug is one asset. Infrastructure can improve hundreds of assets. An agent that helps scientists move from question to hypothesis to experiment in hours instead of weeks doesn't need to be right every time. It needs to be right more often than the human doing it alone.

This is the same distinction that matters in AI more broadly. The companies racing to build AGI want to replace the thinker. The companies building tools that make existing thinkers faster are betting on something humbler and probably more profitable. Copilot didn't replace programmers. It made every programmer more productive, and Microsoft captured the value of that productiveness across the entire software industry.

Ono is a natural early customer. It's an R&D-driven company with deep expertise across oncology, immunology, inflammation, and neurology. Seishi Katsumata, Ono's executive vice president of discovery and research, said their own scientists quickly adopted Biomni and saw its potential. The researchers wanted it before the executives signed the deal. That matters because most enterprise AI deals die in the pilot phase when users find the tool slow, inaccurate, or just not useful enough to change how they work.

Phylo showed one concrete result early on: a collaboration with Ginkgo Bioworks where Biomni accelerated over 10 complex cell-painting and transcriptomic analyses, cutting what would normally take weeks down to hours. The results were validated by Ginkgo scientists and reached publication quality. It's not a drug. It's proof that the tool does actual work.

I suspect the reason Phylo's Ono partnership didn't get the headline treatment it deserved is that it doesn't tell the flashy story. There's no AI-designed molecule, no Phase III program, no $4 billion valuation. There's a company that built a better workbench for scientists and a pharma company that decided to let its scientists use it. The outcome is less cinematic and probably more consequential.

The risk for Phylo is clear. It's still a seed-stage company selling into pharma, where procurement cycles are slow and competitive platforms like Benchling and Cytoreason already occupy the lab-workspace space. And if the big players - Recursion, Insilico, Lilly's TuneLab - end up dominating the molecule-design layer, Phylo could find itself building tools for a workflow that those platforms eventually automate entirely. Infrastructure plays only win if the underlying workflow stays human-in-the-loop long enough to make the tool valuable.

What to watch: whether Phylo lands a second or third pharma partnership beyond Ono, and whether the timelines they're publishing actually hold up under the scrutiny of real drug-discovery projects rather than pilot analyses. The Ginkgo case study is promising but small. Ono's deal is the first real test. The next two will tell you whether this is a category or a niche.