Ono Pharmaceutical is putting AI agents in front of the scientists who run its earliest-stage research. Under a collaboration announced on July 28, 2026, the Osaka drugmaker will deploy Biomni Lab, the agentic research platform built by South San Francisco applied-AI lab Phylo, across its drug discovery organization.

What the platform is meant to absorb, according to the announcement, is the full span of a discovery scientist’s computational day: synthesizing a program’s experimental history, reasoning over internal datasets, designing the next experiment, and executing the computational biology behind it. Phylo says Ono’s own discovery expertise and historical data will be brought into the environment, which is the part that distinguishes this from a software licence. Ono, listed in Tokyo as 4528 and founded in 1717, works in oncology, immunology and inflammation, and neurology, and is best known for the PD-1 antibody nivolumab, sold as Opdivo.

What Biomni was tested on

The agent underneath the product carries more published evidence than most enterprise AI deployments do. Biomni began as an open-source project at Stanford and was published in Science in 2026 as Autonomous biomedical research with an artificial intelligence agent. Its design is unusual: one agent mines tools, databases and protocols out of thousands of papers across 25 biomedical domains to build the environment, and a second plans over that environment, retrieving what a question needs and writing code to execute it instead of following a fixed workflow.

The paper reports the agent generalizing across causal gene prioritization, drug repurposing, rare-disease diagnosis, microbiome analysis and molecular cloning without task-specific tuning, with case studies covering protein-stability optimization and the orchestration of wet-lab instruments. That is a broader remit than the single-purpose models that dominate computational discovery, and closer to the wet-lab-facing co-scientist systems the field has been converging on since Google’s Co-Scientist.

What the benchmark scores show

The most useful read on what Ono is actually buying comes from Phylo’s own evaluation work. On June 30, 2026 the company published DrugDiscoveryBench with Scale Labs: 82 tasks written by working drug discovery scientists across target identification, patent mining and structure–activity analysis, each with a verifiable answer and an expert-written rubric, solved by agents writing code against real biomedical databases.

Frontier models clustered tightly and unimpressively. Phylo reports GPT-5.5 at 51.6%, Gemini 3.5 Flash at 50.0% and Opus 4.8 at 47.2%, with different models leading different task families. Two findings underneath the ranking matter more for a pharma rollout:

  • Holding the model fixed and changing only the surrounding agent architecture moved GPT-5.5 from 40.6% to 51.6%, and GLM 5.2 from 24.4% to 37.8%.
  • Handed a human-written playbook naming the steps and the tools, at least one model passed 76 of the 82 tasks.

Phylo is direct about the failure mode in between: on long workflows the agent quietly loses a constraint and returns a confident wrong answer. In one task it answered a melanoma question with a breast-cancer gene. That is the gap institutional knowledge closes, and it explains why the Ono agreement is framed around the company’s historical data rather than around raw model capability.

Where it fits in Ono’s discovery strategy

Ono has been assembling this for several years. Its research organization states that it uses AI to accelerate the search for discovery seeds and shorten the time to novel compounds, and it runs a deliberately external model of drug discovery: in March 2026 it expanded a computational discovery collaboration with Congruence Therapeutics into neurology and immunology.

“We believe AI will become a core capability for drug discovery,” said Seishi Katsumata, Ono’s corporate officer and executive vice president for discovery and research, who added that Ono’s researchers took up Biomni Lab quickly during evaluation.

For Phylo, the deal extends a customer list assembled in under six months. The company launched Biomni Lab in research preview on February 3, 2026 alongside a $13.5 million seed round co-led by Andreessen Horowitz and Menlo Ventures’ Anthology Fund, and has since published deployment work with Ginkgo Bioworks (DNA ) and agentic clinical-trial emulation with Mount Sinai researchers. It is also arriving in a market where large pharmaceutical companies have decided to buy agent infrastructure rather than build it, as AstraZeneca (AZN ) did when it expanded its Owkin partnership to put agents into development decisions.

Biomni Lab is available now, and the concrete work Ono has signed up for is the harder half: converting three centuries of accumulated discovery practice into procedures an agent can follow. On Phylo’s own benchmark, that is the step that takes an agent from half the tasks to nearly all of them.