For AI evaluators
What BioGPU is, what is measured, and how to test it on your own model.
BioGPU is the cellular biology runtime for AI agents. AI agents run cellular-network experiments on it, in disease-relevant cells including patient-derived cells, and read which molecular activities changed in the network around a target, returned as Cellular Biology Tokens.
Built for AI agents
Built from the ground up for AI agents to experiment in cellular biology and learn from it.
Experiments run in parallel in the same cellular network around a target. They span cells, conditions and timepoints. Many molecular activities are measured together, and the findings return to the agents as Cellular Biology Tokens. Automating individual assays one by one measures each activity in isolation. The biological context between them is lost, and covering the same activities takes many separate runs. BioGPU is designed so that agents keep improving their biological reasoning with each experiment. See it on the homepage →
Measured and illustrative
Every visual on this site is either measured or illustrative.
The Challenge #1 images and results, the two Cellular Biology Token images on the homepage and the observations in the rheumatoid arthritis case. The 98 cellular systems in the library come from Anima's records of what has run on BioGPU.
The homepage BioGPU run, the workflow and recognition schematics and the agent-session example show how BioGPU works, not a result; where one could be read as a result, it carries an "Illustrative" label. In the rheumatoid arthritis case, the mechanism is a hypothesis and the next round is a plan.
The evidence
Two pieces of evidence, with their status.
- 01
BioGPU Challenge #1 · TFEB and Gedatolisib
Where does TFEB go 24 hours after the PI3K/mTOR cancer drug Gedatolisib, in four human cell states? A leading AI model committed its predictions from the literature before seeing any data, then saw BioGPU's data and scored each prediction against it. Anima's experimentally resolved mechanism was held back as the answer key. From the literature the AI said "into the nucleus", with 80% confidence. BioGPU's data showed TFEB on the recycling compartments beside the nucleus, and one cell state not responding at all. 3–4 of 30+ committed predictions held; 9–12 could have misdirected a program.
Status: completed · two internal runs · September–October 2026
See Challenge #1 → - 02
Rheumatoid arthritis program · CK2α1
Patient-derived RA synoviocytes compared with healthy-donor cells, with 17 molecular-specific visualizers placed around CK2α1. SRC built up in the nucleus while its activation mark went down, and CK2α1 moved into the nucleus. The interpretation, a nuclear kinase relay that may reshape splicing, is a hypothesis; the report recommends the next round of experiments.
Source: Anima target-resolution report · July 2026
See the case →
The facts
One set of figures, used across the site.
2B+ images of molecular activities. 1,000+ cellular activities recognized. 150+ cellular systems run on BioGPU; the Cellular systems library documents 98 of them, 10 patient-derived. 25 diseases. 25 drug-discovery programs, pharmaceutical and internal. BioGPU was developed with over $150 million received through strategic pharmaceutical collaborations with Lilly, Takeda and AbbVie. Anima owns the image corpus and the Visual Biology Foundation Model trained on it.
Diligence
What to ask for.
- The Challenge #1 prompt and scoring, to run on your own model.
- The Challenge #1 data, under agreement.
- A walk-through of the rheumatoid arthritis program, from the question to the recommended next round.
Run Challenge #1 on your model.
We share the prompt and scoring; the data under agreement.