BioGPU tech stack

Every layer AI needs to work on cellular biology, built as one system.

BioGPU is the Cellular Biology Runtime for AI Agents. AI agents design and run experiments on it, from the experiment design to the tokens that bring the result back into AI; its robotic lab executes them. Anima built every layer, and each one depends on the layer before it. BioGPU generates the next evidence rather than stopping at an existing dataset: for each target program, agents design and run new experiments on new interventions, across cellular systems and therapeutic modalities. The capability extends from understanding a compound's cellular effects to investigating the biological questions those effects raise.

One integrated stack · design goes down, evidence comes back up

The components

Every cycle tests a network hypothesis in cells and returns evidence for the next decision.

  1. 01

    BioGPU Agent

    Anima's specialized agent and its expert in cellular network experiments. Given a disease, a target and a question, it forms the network hypothesis, designs and runs the cellular network experiments on BioGPU, selects the molecular-specific visualizers and perturbations within them, and designs the next experiment from what comes back. It carries more than a decade of Anima's experience in designing these experiments.

    Designs the experiment · places the visualizers · directs the next step

  2. 02

    Cellular network experiments

    The wet-lab layer. Experiments run in patient-derived and other disease-relevant cells, measuring many molecular activities in parallel across conditions, perturbations and timepoints. Fixed-cell images capture molecular activities at each timepoint.

    150+ cellular systems run on BioGPU

    Explore cellular systems →
  3. 03

    Molecular-specific visualizers

    Each visualizer makes one molecular activity visible inside the cell: a protein interaction, a translocation, translation, RNA localization or a signaling state. Placed across the network, they show how the activities change together.

    1,000+ cellular activities with their own visualizers

  4. 04

    Visual Biology Foundation Model

    The model recognizes the molecular activities in the experimental images, at single-cell resolution, and derives the state of the network. Every finding is a recognition of what was measured.

    Trained on 2B+ proprietary experimental images

  5. 05

    Cellular Biology Tokens

    What the model recognizes is returned as Cellular Biology Tokens. Each token links three parts: the image evidence, the biological meaning and the experimental context. They go straight into an AI agent's context window.

    Image evidence · biological meaning · experimental context

  6. 06

    AI agents

    Your AI agents, connected to BioGPU, and the BioGPU Agent reason across the tokens, refine the network and design the next run. Drug-design AI then tests every molecule it designs against the network.

    Reason · Run · Recognize · Resolve

  7. 07

    BioGPU Program Runtime

    For each target program, the stack is set up as a dedicated runtime: the target, a disease-relevant cellular system and the cellular network resolved around it. Every compound in the program then runs on it.

    One runtime per target program

    See how a program runs →

How we test what the tokens add

  1. PredictThe AI answers from the literature alone, before seeing data.Literature only, before the data
  2. ObserveIt reads the Cellular Biology Tokens from the experiment.The tokens from the run
  3. ScoreThe AI scores each committed claim against the BioGPU data; Anima's experimentally resolved mechanism is kept back as the answer key.Vs. Anima's resolved mechanism

What we offer next: the same AI gets unfamiliar questions about cellular interventions, with and without BioGPU, under the same conditions and time budget, scored against the experimentally resolved mechanism.

Reasoning on the network

With the cellular network in view, AI designs better compounds and sharper hypotheses.

Each finding adds to the picture of the network around the target. With that picture in view, AI agents see connections no single experiment reveals, and use them for two things: designing the next AI compound, or forming the next network hypothesis to test.

Tokens on the network, the agent's context, and the next hypothesis or compound design

Because Anima built every layer, each round sharpens the next: the agent learns what to run.

Questions an AI lab asks

What BioGPU gives an AI model, in plain answers.

What are BioGPU's 2B+ images?
Images of molecular activities in disease-relevant cells. A molecular-specific visualizer, a fluorescent probe for the activity of a protein or mRNA, makes one activity visible, such as a protein interaction, a translocation or a translation event, and selected activities are interrogated together in the same cells, under the same conditions and at the same timepoint.
How does that help an AI model learn biology?
The Visual Biology Foundation Model recognizes 1,000+ cellular activities in these images. Cellular Biology Tokens return each recognized activity to the agent with its image and experimental context, so the agent can test its hypotheses against measured biology and direct the next experiment.
How is this different from morphology profiling?
Morphology profiling images the overall appearance of cells and suits broad screens. BioGPU measures molecular activities chosen for the question, so the result names the activity that changed. The two are complementary.
What does BioGPU give an AI agent?
One runtime combines visualizers, disease-relevant cellular systems, the foundation model and Cellular Biology Tokens. The model recognizes molecular activities; the tokens return those findings to the agent.
Is there evidence that BioGPU data changes an AI's answer?
Yes, in a blinded test: a leading AI model committed to its predictions about a cancer drug from the literature, then answered again with BioGPU data, and the data overturned its answer. See BioGPU Challenge #1 →

Give your AI the biology it is missing.

Pharma team? See Pharma Workflows →

Diagrams on this page are schematic.