Visual Biology Foundation Model
AI that recognizes molecular biology in action.
Trained on proprietary experimental images of cellular molecular activities. Built to recognize biological patterns across new experiments.
Anima has spent more than a decade generating experimental images of molecular activities across cell types, diseases, conditions and perturbations.
From this accumulated Visual Biology, Anima has trained a self-supervised foundation model that learns how molecular activities appear in cellular images.
The model can recognize familiar visual patterns in previously unseen proteins, cellular phenotypes and biological contexts.
This learned visual intelligence is a core technology behind BioGPU, the Cellular Biology Runtime for AI Agents.
01 · Visual language
Teaching AI the visual language of molecular biology.
Molecular activities leave recognizable patterns. Our model learns them.
Inside cells, molecular activities unfold through changes in protein localization, interactions, signaling states, RNA organization and molecular processes.
Visual Biology captures these activities using experimentally defined molecular visualizers.
The foundation model learns visual representations from these images, identifying patterns associated with cellular phenotypes, molecular localization and activity-related changes.
These representations can then be applied to new experimental images without training a separate model for every protein or cellular context.
02 · Proprietary data
A foundation built on proprietary experimental biology.
Billions of experimental images. Diverse cellular activities. One foundation model.
Anima's foundation model is trained on more than 2 billion experimental images spanning 1,000+ cellular activities, 150+ cellular systems and diverse treatments and perturbations.
Its self-supervised architecture learns from the images themselves, without requiring researchers to manually label every cellular phenotype.
The training data captures variation across molecular markers, cellular states, biological conditions and experimental interventions.
This diversity gives the model reusable representations of cellular phenotypes and their visual relationships.
The result is a foundation for analyzing new biological images, including observations outside its original training dataset.
03 · Generalization
Recognizing biology beyond the training data.
The model's knowledge extends beyond the proteins and phenotypes used to train it.
A foundation model becomes powerful when its learned representations transfer to unfamiliar data.
Anima's model can organize previously unseen cellular observations according to visual patterns learned across its training corpus.
For example:
When presented with images of PCNA, a protein absent from training, the model placed its cellular phenotypes near visually similar phospho-eIF2α patterns in its learned representation space.
In fibroblasts displaying a previously unseen phospho-Src phenotype, the model grouped cells sharing similar visual characteristics despite the phenotype and cellular context being absent from training.
These are illustrations of the model's broader ability to transfer visual representations to new experimental observations.
04 · Biological evidence
Turning image recognition into biological evidence.
Every recognized cellular pattern remains connected to the experiment that produced it.
BioGPU analyzes experimental images at single-cell resolution, capturing features such as molecular localization, intensity, staining patterns and differences across cellular populations.
These observations are linked to the molecular visualizer, cellular system, treatment and experimental conditions.
The resulting Visual Biology Tokens preserve:
- Visual evidence: the experimental cellular images.
- Biological meaning: the observed molecular activity or cellular phenotype and its interpretation.
- Experimental context: the conditions under which the observation was generated.
Visual Biology Tokens make this evidence accessible to AI agents without separating biological findings from their experimental origins.
05 · Inside BioGPU
The intelligence that powers BioGPU.
From recognizing individual molecular activities to investigating interconnected cellular networks.
The Visual Biology Foundation Model gives BioGPU a reusable capability to interpret cellular images across experimental contexts.
BioGPU runs cellular network experiments, interrogating selected molecular activities across disease-relevant cells, conditions and timepoints.
The foundation model processes the resulting visual observations, supporting the generation of Visual Biology Tokens.
AI agents reason across these tokens to investigate relationships, identify unresolved biology and design subsequent experiments.
Through repeated experimentation, agents can progressively resolve cellular networks and evaluate how therapeutic molecules affect them.
The foundation model recognizes the evidence. BioGPU makes that evidence available for AI-driven Biology Resolution.