World Models for Biological Design
BioDreamer applies JEPA-based world models and Active Inference to biological design. Instead of querying expensive real environments, the agent learns a latent simulator and plans optimal mutations, force-field changes, and gene perturbations in imagination.
Three Modules, One Architecture
BioDreamer operates across three biological scales (atomic, protein, and cellular), all sharing a common JEPA world-model backbone.
MolDreamer
Learns molecular dynamics in latent space. Predicts binding free energy and stability via SE(3)-equivariant GNN encoding and latent diffusion dynamics.
ProteinDreamerCore
Navigates protein fitness landscapes via dreaming. Plans multi-step mutation strategies using JEPA world models and Active Inference for principled exploration.
CellDreamer
Plans cell reprogramming and perturbation strategies. Uses neural ODE/SDE dynamics on scRNA-seq data to simulate multi-step gene interventions.
Where to start
Read the documentation, browse the codebase, or walk through the architecture and training pipeline.