Features

The full stack for model-based biological design, from JEPA world model training to Active Inference policy learning and candidate ranking.

World Model Planning

Dreams multi-step intervention strategies in latent space before committing to expensive real-world experiments.

JEPA Architecture

Joint-Embedding Predictive Architecture encodes observations into latent states and predicts transitions without a decoder during planning.

Active Inference

Grounded in the Free Energy Principle. Balances exploitation (fitness) and exploration (epistemic uncertainty) automatically via expected free energy minimisation.

Multi-Scale Biology

Operates across atomic (MolDreamer), protein (ProteinDreamer), and cellular (CellDreamer) scales with a unified codebase.

Multiple Backends

Choose from Latent Diffusion JEPA (default) or Energy-Based JEPA with SIGReg regularisation. Both share the same encoder-predictor interface.

Modular Design

Swap any component (encoder, JEPA predictor, reward head, policy) independently. Built with PyTorch and clean abstractions.

Multi-Objective Optimisation

Optimise for stability, binding affinity, catalytic activity, and expression simultaneously with Pareto-based ranking.

Active Learning Loop

Dream, propose, evaluate, update. Iteratively refines the world model with each round of experimental feedback.

Hugging Face Hub

Push and pull pre-trained models via Hugging Face Hub with safetensors serialisation and auto-generated model cards.

Uncertainty Quantification

Ensemble, evidential, and MC-dropout methods provide calibrated uncertainty estimates to guide exploration.

FastAPI Backend

Production-ready REST API with Redis job queue, GPU worker pool, and real-time progress streaming.

Reproducible Science

YAML configs, W&B tracking, deterministic seeding, and versioned model checkpoints for fully reproducible experiments.