Roadmap
Where we are and where we're headed. BioDreamer is an active research project and milestones are updated regularly.
Project Foundation
CompletedQ1 2026- Repository structure and modular architecture
- Core base classes (encoder, dynamics, decoder, reward, policy)
- YAML configuration system
- Project website and blog
ProteinDreamer v0.1, Energy-Based JEPA
In ProgressQ1–Q2 2026- ESM-2 + GVP-GNN dual encoder with FusionMLP
- Energy-Based JEPA predictor with SIGReg
- Fitness reward head (ΔΔG, Kd)
- PPO policy in latent space
- Benchmark on ProteinGym single-mutant landscapes
ProteinDreamer v0.2, Latent Diffusion JEPA
PlannedQ3 2026- Conditional latent diffusion predictor
- Active Inference with EFE-based exploration
- Multi-step mutation planning on evolutionary paths
- Structure-aware encoding via ESMFold
Multi-Objective & Active Learning
PlannedQ4 2026- Multi-objective Pareto ranking (stability + activity + expression)
- Active learning loop (dream, propose, evaluate, update)
- Uncertainty calibration (ensemble, evidential, MC-dropout)
- Hugging Face Hub integration for model sharing
MolDreamer & CellDreamer
Planned2027- MolDreamer: SE(3)-equivariant encoder + latent diffusion dynamics
- CellDreamer: scVI-VAE encoder + neural ODE/SDE dynamics
- FastAPI backend + interactive web interface
- Comprehensive benchmarks and documentation