Scope

Run BioEmu locally on a CUDA-capable NVIDIA GPU to sample protein conformational ensembles from amino acid sequence.
Two supported entry points:
1. CLI — pip install bioemu[cuda] and python -m bioemu.sample directly from the upstream package.
2. Reference app — the React + Flask app in assets/reference-app/. Run its bundled score/ Flask server locally on :5001, point the proxy backend at it via .env, and use the React frontend for visualization.
Analyze trajectories with MDTraj (RMSD, RMSF, Rg, secondary structure) and visualize with Molstar in the reference app.

Prerequisites

Linux (or WSL2 on Windows 11 with NVIDIA GPU passthrough — install WSL per Microsoft's guide and follow NVIDIA's CUDA-on-WSL guide for driver setup) with a CUDA-capable NVIDIA GPU. CPU works only for ~10-residue toy sequences; everything else is unusably slow.
Python 3.10+, ~5 GB free disk for cached weights (AlphaFold2 weights ~3.5 GB + BioEmu checkpoint + working space) at ~/.cache/colabfold/ (CLI) or /app/colabfold_cache (Docker).
Path B only: Docker with GPU support (Docker Desktop on Windows includes this; Linux needs nvidia-container-toolkit), or system Python for a bare-metal score/ run; Node.js 18+ for the frontend.