| name | getting-started |
| description | First-run setup for jpx. Installs prerequisites, downloads data, and launches a notebook. Use when someone says "help me get started", "set up this repo", "how do I run this", or opens the repo for the first time. |
Getting Started with jpx
Prerequisites
- pixi - multi-environment package manager (handles Python, CUDA, conda, and PyPI deps)
- uv - Python package runner (provides
uvx for launching marimo)
- rclone - for downloading data from public S3 (no credentials needed)
just and marimo are installed automatically by pixi install.
On NixOS, pixi is provided by the flake: nix develop or direnv allow.
Setup steps
- Install dependencies and activate the environment:
pixi install
pixi shell
pixi shell puts just, marimo, and other project tools on PATH.
- Choose a data path:
Fast path - download pre-computed results (~12 GB):
just get-results
Full path - download inputs and run the pipeline (~32 GB inputs, ~13 min on 4x H100):
just get-inputs
just redun main
just redun batch_source7
just redun explore
just redun srijit_all
- Launch a notebook to verify:
uvx marimo@latest edit --sandbox notebooks/nb16_ss_target_consistency.py
uvx marimo@latest --sandbox runs notebooks with their PEP 723 inline dependencies in an isolated env, independent of the pixi env.
Remote access (SSH): bind to all interfaces and pick a free port:
uvx marimo@latest edit --sandbox --host 0.0.0.0 --port 3719 notebooks/nb16_ss_target_consistency.py
Then open http://<hostname>:3719?access_token=<token> in your browser (the token is printed at launch). To skip the token: add --token-password ''.
If uv panics with "Too many open files", raise the fd limit first: ulimit -n 65536. This is a known uv issue on systems with a low default limit (1024).
- Install third-party agent skills (marimo notebook authoring, paper implementation, etc.):
npx skills add marimo-team/skills --agent claude-code -y
npx skills add marimo-team/marimo-pair --agent claude-code -y
What to explore first
nb16_ss_target_consistency - target consistency volcano plots and UMAP overlaysnb14_ss_cross_source_reproducibility - how well results replicate across data centersnb19_ss_compound_traits - which chemical properties predict morphological activitynb22_ss_activity_consistency_relationship - how single-compound activity relates to target-level consistency
Once set up, use the compose-notebook skill to compose new analyses from the catalog.
GPU notebooks
Notebooks that need NVIDIA GPUs (nb08, nb09, nb17, nb20, nb25, nb38-nb40) use the rapids pixi environment:
pixi run -e rapids marimo edit notebooks/nb08_ss_umap_seed_scan.py
Do not use --sandbox with GPU notebooks - it shadows conda packages.
Troubleshooting
ModuleNotFoundError on launch: you forgot --sandbox (pure-Python) or used --sandbox with a GPU notebook- Pipeline hangs at copairs: check
CUDA_VISIBLE_DEVICES if GPU memory is contended
pixi install fails on macOS: GPU features (rapids, chemberta, deepchem) are linux-only; the default env works on both platforms- Pipeline segfaults (exit code 139) on first run: cheminformatics and rapids env tasks can crash during first-time environment initialization under redun's multiprocessing executor. Re-run the pipeline - redun caches completed tasks and picks up where it left off.
- uv panics with "Too many open files": the default
ulimit -n 1024 is too low for uv's parallel bytecode compilation. Fix: ulimit -n 65536 before running uvx.
- Port already in use when launching marimo: another user (or a zombie process from a previous crash) may be holding the port. Check with
sudo lsof -i :<port> and pick a different port or fuser -k <port>/tcp to kill your own stale process.
