Spatial Omics Data Analysis: From Raw Data to AI Insights

Physalia course, 14–16 September 2026, 14:00–18:00 Berlin time, online. Instructors: Tim Treis, Robert Gutgesell (Helmholtz Munich).

---

Participants: start here

You need VS Code and the Remote-SSH extension. Nothing else — no Python, no conda, no pixi on your own machine.

1. Connect. VS Code → Remote-SSH: Connect to Host → the host Physalia sent you.
2. Clone this repo into your home directory:

   git clone https://github.com/scverse/2026_09_workshop_physalia.git
   cd 202609_workshop_physalia

3. Open notebooks/00_setup_check.ipynb.
4. Pick the kernel. Top right, choose "Physalia spatial omics". It is already installed — you do not need to create an environment.
5. Run All. The last cell should print SETUP OK.

If anything fails, copy the whole status block into the course Slack. Please do this before day 1, not on the morning of — that way we can fix it in time.

Where things live

Shared data	/opt/workshop/data — read-only, same for everyone
Your work	inside ~/2026_09_workshop_physalia; outputs/ is git-ignored
Privacy	your home directory is private to you — other participants cannot read it

Import paths.py rather than hardcoding either:

import paths
sdata = spatialdata.read_zarr(paths.data("xenium_breast.zarr"))
fig.savefig(paths.OUT / "my_figure.png")

Running it on your own machine later

Everything here is reproducible off the course server. Install pixi, then:

pixi install                 # uses the committed lock, so you get our exact versions
export PHYSALIA_DATA=/where/you/put/the/data
pixi run check

pixi.lock covers linux-64, osx-arm64 and win-64.

---

Instructors

sudo -v && bash scripts/bootstrap.sh     # full server setup, ~20 min, re-runnable
bash scripts/stage_data.sh               # data only
bash scripts/harden_accounts.sh          # private homes + pre-cloned repo, re-runnable
bash scripts/seed_vscode_extensions.sh   # Python + Jupyter into every VS Code server
bash scripts/reset_user.sh user1         # reset one account to its day-1 state
pixi run check                           # verify

seed_vscode_extensions.sh pre-installs the Python and Jupyter extensions into every participant's ~/.vscode-server. Remote-SSH runs extensions on the server, so without this every participant is prompted to install ~300 MB the first time they open a notebook — 25 times over, at the start of day 1.

reset_user.sh wipes a participant's home except .ssh and .vscode-server and restores it to exactly what they see on day 1 — use it to test the real participant experience rather than guessing at it. It takes --dry-run and --all.

bootstrap.sh installs pixi, builds the environment from the committed lock, installs the kernelspec system-wide (--prefix=/usr/local, so it is visible to every account including ones that do not exist yet), and stages the data. Nothing is written into participant home directories, so the setup survives the participant accounts being recreated.

Environment notes

• scvi-tools, squidpy and spatialdata-io come from PyPI, not conda-forge: each feedstock lags PyPI by one patch and this course teaches the current stack.
• cellpose is pinned <4. Unpinned the solver picks 4.x on Linux and 3.0.9 on macOS, so a Mac-authored notebook breaks on the server. Worse, 4.x (Cellpose-SAM) removed the models.Cellpose class and dropped cyto3/nuclei from MODEL_NAMES — passing cyto3 to v4 silently falls back to cpsam_v2 rather than erroring. It is also ~3× slower on CPU at fp32 and ~76× slower at its shipped use_bfloat16=True default, since bf16 is emulated on CPU.
• napari lives in the local feature only (pixi install -e local); it cannot render over Remote-SSH.

Schedule

Day	Sessions
1	Introduction to SpatialData · Quality control and segmentation (SOPA) · SpatialData + Squidpy
2	Spatial structure: clustering and niches · Cell–cell communication · Gene-level spatial analysis
3	Why AI in spatial omics · AI in practice: morphology and resolVI · Synthesis