Status: DEFERRED — Resume after Cytoskeleton Structural Refactor is complete.
Depends on: Cytoskeleton Refactor (assets/, VFS, manifests, adapter system)
Create a reproducible, end-to-end tutorial using cytocast to validate Strix Halo hardware, query the massive semantic KG, and process Brain Cell Atlas single-cell data.
assets/ folder structure finalized and DVC-tracked in cytoskeletoncytos dataset ingestion tooling updated to use new asset/adapter patternscytoskeleton installable as dependency across all reposcytocast Package Creation & Hardware Validationcytocast with data and ml components enabled (which includes scvi-tools, jax, pytorch).
2. Ensure the package correctly identifies strix-halo architecture.
3. Validate locked environments are detected and installed correctly.
4. Run python -c "import scvi" to verify scvi loads.
5. Verify PyTorch and JAX both see the GPU.
6. Run small matrix-vector and matrix-matrix benchmarks to confirm GPU acceleration.
scripts/check_rocm.py logic into src/cytoskeleton/hardware/ module.nox -s hardware_check.cytohost.
cytos is installable via pip install cytos from internal PyPI.
2. Connect to the central Neo4j database on cytohost.
3. Extract the induced subgraph and 1-hop neighborhood of 14 psychiatric disorders:
``python
disorders = [
"MONDO:0007661", "MONDO:0005351", "MONDO:0008114",
"MONDO:0002009", "MONDO:0005618", "MONDO:0005146",
"MONDO:0005258", "MONDO:0007743", "MONDO:0007079",
"MONDO:0005530", "MONDO:0008575", "MONDO:0005015",
"MONDO:0005180", "MONDO:0004975"
]
`
4. Track the extraction via DVC.
Workstream 5C: Brain Cell Atlas Single-Cell Ingestion
Data Source
Brain Cell Atlas — specifically the ADULT BRAIN dataset.
Steps
1. [NEW] cytos/data/brain_cell_atlas.py — Add a function to download datasets from the Brain Cell Atlas, parameterized by dataset type (ADULT_BRAIN, FETAL_BRAIN, TUMOUR, ORGANOID).
2. Track raw downloads via DVC for digestion.
3. Fully harmonize the dataset using cytos tooling:
- Map to HRA/CCF coordinates
- Normalize cell type annotations
- Apply standard QC filters
4. Store the harmonized dataset as a TileDB-SOMA database on the GCP storage bucket.
Workstream 5D: Tutorial Notebooks
Tutorial 1: Hardware & Package Setup
[NEW] cytos/docs/tutorials/01_hardware_and_package_setup.ipynb
Demonstrates: package creation via cytocast, pip install from internal PyPI, environment locking, hardware validation.
Tutorial 2: KG & Single-Cell Workflows
[NEW] cytos/docs/tutorials/02_kg_and_single_cell_workflows.ipynb
Demonstrates: extracting the 14-disorder subgraph from cytohost Neo4j, ADULT BRAIN` TileDB-SOMA ingestion/harmonization.