scANVI Terra Workspace

Workspace	Maintained by	Questions or Feedback
Multiome SCVI and SCANVI	WARP Pipelines	Please file an issue in WARP

Overview

The Multiome SCVI and SCANVI Terra workspace is a ready-to-run home for the scANVI pipeline. It comes preloaded with the scANVI WDL workflow configuration and example Multiome 10k PBMC data, so you can run production-scale cell type label transfer with no additional setup. The workspace also includes an interactive notebook that walks through the same label-transfer logic step by step, using scVI and scANVI (single-cell ANnotation using Variational Inference).

The workspace offers two entry points:

• The scANVI WDL workflow (recommended for production and at-scale runs) — automated, reproducible label transfer via Cromwell/Terra. Start here.
• The "Transfer Labels Using SCANVI Model" notebook (for interactive exploration) — the same logic run cell by cell; see The interactive notebook below.

Running the scANVI WDL workflow

The workspace ships with a preloaded WDL workflow configuration for running scANVI at scale. The workflow is fully automated and suitable for production use, and mirrors the notebook logic.

1. Open the Multiome SCVI and SCANVI workspace on Terra.
2. Navigate to the Workflows tab and select the preloaded scANVI workflow.
3. Set the inputs. At minimum provide input_id plus either the direct file inputs (gex_h5ad and ref_h5ad, and optionally atac_h5ad) or an input_bucket. To run inference only from a previously trained model, supply scanvi_model and set gpu_count = 0 for a cheaper CPU-only run. See the Inputs section of the pipeline overview for the full input list and reference requirements.
4. Launch the analysis. Training runs request GPUs by default (gpu_count = 2); CPU-only prediction runs set gpu_count = 0.

For full documentation — all inputs, outputs, task descriptions, and runtime configuration — see the scANVI pipeline overview.

Example run costs

Representative Terra costs for scANVI runs across three references (mouse hippocampus, human neocortex, and PBMC) in both training and CPU-inference modes. Costs are from single runs on GCP us-central1 (training on 2× NVIDIA T4 GPUs; inference CPU-only via gpu_count = 0) and will vary with region, machine type, preemption, and dataset. "Cells labelled" is the number of query cells annotated (after PreprocessFilter; for multiome, after the GEX↔ATAC shared-barcode intersection); "Reference" is the annotated ref_h5ad.

Dataset (reference)	Mode	Compute	Cost	Cells labelled	Cost / cell	Cost / 1k cells	Query input	Cost / GB (query)	Reference (cells / GB)
Mouse hippocampus (AIT)	train	2× T4 GPU	$26.34	24,078	$0.001094	$1.09	0.53 GB	$49.85	250,734 / 12.87
Mouse hippocampus (AIT)	inference	CPU	$0.56	24,078	$0.0000233	$0.02	1.69 GB	$0.33	250,734 / 12.87
Human neocortex (AIT)	train	2× T4 GPU	$2.51	6,018	$0.000417	$0.42	0.32 GB	$7.98	47,432 / 4.97
Human neocortex (AIT)	inference	CPU	$0.33	6,018	$0.0000548	$0.05	0.32 GB	$1.05	47,432 / 4.97
10k PBMC (multiome)	train	2× T4 GPU	$3.64	911	$0.003996	$4.00	0.04 GB GEX + 0.96 GB ATAC	$100.14 †	33,506 / 2.06
10k PBMC (GEX-only)	train	2× T4 GPU	$3.38	934	$0.003619	$3.62	0.04 GB	$92.98	33,506 / 2.06

† GEX file only; $3.66/GB when counting GEX + ATAC input together.

Takeaways:

• Inference is far cheaper than training. Loading a saved model and predicting on CPU (scanvi_model supplied, gpu_count = 0) cost 8–47× less than the equivalent training run.
• Training cost tracks the reference, not the query. The largest query (1.69 GB, mouse inference) was among the cheapest runs; training cost scales with reference size (mouse 250,734-cell / 12.87 GB → $26.34 vs human 47,432 / 4.97 GB → $2.51).
• Per-cell cost rises as the query shrinks, because the fixed reference-load / training cost is amortized over fewer labelled cells.
• Multiome and GEX-only cost about the same ($3.64 vs $3.38) — the ATAC branch adds little.

The interactive notebook

The workspace also provides the "Transfer Labels Using SCANVI Model" notebook for interactive, step-by-step exploration of the label-transfer process. It focuses specifically on annotating an ATAC query using a gene expression matrix that already carries cell type labels, using human Multiome datasets (10k PBMC) processed through the WARP PeakCalling pipeline alongside the PBMC reference from the scArches tutorial.

The notebook walks through:

1. Downloading input files — retrieves the GEX, ATAC, and reference h5ad files needed for the analysis.
2. Environment setup — installs and configures the required Python packages, including SnapATAC2 and scvi-tools.
3. Data filtering and processing — filters cells, aligns barcodes between the GEX and ATAC modalities, and converts the ATAC cell-by-bin matrix to a gene activity matrix.
4. Label transfer — trains SCVI and SCANVI models and transfers cell type labels from the annotated GEX reference onto the unannotated ATAC query, following the SnapATAC2 label transfer tutorial.

Running the notebook

The workspace is preloaded with the Multiome 10k PBMC data, so no additional setup is required before launching the notebook.

1. Open the Multiome SCVI and SCANVI workspace on Terra.
2. Navigate to the Analysis tab.
3. Select the "Transfer Labels Using SCANVI Model" notebook.
4. Choose a Virtual Machine and Enable GPUs — the SCANVI model training step requires a GPU.
5. Run the notebook cells in sequence to perform the analysis.

GPU requirement

Model training in this notebook requires a GPU-enabled cloud environment. When prompted to select a runtime, make sure Enable GPUs is checked before starting the VM.

Relationship to the WDL pipeline

The notebook walks through the same label-transfer logic implemented in the scANVI WDL pipeline, but interactively — making it a useful companion for understanding the pipeline or for exploratory analysis before running production-scale jobs. See the scANVI pipeline documentation for running at scale via Cromwell or Terra workflows.