| name | scvi-tools-analysis |
| description | Build and evaluate reproducible scvi-tools workflows for single-cell RNA, protein, chromatin, spatial, or multimodal data using scVI, scANVI, totalVI, MultiVI, PeakVI, DestVI, or related models. |
| license | MIT |
scvi-tools Analysis
Gate
Ask once before installing a pinned scvi-tools stack, downloading hub models or
datasets, and using GPU/long compute. Never upload governed human data without
authorization. Follow $cx-compute-environment.
Workflow
- Define assay, biological units, count layers, features, batches/covariates,
labels, controls, missing modalities, and task. Split by donor/sample/well,
never randomly by cells from the same biological unit. Audit whether treatment,
phenotype, cell type, donor, site, and acquisition batch are confounded; stop
when the requested biological effect is not identifiable from the design.
- Select the model for the assay and estimand; pin scvi-tools and dependencies,
AnnData registry fields, likelihood, covariates, architecture, train/validation
split, early stopping, seeds, hardware, and precision.
- Smoke-test data registration and a few training steps. Train into
artifacts/<run-id>/scvi/; retain immutable input/split IDs, setup schema,
model/config/checkpoint, histories, posterior outputs, environment, logs, and
failures.
- Run posterior predictive/model criticism checks. Compare PCA/Harmony or other
simple task-appropriate baselines; report donor-level uncertainty, batch and
biology conservation, subgroup/OOD performance, and sensitivity to covariates.
- Record with
$science-provenance; review with $science-review.
Boundaries
- A visually mixed latent space is not proof of correct integration.
- Never encode treatment or the biological signal as a nuisance covariate merely
to improve batch mixing; require held-out-donor batch
predictability and biology-conservation criteria.
- Do not use test cells to choose highly variable genes, normalization,
covariates, architecture, epochs, or calibration unless declared transductive.
