scvi-model-inputs

scvi-tools model inputs

For every task, apply Caveman. Before broad repository navigation, use CocoIndex Code; for web-page research, use Crawl4AI. Read and follow the repository-root docs/agent-tools.md and its linked skills, and retain the required execution evidence.

The pinned skill says that all scvi-tools models require raw counts. Input requirements are model-specific: SCVI uses count data, while SysVI is designed for approximately normal data. Confirm the model's current documentation and the data's provenance before registration. This focused overlay does not review the other models or validate biological conclusions.

SCVI: registered counts

Choose a count layer that contains the intended measured counts and, if needed, a technical batch_key that exists in adata.obs. None uses no batch key. Do not automatically register donor, condition or other biological variables as nuisance covariates. Complete filtering and feature selection before setup. Supply a chosen integer seed, latent dimension n_latent and dictionary of training_options appropriate to the data and available hardware:

import scvi

if "counts" not in adata.layers:
    raise ValueError("Provide the intended count layer before SCVI setup")
scvi.settings.seed = seed
scvi.model.SCVI.setup_anndata(adata, layer="counts", batch_key=batch_key)
model = scvi.model.SCVI(adata, n_latent=n_latent)
model.train(**training_options)
latent = model.get_latent_representation()

Count-like values alone do not establish provenance. Setup registers the given fields; it does not create a missing count layer or normalize data. If the training object or its feature ordering changes, use the model's documented registration and query-mapping procedures instead of assuming the old registry is still valid.

SysVI: approximately normal input

SysVI's intended input is approximately normally distributed; its documentation uses normalized, log1p-transformed RNA data as an example. Prepare the data appropriate to the experiment, preserve its provenance, and select an existing technical batch_key. The caller supplies seed, n_latent, prior and training_options:

import scvi

scvi.settings.seed = seed
scvi.external.SysVI.setup_anndata(adata, batch_key=batch_key)
model = scvi.external.SysVI(adata, n_latent=n_latent, prior=prior)
model.train(**training_options)
latent = model.get_latent_representation()

The examples are API starting points, not prescribed hyperparameters. Check training and validation behavior, sensitivity to seeds and model choices, and retention of relevant biology. A short synthetic fit does not establish model convergence or biological validity. Normalized expression can be dense and large; request only what the task needs.

See the SCVI API, SysVI model guide and scvi-tools overview. The preserved source entrypoint is in provenance/; the control is not an agent benchmark.