Skip to content

scANVI_develop

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 09 Nov 14:49
· 10 commits to develop since this release
scANVI_develop
43d386b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

3.0.0

2026-11-05 (Date of Last Commit)

Breaking changes

  • Removed the placeholder sample-metadata obs columns (biosample_id, donor_id, species, disease, organ, library_preparation_protocol, their ontology labels, data_modality and sex); sample metadata is joined downstream
  • Changed the sample inputs to arrays paired by position: input_ids, gex_h5ads and atac_h5ads replace input_id, gex_h5ad and atac_h5ad
  • Changed the outputs to arrays with one entry per sample: gex_annotated_h5ads, atac_annotated_h5ads and scanvi_models_out replace gex_annotated_h5ad, atac_annotated_h5ad and scanvi_model_out
  • Removed the scanvi_predictions_h5ad output; its labels and latents are in the annotated matrices
  • Named outputs after their input files (sample.h5ad gives sample_annotated.h5ad and, when training, sample_scanvi_model.tar.gz); input file names must be distinct
  • Added label levels: ref_label_columns (a list) replaces ref_label_column, and a level such as subclass puts its labels in obs['subclass'] instead of obs['final_annotation'] and its confidence in obs['subclass_max_probability']
  • Changed the model archive to hold one scVI model plus one SCANVI model per level; archives in the old single-model form are refused
  • Removed bucket mode (input_bucket, gex_filename, atac_filename and ref_filename)
  • Changed the annotated matrices to the Single Cell Portal form: counts in .raw, log2(CPM + 1) in .X, and per level a 30-dimensional scVI latent (obsm['X_scVI_subclass'] for the subclass level) and a 2-column UMAP in obsm, with X_umap set to the first level's; read counts from .raw
  • Removed the PreprocessFilter task and its nested inputs; preprocessing runs inside the label-transfer task
  • Changed the label-transfer task to run every sample on one machine, by default 16 CPUs and 104 GiB instead of 32 CPUs and 120 GiB

Additional changes

  • Updated the scvi-scanvi image to 1.7.1 (scvi-tools 1.5.1, snapatac2 2.10.0, scanpy 1.12.5, anndata 0.13.4, scipy 1.18.1, numpy 2.5.4, torch 2.15.0, all pinned), which may shift labels and embeddings slightly
  • Added the optional seed input for reproducible training
  • Added scArches query mapping for supplied models that do not know a sample's batch, with the optional query_max_epochs input (default 100)
  • Added gene matching by Ensembl ID: a supplied model whose genes are Ensembl IDs is matched through the sample's ensembl_ids column, and the log reports how many of the model's genes the sample has
  • Made ref_h5ad optional; it is needed only to train
  • Added the PreprocessReference task, which prepares the reference once per training run
  • Added run provenance to each output's uns: scanvi_input_id, scanvi_input_file, scanvi_model, scanvi_levels and scanvi_versions
  • Added input checks that fail the run early when the input arrays do not line up, ids repeat, file names collide or a file is not an h5ad
  • Changed output_max_probability to default to true
  • Changed disk_size to optional; when unset, the disk is sized from the inputs
  • Changed model-archive extraction to Python's safe data filter
  • Added gzip compression to every output h5ad
Assets 5
Loading