v1.0.1

[1.0.1] - 2025-12-10

🎉 Novae is now published in Nature Methods!

Added

• Full dataset added to our Hugging Face Hub
• Official Docker images released - see here
• Support reference=None in fine_tune (random initialization) and use it as default
• Added fast argument in novae.plot.domains to have a quick (but less accurate) rendering of domains.
• Added novae.settings.scale_to_microns if the coordinates are not in microns
• Use fast-array-utils to support multiple backends in adata.X (e.g., dask or backed mode)

Breaking changes

• Remove support for python==3.10

Changed

• Update hf-hub to use xet for faster dataset download from Hugging Face Hub
• Update to spatialdata>=0.5.0 to avoid installation issue related to xarray-dataclasses
• Threshold to decide when to enable lazy loading: based on n_obs * n_vars (i.e., the array size) instead of n_obs

Fixed

• Retrieve mode=multimodal when saving the model and re-loading the model after H&E training (#24)
• Fix tests for recent AnnData versions only supporting CSR/CSC sparse matrices

v1.0.0

[1.0.0] - 2025-08-09

Pre-publication release (for Zenodo DOI creation).

Added

• Support multimodality (H&E + spatial omics). See this tutorial.
• Support multi-references in fine_tune and zero-shot modes
• Added novae.plot.loss_curve for minimal monitoring when not using Weight & Biases

Changed

• fine_tune method: use lr = 5e-4 and max_epochs = 20 as new default values
• Use reference="all" instead of "largest" by default
• Move novae.utils.load_dataset to novae.load_dataset (the old import is deprecated, it will be removed in future versions)
• Move novae.utils.quantile_scaling to novae.quantile_scaling (deprecated, as above)
• Move novae.utils.toy_dataset to novae.toy_dataset (deprecated, as above)
• Migrate to uv as a package manager, and ruff for formatting/linting

Fixed

• Slide-id passed correctly to dataloader for one-adata multi-slide mode
• Auto-detect change in n_hops_{local,view} to re-build graph
• Move representations to numpy when torch is not needed anymore (#19)

v0.2.3

[0.2.3] - 2025-03-21

Added

• New Visium-HD and Visium tutorials
• Infer default plot size for plots (median neighbors distrance)
• Can filter by technology in load_dataset

Fixed

• Fix model.plot_prototype_covariance and model.plot_prototype_weights
• Edge case: ensure that even negative max-weights are above the prototype threshold

Changed

• novae.utils.spatial_neighbors can be now called via novae.spatial_neighbors
• Store distances after spatial_neighbors instead of 1 when coord_type="grid"

v0.2.2

[0.2.2] - 2024-12-17

Added

• load_dataset: add custom_filter and dry_run arguments
• added min_prototypes_ratio argument in fine_tune to run init_slide_queue
• Added tutorials for proteins data + minor docs improvements

Fixed

• Ensure reset clustering if multiple zero-shot (#9)

Changed

• Removed the docs formatting (better for autocompletion)
• Reorder parameters in Novae __init__ (sorted by importance)

v0.2.0

[0.2.0] - 2024-12-03

Added

• novae.plot.connectivities(...) to show the cells neighbors
• novae.settings.auto_processing = False to enforce using your own preprocessing
• Tutorials update (more plots and more details)

Fixed

• Issue with library_id in novae.plot.domains (#8)
• Set pandas>=2.0.0 in the dependencies (#5)

Breaking changes

• novae.utils.spatial_neighbors must always be run, to force the user having more control on it
• For multi-slide mode, the slide_key argument should now be used in novae.utils.spatial_neighbors (and only there)
• Drop python 3.9 support (because dropped in anndata)

v0.1.0

[0.1.0] - 2024-09-11

First official novae release. Preprint coming soon.