Cellbender

This release coincides with the publication of our manuscript in _Nature Methods_. https://doi.org/10.1038/s41592-023-01943-7

Changes from previous versions include:

- Output report in HTML format
- Code produces checkpoints, useful for automatic restarting of WDL workflow without losing progress (enabling use of preemptible GPUs on Terra, and potentially better use of Google Colab GPUs for free)
- Model includes new safeguards on count removal per gene, and treats the construction of the output count matrix as an auxiliary optimization problem
- `--expected-cells` and `--total-droplets-included` input arguments are not typically needed, as default behavior is very much improved
- The full posterior gets saved as an output h5 file
- Smoother learning curves and improved performance on tricky samples
- Bug fixes

Minor bump with a fix for automatic retrying of training with a lower learning rate upon failure.

Minor bump with a fix for setup.py to install sub packages, and new support for anndata inputs.

CellBender v0.2.0 implements a modified model for `remove-background`, the command-line tool used to remove ambient RNA / background noise from scRNA-seq datasets according to a principled, generative model of scRNA-seq count data. The changes to the `remove-background` tool will be documented in an upcoming paper.

Please note that implementation details may be subject to change in future updates.

https://www.biorxiv.org/content/10.1101/791699v1

Please note that the implementation may be subject to change in future updates.