Release: scvi-tools 1.1.0

We’ve released scvi-tools 1.1.0 today - see the release notes for all changes. A highlight of some of the new features:

  • An implementation of contrastiveVI for isolating variations in perturbed cells (e.g. via CRISPR, drug exposure) vs. control cells. See the tutorial here.

  • PoissonVI for modeling quantitative scATAC-seq fragment counts. See the tutorial here on converting 10X matrices to fragment count matrices.

  • A critical bug fix for scANVI. See the corresponding tutorial/benchmark here.

  • Various experimental features such as multi-GPU training, support for custom dataloaders during training, and more efficient transfer of sparse arrays to discrete accelerators such as NVIDIA GPUs.

  • Expanded support for pre-trained models hosted on Hugging Face, including a beta API for uploading and downloading pre-trained models from AWS.

  • Beta Docker images for running and developing in Linux with a CUDA backend.

Feel free to submit any bug reports or feature requests on our repository!

1 Like