scib-rapids

GPU-accelerated metrics for benchmarking single-cell integration outputs using RAPIDS (cuML, CuPy).

This package provides the same metrics as scib-metrics but replaces JAX with RAPIDS (CuPy, cuML) for GPU acceleration. All implementations leverage CuPy for device-level computation on NVIDIA GPUs.

Metrics

• Silhouette: silhouette_label, silhouette_batch, bras
• LISI: lisi_knn, ilisi_knn, clisi_knn
• kBET: kbet, kbet_per_label
• Clustering: nmi_ari_cluster_labels_kmeans, nmi_ari_cluster_labels_leiden
• Graph connectivity: graph_connectivity
• Isolated labels: isolated_labels
• PCR comparison: pcr_comparison

Getting started

Please refer to the documentation.

Installation

You need to have Python 3.11 or newer and a CUDA-capable GPU.

We recommend installing rapids-singlecell and scib-rapids using uv pip rather than conda, as conda often causes dependency conflicts (e.g. cupy vs cupy-cuda12x).

uv pip install rapids-singlecell scib-rapids

Alternatively, install the latest development version:

uv pip install rapids-singlecell git+https://github.com/maarten-devries/scib-rapids.git@main

Release notes

See the changelog.

Contact

For questions and help requests, you can reach out in the scverse Discourse. If you found a bug, please use the issue tracker.

Citation

If you use scib-rapids, please cite the original single-cell integration benchmarking work:

@article{luecken2022benchmarking,
  title={Benchmarking atlas-level data integration in single-cell genomics},
  author={Luecken, Malte D and B{\"u}ttner, Maren and Chaichoompu, Kridsadakorn and Danese, Anna and Interlandi, Marta and M{\"u}ller, Michaela F and Strobl, Daniel C and Zappia, Luke and Dugas, Martin and Colom{\'e}-Tatch{\'e}, Maria and others},
  journal={Nature methods},
  volume={19},
  number={1},
  pages={41--50},
  year={2022},
  publisher={Nature Publishing Group}
}