--- pretty_name: GPSBench license: other license_name: dataset-source-terms license_link: LICENSE language: [en] tags: [biology, spatial-transcriptomics, rna-localization, point-cloud] configs: - config_name: NIH3T3 data_files: - split: train path: data/NIH3T3_train/NIH3T3_train_cells.parquet - split: validation path: data/NIH3T3_val/NIH3T3_val_cells.parquet - split: test path: data/NIH3T3_test/NIH3T3_test_cells.parquet - config_name: U2OS data_files: - split: train path: data/U2OS_train/U2OS_train_cells.parquet - split: validation path: data/U2OS_val/U2OS_val_cells.parquet - split: test path: data/U2OS_test/U2OS_test_cells.parquet - config_name: Lung data_files: - split: train path: data/Lung_train/Lung_train_cells.parquet - split: validation path: data/Lung_val/Lung_val_cells.parquet - split: test path: data/Lung_test/Lung_test_cells.parquet ---

GPSBench: Subcellular Gene Spatial Pattern Classification

GPSBench evaluates gene-level RNA localization from molecular point clouds across seqFISH+, MERFISH, and CosMx imaging data.

Project page · Code

## Dataset overview This snapshot contains **3 datasets, 101,347 cells, and 43,663,730 retained RNA molecules**. | Dataset | Cells | RNA molecules | Genes | | --- | ---: | ---: | ---: | | NIH3T3 | 179 | 2,724,808 | 3,721 | | U2OS | 1,022 | 10,634,467 | 130 | | Lung | 100,146 | 30,304,455 | 960 | Each task has a separate gene vocabulary and label set. Train, validation, and test fields of view are disjoint and shared across methods and seeds. The Hugging Face Dataset Viewer shows one row per cell. [File inventory](https://huggingface.co/datasets/king159/GPSBench/blob/main/manifest.json) · [Download and preparation](https://github.com/king159/GPSNet#-data-preparation) ## Attribution and terms NIH/3T3 and U-2 OS use Bento reference labels and retain molecules from vocabulary-matched cell–gene pairs with finite references and at least one positive Bento pattern. The Bento distributions are available under the MIT license. [Eng et al.](https://doi.org/10.1038/s41586-019-1049-y) · [Mah et al.](https://doi.org/10.1186/s13059-024-03217-7) · [Bento data and license](https://figshare.com/articles/dataset/Bento_spatial_AnnData_formatted_datasets/15109236) **Modified CosMx data.** We modified the NanoString CosMx SMI Lung5 Rep1 data for GPSBench. The Lung split files were organized by field of view. The model input conversion excluded unassigned transcripts and out-of-vocabulary probes, indexed the retained vocabulary, normalized the two spatial coordinates per cell and computed neighbors. The benchmark aggregates the original transcript compartment flags into cell–gene labels and retains the split files’ boundary metadata and original coordinate frames. Use of the Lung subset is subject to the original NanoString data agreement for non-commercial use. [NanoString](https://www.nanostring.com/) · [Lung5 Rep1 source data](https://brukerspatialbiology.com/resources/smi-ffpe-dataset-lung5-rep1-data/) · [He et al.](https://doi.org/10.1038/s41587-022-01483-z) · [Original NanoString data agreement](https://zenodo.org/records/15487520/files/NanoString_Data_License_Agreement.pdf) · [Agreement deposit record](https://zenodo.org/records/15487520) The benchmark inputs derive from the fixed [st-data-distil snapshot](https://huggingface.co/datasets/IcecreamArtist/st-data-distil/tree/999c83feb172fbfcb12df1cb3fcb13f30f01a856).