{"success":true,"data":{"author":"Theis Lab and Gottgens Lab","bucketAccessType":"public","bucketName":"theislab/LabCompass","bucketPrefix":"","category":"Life Sciences","cloudProvider":"huggingface","createdAt":"2026-11-17T14:12:21.482Z","description":"Measurements behind LabCompass, a generative method for experimental design on single-cell data from the Theis lab, with wet-lab experiments and measurements by the Gottgens lab. Spectral flow cytometry (SFC) profiles of in vitro hematopoietic differentiation cultures were collected over successive rounds of a closed design loop: each loop proposes new culture protocols (cytokines such as SCF, TPO, IL-3, GM-CSF, FLT3L, IL-7, M-CSF, small molecules such as SR1, UM171 and retinoic acid, oxygen and hydrogel conditions), runs them at the bench and measures the resulting cells. loops/ holds one AnnData .h5ad per loop (0, 1, 2, 2.5, 3, 4, 4.5, 5), each in a full and a subsampled _500k variant; X is logicle-transformed fluorescence and scatter intensities, and obs carries acquisition metadata, protocol axes with units and cell-type labels. Files are per loop, not cumulative; the theislab/LabCompass GitHub repository has a script that builds the accumulated training sets. checkpoints/ has the flow-matching forward models, cell-type classifiers and protocol priors per loop, with derived/ and solutions/ (designed protocols). About 80 GB in total. Released under CC BY 4.0; cite the LabCompass paper and repository.","displayTitle":"LabCompass: spectral flow cytometry of in vitro hematopoietic differentiation across closed-loop experimental design rounds","huggingFaceUrl":"https://huggingface.co/datasets/theislab/LabCompass","id":"fa35a18c-d966-582a-a8f3-0fccd1dd338c","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/theislab/LabCompass","objectCountEstimate":59,"probeStatus":"ok","referenceUrl":"https://github.com/theislab/LabCompass","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"theislab/LabCompass\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"theislab/LabCompass\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=theislab, license:cc-by-4.0, size_categories:10M<n<100M, downloads=665"],"format":"mixed","generatedAt":"2026-11-17T14:23:00.742Z","layout":{"example":"checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl","truncatedAt":512}],"sampleObjects":["checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl","checkpoints/loop0/eager-brook-26_FlowMap.pkl","checkpoints/loop0/feasible-lake-25_FlowMatchingWithScore.pkl","checkpoints/loop0/likely-donkey-20_FlowMatching.pkl","checkpoints/loop1/copper-thunder-24_FlowMap.pkl"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\ntags:\n  - single-cell\n  - flow-cytometry\n  - spectral-flow-cytometry\n  - haematopoiesis\n  - experimental-design\nsize_categories:\n  - 10M<n<100M\n---\n\n# LabCompass — Spectral Flow Cytometry haematopoiesis dataset\n\nMeasurements underlying **LabCompass**, a method for generative modeling of experimental design in\nsingle-cell data. This dataset contains Spectral Flow Cytometry (SFC) profiles of *in vitro*\nhaematopoietic differentiation cultures, collected over successive rounds of a closed-loop\nexperimental design cycle.\n\nEach round — a **loop** — proposes new culture protocols, runs them at the bench, and measures the\nresulting cells. The measurements from each loop are published here as a separate file.\n\n- **Code and full reproduction pipeline:** <https://github.com/theislab/LabCompass>\n- **Contents:** per-loop measurements in `loops/`, trained models in `checkpoints/`, derived datasets in `derived/`, designed protocols in `solutions/`\n- **Wet-lab experiments and measurements:** Göttgens Lab\n- **License:** CC-BY-4.0\n\n## ⚠️ These files are per-loop, not cumulative\n\n`loops/loop3.h5ad` contains **only the cells measured in loop 3** — not loops 0–3 together. Models in\nthe paper are trained on the *accumulated* data, so a loop's training set is the concatenation of\nevery loop up to and including it:\n\n```\ndataset(N) = concat(dataset(N-1), loopN)\n```\n\nConcatenating them yourself is a few lines of `anndata`, but the exact chain matters (one loop\nintroduces new protocol axes that must be zero-filled on the earlier data — see below). The\nreproduction repository ships a script that does it correctly:\n\n```bash\ngit clone https://github.com/theislab/LabCompass.git\npython scripts/data/build_loop_datasets.py              # downloads from this repo and builds the chain\npython scripts/data/build_loop_datasets.py --variants 500k   # subsampled only: far smaller and faster\n```\n\n## Files\n\nEvery loop is published in two variants: the full measurement set, and a subsampled version\n(`_500k` suffix) intended for fast iteration. The suffix is a naming convention carried over from\nthe source data, not a guaranteed cell count — the subsampled files vary in size.\n\n| Loop | Full | Subsampled | Approx. size (full) |\n| --- | --- | --- | --- |\n| 0 (baseline) | `loops/loop0.h5ad` | `loops/loop0_500k.h5ad` | 36 GB |\n| 1 | `loops/loop1.h5ad` | `loops/loop1_500k.h5ad` | 2.5 GB |\n| 2 | `loops/loop2.h5ad` | `loops/loop2_500k.h5ad` | 3.9 GB |\n| 2.5 | `loops/loop2p5.h5ad` | `loops/loop2p5_500k.h5ad` | 2.7 GB |\n| 3 | `loops/loop3.h5ad` | `loops/loop3_500k.h5ad` | 6.3 GB |\n| 4 | `loops/loop4.h5ad` | `loops/loop4_500k.h5ad` | 0.9 GB |\n| 4.5 | `loops/loop4p5.h5ad` | `loops/loop4p5_500k.h5ad` | 0.5 GB |\n| 5 | `loops/loop5.h5ad` | `loops/loop5_500k.h5ad` | 6.3 GB |\n\nLoop 0 is the baseline screen and is by far the largest. The half-steps (2.5, 4.5) are follow-up\nrounds within a design cycle and accumulate like any other loop, giving the chain\n\n```\nloop0 → loop1 → loop2 → loop2p5 → loop3 → loop4 → loop4p5 → loop5\n```\n\nThe full set is roughly 60 GB; the subsampled set is a few GB.\n\n## Format\n\nEach file is an [AnnData](https://anndata.readthedocs.io/) `.h5ad` object:\n\n- **`X`** — logicle-transformed SFC intensities: fluorescence channels and morphological scatter\n  features, one row per cell.\n- **`obs`** — per-cell metadata, in three groups:\n  - *Acquisition:* `experiment_number`, `experiment_id`, `replicate`, `date`, `well_id`,\n    `cytometer`, `cytometer_serial_no`, `count_beads`, `cell_counts`, `source_id`.\n  - *Protocol axes* — the culture recipe, and the space LabCompass searches over. Cytokines and small\n    molecules carry their units in the column name, e.g. `scf_[ng_ml]`, `tpo_[ng_ml]`,\n    `il3_[ng_ml]`, `gm-csf_[ng_ml]`, `rhflt3l_[ng_ml]`, `ldl_[ng_ml]`, `sr1_[nm]`, `um171_[nm]`,\n    `um729_[µm]`, `butyzamide_[nm]`, `retinoic_acid_[µm]`, `mtg_[µm]`, `740-yp_[µm]`, alongside\n    culture conditions such as `o2_[%]` and `hydrogel_type`.\n  - *Annotation:* cell-type labels, where available.\n\n`experiment","encoding":"utf8","key":"README.md","name":"README.md","size":13922,"truncated":true},{"bytes":"{\n  \"loop0/likely-donkey-20_FlowMatching.pkl\": {\n    \"paths_config\": \"loop0\",\n    \"config_key\": \"perturbation_prediction_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/conditional_model/flow_matching/2026-04-26_19-03-48/likely-donkey-20_FlowMatching.pkl\",\n    \"source_bytes\": 113435296170,\n    \"dest_bytes\": 1013430303,\n    \"ratio\": 111.9,\n    \"dest_sha256\": \"b10c0739fe13151414356ce5b708379b5c4d3ad760f2e1c7f0e924f4e9c392b1\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop0/dandy-vortex-20_TargetPredictionModel.pkl\": {\n    \"paths_config\": \"loop0\",\n    \"config_key\": \"ct_classifier_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/cell_type_classifier/2026-04-28_16-38-47/dandy-vortex-20_TargetPredictionModel.pkl\",\n    \"source_bytes\": 1082156622,\n    \"dest_bytes\": 10341046,\n    \"ratio\": 104.6,\n    \"dest_sha256\": \"00532428a7f72fcf1466dae4ed59f9fb31afb365f0a493e17779e829093913a3\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop0/feasible-lake-25_FlowMatchingWithScore.pkl\": {\n    \"paths_config\": \"loop0\",\n    \"config_key\": \"prior_flow_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/prior/flow_matching/2026-04-27_10-56-42/feasible-lake-25_FlowMatchingWithScore.pkl\",\n    \"source_bytes\": 5523548808,\n    \"dest_bytes\": 361228590,\n    \"ratio\": 15.3,\n    \"dest_sha256\": \"97031d3a9d84ce7c35f0f3d2f3d99db16af2124cf167a8f9a45d2305ac00e4bb\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop0/eager-brook-26_FlowMap.pkl\": {\n    \"paths_config\": \"loop0\",\n    \"config_key\": \"prior_flow_map_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/prior/flow_map/2026-04-27_12-23-55/eager-brook-26_FlowMap.pkl\",\n    \"source_bytes\": 5524672359,\n    \"dest_bytes\": 361958965,\n    \"ratio\": 15.3,\n    \"dest_sha256\": \"5dd5132ba31ca3bd5fed53d10595d8fce82147ebac40d331a55ced7d436ba1e7\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop1/eager-feather-1_FlowMatching.pkl\": {\n    \"paths_config\": \"loop1\",\n    \"config_key\": \"perturbation_prediction_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/conditional_model/flow_matching/2026-04-02_14-16-39/eager-feather-1_FlowMatching.pkl\",\n    \"source_bytes\": 122572866498,\n    \"dest_bytes\": 1063688851,\n    \"ratio\": 115.2,\n    \"dest_sha256\": \"540b2b4100ca05b90f92b2daac415b34215a980ab8cd8419d91d005ad89a6659\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop1/easy-monkey-341_TargetPredictionModel.pkl\": {\n    \"paths_config\": \"loop1\",\n    \"config_key\": \"ct_classifier_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/cell_type_classifier/2026-04-08_00-00-38/easy-monkey-341_TargetPredictionModel.pkl\",\n    \"source_bytes\": 1139827500,\n    \"dest_bytes\": 11669836,\n    \"ratio\": 97.7,\n    \"dest_sha256\": \"f42004c7743dd5c5a1fea2e99026157ea21c46e9125640a439d9489d2c77afa1\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop1/woven-dawn-22_FlowMatchingWithScore.pkl\": {\n    \"paths_config\": \"loop1\",\n    \"config_key\": \"prior_flow_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/prior/flow_matching/2026-04-04_10-19-42/woven-dawn-22_FlowMatchingWithScore.pkl\",\n    \"source_bytes\": 5971588458,\n    \"dest_bytes\": 391236705,\n    \"ratio\": 15.3,\n    \"dest_sha256\": \"abf66dea36045821bb5712f2cc0ba9ecc32b2be2072871b8ef056196d007fcd4\",\n    \"tensors_verified\": true,\n    \"smoke_tested\": true\n  },\n  \"loop1/copper-thunder-24_FlowMap.pkl\": {\n    \"paths_config\": \"loop1\",\n    \"config_key\": \"prior_flow_map_path\",\n    \"source\": \"/ictstr01/home/icb/alessandro.palma/environment/collab-goettgens-SFC/project_folder/output/checkpoints/prior/flow_map/","encoding":"utf8","key":"checkpoints/manifest.json","name":"manifest.json","size":13170,"truncated":true}],"sourceObjectKey":"checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"theislab/LabCompass\", filename=\"checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"checkpoints/loop0/dandy-vortex-20_TargetPredictionModel.pkl","totalObjects":59},"summary":"Measurements behind LabCompass, a generative method for experimental design on single-cell data from the Theis lab, with wet-lab experiments and measurements by the Gottgens lab.","title":"LabCompass: spectral flow cytometry of in vitro hematopoietic differentiation across closed-loop experimental design rounds","updatedAt":"2026-11-17T19:55:32.237Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/theislab/LabCompass"}}