Download README.md from einarolafsson/spacr-example-ops-e2e: direct link, hf CLI and curl.
- Browser
- Download file 12.6 kB
-
https://huggingface.co/datasets/einarolafsson/spacr-example-ops-e2e/resolve/main/README.md
- Command line
-
hf download hf://datasets/einarolafsson/spacr-example-ops-e2e/README.md
-
curl -L -o README.md https://huggingface.co/datasets/einarolafsson/spacr-example-ops-e2e/resolve/main/README.md
license: cc-by-4.0
pretty_name: 'spaCR example: OPS end-to-end subset (Funk et al. 2022, plate A, well A1)'
tags:
- microscopy
- optical-pooled-screen
- spacr
size_categories:
- n<1K
spaCR example: OPS end-to-end subset, Funk et al. 2022 plate A, well A1
A small raw subset of one optical pooled screen (OPS) well that runs every OPS step in spaCR: stitch, align, genotype (decode) and phenotype. It is the test data for spaCR's end-to-end OPS test and the default example for OPS live mode.
This is the core tier (about 190 MB, for CI). Every file has the published name, shape, dtype,
axes and ImageJ metadata. Pixels inside a region of interest (ROI) are byte-identical to the archive; every
other pixel is 0. The ROI of each file is listed in roi_manifest.json. Phenotype site 1280 is an all-zero
placeholder: it keeps the 1,281-field round-well layout and tests the field-without-nuclei path.
Source and citation
- Source: BioImage Archive S-BIAD394, "The phenotypic landscape of essential human genes", released
2022-03-31. Files: https://ftp.ebi.ac.uk/biostudies/fire/S-BIAD/394/S-BIAD394/Files/
(
screenA/20200202_6W-LaC024A). The SHA-256 of every original file is inroi_manifest.json(source_sha256). - Citation: Funk L, Su KC, Ly J, Feldman D, Singh A, Moodie B, Blainey PC, Cheeseman IM. The phenotypic landscape of essential human genes. Cell 185(24):4634–4653.e22 (2022). doi:10.1016/j.cell.2022.10.017, PMID 36347254.
- Library:
library/pool10_design.csv, the pool10 design from the authors' code repository (OpticalPooledScreens, MIT licence, Copyright (c) 2017 David Feldman, Avtar Singh, Rebecca Carlson, Paul Blainey).library/pool10_prefixes.csvis derived from it: one columnprefix=sgRNA[:prefix_length]for every design row, in file order (20,445 prefixes of 11 bases). spaCR reads this file (ops_library). - Licence: this subset is redistributed under CC-BY-4.0; please cite Funk et al. 2022. The source record S-BIAD394 carries no licence attribute; the BioImage Archive releases direct submissions under CC0 or CC-BY-4.0, and both allow redistribution.
What is included
| Part | Files |
|---|---|
| Sequencing (SBS, 10X, 1,480 × 1,480 px, uint16) | tiles (sites) 321, 322, 331, 332, all 11 cycles: c1 5-plane stack DAPI-CY3-A594-CY5-CY7 (axes CYX), c2–c11 one file per base channel CY3, A594, CY5, CY7 (axes YX). 44 files. |
| Phenotype (20X, 2,960 × 2,960 px, 4 focal planes, uint16) | fields 1255, 1264, 1280: DAPI-GFP (axes ZCYX, 4 × 2), A594 and AF750 (axes ZYX). 9 files. |
library/ |
pool10_design.csv (original), pool10_prefixes.csv (what spaCR reads). |
settings/ |
ops_testdata_settings.csv (ops_align_model = tile_affine) and ops_testdata_settings_similarity.csv, written by spaCR's settings writer, with .json twins. Paths are relative: run spaCR with this folder as the working directory. |
reference/inputs/ |
ops_ph_placement_seed.json (phenotype placement seed, schema below) and flatfield/ (the plate's stored spaCR flat-field maps: flat_info.json, fields.json, one flat_20200202_6W-LaC024A_<channel>.npy per channel). |
reference/ |
spaCR reference outputs (see reference/PROVENANCE.json). |
roi_manifest.json |
per file: source SHA-256 and size, ROI(s) in file coordinates (core tier), the canvas ROI and the production placements. |
SHA256SUMS |
SHA-256 of every file in this tier. |
Why these sites. spaCR takes a well's layout from its highest site number, so the subset holds SBS site 332 (333-tile round well) and phenotype site 1280 (1,281-field well). The stitch keeps only the largest connected group of tiles, so the subset is one 2 × 2 block at the well's right rim (with all 4 overlaps). Field 1264 sits on the block's 4-tile junction and field 1255 is its left neighbour in another tile group. No included file is on the plate's broken-input lists (phenotype fields 598, 757, 860, 864, 867, 871, 909, 988, 1133, 1136; SBS c11 CY5 Site-56).
Folder layout mirrors the plate as spaCR's OPS reader indexes it:
ops_testdata_funk2022_A1/
20200202_6W-LaC024A/
sequencing/images/input/c1 … c11/10X_c{N}_A1_<channel>_Site-<site>.tif
phenotype/images/input/DAPI-GFP-A594-AF750/20X_DAPI-GFP-A594-AF750_A1_<channel>_Site-<site>.tif
library/ settings/ reference/ roi_manifest.json README.md SHA256SUMS
What was cut (core tier)
- SBS ROI, in the production canvas of well A1 (cycle-1 stitch): y 14,750–15,750, x 24,800–25,800. It holds the 4-tile junction, both overlap strips across it and the SBS area under the two phenotype ROIs. In each file it is the canvas ROI minus that tile's production placement, rounded outward and padded on every side by |cycle shift| + max |channel shift| of that cycle + 4 px (shifts from the production decode), so the registered cycles still cover it.
- Cycle-1 DAPI also keeps each tile's whole overlap with the other three tiles (+ 8 px). Without it a zeroed edge crossing an overlap strip pulled the stitch by 1 px; with it the core and full stitches agree exactly.
- Phenotype ROIs, 1,200 × 1,200 field px, all focal planes and channels: field 1255 rows 952–2152, columns 1760–2960 (touches the right border); field 1264 rows 952–2152, columns 0–1200 (touches the left border). Both lie inside the SBS ROI. Field 1280 is all zero.
- Storage: lossless deflate (zlib) with the horizontal predictor. tifffile decodes it with the standard library; imagecodecs is not needed. Every written file was read back and compared with the source: inside the ROI bit-identical, outside 0 (core); everything bit-identical (full).
What this subset can and cannot test
- Core tier: the reader and layout, the stitch (4 overlaps and a junction) and, after item 940, alignment on 2 tile groups and phenotype measurement per cell. Full tier: all of these plus decode (all rules, decoy counting) and determinism.
- It cannot estimate a false-discovery rate (FDR). spaCR's FDR is decoy calls / real calls × 1.268 / 4 (81,780 decoys). The core tier expects about 1 decoy call (Poisson 0–3) and the full tier about 6; even 6 decoys leave the FDR uncertain by more than a factor of 2. FDR calibration stays with full-well runs.
- Flat-field fitting needs many fields; the stored plate maps are used instead.
Phenotype placement seed (reference/inputs/ops_ph_placement_seed.json)
A partial well cannot supply the whole-well anchor fields spaCR's phenotype phase aligns, so the phenotype
placement of well A1 (production spaCR run) is stored as a file. spaCR item 940 adds the setting that reads it.
Schema spacr.ops_ph_placement_seed, schema_version 1 (JSON object):
| Key | Type | Meaning |
|---|---|---|
schema, schema_version |
str, int | "spacr.ops_ph_placement_seed", 1. |
plate, well |
str | "20200202_6W-LaC024A", "A1". |
phenotype_fields |
int | fields of the phenotype acquisition (1281); the layout is spacr.ops_layout.round_well_layout(phenotype_fields). |
sbs_tiles |
int | tiles of the sequencing acquisition (333). |
frame.kind |
str | "sbs_tile_anchored": every (y, x) is in the stitched cycle-1 sequencing well frame of the production run (ops_geometry y/x). |
frame.reference_tiles |
{site: {y, x}} | production placements (top-left, px) of SBS tiles 321, 322, 331, 332 in that frame. |
frame.tile_shape |
[h, w] | SBS tile size, [1480, 1480]. |
raster.origin, raster.column_step, raster.row_step |
[dy, dx] floats | centre(site) = origin + column × column_step + row × row_step, (column, row) = layout.position(site); the same raster spacr.ops_phenotype.phenotype_centres fits from anchors. |
transform.scale |
float | well px per field px (0.24994). |
transform.rotation_deg |
float | field rotation in the well frame (−0.002). |
transform.field_shape |
[h, w] | phenotype field size, [2960, 2960]. |
fields[] |
list | for each included field: site, grid_column, grid_row, centre_y, centre_x, sbs_site, offset_y, offset_x (derived; for checking). offset = centre − scale × R(rotation) @ (h/2, w/2) (spacr.ops_merge.field_transform). |
provenance |
object | source database and table, the anchor fields with their raster residuals, and the largest difference between the raster and any stored production centre (≤ 1e-6 px). |
Using it on another stitch of these tiles (for example this subset's own stitch, whose origin differs):
offset = mean over the reference tiles present of (new placement − stored placement); add offset to
raster.origin and to every centre; the steps, scale and rotation stay. On this subset all 4 tiles are present.
Status of the reference outputs (2026-11-13, spaCR nightly 6afe2e8f8c)
- Run and reproducible: stitch (core and full agree exactly; full is within 0.15 px of production), SBS objects (Cellpose-SAM, CPU), decode (rule (a) + S2; reruns are identical).
- Full tier vs production (gate (c), PASS): the rule is at least 99 % same guide where both called, for
nuclei more than 100 px from the subset edge: 99.01 % on 13,705 called nuclei. Agreement where either side
called (always reported beside it): 94.1 % overall, 95.8 % beyond 100 px. Over all nuclei, 98.9 % same guide
where both called, and 735 nuclei called on one side only (
reference/provenance_call_differences.parquet). The full tier's cycle and channel shifts equal production's exactly. - Scope of the core tier (maintainer decision, 2026-11-13): the core tier is for the reader and layout, stitch and (after item 940) phenotype tests only. Decode tests use the full tier. The core tier's decode outputs are kept for the record but are not a test reference: spaCR's whole-field cycle registration locks onto the zeroed ROI edges, so its cycle shifts are wrong.
- Not yet run: phenotype placement, segmentation, measurement and merge (similarity and tile_affine).
The phenotype phase refuses without whole-well anchors; spaCR item 940 will place the fields from
reference/inputs/ops_ph_placement_seed.json. - Details and gates:
reference/PROVENANCE.json.
How to regenerate
- Data:
build_ops_testdata.py <out_root>reads the 53 source files from an S-BIAD394 mirror (the NAS copy is byte-identical to the archive) and writes both tiers;write_placement_seed.pywrites the seed. - Reference: with this folder as the working directory, run spaCR on
settings/ops_testdata_settings.csv(spacr.ops_engine.run_ops(settings, wells=["A1"], phases=...)); seereference/PROVENANCE.jsonfor the commit, the phases run, the environment lock and the CPU.
Repository layout: two tiers
ops_testdata_funk2022_A1_core.tar(about 194 MB): the core tier, the smallest subset that runs the OPS end-to-end test.ops_testdata_funk2022_A1_full.tar(about 1.25 GB): the full tier, with the complete set of inputs for plate A well A1.SHA256SUMS_tars.txt: sha256 of both tar files (sha256sum -c SHA256SUMS_tars.txt).
Each tar unpacks to a top folder ops_testdata_funk2022_A1/ that holds README.md, roi_manifest.json, SHA256SUMS (check with sha256sum -c SHA256SUMS inside the folder), settings/, reference/ and library/.
Revision 2: phenotype and merge references
Revision 2 adds reference outputs for the phenotype and merge phases. Every revision 1 file is unchanged.
ops_testdata_funk2022_A1_rev2_core.tar(4.2 MB) andops_testdata_funk2022_A1_rev2_full.tar(26.6 MB): unpack each over its tier's folder. They add only new paths:reference/phenotype/;reference/merge/tile_affine/andreference/merge/similarity/;reference/PROVENANCE_rev2_addendum.json;SHA256SUMS_rev2.
SHA256SUMS_rev2.txt: sha256 of both revision 2 tars.- How the references were made: spaCR's phenotype phase, with phenotype placement seeded from
reference/inputs/ops_ph_placement_seed.json(settingops_ph_placement_seed), then the merge phase with both alignment models. The revision 1ops_objectsandops_barcodeswere used as fixed inputs. The commit, settings and timings are in the addendum. - Merge results:
- full tier: 4,531 phenotype cells, of which 4,106 merged with tile_affine and 4,105 with similarity; median distance 0.64 px and 0.88 px;
- core tier: 490 cells, 392 and 391 merged; median distance 0.38 px and 0.53 px.
- Licence and citation are as above (CC-BY-4.0; Funk et al. 2022).