feat: rewrite the CosMx reader (stitching, multimodal, co-registration)#406
Draft
timtreis wants to merge 9 commits into
Draft
feat: rewrite the CosMx reader (stitching, multimodal, co-registration)#406timtreis wants to merge 9 commits into
timtreis wants to merge 9 commits into
Conversation
Replace the CosMx reader with a substantially expanded implementation ported from Nanostring-Biostats/spatialdata-io-ns: multi-FOV stitching, multimodal RNA+Protein, polygon<->label<->table co-registration, skip_empty_fovs, and opt-in per-channel percentile image normalization. - 5 modules: cosmx.py + _cosmx_io/_cosmx_stitching/_cosmx_discovery/_cosmx_utils - reuses the existing CosmxKeys constant - adds runtime deps: polars, dask, zarr>=3 - synthetic regression test suite (tests/test_cosmx_regression.py) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Match the upstream reader convention: decorate cosmx() with @inject_docs and document the recognized CosMx file suffixes from CosmxKeys in the docstring.
readers/cosmx.py + the 4 _cosmx_* helper modules -> a readers/cosmx/ package (_reader/_io/_stitching/_discovery/_utils), with relative intra-package imports. No behavior change; all 86 tests pass.
The dev-era suite (82 tests) was ~2x upstream's heaviest reader test. Drop the brittle snapshot tests and the pure internal-helper unit tests (header matching, tile clipping) that are already covered end-to-end, and thin the remaining helper-unit classes to their highest-value cases. Keeps all end-to-end, feature (skip_empty_fovs, normalization, flip), and co-registration coverage. 56 tests (60 incl. parametrize), all passing.
Extract repeated boilerplate in _reader.py into small helpers, no behavior change: - _global_id_series: the AnnData unwrap + categorical->Int64 coercion (was in _collect_global_cell_ids_from_tables and _filter_tables_to_ids) - _parse_cell_table: the identical TableModel.parse(... region_key/instance_key/ overwrite_metadata) call at 3 sites - _parse_labels: the identical Labels2DModel.parse(... dims/translation) at 3 sites - _finalize_transcripts: the duplicated PointsModel.parse tail of the parquet and csv transcript readers - move the self-contained obs column-sanitization out to _utils._sanitize_obs_columns All 60 tests pass.
…hored) Make the single-modality path label-anchored, matching the multimodal path, so both follow one rule: the segmentation is the ground truth — rasterise every cell, then filter the table(s) to the cells that rasterised. Decided from real data: in multimodal runs the protein-derived segmentation is ground truth and modalities quantify DIFFERENT cell subsets (breast: RNA covers 99.1%, protein 100% of segmentation), so labels cannot be anchored on any single table. - single path: drop the table->polygon->label 'allowed_global_ids' cascade and the final-pass table filter; read_labels() rasterises all polygons, then the table is filtered to the label IDs (was: only rasterise polygons with a table row, dropping segmented-but-unquantified cells) - remove the now-dead allowed_global_ids param (read_labels/_labels_from_polygons) and _collect_global_cell_ids_from_tables - fix _prescan_max_cell_id: it read column 'cell_ID' but polygon CSVs use 'cellID', so polygon-only high cell IDs (exactly the orphan cells) were missed and mis-locked max_cell_id — now reads either column Validated on real data (breast multiomics, FOV 128): labels=649 (all polygons), RNA table=541, Protein table=649 — the 108 cells RNA lacks are kept in the labels. All 60 synthetic tests pass.
…scan A polygon cell with no expression row (highest cell_ID) must stay in the labels while the table holds only quantified cells; the high cell_ID also exercises the max_cell_id prescan reading the polygon 'cellID' column.
The pre-scan that locks max_cell_id matched the id column against a hardcoded
("cell_ID", "cellID") pair and swallowed every failure with a bare
`except: continue`. Both could silently drop a file and mis-lock the max — the
exact failure class behind the polygon-orphan mis-lock fixed in 190b121.
- Extract `_match_canonical(hdr, canon)` from `_match_header` so a single
alias table backs both; route the pre-scan's id- and fov-column lookups
through it (now also matches `cell_id`/`object_id`/`roi`/`fov_id`,
case-insensitive) instead of a narrower ad-hoc set.
- Replace the blanket silent `except` with per-step warnings: a file that
exists but can't be scanned, has no recognizable cell_ID column, or whose
max can't be read is logged and skipped — never silently dropped.
Add regression tests for aliased id/fov columns and the no-silent-drop warning.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drop three re-implemented helpers in favour of the package/ecosystem equivalents, and attach the standard reader provenance every other reader sets. Net -50 LOC in the cosmx package; behaviour for valid inputs is unchanged (full regression suite + zarr round-trip green). - `_deduplicate_names` -> `anndata.utils.make_index_unique` for channel names (suffix format changes from "x (1)" to "x-1", the anndata standard). - `_sanitize_obs_columns` -> `spatialdata.sanitize_table`, the blessed zarr-safe key sanitiser (case-insensitive uniqueness across obs/var keys). - Add `_set_reader_metadata(sdata, "cosmx")` in `_assemble_sdata` so the output carries `spatialdata_io_reader`/`_software_version` like the other readers (cosmx previously set neither). Kept bespoke (no core equivalent): per-channel image normalization, categorical->string coercion, and the CSV-header canonicaliser. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This replaces the CosMx reader with a substantially expanded implementation developed in
Nanostring-Biostats/spatialdata-io-ns. The current reader (~280 lines, single file) is replaced by a 5-module reader (~3.8k lines) that adds:global_cell_idacross elements).skip_empty_fovs— drops phantom FOVs listed in the positions file but absent on disk (avoids an inflated/desynced canvas).image_normalization_percentile, defaultNone= legacy dtype-max) — recovers low-signal uint16 channels that otherwise render near-black; scale-only and reversible.CellStatsDir, px-only / mm-only / px+mm FOV positions).Reuses the existing
CosmxKeysconstant (values unchanged). Adds runtime depspolars,dask,zarr>=3. Other readers are untouched.The public
cosmx()signature changes:transcripts=Trueread_transcripts=True(+read_images/labels/proteins/polygons/gexp)imread_kwargs: Mapping = MappingProxyType({})imread_kwargs: dict | None = Noneimage_models_kwargs: Mapping = MappingProxyType({})image_models_kwargs: dict | None = Nonefovs,channels,n_workers,flip_image,image_normalization_percentile,polygons_as_labels,keep_polygons_after_rasterize,align_rasters_to_polygons,add_fovs_as_shapes,skip_empty_fovs,preview_fovsNew public objects:
CosMxDataset(frozen dataclass describing the on-disk layout) andCosMxDatasetReader.