# SpatialDataAgent-DB sample metadata v2 Schema `sda-sample-metadata-2`, version 2.2.1 (2026-11-06), built from release v3 (`sda-unified-20261010-v3`, 12,163 samples). No expression, coordinate or image file changes; v2 is a new metadata table that sits next to the release. ## Files | File | Content | |---|---| | `samples_metadata_v2.parquet` | one row per sample, list columns kept as lists (recommended) | | `samples_metadata_v2.csv` | the same table; lists joined with `;`, confidence as JSON | | `studies_v2.csv` | one row per study (1,265): title, species, platforms, assays, organs, diseases, PubMed IDs, split | | `schema_v2.json` | definition of every column | | `changes_v2.csv` | every value that differs from a v3 column, with its basis (204 platform changes, 32 manual corrections in 14 samples) | | `dropped_disease_terms.csv` | disease terms removed by the short-synonym filter (331) | | `qa_v2.json` | counts and consistency-flag totals | ## What is new for users - **Organ and tissue.** Each sample has an organ from a fixed list of 43 (a metastasis is placed at its site), and 11,605 samples (95%) have a UBERON tissue term (279 distinct terms). The GEO text stays in `tissue_raw`. - **Disease.** Each sample has a `disease_status`: disease 7,023, normal 3,948, perturbation 1,125, not stated 67. "Perturbation" covers interventions that are not a named disease, such as a knockout without a disease name, a drug, mechanical loading or regeneration. 6,479 samples have MONDO terms (404 distinct); a sample can have more than one, for example IPMN with co-occurring PDAC. The control arm of a disease or tumour model counts as disease, not normal. - **Sample kind.** primary tissue 8,574, tumour tissue 2,754, metastasis 352, organoid 212, xenograft 185, whole embryo/organism 57, cell culture 17, engineered tissue 10, unknown 2. - **Donor.** - `life_stage`: embryonic, fetal, neonatal or infant, juvenile, adult, aged, or not stated. 5,103 samples have a stated stage. Where the age notation is unambiguous, the stage comes from a fixed age rule (3,752 samples); otherwise from the model. Rule for mouse and rat: neonatal up to P21, juvenile under 6 weeks, adult under 18 months, aged from 18 months. For human, ages in years or post-conception weeks. - `sex`: male 2,689, female 2,012, mixed 21, unknown 7,441 (77 taken from the words male / female in the sample's own title or characteristics). Coded values without a key, such as 1/2, stay unknown. - `donor_key` = `:`, for donor-level splits (3,391 samples). - **Species.** - NCBI Taxonomy ID and common name for all 42 host species. - Mixed entries are split. 78 samples list several species and 36 are xenografts; pathogens such as *Plasmodium berghei* go in `co_detected_organisms`. - Two series that GEO lists only as the genus *Mus* are set to house mouse. The evidence for each is recorded: a C57BL/6J strain in one, ENSMUSG gene IDs in the other. One series that GEO lists only as Hepatitis B virus is set to human, from its gene IDs. - `feature_species` gives the species implied by the Ensembl gene IDs (8,032 samples). It differs from `organism` for 47 samples (flag `feature_species_differs`). Most are macaque or rat tissue mapped to a human or mouse reference, or mouse hosts carrying human grafts. In one study (GSE313530) the macaque and baboon samples carry each other's gene IDs; check it before any cross-species use. - **Assay.** `assay_name` gives the specific method (Visium CytAssist, Slide-seqV2, Open-ST, Pixel-seq, Decoder-seq, ...), and `assay_type` says whether it is imaging-based or sequencing-based. - Of the 299 v3 samples labelled "Other", 197 now belong to a named platform family: DBiT-seq 100, BMK S1000 51, Molecular Cartography 23, Visium 18, Slide-seq 5. - Nova-ST (7 samples) was labelled Stereo-seq in v3 and is now Other. - All are listed in `changes_v2.csv`. The assay of each "Other" study was checked by hand against the GEO text. - **Preservation.** FFPE 2,839, fresh frozen (including OCT-embedded) 2,327, fixed frozen 40. - **Publications.** PubMed IDs for 8,479 samples, from GEO series and sample records. - **Review fields** are in English. ## How the annotations were made 1. **Input.** For each sample: the release tissue, organism and platform, plus the GEO sample title, source and characteristics and the series title. 2. **Model.** Official GPT-6 Luna Decisions (`api.openai.com/v1/decisions`). This API answers only multiple-choice and yes/no questions, so the model chooses from fixed lists and never writes free text or ontology IDs. - Step 1: organ, sample kind, disease status and life stage. - Step 2: one UBERON term from candidates. Each candidate MONDO term is a separate yes/no question. 3. **Candidates.** A script finds candidate terms in the text by exact name or synonym match. It handles organ adjectives (peritoneal → peritoneum), plurals, British spelling, possessives, common disease abbreviations and standard disease-model names (5xFAD, R6/2, SOD1-G93A, mdx …). The direct parents of each matched term are added. All IDs come from the ontology files (UBERON basic and MONDO, downloaded 2026-11-04). 4. **Fixed rules after the model.** - The most specific true MONDO terms are kept. - A true disease term sets the status to disease. - Disease terms that matched only through a short upper-case synonym (e.g. "CHS", "SARS" inside "SARS-CoV-2") are dropped. - The organ follows the UBERON term's part-of chain when that chain names a different listed organ. - Life stage comes from the age rule where the notation is unambiguous. 5. **Versions.** Each row records its `annotation_version`. - 2.3 is the main run. - 2.4 re-asked the 842 non-human, non-mouse samples (species-aware life stage) and the samples that mention SARS-CoV-2. - 2.5 re-asked 303 samples whose text contains a disease adjective such as "diabetic". 6. **Confidence.** Luna's confidence is stored per field. Fields below 0.6 are listed in `annotation_low_confidence` (3,119 samples, mostly disease status). ## Accuracy The checks ran on 300 random samples never used to develop the method. Gemini 3.8 Flash judged each field against the same source text, and every "wrong" verdict was then read by hand (Claude). | Field | Judge "wrong" rate (300 random samples, final build) | |---|---| | organ | 0.3% | | tissue term | 0.7% | | sample kind | 0.7% | | life stage | 1.7% | | disease | 3.0% | - **Disease.** Most "wrongs" are a study-level disease given to a control (see limitations), or a missing term where the status is right. - **Life stage.** The remaining "wrongs" are ages the rule leaves to the model, such as "0d" or an unlabelled number in an embryo study. - **"Acceptable"** (true but less specific) was 1–13% per field. ## Known limitations - The study-level disease sometimes reaches a control whose own record does not say it is a control (about 1.5% of samples in QA). For example, a "WT" sample in a pulmonary hypertension study got pulmonary hypertension. Check `disease_status_source` and the raw fields when the control/disease split matters. - Vagina and pleura are not in the organ list, so those samples are "other" or "lung". - Disease abbreviations outside the curated list (e.g. "MN" for membranous nephropathy) give a disease status without a MONDO term. - `life_stage` is "not stated" for 7,060 samples: most GEO records give no age. - 20 samples are flagged `normal_but_disease_text`: the raw disease field names a disease but the sample was judged normal. Most are tissue adjacent to a tumour, which is correct; the rest are unclear. ## Reproduce Scripts are in `research/metadata-v2-20261104` and `research/tissue-ontology-pilot-20261104` of the project repository: - `map_decisions.py --all`: annotation (resumable); - `assay_decisions.py`: assay of the "Other" studies; - `build_metadata_v2.py`: this table; - `qa_full.py`: QA sample and judge. ## Versions - **2.2.1** (tag `metadata-v2.2.1-20261106`): a field corrected by hand no longer carries the model's confidence or low-confidence flag, so `annotation_low_confidence` filters keep human-checked samples. The schema text now lists every `life_stage_source` value. No label changed. - **2.2** (tag `metadata-v2.2-20261106`): corrections from a human reader's review of the 200-sample pack. - Rules applied to every sample: - MONDO "..., non-human animal" terms are removed from human samples (16 samples). - Organoids, cultures and engineered tissue get life stage "not stated", because culture time is not the donor's age. - Sex is read from the words male / female in the sample's own text when the extracted field is empty (77 samples). - Mouse and rat ages in days within an age series (e.g. "0d") are read as postnatal days. - Manual corrections in 10 samples, each with its basis in `overrides_v2.csv`: - actinic keratosis for the 5 GSE294844 samples (the series design says so; the sample records do not); - "HD" read as healthy donor; - a diverticular fistula read as diverticulitis, not IBD; - Cre-only and wild-type controls set to normal; - a graft with no stated site set to organ "other". - New column `sex_source`. - **2.1** (tag `metadata-v2.1-20261105`): each assay now has one name (for example "Resolve Molecular Cartography" became "Molecular Cartography", and "Visium (FFPE)" became "Visium", since preservation is its own field). `um_per_unit` and `image_um_per_px` are copied exactly from the release parquet. Organ, tissue, disease, life stage and the other annotations did not change. One derived field did: `assay_type` of the 6 ISS (Cartana) samples of GSE227046 is now imaging-based (it was sequencing-based). - **2.0** (tag `metadata-v2-20261105`): first release.