Where the values come from

Sources other than the archive supply at least one fact beyond location and collection date for 294,648 of the 573,001 catalog samples (51.4%) in 1,726 of 2,798 projects. Not counted: 427,176 subject identifiers (they link samples of one person but say nothing about the person) and derived columns (age category, detailed location). Human verification adds no values of its own.

Bar width = share of the concept's facts from each source; samples = catalog samples with at least one fact for it, and their share of all samples. Hover a segment for counts, the concept name for its columns.

Sequence archive   Supplementary tables   Full text   Abstract / description   External resources

Subjects. 343,328 samples carry a subject identifier read from a source (BioSample attributes such as host_subject_id, sample names decoded by two agreeing rules, participant columns of supplementary tables). 83,848 more carry a generated identifier (the sample accession, marked generated on the sample page) because the paper or archive states that every sequenced stool sample comes from a different person — for example "a single stool sample was collected from each participant", or a stated number of participants equal to the number of samples with no mention of repeated sampling. Two independent model readings must agree, and it is never done for longitudinal, before–after, multi-visit, replicate or pooled designs.

Longitudinal projects (1,117, with 56,251 people sampled more than once and 223,100 samples; plus 207 projects whose source describes repeated sampling but whose samples cannot yet be assigned to people). A person's samples are ordered on one time axis, chosen in this order: age at collection (when every sample has it and it varies), full collection dates, then timepoint labels converted to days ("week 4" = 28 d, "month 6" = 183 d, "baseline" and "birth" = 0). From these the project page shows the number of people sampled repeatedly, samples per person, follow-up length (first to last sample per person; median and maximum) and the median interval between consecutive samples, with a timeline of up to 36 people. The sampling schedule and time origin quoted from the paper are shown alongside (time axes used: date 296, age 135, timepoint 107, timepoint_mapped 58, date_month 23, timepoint_stated_unit 21 projects).

Self-contradictory metadata (red ! on the project sheet and project pages). When two sample-level sources give incompatible values for the same sample — different sex or country, antibiotics yes vs no, ages more than 30 days and 20 % apart, BMI more than 2 apart, or healthy control vs a disease — or when the samples of one subject id disagree on sex, the catalog publishes neither value: the field is left empty and both sides, with sources and quotes, go to gut_metadata_contradictions.parquet. A subject id shared by samples of both sexes in a larger series is treated as not identifying one person and is withheld instead. Differences in formatting, rounding or specificity (e.g. Crohn's disease vs IBD) are not contradictions; the source precedence decides those.

Resolved vs. not resolved. A longitudinal project is resolved (solid sheet tag L, badge "longitudinal") when its repeatedly sampled people have samples placed on a time axis, so the gap between samples is known and the per-person timeline can be drawn; otherwise it is not resolved (dotted L, badge "longitudinal — not resolved"): the source describes repeated sampling but the samples lack ages, full dates or convertible timepoint labels, or cannot be assigned to people. Relatedness is resolved (solid R) when related or co-located people are linked to their samples — families or shared settings with at least two catalogued people, or explicit pairwise relations such as FMT donor → recipient — so the relationship graphs can be drawn; a dotted R means the paper describes them but the link to samples is not resolved yet. Both flags are columns of gut_studies (longitudinal_resolved, relatedness_resolved). An intervention project (randomised or non-randomised controlled, crossover, single-arm before/after, or observational with a procedure) gets a solid I when at least 80 % of its samples carry a per-sample arm (the intervention column) — the project page then draws the arms by design — and a dotted I otherwise.

Relatedness and shared settings (335 projects; 5,962 families, 4,008 of them with a sampled parent and child; 610 shared settings). Four fields record it per sample: family_id (the source's family, dyad, household or twin-pair id), family_role (mother, father, parent, child, twin, sibling, partner, grandparent, other relative, household member), and shared_setting_type / shared_setting_id (a household, village, tribe, daycare or school, care home, ward, military unit, team or workplace that at least two sampled people share — never the whole cohort or the city of recruitment). They come from archive attributes, sample-name tokens whose meaning the source states, supplementary tables and paper text, each with its quote; relatedness is never inferred from similar ages or a shared country. Families are summarised as structures (e.g. "mother + child×2"), drawn as small diagrams on the project page, and listed in gut_relatedness.parquet (one row per family or shared setting); relations that roles cannot express — twin zygosity, sibling pairs, FMT donor → recipient — are in gut_relations.parquet. Designs found: fmt donor recipient 95 · mother infant 89 · households 50 · families 38 · twins 31 · community 24 · siblings 21 · parent child 20 · partners 17 · tribe 11 · hospital ward 10 · workplace 8 · care home 6 · father child 5 · team 4 · daycare school 2 · military 2. The Project sheet filter "Sampling" selects longitudinal projects and each kind of relatedness; tags L and R mark them in the table.