domifuzzz/statLens-Realworld-Benchmark-284

statLens Real-World Benchmark (284 studies)

This dataset holds 284 public omics studies. Each study is packaged the same way: a feature-by-sample matrix, sample metadata, and a study abstract. It also includes a design card for every study, which is a structured JSON description of the study's experimental design.

The repository has two top-level folders:

Folder	Content	Files	Size
Realworld-Benchmark/	284 study folders, each holding 3 files	852	~4.0 GB
real-world-design-card/	One design card per study, plus the input bundles and records that go with it	1,148	~14 MB

Studies are identified by their source-repository accession. The same ID names the study folder in Realworld-Benchmark/<ID>/ and the files real-world-design-card/cards/<ID>.json and real-world-design-card/bundles/<ID>.json.

---

1. Realworld-Benchmark/

Realworld-Benchmark/
├── GSE101491/
│   ├── study_abstract.md
│   ├── metadata.csv
│   └── omics_feature_matrix.csv
├── ST004675/
│   └── ...
└── ... (284 study folders)

1.1 Files in each study folder

omics_feature_matrix.csv: the quantitative data, with one row per feature and one column per sample.

• The first column is always feature_id. It holds a gene symbol, Ensembl ID, probe ID, metabolite name, taxon, protein, and so on, depending on the study.
• Every other column is a biological sample. The sample columns match metadata.csv sample_id, in the same order, in all 284 studies.
• Values are the processed values the depositors reported: counts, TPM/FPKM, microarray intensities, peak areas, and so on. They have not been renormalized, imputed, or filtered.
• Missing values are empty cells.

metadata.csv: one row per sample. In every study the first five columns are the same:

Column	Meaning
sample_id	Sample identifier; matches a column name in the matrix
group	Biological group or condition. Filled for every sample in every study
block	Subject, pairing, or blocking ID. Filled in 16 studies, empty otherwise
time	Time point or ordered stage. Filled in 35 studies, empty otherwise
batch	Technical batch. Filled in 29 studies, empty otherwise

The remaining columns differ from study to study. Examples are organism, tissue, specimen, cell_type, treatment, genotype, disease, platform, assay, instrument, original_sample_id, sample_description, characteristics, source_file, and study_id. Each file has between 8 and 31 columns.

study_abstract.md: a Markdown description of the study with three parts.

1. A header list with these fields: study accession, repository, organism, tissue/specimen, disease or biological context, omics modality, assay/platform, measurement type, number of samples and number of features in the final matrix, primary experimental factors, primary comparison, source files for the matrix and the metadata, repository URL, and associated publication.
2. Study summary: the depositors' study description.
3. Data-processing notes: the sources that were checked and how the matrix was taken from the deposited files.

1.2 Composition

Source repository and omics modality (from study_abstract.md)

Repository	Modality	Studies
GEO (GSE*)	Transcriptomics (RNA-seq and microarray)	208
GEO	Microbiome amplicon (16S / ASV / OTU)	5
GEO	Small RNA / miRNA expression	3
GEO	Epigenomics / DNA methylation	2
GEO	Proteomics	1
Metabolomics Workbench (ST*)	Metabolomics	42
MetaboLights (MTBLS*)	Metabolomics	21
PRIDE (PXD*)	Proteomics	2
Total		284

Measurement type (top categories, from study_abstract.md)

Measurement type	Studies
Gene-level counts (GEO supplementary table)	104
Microarray processed intensities (GEO series matrix)	48
Other processed values (GEO supplementary table)	30
Metabolite peak intensities / abundances (MetaboLights MAF)	21
Normalized expression values (GEO supplementary table)	15
FPKM / RPKM	12
TPM	10
Metabolite abundances (Metabolomics Workbench REST; peak area, intensity, etc.)	42
Protein intensities / abundances (PRIDE)	2

Organism

Organism	Studies
Homo sapiens	153
Mus musculus	80
Rattus norvegicus	8
Sus scrofa	3
Other organisms: other animals, plants, bacteria, fungi, parasites, metagenomes, and one rat + mouse study	40

There are 45 distinct organism strings in total. The Organism field in study_abstract.md is kept exactly as deposited.

Matrix size per study

	Min	Q1	Median	Q3	Max	Total
Samples (matrix columns)	5	10	20	48	1,254	15,910
Features (matrix rows)	11	1,785	29,349	55,536	2,749,693	12,260,398

The median proportion of exact-zero cells per matrix is 0.108.

Feature identifier type (feature_id column)

Type	Studies
Gene symbol	72
Ensembl gene ID	71
Metabolite name	63
Array probe ID	30
Numeric / row-index-like	13
Ensembl transcript ID	5
Taxon	3
Protein / composite	2
Other	25

Number of groups (distinct group levels)

Groups	2	3	4	5	6	7	9	10	12
Studies	161	45	44	14	10	6	2	1	1

---

2. real-world-design-card/

real-world-design-card/
├── output/
│   ├── design_cards.jsonl         # 284 design cards, one per line
│   ├── design_card_review.jsonl   # 284 per-study records; maps line number ↔ study ID
│   └── gt_agreement.json          # per-study and summary agreement with ground-truth labels
├── cards/<ID>.json                # the same 284 cards, one file per study (keyed by ID)
├── bundles/<ID>.json              # 284 study bundles (per-study input summaries)
├── prompt/system_prompt.txt
├── pipeline/                      # Python scripts and GENERATOR_BRIEF.md
├── state/                         # manifest, batches, per-study diagnostics/ and attempts/, summary
└── README.md

2.1 Design card (cards/<ID>.json, output/design_cards.jsonl)

Each card is a flat JSON object with 57 design fields plus field_provenance, 58 keys in total. The cards in design_cards.jsonl do not include the study ID. To find which study a line belongs to, use design_card_review.jsonl (study_ID ↔ design_cards_output_line) or read cards/<ID>.json.

Aspect	Fields
Study	study_context, biological_or_clinical_question, study_hypothesis, organism, tissue_or_specimen
Data & measurement	data_modality, assay_type, platform, feature_type, feature_granularity, measurement_family, measurement_scale, expected_zero_proportion, zero_inflation_expected
Groups & contrast	design_factors, factor_levels, number_of_groups, group_levels, group_sample_counts, reference_level, primary_factor, primary_comparison, primary_contrast_type, primary_contrast_levels
Sample size & units	study_population_sample_count, total_biological_units, total_assayed_samples, target_sample_count, experimental_unit, sampling_unit, replicate_unit, technical_replicates_present
Dependence structure	independent_groups, clustered_design, cluster_id_variable, paired_or_repeated, paired_across_groups, subject_id_available, repeated_measure_unit, samples_per_subject, within_subject_factor
Time	ordered_time_factor, number_of_timepoints, time_levels, samples_per_timepoint, same_subjects_across_time, test_group_time_interaction
Batch & covariates	batch_covariate_recorded, batch_variable_name, batch_levels, other_covariates, covariate_types, covariate_levels_or_range
Features	analysis_subset_description, reported_raw_feature_count, reported_filtered_feature_count, target_feature_count
Provenance	field_provenance: a tag for each field, one of article, derived, or simulated

Card-level composition

primary_contrast_type	Studies
two_group	137
multi_group	97
paired	31
time	11
group_time_interaction	5
trend	2
other	1

measurement_family	Studies
continuous	165
count	119

data_modality	Studies
transcriptomics	206
metabolomics	63
microbiome	6
proteomics	4
genomics	1
other	4

2.2 Study bundle (bundles/<ID>.json)

A per-study JSON summary of the three benchmark files. All counts and statistics in it were computed from the files.

Key	Content
study_id	Accession
study_abstract	Full text of study_abstract.md
metadata	n_rows, columns, and column_summary (per-column distinct values and counts). Studies with ≤ 100 samples also include all metadata rows
matrix_profile	n_features, n_sample_columns, sample_columns_match_metadata, feature_id_examples, feature_id_pattern, duplicate_suffix_ids, value_summary (integer_valued, has_negative, min, median, max, missing_proportion, zero_proportion), suspected_non_sample_columns, technical_suffix_groups
design_ledger	n_samples; group (levels, counts, placeholder values); block, time, batch (filled status, levels, counts, block sizes); crosstab_group_by_time; crosstab_group_by_batch; partition_equalities; abstract_count_check (sample/feature counts from the abstract compared with the files)

2.3 Other files

File	Content
output/design_card_review.jsonl	One record per study: input_order, study_ID, design_cards_output_line, prompt_sha256, bundle_sha256, validator_attempts, first_attempt_passed, schema_violations, warnings, ledger_flags, unresolved_issues, raw_file_checks, status
output/gt_agreement.json	per_study: a card-vs-ground-truth comparison for five items (primary_contrast_structure, subject_dependence, batch_effect, other_covariates, statistical_model). summary: agreement counts, accuracy, and confusion tables
prompt/system_prompt.txt	The design-card specification text
pipeline/	build_bundles.py, validate_rw.py, assemble.py, compare_gt.py, GENERATOR_BRIEF.md
state/	manifest.json (sha256 of the prompt and each bundle), batches.json, summary.json, generator_overrides.json, build.log, diagnostics/<ID>.json (per-study file checks), attempts/<ID>.jsonl (per-study validation log)
README.md	Folder documentation (in Chinese)