---
pretty_name: RNA-seq Subsets Derived from the 1M Human SRA Metadata Dataset

language:
  - en

license: cc-by-4.0

task_categories:
  - tabular-classification
  - text-classification

tags:
  - biology
  - rnaseq
  - sra
  - metadata
  - annotations
  - bioinformatics
---

# RNA-seq Subsets Derived from the 1M Human SRA Metadata Dataset

Extracted from the 1M annotated sample: chumphati/1M_RNAseq_SRA_samples_annotations.
Annotated with Metappuccino *(Fiona Hak, Camille Marchet, Daniel Gautheret, Mélina Gallopin, Metappuccino: large language model-driven reconstruction of sequence read archive metadata for cancer research, Bioinformatics, Volume 42, Issue 5, May 2026, btag166, https://doi.org/10.1093/bioinformatics/btag166)*.

## CCLE-like
*ccle_cellosaurus_conf35.v1.0.csv* : Diverse, high-confidence collection of human cell-line RNA-seq samples matched to unique Cellosaurus entries. Runs were retained only when both the `cell_line` confidence score and the overall metadata confidence score were at least 3, and when the reported cell-line name could be unambiguously matched to a Cellosaurus identifier using its canonical name or a non-conflicting synonym after normalization. Primary-tissue annotations, unknown cell lines, samples describing multiple cell lines, unmatched names and aliases shared by several Cellosaurus entries were excluded to ensure that each selected run represented a clearly identified cell line. Small-RNA and hybrid-capture libraries were also excluded, while the remaining libraries were grouped as poly(A), rRNA-depleted, alternative protocols or unspecified protocols. For each Cellosaurus identifier and library group, only the highest-quality run was retained, prioritising samples with the most complete metadata, parsable age, known sex, organ, biopsy site and cell type, and the largest base count. A single representative was normally selected per Cellosaurus cell line, with a second representative permitted only when both a poly(A)-selected and an rRNA-depleted sample were available, thereby preserving protocol diversity without allowing repeated runs from the same cell line to dominate the final dataset. When neither of these two preferred protocols was available, another explicitly reported library-selection protocol was prioritised, and an unknown or `other` protocol was used only as a last resort.

## GTEX-like

Contains four GTEx-like datasets of high-confidence, non-cancerous, untreated bulk RNA-seq samples derived from primary human tissues rather than cell lines or cultured material. Eligible runs were required to have valid run and study accessions, `conf_organ ≥ 3`, `row_confidence_1_5 ≥ 3`, a single valid UBERON organ code, a known organ name, and a sufficiently confident non-cancer annotation; disease status had to be explicitly healthy with `conf_disease ≥ 3`, or unknown only when `conf_disease = 5`. Small-RNA and hybrid-selection libraries, single-cell samples, explicit cell lines, organoids, transformed cells, treated samples, cancer samples and ambiguous organ annotations were excluded. The remaining samples were divided according to both organ reproducibility and library protocol: `gtex_like_all.v1.0.csv` contains organs represented by at least three independent studies and includes all allowed non-small-RNA, non-hybrid protocols; `gtex_like_all_polya.v1.0.csv` applies the same criteria but retains only poly(A)-selected libraries; `gtex_like_rare_organ_all.v1.0.csv` contains rarer organs represented by exactly two independent studies across all allowed protocols; and `gtex_like_rare_organ_polya.v1.0.csv` contains those rare organs using poly(A)-selected libraries only. Within each file, samples were selected in a balanced round-robin manner, with a maximum of 500 runs per organ, 20 runs per study and organ, 100 runs per study overall, and no single study contributing more than 50% of the samples for an organ.

## TCGA-like

Contains six TCGA-like datasets of high-confidence bulk RNA-seq cancer samples. Eligible runs were required to have `conf_organ ≥ 3`, `conf_disease ≥ 3`, `conf_is_cancer ≥ 3` and `row_confidence_1_5 ≥ 3`, a single valid UBERON organ code, a single valid Disease Ontology code, and an explicit cancer status. Cell lines, organoids, PDX models, xenografts, cultured or transformed cells, adjacent or matched-normal tissues, benign and precancerous conditions, ambiguous composite annotations, small-RNA libraries and hybrid-selection protocols were excluded. Cancers were defined by the combination of organ UBERON code and disease DOID and were retained only when represented in at least three independent studies. `tcga_like_primary_all.v1.0.csv` contains primary tumours from all permitted library protocols, whereas `tcga_like_primary_polya.v1.0.csv` contains only their poly(A)-selected libraries; blood, bone marrow and other liquid-biopsy material were excluded from both. `tcga_like_metastasis_all.v1.0.csv` contains tissue metastases from all permitted protocols, while `tcga_like_metastasis_polya.v1.0.csv` is restricted to poly(A)-selected metastatic samples; blood, plasma, serum, circulating tumour cells, platelets and exosomes were excluded. `tcga_like_hematological_all.v1.0.csv` contains haematological malignancies from all permitted protocols, and `tcga_like_hematological_polya.v1.0.csv` contains their poly(A)-selected subset; this category includes explicitly annotated leukaemias, lymphomas, myelomas and related malignancies, including mycosis fungoides, as well as generic cancer annotations when the organ is blood or bone marrow. Within each file, samples were selected by a balanced round-robin procedure, with a maximum of 500 runs per cancer type, 1000 per organ, 20 per study and cancer, and 100 per study overall, while preventing any single study from contributing more than 50% of one cancer category; untreated samples and rows with known age and sex were prioritised, although treated samples or missing demographic information were not automatically excluded.

## Others

- *pbmc_blood_mononuclear_age_balanced_conf35_primary_multibin_dataset.v1.0.csv* : Age-balanced collection of healthy primary PBMC-related bulk RNA-seq samples. Samples were selected when PBMCs were explicitly identified in the `organ` or `biopsy_site` annotation, or when the material was annotated as blood and the reported `cell_type` corresponded to a PBMC-compatible mononuclear population, such as T cells, B cells, NK cells, monocytes, dendritic cells or related CD-marker-selected populations. Red blood cells and erythroid cells were excluded, as were platelets, megakaryocytes, granulocytes, neutrophils, eosinophils, basophils, plasma, serum and blood-vessel samples. Selected runs also had to represent healthy, primary tissue, have an exact parsable age between 0 and 110 years, an organ confidence score and overall row confidence of at least 3, and no small-RNA or hybrid-selection protocol. Ages were divided into ten-year bins, with the selection aiming for approximately equal representation across bins and an overall balance of 5000 samples below age 50 and 5000 samples aged 50 or older. To maximise study diversity and reduce domination by large cohorts, the dataset retains only studies containing eligible samples in at least two different age bins, required each selected study to contribute to at least two bins, limited each study to 25 samples per age bin and 100 samples overall, and filled the dataset using a quality-ranked round-robin procedure across studies and age groups. Rows not included therefore either did not unambiguously represent PBMCs or PBMC-compatible blood cells, lacked the required age, health, tissue, protocol or confidence information, or could not be retained under the age-balance and study-diversity constraints. Starting from the original 1037214 RNA-seq runs, 946130 remained after requiring conf_organ ≥ 3, 887388 after additionally requiring row_confidence_1_5 ≥ 3, 104733 after selecting explicit PBMC samples or blood samples with a PBMC-compatible cell type, 32425 after requiring a precise and parsable age, 7184 after restricting the dataset to healthy samples, 2956 after requiring cell_line = primary tissue, and finally 2936 after excluding small-RNA and hybrid-selection libraries.
- *melanoma_set.csv* : Contains all high-confidence melanoma RNA-seq runs identified in the source dataset. Selection was based directly on the disease text and did not require a Disease Ontology code. No additional restrictions were applied to the organ, biopsy type, treatment status, study, library-selection protocol, or sequencing source: the file may include primary tumours, metastases, cell lines, bulk RNA-seq and single-cell RNA-seq whenever they satisfy the melanoma-specific criteria.