By requesting access you agree to the Data Use Agreement: noncommercial use only and no reverse engineering of the generator. Full text: https://huggingface.co/datasets/NandaAditya/rna-twin#license-and-data-use-agreement

rna-twin

Synthetic bulk RNA-seq cohorts with linked clinical data, each released with a benchmark report that shows what it preserves.

This dataset holds synthetic versions of three public cohorts: pediatric inflammatory bowel disease (RISK, GSE57945), sepsis (GSE184900), and TCGA lung adenocarcinoma. Every synthetic sample has full-transcriptome expression and clinical variables, generated with dbTwin. For each cohort you get one full synthetic cohort, five training-fold cohorts, and a PDF report comparing the synthetic data with the real data.

Sepsis cohort. (A) Real and synthetic samples embedded in a single UMAP fit. (B) Correlations among the 500 most variable genes, real below the diagonal and synthetic above.

Benchmark code · Preprint (forthcoming) · HF collection · Data Use Agreement

Get the data

Anyone can download the files after accepting the Data Use Agreement. In short: noncommercial use only, and no reverse engineering of the generator.

Log in to Hugging Face, fill in the access form on this page, and click Agree and access.
Log in from the terminal with the same account: hf auth login. If the CLI isn't installed, use uvx --from huggingface-hub hf auth login.
Load a cohort. Each config (pediatric_risk, sepsis, tcga_luad) has a full split and fold1_synthetic through fold5_synthetic.

from datasets import load_dataset

full = load_dataset("NandaAditya/rna-twin", "sepsis", split="full",
                    revision="v1.0.0", token=True)
fold1 = load_dataset("NandaAditya/rna-twin", "sepsis", split="fold1_synthetic",
                     revision="v1.0.0", token=True)

Or with pandas:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("NandaAditya/rna-twin", "sepsis/syn_sepsis.csv",
                       repo_type="dataset", revision="v1.0.0")
df = pd.read_csv(path, index_col=0)

To rerun the benchmark yourself, clone the GitHub repo and run uv run python get_data.py. It walks you through the login and agreement steps and downloads these files into datasets/<cohort>/dbTwin/. Then uv run python scripts/run_pipe.py sepsis reruns the sepsis benchmark.

Files

<cohort>/syn_<cohort>.csv                        full synthetic cohort
<cohort>/artifacts/syn_<cohort>_fold_<1-5>.csv   synthetic training folds
reports/<cohort>.pdf                             benchmark report
assets/qa_preview.png
LICENSE                                          CC BY-NC 4.0
DUA.md                                           Data Use Agreement

Each <cohort>/ folder mirrors datasets/<cohort>/dbTwin/ in the benchmark repo.

Each row is one synthetic sample. The first, unnamed column is a row index, not a gene or a patient ID, so read it with index_col=0. The remaining columns are expression values and clinical variables under their source names. Genes are Ensembl IDs, plus some gene symbols in TCGA-LUAD. Clinical values can be missing. RISK and sepsis expression values are read counts. TCGA-LUAD values are not read counts: they are integers on the scale of the source data, upper-quartile-normalized RSEM, where each sample's 75th percentile over non-zero genes is about 1,000.

Cohort	Full rows	Fold rows (1-5)	Genes	Clinical columns	Prediction target	Anchor columns
`pediatric_risk`	322	257, 257, 258, 258, 258	43,790	4	`sample_disease`	disease, age, sex
`sepsis`	152	121, 121, 122, 122, 122	27,097	4	`Classification`	classification, age, sex
`tcga_luad`	510	406, 406, 406, 407, 407	19,719	10	`clinical_stage_clean`	EGFR, KRAS, stage, age, sex

The clinical columns are:

RISK: sample_disease, Sample Characteristic[sex], Sample Characteristic[age at diagnosis years], Sample Characteristic[organism part]
Sepsis: age, sex, ethnicity, Classification
TCGA-LUAD: clinical_os_months, clinical_os_status, clinical_age, clinical_sex, clinical_egfr_mut, clinical_kras_mut, clinical_alk_mut, clinical_tp53_mut, clinical_stk11_mut, clinical_stage_clean

Which file to use:

The full file is the one to use for analysis. The benchmark uses it for the clinical, co-expression, UMAP, and DE comparisons.
Each fold file was generated from one real training split only, so compare it only with its matching held-out real split. The split manifests are in the GitHub repo. Don't pool the five folds into one cohort.

How the data was made

dbTwin is a proprietary generator that does not use deep learning. It takes a real expression matrix, the sample metadata, and an optional set of anchor columns. Anchor columsn are clinical variables whose relationships with expression matter most. Synthetic cohorts include both de-novo expression tables and linked clinical metadata. The anchor columns for each cohort are listed in the table above. The full cohort was generated from all real samples. Each fold cohort was generated from one training split, with that split's held-out patients left out. The generator itself is not included here or in the GitHub repo.

Benchmark results

All numbers come from the reports in reports/, generated on November 2 and 3, 2026. Co-expression, UMAP, and DE compare the full synthetic cohort with the full real cohort once. Values with ± are the mean ± SD over five outer splits. ✅ marks a pass against the report's fixed thresholds.

	Pass if	Pediatric RISK	Sepsis	TCGA-LUAD
Co-expression Spearman ρ	≥ 0.8	0.996 ✅	0.965 ✅	0.940 ✅
UMAP MMD	≤ 0.1	0.022 ✅	0.037 ✅	0.020 ✅
ROC-AUC, real-trained (TRTR)	reference	0.880 ± 0.032	0.894 ± 0.032	0.573 ± 0.036
ROC-AUC, synthetic-trained (TSTR)	≥ TRTR − 0.05	0.878 ± 0.035 ✅	0.893 ± 0.027 ✅	0.573 ± 0.044 ✅
SHAP Pearson r	≥ 0.8	0.949 ± 0.012 ✅	0.881 ± 0.043 ✅	0.828 ± 0.056 ✅

TRTR models are trained on real data and TSTR models on the matching synthetic fold. Both are scored on the same held-out real patients, whom the generator never saw. Synthetic-trained classifiers matched real-trained ones in all three cohorts and relied on the same genes. Tumour stage in TCGA-LUAD is hard to predict even from real data, and the synthetic cohort reproduced that ceiling rather than inflating it.

Differential expression. Real DE genes have padj < 0.05 and |log2FC| > 0.5. Log2FC r and slope compare synthetic with real fold changes over all real DE genes; a slope below 1 means synthetic effects are shrunk toward zero. Recall is the share of real DE genes that are also significant in the synthetic data at a relaxed rule (padj < 0.075, |log2FC| > 0.4).

Cohort	DE contrast	Real DE genes	log2FC r (≥ 0.8)	Slope	Recall (≥ 80%)
Pediatric RISK	Crohn's disease vs control	5,641	0.993 ✅	1.00	96% ✅
Pediatric RISK	Ulcerative colitis vs control	1,169	0.980 ✅	0.95	82% ✅
Sepsis	SepsisBSI vs No-Sepsis	2,743	0.994 ✅	1.01	97% ✅
TCGA-LUAD	KRAS, 0 vs 1	1,121	0.961 ✅	0.94	92% ✅
TCGA-LUAD	EGFR, 0 vs 1	1,369	0.970 ✅	0.92	93% ✅
TCGA-LUAD	STK11, 0 vs 1 (not an anchor column, not graded)	2,586	0.883	0.70	69%

Effect sizes kept their real magnitude in every anchored contrast (slope 0.92 to 1.01). STK11 is the instructive exception, and the only contrast on a column not declared as an anchor column. Its effect sizes still correlated with the real ones (r = 0.88) but shrank by about 30%, and recall fell to 69%. Associations with anchor columns are preserved with high fidelity, and associations with other columns only partially.

Privacy.

	Pediatric RISK	Sepsis	TCGA-LUAD
DCR 5th-percentile ratio	0.916 ± 0.011	1.056 ± 0.035	1.047 ± 0.012
Excess training success, attacker knows age and sex	0.029	0.066	0.028
Excess training success, attacker also knows 100 genes	0.094	0.112	0.143
Secrets with excess > 0.2 (age and sex / + 100 genes)	0 / 1	0 / 0	0 / 3

DCR checks for memorization. A ratio near 1 means synthetic samples sit no closer to training patients than training patients sit to each other. Attribute inference, run with Anonymeter, asks whether the synthetic data helps an attacker learn something new about a real training patient. Excess training success is the attack's success on training patients minus its success on held-out patients, and the two excess rows report the median of the 10 highest per-secret means. An attacker who knows only age and sex gains little (0.028 to 0.066); one who also holds 100 genes measured on the patient's own sample gains more (0.094 to 0.143).

Uses and limitations

The data is meant for noncommercial method development, benchmarking, teaching, pipeline prototyping, and hypothesis generation. Confirm anything that matters on real data. The data is not validated for diagnosis, treatment decisions, or clinical or regulatory evidence.

The results above cover only the tested targets, contrasts, gene subsets, and classifier. We have not tested survival models, pathway enrichment, or external cohorts. The cohorts are small, so subgroup estimates can be unstable, and biases in the source studies carry over into the synthetic data.

The records are synthetic and were generated from public, de-identified data. We measured disclosure risk with DCR and two attribute-inference attacks. These are empirical tests.

The real data comes from GEO (GSE57945, GSE184900) and TCGA (cBioPortal PanCancer Atlas). Its original terms still apply, so please also cite the original studies, listed in each cohort's README.md in the GitHub repo. The TCGA results shown here are in whole or part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga.

Citation

If you share the data or anything made from it, include this attribution and say what you changed:

rna-twin v1.0.0 © 2026 dbTwin, Inc., CC BY-NC 4.0, https://huggingface.co/datasets/NandaAditya/rna-twin

Please cite both the benchmark repository and the dataset:

@software{nanda2026rnatwin,
  title  = {rna-twin: synthetic bulk RNA-seq cohorts and benchmark},
  author = {Nanda, Aditya},
  year   = {2026},
  url    = {https://github.com/Nanda-Aditya/rna-twin}
}

@misc{rnatwin2026,
  title     = {rna-twin: synthetic bulk RNA-seq cohorts},
  author    = {Nanda, Aditya},
  publisher = {Hugging Face},
  year      = {2026},
  version   = {v1.0.0},
  url       = {https://huggingface.co/datasets/NandaAditya/rna-twin}
}

Changelog

v1.0.0, November 3–4, 2026: first public release. These CSVs are the exact files the benchmark reports in reports/ were computed from.

The code examples above pin revision="v1.0.0", so they keep loading this version after later releases.

License and Data Use Agreement

Version 1.0, October 30, 2026

This agreement is between you and dbTwin, Inc. It covers the synthetic CSVs in the Hugging Face dataset NandaAditya/rna-twin, both the full cohorts and the training-fold datasets. You accept it by clicking Agree and access on the dataset page.

The synthetic data is © 2026 dbTwin, Inc. and licensed under CC BY-NC 4.0. The full text is in the LICENSE file. The terms below are in addition to that license. They don't change it, and they don't take away any rights it gives you.

Use the data only for noncommercial purposes, as CC BY-NC 4.0 defines them, unless dbTwin has given you a commercial license.
Don't try to reconstruct dbTwin's generation method, or use the data to train a model meant to imitate it, without our written permission. Training other models, benchmarking, comparing generators, running privacy tests, and publishing your results, good or bad, are all fine.
The data is provided as is, without warranties of any kind. It isn't validated for diagnosis, treatment, or other clinical decisions.
We may revoke your Hugging Face access if you break these terms.
If we update this agreement, the new version applies only to access granted after the update.

Hugging Face shares your account details and form answers with dbTwin. We use them to manage access and to understand how the data is used, and we keep them only as long as we need them for that. We'll email you about new releases only if you chose Yes on the form. To see, correct, or delete your information, or to change that choice, email aditya@dbtwin.ai.

Questions and commercial licensing: aditya@dbtwin.ai.

Downloads last month: 1,112

Total file size:

445 MB