Biomarker discovery on synthetic patient transcriptomes: three public cohorts and a benchmark

Background : Patient-level clinico-transcriptomic cohorts are used for biomarker discovery, patient stratification, validation and model training. However, most transcriptomic data remains siloed in hospitals, biopharma, and consortia because expression profiles and linked clinical metadata carry re-identification risk. Synthetic RNA-seq cohorts that capture population structure without retaining individual records could be shared in their place, but only if they accurately reproduce the results of standard translational analyses on the real data. We test this with dbTwin, a proprietary generator, on three public cohorts, so that every synthetic result can be checked against its real source, and release the synthetic cohorts with the benchmark used to evaluate them. Results: Synthetic versions of TCGA lung adenocarcinoma (510 samples; 19,719 genes), whole-blood sepsis (152; 27,097), and pediatric inflammatory bowel disease (322; 43,790) recovered most of the differential-expression (DE) signals tested. Across five contrasts on clinical variables declared to dbTwin as anchor columns, synthetic effect sizes kept their real magnitude (log2 fold-change slope 0.92–1.01), and 71.9–90.8% of real DE genes were recovered (81.7–97.2% at a relaxed screening threshold). For STK11 mutation (non-anchor column), effect sizes shrank (slope 0.70) but 96.8% of DE genes kept their direction. In five-fold cross-validation, classifiers trained on synthetic data were within 0.002 ROC-AUC of real-trained ones on held-out real patients and relied on largely the same driver genes (SHAP Pearson r = 0.808–0.975). Synthetic cohorts also retained gene–gene co-expression (Spearman ρ = 0.940–0.996), and 0 of 22 clinical comparisons exceeded sampling noise. Synthetic cohorts showed no sign of gross memorization by a distance-based check (fifth-percentile distance ratio 0.916–1.056), and attackers knowing a patient’s age and sex gained little on training patients over unseen ones (excess success 0.028– 0.066); those also knowing the patient’s expression gained more, mainly on clinical labels (up to 0.518 for tumor stage), not on expression values (per-gene maximum 0.092–0.177). Conclusions: On standard DE and machine-learning analyses, synthetic RNA-seq cohorts reproduced most real-data results on these three cohorts while retaining the joint structure of gene expression and clinical metadata. This suggests a possible route around data silos: a data custodian can generate and share a synthetic version of its cohort, and recipients can run their usual stratification and biomarker analyses on it and confirm candidates on the governed real data. The benchmark is generator-agnostic; its code and the real cohorts are publicly available at github.com/Nanda-Aditya/rna-twin, and the synthetic cohorts are freely and publicly available at huggingface.co/datasets/NandaAditya/rna-twin.

Authors

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11732548/v1
Primary Topic
Gene expression and cancer classification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Biomarker discovery on synthetic patient transcriptomes: three public cohorts and a benchmark

Aditya Nanda
Research Square
Gene expression and cancer classification
preprint

Biomarker discovery on synthetic patient transcriptomes: three public cohorts and a benchmark

Aditya Nanda
preprint en

Abstract

Background : Patient-level clinico-transcriptomic cohorts are used for biomarker discovery, patient stratification, validation and model training. However, most transcriptomic data remains siloed in hospitals, biopharma, and consortia because expression profiles and linked clinical metadata carry re-identification risk. Synthetic RNA-seq cohorts that capture population structure without retaining individual records could be shared in their place, but only if they accurately reproduce the results of standard translational analyses on the real data. We test this with dbTwin, a proprietary generator, on three public cohorts, so that every synthetic result can be checked against its real source, and release the synthetic cohorts with the benchmark used to evaluate them. Results: Synthetic versions of TCGA lung adenocarcinoma (510 samples; 19,719 genes), whole-blood sepsis (152; 27,097), and pediatric inflammatory bowel disease (322; 43,790) recovered most of the differential-expression (DE) signals tested. Across five contrasts on clinical variables declared to dbTwin as anchor columns, synthetic effect sizes kept their real magnitude (log2 fold-change slope 0.92–1.01), and 71.9–90.8% of real DE genes were recovered (81.7–97.2% at a relaxed screening threshold). For STK11 mutation (non-anchor column), effect sizes shrank (slope 0.70) but 96.8% of DE genes kept their direction. In five-fold cross-validation, classifiers trained on synthetic data were within 0.002 ROC-AUC of real-trained ones on held-out real patients and relied on largely the same driver genes (SHAP Pearson r = 0.808–0.975). Synthetic cohorts also retained gene–gene co-expression (Spearman ρ = 0.940–0.996), and 0 of 22 clinical comparisons exceeded sampling noise. Synthetic cohorts showed no sign of gross memorization by a distance-based check (fifth-percentile distance ratio 0.916–1.056), and attackers knowing a patient’s age and sex gained little on training patients over unseen ones (excess success 0.028– 0.066); those also knowing the patient’s expression gained more, mainly on clinical labels (up to 0.518 for tumor stage), not on expression values (per-gene maximum 0.092–0.177). Conclusions: On standard DE and machine-learning analyses, synthetic RNA-seq cohorts reproduced most real-data results on these three cohorts while retaining the joint structure of gene expression and clinical metadata. This suggests a possible route around data silos: a data custodian can generate and share a synthetic version of its cohort, and recipients can run their usual stratification and biomarker analyses on it and confirm candidates on the governed real data. The benchmark is generator-agnostic; its code and the real cohorts are publicly available at github.com/Nanda-Aditya/rna-twin, and the synthetic cohorts are freely and publicly available at huggingface.co/datasets/NandaAditya/rna-twin.

Research Square
Gene expression and cancer classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon