IBDex: A Frozen Patient-Scale Atlas for Bulk Transcriptomics in Inflammatory Bowel Disease
- 1University of Birmingham
Abstract
Background & AimsBulk mucosal RNA sequencing in inflammatory bowel disease (IBD) is analyzed one cohort at a time, so each study rebuilds its own molecular scale. We aimed to build a frozen atlas that new cohorts could use without retraining.
MethodsWe built IBDex from 3,168 intestinal biopsies from 1,680 patients with Crohns disease, ulcerative colitis or no IBD, drawn from four cohorts. Patients were kept within a single training, validation or test partition. A tissue-aware conditional variational autoencoder with adversarial suppression of cohort identity was trained on a 3,000-gene interface. The scaling, encoder, decoder, pathway basis and phenotype directions were then frozen. Phenotype-anchored counterfactual perturbation (PACP) produced anatomy-conditioned gene-response maps. We tested the atlas without parameter updates in three unseen IBD cohorts, a longitudinal olamkicept trial, and cohorts with primary sclerosing cholangitis (PSC) or colorectal cancer (CRC).
ResultsIn held-out data, dataset identity fell to near chance (balanced accuracy 0.307 vs 0.250). Anatomy, diagnosis and inflammation remained recoverable (inflammation AUC 0.82). Marginal differential expression conflated inflammation and disease state ({rho} = 0.834). PACP separated them ({rho} = -0.161), with inflammation driven by TNF-NF-{kappa}B, and disease state by proliferative, type I interferon and metabolic programs. In 93 patients in combined endoscopic and histological remission, inflammation approached controls (AUC 0.58), but disease state stayed displaced (AUC 0.725, P = 1.8 x 10{square}1{square}). The frozen atlas recovered clinical, single-cell-defined and disease-behavior structure in unseen cohorts. It also tracked responder trajectories during olamkicept therapy and separated PSC (patient-level AUC 0.794) and CRC (AUC 0.938- 1.000). Where de novo differential expression became unreliable, transferred response maps stayed reproducible ({rho} [≥] 0.89).
ConclusionsIBDex turns historical bulk biopsy data into a reusable reference. It shows that deep remission moves inflammation toward control levels but leaves a persistent molecular disease state.
Related
- Characterisation of Genetically-Derived Patient Clusters Highlights Genetic Heterogeneity of Inflammatory Bowel Disease
- 5-Aminosalicylic Acid-induced Colitis in a Patient with Rheumatoid Arthritis without Inflammatory Bowel Disease
- Aligning patient and planetary health: planetary health diet adherence and inflammatory bowel disease prognosis across Western and Eastern cohorts
- Fatigue burden and patient attribution across remission states in inflammatory bowel disease: a longitudinal cohort study.
- Genome-Wide Transcriptomics Uncovers a Lung-to-Aorta Inflammatory Program Driving Influenza-Induced Atherosclerosis Exacerbation
BibTeX
@misc{madeswaran2026ibdex,
title = {{IBDex: A Frozen Patient-Scale Atlas for Bulk Transcriptomics in Inflammatory Bowel Disease}},
author = {A. K. Madeswaran and A. Acharjee},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.20.26365993},
url = {https://www.medrxiv.org/content/10.64898/2026.10.20.26365993v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - IBDex: A Frozen Patient-Scale Atlas for Bulk Transcriptomics in Inflammatory Bowel Disease
AU - A. K. Madeswaran
AU - A. Acharjee
PY - 2026
DA - 2026/10/20
PB - medRxiv
DO - 10.64898/2026.10.20.26365993
UR - https://www.medrxiv.org/content/10.64898/2026.10.20.26365993v1
AB - Background & AimsBulk mucosal RNA sequencing in inflammatory bowel disease (IBD) is analyzed one cohort at a time, so each study rebuilds its own molecular scale. We aimed to build a frozen atlas that new cohorts could use without retraining. MethodsWe built IBDex from 3,168 intestinal biopsies from 1,680 patients with Crohns disease, ulcerative colitis or no IBD, drawn from four cohorts. Patients were kept within a single training, validation or test partition. A tissue-aware conditional variational autoencoder with adversarial suppression of cohort identity was trained on a 3,000-gene interface. The scaling, encoder, decoder, pathway basis and phenotype directions were then frozen. Phenotype-anchored counterfactual perturbation (PACP) produced anatomy-conditioned gene-response maps. We tested the atlas without parameter updates in three unseen IBD cohorts, a longitudinal olamkicept trial, and cohorts with primary sclerosing cholangitis (PSC) or colorectal cancer (CRC). ResultsIn held-out data, dataset identity fell to near chance (balanced accuracy 0.307 vs 0.250). Anatomy, diagnosis and inflammation remained recoverable (inflammation AUC 0.82). Marginal differential expression conflated inflammation and disease state ({rho} = 0.834). PACP separated them ({rho} = -0.161), with inflammation driven by TNF-NF-{kappa}B, and disease state by proliferative, type I interferon and metabolic programs. In 93 patients in combined endoscopic and histological remission, inflammation approached controls (AUC 0.58), but disease state stayed displaced (AUC 0.725, P = 1.8 x 10{square}1{square}). The frozen atlas recovered clinical, single-cell-defined and disease-behavior structure in unseen cohorts. It also tracked responder trajectories during olamkicept therapy and separated PSC (patient-level AUC 0.794) and CRC (AUC 0.938- 1.000). Where de novo differential expression became unreliable, transferred response maps stayed reproducible ({rho} [≥] 0.89). ConclusionsIBDex turns historical bulk biopsy data into a reusable reference. It shows that deep remission moves inflammation toward control levels but leaves a persistent molecular disease state.
ER -