IBDex freezes 3,168 gut biopsies into one reference atlas: deep remission quiets inflammation but leaves a disease state behind
Synopsis
Bulk mucosal RNA sequencing in inflammatory bowel disease has been analyzed one cohort at a time, so historical data accumulate without becoming a reusable reference. IBDex trained a tissue-aware conditional variational autoencoder on 3,168 biopsies from 1,680 patients across four cohorts and suppressed cohort identity adversarially, then froze the gene interface, encoder, decoder, pathway basis and phenotype directions. After freezing, dataset identity fell to balanced accuracy 0.307 against chance 0.250 while inflammation stayed recoverable (AUC 0.82); in 93 patients with concordant endoscopic and histological remission, inflammation approached controls (AUC 0.58) but disease state stayed displaced (AUC 0.725).
Interpretation
Bulk intestinal transcriptomes can now be read against a frozen reference atlas instead of rebuilding normalization, batch correction and pathway scale for every new cohort. Previously each study rebuilt its own molecular scale within its own cohort, so the scale belonged to the cohort rather than the tissue, and historical data accumulated without becoming a reference. On the held-out test partition, dataset identity reached balanced accuracy 0.307 against chance 0.250, while colon versus ileum, diagnosis and inflammatory activity remained decodable in the same representation, and the inflammation direction reached AUC 0.82 (95% confidence interval 0.78–0.85).
The frozen atlas separates inflammation from a broader disease-associated molecular state that conventional marginal differential expression conflates. On the same 3,000 genes and the same samples, the inflammation and disease-state differential-expression contrasts correlated at ρ = 0.834 and shared 51 of their top 100 genes. On those same samples, the two response maps from phenotype-anchored counterfactual perturbation correlated at ρ = −0.161 and shared five top-100 genes, with inflammation driven by TNF–NF-κB and disease state by proliferative, type I interferon and metabolic programs.
Deep remission with concordant endoscopic and histological healing brings inflammation back near controls without bringing disease state back. Among 93 patients in concordant remission versus 241 controls, the inflammation coordinate reached AUC 0.58 (P = 0.021) and the disease-state coordinate AUC 0.725 (P = 1.8 × 10⁻¹⁰). Against 252 patients with concordantly active disease, remission separated strongly on inflammation (AUC 0.165, P = 1.19 × 10⁻²¹) but not on disease state (AUC 0.518, P = 0.600).
Perspective
The demonstrated scope is 3,168 adult mucosal biopsies from four construction cohorts, plus frozen projection onto three unseen IBD cohorts, a longitudinal olamkicept trial, and external PSC and colorectal cancer cohorts; the atlas is restricted to its fixed gene interface, represented anatomical conditions and data compatible with its preprocessing assumptions.
PACP is a directional response of a fitted decoder, not a causal intervention, and external disease directions require labels fitted after projection. Reliability varies by pathway and compartment, so weak or untestable coordinates are explicitly flagged rather than treated as equivalent measurements; in the olamkicept trial only seven paired responders and two non-responders contributed, so those trajectories are descriptive.