Crohn's Anti-TNF Nonresponse Signature

What this research found

Anti-TNF therapy fails in a substantial share of Crohn's disease patients. A 26-gene, cell-state-informed signature of nonresponse was derived from single-cell ileal tissue, locked before any contact with an independent 199-sample bulk RNA-sequencing cohort, and then scored there against a surrogate endpoint, because true response labels turned out not to be publicly retrievable. The signature carries a real signal that survives adjustment for baseline inflammation, with an odds ratio of 1.93 per standard deviation and an AUROC of 0.685, but its probabilities do not survive the move between platforms — which led to a verdict of redesign rather than advance or abandon.

• The signal is real but modest, and not simply a restatement of how inflamed the tissue is. Adjusted for a baseline mucosal-inflammation proxy, the signature gives an odds ratio of 1.93 per standard deviation (p = 0.012), with discrimination of AUROC 0.685 (95% CI 0.603 to 0.760) and AUPRC 0.672 against a 50% prevalence baseline.
• Direct transfer between the two cohorts is blocked by an almost complete, platform-driven domain shift: a squared maximum mean discrepancy of 0.88 and a Jensen-Shannon divergence of 1.00 on the first principal component. Mean expression profiles still correlate across cohorts at Pearson r = 0.82, so the gap is largely one of measurement scaling rather than biology.
• Three of the 26 locked genes cannot be measured on the bulk platform at all: IGHG1 and IGHG3, which carry the two largest positive weights, along with MIR4435-2HG.
• Rank ordering transfers but absolute probabilities do not. The calibration slope is 3.34 (95% CI 1.66 to 5.53, excluding 1), meaning predicted probabilities are severely underconfident and compressed into an observed range of 0.40 to 0.72, with a Brier score of 0.241.
• Discovery-cohort cross-validation was close to chance, at a leave-one-out ROC-AUC of 0.529 and a five-fold value of 0.564, a reflection of how few patients single-cell cohorts contain — 22 ileal samples here.
• Weighing those results against each other, the recommendation was to redesign rather than stop or advance: move to rank-based scoring that is robust to platform, replace the genes the bulk platform cannot see, and recalibrate on a pilot cohort carrying real clinical response labels.

How it was done

Twenty-two ileal single-cell samples were quality-controlled from 159,677 candidate cells down to 112,014, clustered, and annotated for the inflammatory macrophage and activated fibroblast states of the GIMATS module before aggregation to patient-level pseudobulk. An elastic-net logistic model produced a 26-gene panel, locked and held apart from the validation data to prevent leakage. Domain shift between cohorts was quantified with Wasserstein distance, Jensen-Shannon divergence and maximum mean discrepancy over a 17,075-gene common universe, after which the locked score was applied to 199 treatment-naive paediatric biopsies and tested against a surrogate endpoint: a median split on an independent 27-gene mucosal inflammation module deliberately chosen not to overlap the signature, with baseline activity adjusted via a separate S100A8 and S100A9 proxy. Discrimination and calibration were bootstrapped over 1,000 resamples, and the resulting audit fed a stop, advance, or redesign decision framework.

Data sources

• GEO accession GSE134809 — single-cell RNA sequencing of ileal Crohn's disease resections; 22 ileal samples, 112,014 cells after quality control
• GEO accession GSE134881 (SRP216403) — RISK treatment-naive paediatric Crohn's bulk RNA sequencing, 199 ileal biopsies across 22,570 genes
• Martin et al., Cell 178:1493 (2019) — source of the GIMATS cell-state module

Figures from this analysis