A stability-selected seven-gene transcriptomic panel in ulcerative colitis: Nested discovery and multicohort external evaluation
Jianyun Wang, Ru Xu, Ruofeng Liu, Wenjie Li, Yang Yang, Jie Liang, Xiaoshuang Shi, Yuan Gao, Baoxiang Wang, Qiong ChenObjective
To derive a stable mucosal transcriptomic panel for ulcerative colitis (UC) using leakage-resistant feature selection and to evaluate its performance across independent cohorts.
Methods
This secondary model-development study used human colonic mucosal transcriptomic data from GSE206171 (59 UC and 19 controls) for discovery. Twenty repetitions of five-fold outer cross-validation repeated differential-expression screening, candidate construction, feature ranking, scaling, and inner-CV tuning within each training fold. Genes selected in at least 70% of 100 outer folds defined the high-stability panel. Fixed coefficients were evaluated in GSE47908, GSE75214, GSE16879, and GSE243625. Secondary analyses examined published and inflammatory comparators, disease activity, immune enrichment, and donor-level SCP259 expression.
Results
The fully nested pipeline yielded a mean AUC of 0.832 (SD 0.052; repeat-level 2.5th-97.5th percentiles 0.741-0.898). Seven genes met the 70% stability threshold: CASP1, SAA1, LCN2, DUOX2, LPCAT3, IDO1, and DEFA6. AUCs were 0.944-0.971 for disease or activity contrasts in GSE47908 and GSE75214. In GSE243625, participant-aware AUCs were 0.951 for inflamed UC mucosa versus healthy controls and 0.964 for paired inflamed versus macroscopically non-inflamed UC mucosa, but 0.529 for non-inflamed UC mucosa versus healthy controls; all 15 paired participants had higher scores in inflamed mucosa (P = 6.10 × 10 -5 ). Baseline infliximab-response discrimination was poor (AUC 0.469). No AUC difference versus inflammatory comparators or three published UC signatures remained significant after multiple-testing correction in independent-sample comparisons. The panel-Mayo correlation (rho=0.714) was weaker than that for CXCL8 (rho=0.791).
Conclusions
Repeated nested selection identified a seven-gene panel with consistent discrimination for active UC mucosal contrasts. Near-chance separation of macroscopically non-inflamed UC from healthy mucosa and the negative treatment-response result indicate that the score mainly reflects active mucosal inflammation and requires prospective fixed-reference evaluation before clinical use.