DOP097Discovery and external validation of machine learning-derived blood gene signatures for predicting endoscopic disease severity in Ulcerative Colitis

Abstract Background Non-invasive prediction of endoscopic severity in ulcerative colitis (UC) remains challenging, as current clinical indicators often fail to accurately distinguish patients with inactive/mild from moderate-to-severe endoscopic disease, an important distinction for guiding therapeutic decisions and clinical trial design. Methods In a prospective cohort of UC patients (THAMES-IBD; n = 182), peripheral blood transcriptomes were compared between moderate-to-severe (n = 89) versus inactive/mild (n = 93) endoscopic disease, revealing differentially expressed genes (DEGs) enriched for B cell-mediated humoral immunity and type 1 interferon response (Figure 1). Machine learning (ML) models (LASSO, Random Forest, Elastic Net, and XGBoost) were used for feature selection to identify DEGs predictive of endoscopic severity. Random Forest classifiers trained on these DEGs evaluated predictive performance using 10-fold cross-validation, with discriminative ability measured by area under the receiver operating characteristic curve (AUC), both alone and in combination with validated clinical symptom scores. In a subset of patients with faecal calprotectin (FC) data (n = 132), models comparing gene signatures alone, clinical symptoms plus FC, and all three in combination were evaluated. External validation was performed in a large independent cohort of UC patients1 (Mount Sinai; n = 386), using logistic regression of gene set variation analysis (GSVA) scores of the discovered peripheral blood gene signatures together with clinical symptom scores. Results In the THAMES-IBD discovery cohort, ML-derived blood gene signatures robustly discriminated moderate-severe from mild-inactive endoscopic disease activity in UC patients (AUC 0.81–0.87), outperforming symptom-based models alone (AUC 0.77). In the cohort subset with FC data, gene signatures alone (AUC 0.82-0.85) and gene signatures together with clinical symptoms + FC (AUC 0.82–0.88) exceeded the predictive performance of the clinical symptoms + FC model (AUC 0.66). External validation in the Mount Sinai cohort confirmed these findings: GSVA scores of the gene signatures achieved AUC 0.75-0.77, and combined gene signature plus clinical symptom models achieved AUC 0.80-0.81, both outperforming clinical symptoms alone (AUC 0.73) (Figure 2). Conclusion ML-derived blood gene signatures reliably predict endoscopic severity in UC and are reproducible across cohorts. Integration of blood transcriptomics with routine clinical indices enhances non-invasive assessment of endoscopic disease severity. These findings reflect peripheral blood immune responses associated with mucosal inflammation and support the development of blood-based biomarkers for precision monitoring in UC. References: 1.Argmann C, Hou R, Ungaro RC, et al. Biopsy and blood-based molecular biomarker of inflammation in IBD. Gut. 2023;72(7):1271-1287. doi:10.1136/GUTJNL-2021-326451. Conflict of interest: Thomas, John P: Research support from AstraZeneca and Roche Liu, Yufan: No conflict of interest Cozzetto, Domenico: No conflict of interest Matadamas Guzman, Meztli: Employed by AstraZeneca Malas, Sadek: No conflict of interest Saifuddin, Aamir: Personal Fees: I have received speaker fees from Galapagos (now, Alfasigma) and Ferring. I have received travel support from Galapagos (now, Alfasigma), Janssen Pharmaceuticals and Dr Falk Pharma. Dixit, Dimple: No conflict of interest Laszczak, Isabel: No conflict of interest Song, Yang: Employed by AstraZeneca Jugder, Bat-Erdene: Employed by AstraZeneca Korcsmaros, Tamas: Grant: Unilever, Roche Hess, Sonja: Employed by AstraZeneca Bednar, Kyle: Employed by AstraZeneca Powell, Nick: Grant: Takeda, BMS, Pfizer, Astra-Zeneca Personal Fees: Abbvie, Abivax, Allergan, Astra-Zeneca, Bristol-Myers Squibb, Celgene, Celltrion, Dr Falk Pharma UK Ltd, Ferring, Galapagos, GSK, Janssen, MSD, Roche, Pfizer, Sobi, Takeda, Tillotts