{
  "abstract": "Background Molecular subtyping of IBD currently requires genome-wide transcriptomic profiling, limiting clinical applicability. We aimed to develop a minimal gene classifier capable of assigning IBD molecular subtypes from the smallest possible gene set, determining whether a clinically deployable panel could achieve diagnostic-grade accuracy.Methods Transcriptomic data from 544 IBD samples (UC=215, CD=329) across four GEO cohorts (GSE75214, GSE179285, GSE16879, GSE36807) were integrated using ComBat batch correction. Unsupervised consensus clustering (k=3) identified three molecular subtypes: Metabolic-Absorptive (C1), Hypoxia-Metabolic (C2), and Inflammatory-EMT (C3). A LASSO-penalised multinomial logistic regression classifier was trained on the top 2,000 most variable genes with L1 regularisation swept across 13 values (C=0.003–0.500), selected at minimum gene count achieving ≥85% cross-validated accuracy. Performance was evaluated by 5-fold stratified cross-validation.Results A 10-gene classifier achieved 92.3% cross-validated accuracy (macro AUC=0.980) - with accuracy plateauing rapidly beyond this point: 172 genes achieved only 96.1%, a 3.8 percentage point gain for a 17- fold increase in gene count (IDDF2026-ABS-0259 Figure 1. Accuracy vs gene count curve). The confusion matrix confirmed robust three-class discrimination with minimal off-diagonal misclassification (IDDF2026-ABS-0259 Figure 2. Confusion matrix). The 10-gene panel comprised APOB, APOC3, MZB1, DAK, TGM2, KYNU, TIMP1, RAB8B, ARFGAP3, and DHRS11, with LASSO coefficients revealing distinct biological signatures per subtype (IDDF2026-ABS-0259 Figure 3. LASSO coefficient heatmap for all 10 classifier genes per subtype): C1 defined by positive APOB/APOC3 (lipid-absorptive function); C3 by positive KYNU, TIMP1, and MZB1 (tryptophan catabolism, fibrosis, plasma cell infiltration); C2 by negative RAB8B/ARFGAP3 distinguishing it from both phenotypes. Mean gene expression per subtype confirmed clean transcriptomic separation across all 10 genes (IDDF2026-ABS-0259 Figure 4. Mean z-score expression of all 10 classifier genes per molecular subtype). Per-class F1-scores were 0.96 (C1), 0.91 (C2), and 0.90 (C3) (IDDF2026-ABS-0259 Figure 6. Per class precision and recall and F1 score). Applied to all 616 samples including controls, only 20 were misclassified (96.8% accuracy), all clustering at biologically ambiguous subtype boundaries on UMAP projection (IDDF2026-ABS-0259 Figure 5. UMAP projection).Conclusions IBD molecular subtype can be predicted with 92.3% accuracy and macro AUC=0.980 from just 10 genes using a clinically interpretable LASSO classifier. The panel captures core biological axes - lipid metabolism, tryptophan catabolism, fibrosis, and plasma cell infiltration compatible with NanoString panels or RT-qPCR. Prospective validation could enable point-of-care molecular subtyping to guide biological therapy selection in IBD.Abstract IDDF2026-ABS-0259 Figure 1Abstract IDDF2026-ABS-0259 Figure 2Abstract IDDF2026-ABS-0259 Figure 3Abstract IDDF2026-ABS-0259 Figure 4Abstract IDDF2026-ABS-0259 Figure 5Abstract IDDF2026-ABS-0259 Figure 6",
  "authors": [
    {
      "affiliations": [
        "New York Medical College, United States"
      ],
      "name": "Jeril Lasington"
    },
    {
      "affiliations": [
        "Rutgers University, United States"
      ],
      "name": "Lawin Steve Mathew Lasington"
    },
    {
      "affiliations": [
        "Boston University, United States"
      ],
      "name": "Swamynathan Umamaheshwaran"
    }
  ],
  "title": "IDDF2026-ABS-0259 The biopsy report of the future: a minimal gene classifier for IBD molecular subtype prediction",
  "uid": "097fdcf1-20f0-5c6f-9e20-c77fdcd6a283"
}