Proteogenomic Classification of IBD Subtypes
Summary¶
About 15% of inflammatory bowel disease (IBD) cases are difficult to classify as Crohn's disease (CD) or ulcerative colitis (UC) because the two share overlapping symptoms, endoscopic features, and histopathology, which complicates management decisions such as colectomy. Khunsriraksakul, Zhang, et al. (2025) built a 12,194 CD vs. 12,366 UC GWAS with CC-GWAS, layered transcriptome-wide and proteome-wide association study (TWAS/PWAS) signals from a large cross-platform pQTL atlas on top of it, and showed that combining a polygenic risk score (PRS) with triangulated plasma protein levels meaningfully improves discrimination between the two subtypes in UK Biobank.
Genetic and Molecular Signal¶
- The direct CD-vs-UC GWAS found 15 divergent genetic loci, including well-established IBD genes NOD2, HNF4A, PTGER3, and PTGER4.
- TWAS (using PUMICE) identified 27 additional divergent loci, most notably NFKB1.
- PWAS uncovered 32 further divergent loci, including CD8A, IL34, PLA2G10, and TFF3.
- Across GWAS, TWAS, and PWAS layers, NF-κB signaling repeatedly emerged as a convergent hub differentiating CD from UC: NOD2 mutations inappropriately activate NF-κB in monocytes, PTGER4 attenuates macrophage activation via the NF-κB pathway, and IL34, PLA2G10, and TFF3 each modulate NF-κB-linked inflammatory signaling in the intestinal epithelium.
- Triangulating TWAS and PWAS signals for predictive differential regulators nominated eight proteins — AGER, APOM, ATP6V1G2, DDR1, HSPA1A, LTBR, RNASET2, and TNXB — several of which (AGER/RAGE, DDR1, HSPA1A, RNASET2) have independent prior evidence in IBD pathophysiology.
Predictive Modeling¶
Using 565 UK Biobank IBD patients (193 CD, 372 UC) split into training (87 CD/171 UC) and testing (106 CD/201 UC) sets by timing of protein measurement relative to diagnosis, several nested models were compared by AUC in the test set:
| Model | Inputs | Median AUC | Top vs. bottom quintile OR |
|---|---|---|---|
| BASE | Sex, year of birth, 20 PCs | 0.52 | 1.23 [0.59–2.58] |
| PRS | BASE + polygenic risk score | 0.60 | 2.82 [1.35–6.06] |
| SPARC | 13 proteins from the SPARC IBD cohort | 0.56 | 2.08 [1.02–4.34] |
| PRS–SPARC | PRS + SPARC proteins | 0.61 | 2.75 [1.33–5.82] |
| OMICS | The 8 triangulated TWAS/PWAS proteins | 0.56 | 1.99 [0.96–4.20] |
| PRS–OMICS | PRS + OMICS proteins | 0.63 | 2.82 [1.36–6.06] |
| PRS–SPARC–OMICS | All predictors combined | 0.64 | 3.35 [1.62–7.14] |
The combined PRS–SPARC–OMICS model performed best. The authors interpret the germline-PRS-derived and pQTL-triangulated OMICS proteins as capturing stable, genetically-influenced signal present even in healthy individuals, complementary to the SPARC proteins, which likely reflect transient, active-disease-state biology — motivating the combination of both marker classes for earlier subtype differentiation.
Data Source¶
The TWAS and PWAS signals used for triangulation were derived from the cross-platform pQTL atlas across Olink and SomaScan platforms built in the same study, meta-analyzing over 90,000 individuals.
Limitations¶
- All pQTL and GWAS data are restricted to European-ancestry individuals, limiting generalizability.
- The number of IBD cases with paired genomic and proteomic data available for the classification analysis was small (565 total).
- The comparison is cross-sectional; the source is a preprint, not yet peer-reviewed.
See Also¶
- Plasma Proteogenomics — the broader field of protein-genetics-phenotype integration, including the pQTL atlas this analysis triangulates against.
- Polygenic Risk Scores — the PRS component of the combined predictive model.