Evidence of a genetically driven metabolomic signature in actively inflamed Crohn's disease

Abstract

Crohn's disease (CD) is characterised by chronic inflammation. We aimed to identify a relationship between plasma inflammatory metabolomic signature and genomic data in CD using blood plasma metabolic profiles. Proton NMR spectroscopy were achieved for 228 paediatric CD patients. Regression (OPLS) modelling and machine learning (ML) approaches were independently applied to establish the metabolic inflammatory signature, which was correlated against gene-level pathogenicity scores generated for all patients and functional enrichment was analysed. OPLS modelling of metabolomic spectra from unfasted patients revealed distinctive shifts in plasma metabolites corresponding to regions of the spectrum assigned to N-acetyl glycoprotein, glycerol and phenylalanine that were highly correlated (R² = 0.62) with C-reactive protein levels. The same metabolomic signature was independently identified using ML to predict patient inflammation status. Correlation of the individual peaks comprising this metabolomic signature of inflammation with pathogenic burden across 15,854 unselected genes identified significant enrichment for genes functioning within 'intrinsic component of membrane' (p = 0.003) and 'inflammatory bowel disease (IBD)' (p = 0.003). The seven genes contributing IBD enrichment are critical regulators of pro-inflammatory signaling. Overall, a metabolomic signature of inflammation can be detected from blood plasma in CD. This signal is correlated with pathogenic mutation in pro-inflammatory immune response genes.

Figure 1

NMR and genomic data integration. Phase (I) NMR spectra and patient CRP data were input to the RF model using (a) RFECV and (b) cross-validated methods to select spectral regions discriminating non/inflamed patients. Phase (II) Informative data-points were clustered and peaks reduced to a single eigenvector. Phase (III) Eigenvectors for each peak were individually correlated against all genes and tested for enrichment. Created with BioRender.com.

Figure 2

CRP prediction and spectra deconvolution. (A) OPLS scores plot. Each point represents one patient spectrum, colour-coded according to CRP levels. Strong correlation between T and Tcv indicates a robust model. (B) Loadings plot; colour-scale indicates the correlation magnitude of metabolites with the model scores (r²).

Figure 3

Machine learning classification of patients using NMR data. (A) Most informative regions selected by RF model to discriminate patient inflammation status. (B) PCA of patient’s spectra using 258 most informative NMR datapoints. (C) Distribution of the selected most informative datapoints by their shift δ (ppm) and importance. Green and red dashed lines indicate the start and end of a peak.