Immune and inflammatory resolution pathways through multi-omics using the AI-based Network Integration

Multi-omics integration faces several challenges, including heterogeneity across platforms, high dimensionality, and reduced effective sample size when all omics types are not available for every participant. We propose the AI-based Network Integration approach, which models connectivity within each omics layer as shaped by the broader biological system, including structured influences from other omics layers. We conducted a cross-sectional multi-omics study of genetic, glycomic, and lipidomic profiles related to immune and inflammatory pathways. We integrated N-glycans of immunoglobulin G (glycans) and specialized pro-resolving lipid mediators (SPMs) to advance the understanding of immune regulation and inflammation resolution, analyzing 24 glycans and 14 SPMs. The sample size decreased from N = 456 (with glycomics) to 368 (with both lipidomics and glycomics), and further to 266 (with genetic data, lipidomics, and glycomics). Pairwise association analysis revealed 10 statistically significant glycan-SPM associations (FDR < 0.05). Using the AI-based Network Integration, we found that these associations disappeared when conditioned on two bridging molecules, 5-Hydroxyeicosapentaenoic acid (5-HEPE) SPM and a glycan peak (GP 21) corresponding to disialylation, suggesting they may be mediated or confounded by a subset of these two molecules. Principal component analysis was applied to the previously identified associated genetic variants within each omics layer to derive polygenic factors, composite components summarizing shared genetic variation across variants. These polygenic factors were used in the Granularity Directed Acyclic Graph (G-DAG) algorithm to identify valid instrumental variables for the identification of molecular causal networks based on the principles of Mendelian randomization. This approach improved robustness and also identified polygenic molecules. Both bridging molecules, 5-HEPE and the GP21 glycan, were under polygenic control. 5-HEPE acted as a receiver (influenced by other SPMs), while GP21 acted as a broadcaster (influencing other glycans). The results suggest that variations in SPM lipidomics are linked to changes in glycosylation. These findings highlight a potential biological link between lipid mediators and glycan-mediated immune regulation and inflammation resolution. The AI-based Network Integration approach turns noisy associations into a small, interpretable, and biologically meaningful connectivity that cannot be explained by any other component in the study.