# Multi-omics signatures of chronic inflammation across immune-related disease states

## TL;DR

This study uses multi-omics data and deep learning to uncover immune and metabolic patterns in chronic diseases, linking them to mortality and validating key immune mediators.

## Contribution

The novel contribution is integrating multi-omics deep learning with competing-risks modeling to decode immune-metabolic communication across chronic diseases.

## Key findings

- Omics-augmented deep learning models outperformed clinical-only models in predicting chronic disease risk.
- Machine-learning risk scores showed strong associations with cancer and non-cancer mortality.
- In vitro validation confirmed the role of key immune mediators like BAFF, GDF15, IL-15, and CD276.

## Abstract

Chronic inflammation and immune cell communication underpin a wide range of chronic diseases, yet population-scale maps integrating systemic inflammatory, metabolic and proteomic signals across multiple disease states are scarce.

Using UK Biobank, we classified participants into six baseline groups—healthy controls, cancer, autoimmune, infectious, metabolic diseases, and multiple comorbidities. We profiled clinical and hematological indices, NMR-based metabolites and Olink proteomics, and trained four multi-class deep learning models (clinical/inflammatory only; +NMR; +Olink; three-tower multi-omics) with 10-fold cross-validation. Out-of-fold predicted probabilities were combined in a stacking meta-model to derive machine-learning risk scores for “any chronic disease.” Shapley value analyses were used to identify key features reflecting systemic immune and metabolic communication. Cause-specific cumulative incidence and Fine–Gray competing-risks models evaluated associations between these risk scores and cancer-related and non-cancer mortality, adjusting for conventional risk factors. To provide biological validation of model-prioritized immune mediators (BAFF [TNFSF13B], GDF15, IL-15 and CD276), we performed in vitro stimulation of healthy-donor PBMCs by ELISA, flow cytometry, and qPCR.

We observed pronounced and pathway-specific heterogeneity of inflammatory markers, lipid-related metabolites and immune–inflammatory proteins across disease groups. Omics-augmented deep learning models outperformed the clinical-only model, and the stacking ensemble achieved the best accuracy, macro-F1 and multi-class AUC. Machine-learning–derived risk scores showed monotonic gradients in cancer and other-cause death and remained independently associated with several cause-specific outcomes. In vitro validation supported myeloid inflammatory inducibility of model-highlighted mediators.

By integrating multi-omics deep learning with competing-risks modelling, this study decodes population-level immune–metabolic communication patterns across chronic disease states, linking shared inflammatory and proteomic signatures to long-term mortality and providing a quantitative framework to support future, mechanism-focused and immunologically informed risk stratification.