This study employed a cross-sectional and prospective design aimed at exploring the baseline differences in clinical characteristics, inflammation/hematological markers, metabolomics (NMR), and proteomics (Olink) across immune-related disease states, as compared with healthy controls. Participants were classified into six mutually exclusive disease groups using ICD-10 codes recorded prior to baseline. We defined four single-disease categories (Cancer, autoimmune diseases [AD], infectious diseases [ID], and metabolic diseases [MD]). If a participant met the criteria for ≥2 of these categories, they were assigned to the ‘Multiple’ group; otherwise, they were assigned to the corresponding single-disease group. Participants who did not meet any of the four disease-category criteria were assigned to the Control group. A schematic flowchart summarizing the assignment logic is provided in Figure 1. Data for this study were sourced from the UK Biobank (29). Data usage and analysis were approved by the UK Biobank, with project ID 194370. The study protocol was ethically reviewed and approved by the UK Biobank ethics committee.

The study variables encompass baseline clinical characteristics, inflammation/hematological markers, metabolomics data, proteomics data, and survival outcomes. Initially, the clinical variables include age, sex, ethnicity, educational level, smoking status, alcohol consumption status, sleep duration, physical activity level, and body mass index (BMI). In addition, inflammation/hematological markers include white blood cell count (WBC), hematocrit (HTC), platelet count (PLT), red cell distribution width (RDW), hemoglobin concentration (Hb), C-reactive protein (CRP), among others. The metabolomics data were derived from NMR-based metabolic profiling, with 251 metabolites analyzed, all of which underwent standardization. The NMR metabolomics data were further categorized into 16 official pathways and 7 broader super pathways. This categorization aligns with the hierarchical structure of biological pathways, facilitating a comprehensive understanding of metabolic alterations in various disease states. Proteomics data were obtained using the Olink platform, encompassing over 2,900 proteins, which were also standardized for further analysis. The Olink proteomics measurements are directly mapped to corresponding gene data, providing a strong basis for subsequent gene functional and pathway analyses. The study outcomes include overall survival status, time to death, and cause of death, which will be used in subsequent survival analysis and mortality prediction models.

To evaluate the differences across the six disease groups, various statistical and visualization methods were employed. For continuous variables, the Kruskal-Wallis test was performed to identify significant differences among the groups. Pairwise comparisons were subsequently conducted using the Mann-Whitney U test to pinpoint specific group differences. For categorical variables, Pearson’s chi-squared test was applied. Additionally, multiple approaches were utilized to further explore the difference, including heatmaps for the expression patterns of clinical, inflammatory, and omics markers across disease groups. GO enrichment analysis was performed using bubble charts to reveal the biological significance of differential protein expression across groups. Volcano plots were used to visualize the magnitude and significance of metabolic and proteomic differences between disease states, while Sankey diagrams were generated to illustrate the relationships between disease groups, specific proteins, and their associated biological pathways. Furthermore, protein-protein interaction networks were constructed to examine the interactions of key proteins identified in the differential analysis. Lastly, k-means clustering (K = 6) was performed at the feature level to group metabolites or proteins with similar standardized profiles across participants (i.e., clustering rows of the feature × participant matrix). PCA and UMAP were used solely for low-dimensional visualization to illustrate the geometry of feature-level clustering in two dimensions, while disease-group distributions were summarized post hoc descriptively.

To assess the predictive value of baseline clinical characteristics, inflammatory/hematological indices, NMR-based metabolomics, and Olink proteomics for the six disease states, we primarily used multi-class deep learning models. Four architectures were specified. Model 1 was a single-tower fully connected network using only clinical and inflammation variables. Models 2 and 3 adopted a two-tower structure in which one tower encoded clinical/inflammatory variables and the other encoded either NMR metabolites (Model 2) or Olink proteins (Model 3). Model 4 was a three-tower network trained in the subset of participants with all three data layers available, combining clinical/inflammatory, NMR, and proteomic representations. Each tower consisted of dense layers with rectified linear unit activation, L2 regularization, batch normalization, and dropout, followed by a shared fully connected block and a 6-node softmax output layer for the six disease categories.

To mitigate class imbalance, inverse-frequency class weights were applied in the categorical cross-entropy loss, and class-specific decision thresholds were tuned on validation data. All models were trained with the Adam optimizer in 10-fold stratified cross-validation, and out-of-fold predicted probabilities were stored. A stacking ensemble was then built using these probabilities (four models × six classes) as meta-features in a multinomial logistic regression meta-learner, yielding final class probabilities and a continuous machine-learning–derived risk score for “any chronic disease” (1 − predicted probability of being in the control group). Model performance was summarized by accuracy, macro-average F1 score, multi-class ROC-AUC (Hand–Till method), and class-specific sensitivity and specificity. For interpretability, surrogate gradient-boosting models were fitted within each data layer and Shapley value–based importance metrics were used to rank clinical, metabolomic, and proteomic features.

To quantify cause-specific mortality risks across disease groups and machine-learning–derived risk strata, we used a Fine–Gray competing risks framework. Cause-specific cumulative incidence functions were first estimated for each type of death in the overall cohort and within each baseline disease group (30). Subsequently, Fine–Gray subdistribution hazard models were fitted separately for each cause of death, treating other causes as competing events. Conventional risk factors included age, sex, body mass index, smoking status, and Townsend deprivation index. In addition, the standardized ML-derived risk scores from the clinical/inflammatory model (ML-risk A), NMR model (ML-risk B), and Olink model (ML-risk C) were incorporated as continuous covariates to capture the aggregated contribution of multi-omics information. Subdistribution hazard ratios (SHR) and 95% confidence intervals were reported for each covariate, and model performance was further examined through assessment of discrimination, calibration, and decision-analytic net benefit for all-cause and cancer-specific mortality.

To biologically ground the model-identified immune-communication signals, we performed an in vitro validation focused on four representative mediators highlighted by the stacking model interpretability analyses: GDF15, BAFF (TNFSF13B), IL-15, and the myeloid surface checkpoint CD276. Peripheral blood mononuclear cells (PBMCs) from healthy donors were isolated using density-gradient centrifugation and cultured under standard conditions. Cells were stimulated with canonical inflammatory and polarization cues (Control, IL-4, LPS, IFN-γ, poly(I:C), and LPS+IFN-γ) to emulate distinct immune-activation states relevant to chronic inflammation and immune cell communication. Culture supernatants were collected for ELISA-based quantification of BAFF, GDF15, and IL-15, while matched cellular pellets were harvested for RNA extraction and qPCR quantification of TNFSF13B, GDF15, IL15, and CD276 transcripts. In parallel, surface CD276 expression was assessed by flow cytometry within CD45⁺CD14⁺ monocytes reporting both positivity and distributional shifts. Group comparisons for experimental readouts were conducted using appropriate parametric or non-parametric tests depending on distributional assumptions, with multiplicity controlled for post-hoc comparisons.

All data analyses were conducted in the R (version 4.5.2) programming environment, with statistical analyses and visualizations performed using a comprehensive suite of R packages, including tidyverse, data.table, caret, heatmap, pROC, survival, cmprsk, ggplot2, xgboost, keras3, shapviz, and tensorflow. Continuous variables were expressed as mean ± standard deviation or median (interquartile range), and categorical variables were expressed as counts (percentages). Group comparisons were made using the Kruskal-Wallis test (for continuous variables) and Pearson’s chi-squared test (for categorical variables). Multi-omics differential analysis was performed using linear regression models for metabolites and proteins, with multiple testing correction applied using the Benjamini-Hochberg method. A P-value < 0.05 was considered statistically significant.

The baseline characteristics of participants across six disease groups were examined, with key demographic and clinical measures displayed in Table 1, Figure 2A. The demographic distribution showed notable differences across groups in terms of age, sex, and other health indicators. The age at recruitment varied significantly across groups, with Cancer, MD, and Multiple groups having a higher mean age of 60 ± 7 years, compared to the Control and ID groups, which had a mean age of 56 ± 8 years (p < 0.001). Regarding sex, Cancer and AD groups had a higher proportion of females (64% and 61% respectively), while the MD group had a higher proportion of males (61%). For the Townsend deprivation index, the MD group showed a less negative mean of -0.25 ± 3.43, compared with the Control group of -1.33 ± 3.07 (p < 0.001), suggesting relatively higher deprivation in the MD group. The smoking status indicated a high proportion of current smokers in the Cancer and Multiple disease groups (80% and 79% respectively) compared to the Control group (76%), and this difference was statistically significant (p < 0.001). Regarding alcohol consumption, the Control group had the lowest proportion of current drinkers (3.4%), while the other groups, especially MD and Multiple, had higher proportions of current drinkers (9.6% and 9.4% respectively), with significant differences observed (p < 0.001). Physical activity levels also differed significantly between groups, with the MD group showing the highest percentage of individuals with low physical activity (29%), while the Control group had the highest percentage of individuals with high physical activity (31%) (p < 0.001). The Body Mass Index (BMI) was highest in the MD group (31.3 ± 6.1), reflecting higher obesity levels compared to the other groups, with the Control group having a lower BMI (27.3 ± 4.7) (p < 0.001).

To analyze the differences in inflammation and hematological indicators across six disease groups, heatmap was firstly used in comparing the expression levels of various inflammatory markers between the Control group and the five disease groups (Figure 2B). The Cancer group exhibited significantly higher levels of CRP and WBC, indicating a heightened inflammatory state, while the MD and Multiple groups demonstrated lower levels of these markers. Elevated levels of certain inflammatory markers, particularly CRP, were also observed in the AD group. Volcano plots were used to highlight the differential regulation of these markers, emphasizing the magnitude and statistical significance of the differences between each disease group and the Control group (Figure 2C). CRP and WBC were consistently upregulated in the Cancer and AD groups, while Hb and MCV showed significant downregulation in the MD and Multiple groups, suggesting possible anemia-related effects in these diseases. Raincloud and boxplots were used to visualize the distribution of 15 inflammatory and hematological markers across the six groups (Figure 2D). These plots revealed significant differences in CRP and WBC between the groups, especially between Cancer and Control. The Cancer and AD groups exhibited the highest variability in several markers, while the Control group showed minimal fluctuation, reinforcing its role as a baseline comparator.

To investigate the metabolic differences across various disease groups, NMR metabolomics techniques were employed, and multiple analytical methods were utilized. The heatmap depicts the distribution of 251 metabolites across 16 official pathways in the UKB dataset (Figure 3A). It is observed that the Cancer group exhibits significantly higher metabolite levels in pathways such as Fatty Acid and Triglycerides, while the Multiple group shows relatively lower levels in these pathways. The AD group demonstrates substantial abnormalities in lipid-related pathways, particularly in Phospholipids and Total Lipids. To further confirm these differences, enrichment analysis results are presented through a bubble plot (Figure 3B), highlighting the enrichment of 16 official pathways and 7 super-pathways in different disease groups. Notably, the Cancer and AD groups exhibit significant enrichment in Fatty Acid and Triglycerides pathways. We first applied k-means clustering (K = 6) in the metabolite feature space using the standardized metabolite × participant matrix to group metabolites with similar abundance patterns across participants. PCA and UMAP were then used only to visualize and illustrate the feature-level structure of these clusters in two dimensions (Figure 3C). Finally, we descriptively summarized the cluster composition by baseline disease group (Figure 3D) to provide a post hoc overview of how disease-group–associated metabolite alterations distribute across the six clusters. The Cancer and Multiple groups predominantly occupy Cluster 1 and Cluster 6, while the AD and ID groups are more concentrated in Cluster 2 and Cluster 4.

To explore the differences between disease groups at the protein level, nearly 3,000 Olink proteomics data measured by the UK Biobank were analyzed. Heatmap was used to display the expression differences of the top 80 differential proteins (FDR < 0.05) between the five disease groups and the Control group (Figure 4A). In the Cancer group, several proteins involved in immune response, cell adhesion, and inflammation were significantly upregulated, indicating a heightened inflammatory state. In contrast, these markers were relatively lower in the MD and Multiple disease groups, suggesting a weaker immune response in these groups. The AD group showed elevated levels of specific inflammatory markers. The GO enrichment analysis bubble plot (Figure 4B) displayed the functional enrichment of genes associated with these differential proteins. Notably, immune response-related genes were significantly enriched in the Cancer group compared to other groups, while the MD and ID groups showed lower enrichment in immune-related functions. Volcano plots illustrated the differential regulation of proteins between each disease group and the Control group (Figure 4C), where the Cancer and AD groups consistently upregulated many differential proteins, while the Multiple group showed relatively smaller differences. The Sankey diagram (Figure 4D) showed the relationships between disease groups, differential proteins, and associated pathways, emphasizing the role of immune response and cell adhesion pathways in the Cancer group, particularly associated with proteins such as TNF and CRP. The protein-protein interaction network (Figure 4E) revealed strong interactions between proteins related to diseases such as AD and Cancer, particularly in immune and inflammatory pathways, where these proteins formed highly concentrated nodes in the network. Clustering analysis using K-means (k=6) further grouped the differential proteins, and the heatmap (Figure 4F) revealed the expression patterns of these proteins across the disease groups. PCA and UMAP (Figure 4G) were further used to visualize and illustrate the feature-level clustering structure between clustering and disease groups. Finally, the stacked distribution percentage plot (Figure 4H) displayed the distribution of each disease group across different protein clusters, reinforcing the higher proportion of Cancer and AD groups in specific protein clusters.

To comprehensively evaluate disease-state classification, we constructed five deep learning–based models: a clinical plus inflammatory model (Model 1), a clinical+inflammation+NMR model (Model 2), a clinical+inflammation+Olink proteomics model (Model 3), a three-tower multi-omics network (Model 4), and a stacking meta-model (Model 5) that takes the out-of-fold predicted probabilities from the four base models as inputs. Model comparison was based on 10-fold cross-validated out-of-fold predictions (Figure 5). The performance landscape showed that Model 1 provided a basic level of discrimination but was inferior to the omics-augmented models in terms of accuracy, macro-average F1, and multi-class AUC. Adding NMR or Olink data (Models 2 and 3) substantially improved overall performance, with the NMR-based model yielding the highest macro-AUC (0.723) among single-block models. The three-tower multi-omics model (Model 4) showed slightly less stable performance, likely due to reduced sample size after requiring complete multi-omics data. In contrast, the stacking model (Model 5) achieved the best accuracy (0.848), macro-F1 (0.240), and mAUC (0.720) across all configurations, indicating that integrating heterogeneous information at the meta-learning level effectively combines clinical, metabolic, and proteomic signals (Figures 5A, D).

Class-specific ROC curves for the stacking model demonstrated overall improvement over the baseline model for MD and Multiple disease groups, with stable AUCs for AD and Cancer and high specificity for the Control group (Figure 5B). The confusion matrix further showed that most individuals were correctly classified, such as AD versus ID and MD versus Multiple (Figure 5E). Precision–Recall curves and calibration analysis indicated that the stacking model maintained favourable precision and recall in the medium-to-high risk range, with predicted probabilities closely aligned with observed risks and only mild underestimation at the extreme high-risk tail (Figures 5F, G). Within the second-layer multinomial meta-model, the contributions of base-model outputs were heterogeneous. Meta-feature importance revealed that the NMR model probabilities for Multiple and Control, together with the three-tower model outputs for MD and Cancer, carried the largest weights, suggesting that these signals act as key anchors in the final decision (Figure 5C). Shapley-value–based interpretation of the stacking-derived risk score identified the top 50 contributing features across clinical/inflammatory markers, NMR metabolites, and Olink proteins. Age, BMI, CRP, and several blood cell indices were among the leading clinical and inflammatory contributors; lipoprotein and lipid-related NMR metabolites (multiple HDL, LDL, and triglyceride fractions) played major roles in differentiating disease patterns; and proteins such as GDF15, CD276, and TNFSF13B, involved in immune and inflammatory pathways, ranked highly among proteomic predictors (Figure 5H). Together, these findings indicate that multi-omics deep learning combined with stacking not only improves multi-class classification performance but also delineates the layered contributions of diverse biomarkers, providing a quantitative basis for subsequent mechanistic analyses and risk stratification.

To characterize long-term cause-specific mortality, we applied Fine–Gray competing risks models. In the overall cohort, the cumulative incidence of cancer-related death increased rapidly early during follow-up and remained the highest across the entire period, whereas AD-, ID-, and MD-related deaths showed much lower cumulative risks (Figure 6A). At fixed time points of 60 and 120 months, cause-specific cumulative incidences further confirmed that Cancer and Multiple groups had the greatest burden of cancer mortality, whereas the MD group showed progressively increased risks of metabolic/cardiovascular and other-cause deaths (Figure 6B). In Fine–Gray models including only conventional risk factors, older age, former or current smoking and higher deprivation were associated with increased subdistribution hazard for several causes of death, while higher BMI showed neutral or mildly protective effects for some outcomes (Figures 6C, D). We then incorporated the standardized multi-disease risk scores derived from deep-learning Models 1–3 as composite covariates. Outcome composition across quintiles of each ML risk score showed that, from Q1 to Q5, the proportions of cancer and other-cause deaths increased monotonically, whereas the proportion of censored individuals decreased, indicating that ML-derived scores effectively stratified long-term mortality risk (Figure 6G). Cause-specific distributions of the scores further demonstrated that individuals who died from cancer consistently had the highest ML risk across all three models, followed by AD and MD deaths, whereas censored and non-fatal events clustered in the low-risk range (Figure 6H). In extended Fine–Gray models including the ML risk scores, all three models showed robust positive associations with multiple causes of death, with subdistribution hazard ratios largely between 1.3 and 1.8 and the strongest effects for cancer and other-cause mortality (Figures 6E, F). These findings suggest that deep-learning–integrated multi-omics signatures provide additional prognostic value for long-term survival and the distribution of causes of death beyond traditional risk factors.

To biologically validate key mediators highlighted by the multi-omics stacking framework, we stimulated PBMC cultures with inflammatory cues and quantified BAFF, GDF15, and IL-15 at the protein level and TNFSF13B, GDF15, IL15, and CD276 at the transcript level (Figures 7A, C). Across conditions, LPS-based stimulation produced the most consistent induction of all three secreted mediators, with the combined LPS+IFN-γ condition yielding the highest concentrations for BAFF, GDF15, and IL-15 (Figure 7A). Poly(I:C) and IFN-γ alone showed intermediate upregulation patterns, whereas IL-4 exerted comparatively modest effects. Concordantly, qPCR measurements demonstrated significant increases of TNFSF13B, GDF15, and IL15 transcripts under LPS and LPS+IFN-γ stimulation, mirroring the protein-level directionality (Figure 7C). For the surface immune checkpoint CD276, flow cytometry within CD45⁺CD14⁺ monocytes revealed a clear rightward shift of CD276 signal following LPS exposure relative to control, indicating increased surface expression at the myeloid compartment (Figure 7B). Consistent with these flow findings, CD276 transcript levels were also elevated in LPS-driven conditions, with the strongest induction observed in the LPS+IFN-γ group (Figure 7C).