Network-based multi-omics integration reveals metabolic at-risk profile within treated HIV-infection

Abstract

Multiomics technologies improve the biological understanding of health status in people living with HIV on antiretroviral therapy (PWH). Still, a systematic and in-depth characterization of metabolic risk profile during successful long-term treatment is lacking. Here, we used multi-omics (plasma lipidomic, metabolomic, and fecal 16 S microbiome) data-driven stratification and characterization to identify the metabolic at-risk profile within PWH. Through network analysis and similarity network fusion (SNF), we identified three groups of PWH (SNF-1-3): healthy (HC)-like (SNF-1), mild at-risk (SNF-3), and severe at-risk (SNF-2). The PWH in the SNF-2 (45%) had a severe at-risk metabolic profile with increased visceral adipose tissue, BMI, higher incidence of metabolic syndrome (MetS), and increased di- and triglycerides despite having higher CD4⁺ T-cell counts than the other two clusters. However, the HC-like and the severe at-risk group had a similar metabolic profile differing from HIV-negative controls (HNC), with dysregulation of amino acid metabolism. At the microbiome profile, the HC-like group had a lower α-diversity, a lower proportion of men having sex with men (MSM) and was enriched in Bacteroides. In contrast, in at-risk groups, there was an increase in Prevotella, with a high proportion of MSM, which could potentially lead to higher systemic inflammation and increased cardiometabolic risk profile. The multi-omics integrative analysis also revealed a complex microbial interplay of the microbiome-associated metabolites in PWH. Those severely at-risk clusters may benefit from personalized medicine and lifestyle intervention to improve their dysregulated metabolic traits, aiming to achieve healthier aging.

Keywords: HIV; aging; computational biology; infectious disease; metabolomics; microbiology; microbiome; systems biology; viruses.

Figure 1.. Similarity network fusion-based PWH stratification using lipidomics, metabolomics, and microbiome integration.

(A) Scatter plot showing the maximization of Eigen gap and the minimization of rotation cost for optimizing the number of clusters. (B) Concordance matrix between the combined network (SNF) and each omics network based on NMI calculation (0=no mutual information, 1=perfect correlation). (C) SNF-combined similarity network colored by clusters (SNF-1/HC-like=blue, SNF-2/severe at-risk=yellow, SNF-3/mild at-risk=grey) obtained after spectral clustering. Edges' color indicates the strength of the similarity (black = strong, grey = weak). (D) PCA plot of samples based on fused network. Samples are colored by condition.

Figure 1—figure supplement 1.. PCA plot of samples after prior standardization based on (a) Lipidomics (b) Metabolomics (c) Microbiome.

Variance proportions are written on each component axis. Samples are colored by condition (Ctrl = green, SNF-1=blue, SNF-2=yellow, SNF-3=grey).

Figure 2.. Lipidomics and metabolomics, characterization of the PWH clusters.

(A) Boxplots of DAG from untargeted lipid classes separated by groups. Significant stars are displayed for each comparison with *FDR <0.05, **FDR <0.01, ***FDR <0.001 (limma). (B) Boxplots of TAG from untargeted lipid classes separated by groups. (C) PCA plot of samples after prior standardization based on significant metabolites between at least one pairwise comparison (limma, FDR <0.05). Variance proportions are written on each component axis. Samples are colored by condition. (D) Circular heatmap of the top 159 metabolites (FDR <0.005). Metabolites are represented as slices and labeled around the plot. LogFold Change from significant metabolites between groups is displayed in the first six outer layers. The 7th to 9th layers represents the coefficient of correlation between metabolites and BMI, metabolites and age (Spearman, p <0.1, absolute R>0.15) and the p-value from significant associations between metabolites and gender (χ², p<0.1). The inner layer represents the pathway of each metabolite. (E) PCA plot based on metabolites differing clusters adjusted for transmission mode and CD4 count.

Figure 2—figure supplement 1.. Boxplots of untargeted lipid classes separated by groups.

Color is based on groups (Ctrl = green, SNF-1=blue, SNF-2=yellow, SNF-3=grey). p Values are displayed for each comparison (Mann Withney U Test).

Figure 3.. Transmission mode drove cluster differences in microbiome data.

(A) Boxplots of alpha diversity indices (Observed, ACE, Chao1, Fisher) separated by HIV cluster. Significant stars are shown for each comparison (Mann-Whitney U test). (B) Non-metric multidimensional scaling (NMDS) plot of Bray-Curtis distances. Samples are colored by clusters. Boxplots based on NMDS1 and NMDS2 are represented. (C) Barplot represents the relative abundance of bacteria at the family level for each patient. Patient information is displayed above the barplot, including cluster, metabolic syndrome (MetS: yes/no), hypertension (yes/no), transmission mode, and gender. (D) Barplot showing the top microbial families by representing their coefficient from PERMANOVA between SNF-1 and SNF-2. (E) Barplot showing the top microbial families between SNF-1 and SNF-3. (F) LEfSe cladogram representing cluster-specific microbial communities to HC-like and to at-risk groups (SNF-2/SNF-3). Top families from PERMANOVA are labeled. (G) Boxplot of relative abundance at family level for Bacteroides (top) and Prevotella (bottom). Significant stars are shown for significant comparisons (Mann-Whitney U test).

Figure 3—figure supplement 1.. Boxplots of alpha diversity indices (se.chao1,Simpson, Shannon, se.ACE, InvSimpson) separated by HIV-cluster.

Color is based on groups (Ctrl = green, SNF-1=blue, SNF-2=yellow, SNF-3=grey).

Figure 3—figure supplement 2.. Non-metric multidimensional scaling (NMDS) plot of Bray-Curtis distances.

Samples are colored by (A) Transmission mode (B) Central obesity (C) Metabolic Syndrome (D) Hypertension.

Figure 4.. Factor analysis highlights the essential features for cluster separation and potential microbiome-derived metabolites importance (A) Barplot of total variance explained by MOFA model per view.

(B) Variance decomposition plot. The percentage of variance is explained by each factor for each view. (C) External covariate association with factors plot. Association is represented with log10 adjusted p-values from Pearson correlation. (D) Heatmap representing levels of microbial communities, metabolites, and lipids with the higher absolute weight in MOFA factors associated with cluster (F1, F2, F3, F5, F8). Samples are labeled according to the study groups. Data were Z-score transformed. The type of data (lipid, metabolite, microbe) is displayed on the right. (E) Top 20 features with higher absolute weight in MOFA factors associated with cluster (F1, F2, F3, F5, F8) from lipidome, metabolome, and microbiome. Microbiome-derived metabolites and bacterial phylum of interest are colored in blue and red, respectively. (F) MOFA features differing clusters and interactions extracted from the three-layers consensus co-expression network. Microbiome-derived metabolites are labeled.

Figure 4—figure supplement 1.. Correlation matrix of MOFA factors.

Size and transparency are proportional to the absolute coefficient of correlation. Color is displayed as a gradient-based coefficient of correlation from –1 (red) to 1 (blue).

Figure 4—figure supplement 2.. Cytoscape consensus co-expression network.

Color and label are based on communities.

Figure 5.. Microbiome-associated metabolites are affected in HIV clusters (A) Heatmap representing abundances of microbiome-derived metabolites differing in at least one comparison.

Data were Z-score transformed. Significant logFC (limma, FDR <0.05) of pairwise comparisons between conditions, groups, and under groups of microbiome-derived metabolites are displayed on the right. (B) Cytoscape network showing significant positive and negative associations between clinical parameters and microbiome-derived metabolites (univariate linear regression, FDR <0.05). Clinical parameters are colored based on categories. (C) Co-expression network of metabolomics data in PWH. Metabolites are grouped by communities, and microbiome-derived metabolites are labeled and colored based on the subgroup. (D) The subset of microbiome-derived metabolites from the co-expression network. Non-significant metabolites in all comparisons are displayed with transparency. Significant microbiome-derived metabolites between at least two conditions are labeled.

Author response image 1.. Venn diagram of metabolites significantly different between clusters in noncorrected limma model (left) and model corrected for transmission mode and CD4 count (right).

Author response image 2.. Venn diagram of lipids significantly different between clusters in noncorrected limma model (left) and model corrected for transmission mode and CD4 count (right).