Extreme phenotypes define epigenetic and metabolic signatures in cardiometabolic syndrome.

Providing a molecular characterisation of cardiometabolic syndrome (CMS) could improve our understanding of its pathogenesis and pathophysiology, and provide a step toward the development of better treatments. To this end, we performed a deep phenotyping analysis of 185 blood donors, 10 obese, and 10 lipodystrophy patients. We analysed transcriptomes and epigenomes of monocytes, neutrophils, macrophages and platelets. Additionally, plasma metabolites including lipids and biochemistry measurements were quantified. Multi-omics integration of this data allowed us to identify combinations of features related to patient status and to order the donor population according to their molecular similarity to patients. We also performed differential analyses on epigenomic, transcriptomic and plasma proteomic data collected from obese individuals before and six months after bariatric surgery. These analyses revealed a pattern of abnormal activation of immune cells in obese individuals and lipodystrophy patients, which was partially reverted six months after bariatric surgery.

Introduction with extreme phenotypes (obese referred for bariatric surgery and lipodystrophy patients) to be discriminated from lean, metabolically healthy individuals. Predictions made using the disease signatures were generally in good agreement with predictions made on the basis of plasma biochemistry markers only, although there was stark disagreement for some individuals. The approach was validated by training our classifier using data from one patient group, and making predictions for the other group. Moreover, we found that features identified in the lipidomic layer were associated with known risk factors in a large independent cohort.
As a result of our integrative approach, we also obtained the molecular signatures at transcriptional, chromatin (H3K27ac), and DNA methylation levels. These helped to shed light on the molecular events that determine the changes in the functional phenotypes of platelet, neutrophils, monocytes and macrophages.
Our approach was deployed to characterise the same individuals six months after bariatric surgery which is an effective therapy to reduce weight and improve wellbeing for morbidly obese individuals 54,55 . Major benefits are a substantial weight loss in the first year following the intervention, with maximum loss reached around 6-8 months after surgery 56 , and improvement of several clinical parameters (HbA1c, glucose and cholesterol levels, insulin resistance, and modulations of gut hormones 57 ). It has also been shown that bariatric surgery modifies metabolite abundance in the first year following surgery 58 . Moreover, bariatric surgery has been shown to affect DNA methylation patterns and gene expression 59 . Additionally, to reduce stomach size, bariatric surgery also reduces inflammatory markers 60,61 . We investigated and determined the changes that occur in the neutrophils, monocytes, macrophages and platelets. While the gene expression showed dramatic changes, especially in neutrophils and platelets more modest differences were observed in regulatory elements and almost none in methylation profiles. Plasma proteome analysis allowed us to have some insight on the changes in other tissues and organs whilst neutrophils and platelets cell function assays results indicated reduced ability to adhere, the key initial step for their activation.

Metabolic signatures in obese individuals and lipodystrophy patients.
To determine the metabolic health of the different groups, we collected anthropometric characteristics: age and body weight (BW) and performed plasma biochemistry assays for the following: leptin, adiponectin, insulin, free fatty acids (FFA), glucose (GLC), triglycerides (TG), total cholesterol (TC), high density lipoprotein cholesterol (HDL-C), low-density lipoprotein (LDL-C)), activity of alanine and aspartate amino-transferases (ALT and AST, respectively) and high-sensitivity C-reactive Protein (hsCRP). Additionally, we computed leptin-adiponectin ratio (LAR), Homeostatic Model Assessment for Insulin Resistance (HOMA-IR) and adipose tissue insulin resistance (AT-IR) indexes ( Table 1; Table S1 ). To investigate the combination of these parameters, we performed a principal component analysis (PCA) to reduce a large set of variables into a smaller set that still contains most of the information of the larger set. PCA showed that obese and lipodystrophy patients and WP10 individuals are distributed over distinct but partially spanning dimensions ( Fig. 1A ). The first two components (PC1 and PC2) distinguished well the different groups; in particular obese individuals were separated from WP10 participants along PC1 and lipodystrophy patients were separated from WP10 participants along PC2. Analysis of loading and contribution for each parameter analysed ( Fig. 1B ) indicated that weight contributed to both global variance (59.2%, Fig. 1B color scale) and to PC1 ( Fig. 1B, arrow length and direction), as expected, followed by AGE (12.6%), LAR (7.4%) and HOMA-IR score (5.8%). ALT and AST explained most of the variance along PC2 (45.6% and 29.8% respectively), followed by HOMA-IR (8.6%) and AT-IR (4.2%) indexes, AGE (1.6%) and GLC (2.7%).
Obesity has been shown to have a profound impact on plasma metabolites 62,63 . Previous studies have also shown that metabolite patterns were affected in patients suffering from lipodystrophy 64,65 . To determine which metabolites were present, all plasma samples were analysed on the Metabolon platform (Material and Methods).
This led to the identification and quantification of 988 metabolites. To identify and characterise groups of metabolites whose levels were correlated across samples, we performed a weighted co-expression network consensus analysis (WGCNA) 66 . Using the entire dataset, we identified 16 consensus modules (M1 to M16; Table S2 ). We did not identify a significant enrichment for a specific biological pathway for any of author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint these modules. Next, we determined if there were any associations between these modules and the results of the plasma biochemistry assays. Of the 208 tested associations, 11 showed significant correlation with the results of the plasma biochemistry assays in the combined patient group (FDR adjusted Fisher p values < 5%; 6 were positively correlated, 5 were negatively correlated; Fig. 1C ) and 51 showed correlation with the results of the plasma biochemistry assays in the WP10 cohort (FDR adjusted Fisher p values < 5%; 27 were positively correlated, 24 were negatively correlated; Fig. 1D ). Of these, 3 were shared: M6 was positively correlated with module TG both in patients and donors. whereas module M2 was negatively correlated with TG and positively correlated with HDL-C in patients and reversely in donors. This suggests that, depending on the individual status, M2 metabolites participate differently to TG and HDL-C concentrations.

Transcriptional and epigenetic signatures in obese and lipodystrophy patients for 3 innate immune cell types and platelets.
Next, we sought to determine if the metabolic changes observed in plasma were accompanied by changes in the transcriptional and epigenetic signatures of innate immune cells (neutrophils, monocytes macrophages) and platelets ( Fig. 2A ). We characterised the transcriptome by ribo depleted RNA-sequencing (RNA-seq), the genome-wide distribution of histone 3 lysine 27 acetylation (H3K27ac), a marker of active promoters and transcriptional enhancers, by chromatin immunoprecipitation (CHIP-seq) and DNA methylation by reduced representation bisulfite sequencing author/funder. All rights reserved. No reuse allowed without permission.
The comparison between lipodystrophy patients and lean individuals ( Fig. S2C) identified 125 DEG in macrophages (115 up and 10 down regulated in lipodystrophy,  S2C and Table S12 ), 1 of these was located on gene promoter, 12 (70)% were intergenic and 4 (23%) in introns and one in the latter (down regulated; intergenic). No DAcR were identified in neutrophils. DNA methylation analysis found 20 differentially methylated CpG islands in macrophages ( Table S15 ); 60 in monocytes ( Table S16 ) and 44 in neutrophils ( Table S17 ).
The comparison between obese individual and lipodystrophy patients ( Fig. S2B) identified 4 DEG in macrophages (2 up and 2 down regulated in lipodystrophy, Table   S8 ), 40 in monocytes (22 up and 18 down regulated, Table S9 ), 1 upregulated gene each in neutrophils ( Table S10 ) and in platelets ( Table S11 ). We observed 1,764 DAcR in macrophages (all hyper-acetylated in lipodystrophy; Fig. S2B and Table  S12 ), 22% of these were located on gene promoters or gene body, 33% were intergenic and 45% in introns. We also observed 1,766 DAcR in monocytes (1,098 up and 668 down; Fig. S2B and Table S13 ), 38% of these were located on gene promoters or gene body, 13% were intergenic and 49% in introns. Of these, only 50 overlapped with those observed in macrophages.
To gain insight into the changes observed in gene expression, we performed functional annotation by gene ontology (GO) terms enrichment analysis. In the comparison between obese and lean individuals ( Fig. 2C ), we found an enrichment for GO terms related to phagocytic and degranulation activities, as well as markers of cardiometabolic risk in macrophages ( Table S18) . In monocytes, down-regulated DEG were enriched in GO terms related to programmed cell death and ion homeostasis; whereas up-regulated DEG were enriched for GO terms related to inflammatory response ( Table S19 ). In neutrophils, down-regulated DEG showed enrichment for GO terms related to vesicle trafficking, protein sulfation and regulation of lipid droplet morphology and activity ( Table S20 ). In platelets, down-regulated DEG showed enrichment for GO terms related to cholesterol biosynthesis and C-type lectin receptor signaling pathway; whereas up-regulated DEG were enriched for GO terms related to bicellular tight junction mechanism and palmitoylation in platelet activation and thrombus formation ( Table S21 ). Genes associated to DAcR with an increased acetylation in macrophages showed an enrichment in GO terms related to response to inflammation, infection and cell adhesion ( Table S22 ). No functional enrichment was performed for monocytes and neutrophils due to the low numbers of genes associated with DAcR.

Multi omic signatures of obesity and partial lipodystrophy and their use in prediction of cardiometabolic risk
Multivariable regression approaches for variable selection have provided an effective means to integrate multiple omics layers and elucidate disease signatures 67,68 . We used one such approach to integrate RNAseq H3K27ac histone modification, DNA author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint methylation, metabolic and lipidomic datasets. We identified a training set comprising 6 donors from the WP10 group with values below this combination of thresholds: BMI <25, GLC <5.4 mmol/L, TG <1.7 mmol/L, LDL-C <2.59 mmol/L, HDL-C >1 mmol/L for men and >1.3 mmol/L for women, HOMA-IR index <2.2, 6 obese individuals and 10 lipodystrophy patients for which we had complete measurements on all omic data layers, in monocytes and in neutrophils. Using this training set, we used elastic net penalised logistic regression to identify putative signatures associated with an increased probability of belonging to the obesity and/or lipodystrophy groups ( Fig. 3A ). The values taken by the variables selected into each signature defined patterns characterising the different groups (WP10, all patients, obese and lipodystrophy) ( Fig. 3B ; Table S33 ). When comparing features selected for each comparison, i.e. obese versus lean and lipodystrophy versus lean, 27 genes in monocytes (pvalue=1.5*10 -26 , hypergeometric test) and 6 genes (p value=1*10 -3 , hypergeometric test) in neutrophils were also differentially expressed between obese and lean. Two and one selected genes were also differentially expressed between lipodystrophy and lean in monocytes and in neutrophils respectively. Genes within 10 kilobases of a H3K27ac peak or within 10 kilobases of a DNA methylation site identified using the variable selection did not return significant enrichment for gene ontology terms (above 5% FDR threshold). The variable selection analysis identified groups of features that were, together, predictive of patient status, whereas the differential analysis identified individual variables that were different between each of the patient groups and the lean individuals.
We used the variables selected for each signature, together with the biometric variables, to construct multivariable logistic regression models to predict whether an individual was a patient or donor ( Fig. 3C ). Although rigorous validation of the full predictive model was hampered by a paucity of other cohorts for which multiple omics datasets were available, we observed that training the model using just the obese and lean groups allowed us to correctly identify the lipodystrophy patients as having a CMS profile ( Fig. S3 ); and vice versa ( Fig. S4 ). Further validation was obtained using the lipidomic layer signature. We identified selected lipids that were also measured in two other studies: a subset of 1,507 participants of the Fenland study 69  To assess the specificity of the selected lipids, we repeated the analysis with 10 lipids that were not selected into the signature ( Fig. 3E ). We found far fewer associations were found to be significant.
We performed the same analysis using data from the NASH cohorts ( Fig. S5 ), as well as data from the present study ( Fig 3F ). Although these studies were insufficiently powered to allow us to identify significant associations after correcting for multiple testing, the effect estimates were similar to those obtained in the Fenland study. In summary, we managed to define molecular signatures representing CMS abnormalities overlapping components. We next sought if these signatures were reversible if obese individuals underwent important surgery which led stomach reduction.

Effect of bariatric surgery on transcriptional profile, epigenetic landscape and cell functions .
Bariatric surgery is an effective option for the management of extreme obesity and its comorbidities, including CMS risk 71 , with well established long term benefits on weight loss, diabetes, hypertension and dyslipidemia 72 . Here, we investigated the effects of weight loss by bariatric surgery on the transcriptional and epigenetic profiles of innate immune cells and platelets, and on plasma proteins. To this end, a second blood sample was obtained six months after bariatric surgery and subjected to the same assays. Pairwise comparison of each biochemical parameter showed a decrease for LAR, TG, hsCRP, AT-IR and AST and an increase of HDL-C level (p values of: 7.22*10 -6 , 2.63*10 -9 , 4.98*10 -4 , 2.51*10 -2 , 1.48*10 -3 and 1.86*10 -3 respectively ; conditional multiple logistic regression ; Table S1 for other author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint comparisons ). We next compared the transcriptional and epigenetic profiles in monocytes, neutrophils, macrophages and platelets before and after bariatric surgery. We identified, using paired analysis, 713 DEG in macrophages (403 down and 310 up regulated; GO terms enrichment for DEG identified the following processes: ribosome formation and translation, platelet activation, fatty acids synthesis, lipids metabolism and transport, several immune related pathways ( Table S18 , S19 , S20 and S21 for macrophages, monocytes, neutrophils and platelets, respectively). The results of the other possible comparisons ( Fig. S2 ) are available in Tables S8 to S17 . Next, we searched for those genes whose expression changed in obese individuals and reverted to the level observed in lean individuals after bariatric surgery. In macrophages we found 3 genes that after having been upregulated in obese individuals compared to lean individuals, returned to the same expression level after weight loss (MCEMP1, PHACTR1 and CTD-2135D7.2; p value=1*10 -2 , hyper-geometric test). In monocytes, we found 7 genes with down-up-down profile (CLASP1, ATP11C, RALGAPA2, MTHFD2, GPRIN3, FAM129A and MCTP2; p value=1*10 -2 , hyper-geometric test) and 4 genes with a up-down-up profile (HLA-DRB9, RMRP, LINC00899 and APEX1; p value=5*10 -2 , hyper-geometric test).
No genes with a down-up-down profile were found in neutrophils and only PCGF6 was found with a up-down-up profile (p value=9*10 -2 ; hyper-geometric test). In  In addition, overlap between DEG and DAcR was found for neutrophils in obese versus post surgery (208 genes, p value=4.9*10 -30 , hyper-geometric test) and lean versus post surgery (8 genes, p value=7*10 -3 , hyper-geometric test). To obtain further insight into how bariatric surgery affects gene expression and signaling pathways in other tissues and organs, we investigated plasma protein levels before and after surgery. We quantified 3,098 plasma proteins; 604 of which were found to be differentially abundant (DAP, Fig. 4C and Table S28 ) above ordinal Q-value of 1*10 -3 . Proteins whose levels increased after bariatric surgery (72) were enriched in GO terms related to tight junction, WNT signalling, PI3K/AKT signalling, and sphingolipid signalling. Instead, proteins whose abundance decreased after surgery (532) were enriched in the following GO terms: cell cycle and DNA repair, ribosomal RNA metabolism and cell senescence, phagocytosis and T cell receptor signalling as well as FGF, IL2, VEGF and insulin signaling pathways ( Table S29 ) in agreement with Albrechtsen and colleagues 73 . Plasma proteins can have different origins; to determine if any of the proteins identified could be linked to a specific tissue, we curated the GTEx project 74 database to extract tissue specific genes, these range from 286 in the heart left ventricle to 1286 in the spleen ( Table S30 and methods ).
The data generated in monocytes and macrophages also allowed us to explore the effect of bariatric surgery on trained innate immunity 75 , as it has been shown that trained innate immunity could play a role in atherosclerosis 76,77 . We found two distinct overlaps: one (p value 5*10 -2 ; t-test) between the genes associated with the top 500 regions with a gain in histone 3 lysine 4 trimethylation (H3K4me3) after β -glucan author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint treatment in Quintin and colleagues 75 and the DEG found comparing lean individuals and obese individuals. The second overlap (p value 2*10 -4 ; t-test) between the genes associated with the top 500 regions with a gain in histone 3 lysine 4 trimethylation (H3K4me3) without treatment and the DEG found comparing obese individuals before and after bariatric surgery. All other overlaps were non-significant ( Table S31 ).
We performed functional tests on neutrophils and platelets to determine if the changes observed at molecular levels resulted in changes in the functional phenotypes of these cells. After bariatric surgery, neutrophils showed a reduction in their ability to adhere when unstimulated and when subjected to a variety of stimuli (DTT, LBP, PAM3, PAF and fMLP) but not when treated with TNFalpha or PMA.
These results were accompanied by a reduction in the cell surface levels of CD16 and CD32 but not CD66b, CD63, CD62L or CD11b (paired t-test, all result in Table   S32 ). Alongside, we performed platelet functional tests which showed a reduction in P-selectin upon collagen stimulation, but not upon ADP or thrombin stimulations. These results were accompanied by a reduction in the cell surface levels of fibrinogen receptor (both CD61 and CD41a but not CD41b) and CD36, whereas no changes were observed for CD49b, CD42a, CD42b, CD29 and CD9 (paired t-test, all result in Table S32 ).

Biochemical and Metabolic signatures.
We considered two groups of patients with extreme phenotypes associated with cardiometabolic syndrome. These two groups can be distinguished amongst the general population using anthropometric and plasma biochemistry parameters. PCA of these parameters showed that obese individuals were separateded along PC1, with weight explaining most of the variance. ALT and AST activities explained most of the variance along PC2 (45.6% and 29.8% respectively). These transaminases are known to be elevated in lipodystrophy patients 78 . Some WP10 donors overlapped either with obese or lipodystrophy patients (Fig. 1A) due to similarities either in weight or plasma biochemistry profile.
author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint We used a network-based approach to determine the differences in metabolite abundances in the different groups. To increase statistical power, we merged the patient groups under the assumption that they share similar associations of metabolites and phenotypic traits. In the consensus analysis, we used a conservative approach, considering 988 metabolites Of these, 375 were assigned to 15 different modules and the remaining 613 were put in an ad hoc extra module because they did not show any correlation. Analysis of the correlation module matrix divided the modules into two different clusters for patients, C1 and C2. Together with the analysis performed using the results from PC1, in Fig.1E, we showed that these clusters represented each a patient group. However, these separations were based on CMS features rather than obesity and lipodystrophy specific features. C1 showed an enrichment for modules involved in alanine, aspartate and glutamate metabolism known to be associated with NAFLD 79,80 , but they also showed enrichment for nitrogen and phenylalanine metabolisms, previously described in obesity 81,82 . The same functional annotations were retrieved in a large study on the effects of bariatric surgery on the metabolome 83 . C2 showed an enrichment for cysteine and methionine metabolism, and glycine, serine and threonine metabolism. These have been associated with NAFLD 84 and have also been shown to play a role in other CMS associated diseases such as obesity and T2D 85 .

Transcriptional and epigenetic signatures.
The comparisons between lean/obese and lean/lipodytrophy ( Fig. 2A ) found a modest number of changes in terms of gene expression, active chromatin and DNA methylation ( Table S7 ), however these were enough to clearly highlight inflammatory response and platelet activation related terms in GO enrichment analysis, thus confirming that these conditions modify the molecular phenotype of cell types involved in the development atherosclerosis and in thrombus formation. Similar results have already been reported for blood cells DNA methylation 86 , while more extensive changes have been observed in adipose tissue 59 .
The limited changes we observed can be, at least in part, explained by the absence of acute challenge when the samples were collected, as previously shown 87 . The largest number of changes observed in active chromatin, DAcR, was found, both in author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . monocytes and in macrophages, in the comparison between obese individuals and lipodystrophy patients, where the latter had in excess of 1,700 transcriptional enhancers ( Fig. S2B and Table S12 ). However, this was not associated with changes of the same magnitude at transcriptional level, suggesting that the same transcriptional output can be achieved with different regulatory landscapes, as previously shown 88 , but also that the same cell types could be primed differently in the two groups.

Multi omic signatures
We used variable selection approaches to identify signatures of disease, which we incorporated into a predictive mode. Within-study validation demonstrated the utility of the model for out of sample prediction ( Fig. S3 and S4). We moreover showed that prioritized lipid species are associated with major cardiometabolic risk factors in the Fenland study. Among the selected lipids, previous studies have shown PC (38:6) to be reduced in models of liver damage, PC (36:2) to be reduced in obese mice livers 89 , TG (50:1) and TG (52:2) 89 to be a product of de novo lipogenesis increased in NAFLD 69 and NASH 70 . Although we believe this represents the limits of the validation that can be performed using present datasets, we acknowledged that further multi-omic studies will be required to evaluate this model in external cohorts

Effect of bariatric surgery.
We found that bariatric surgery has remodelled plasma biochemistry results. In particular, the decrease of TG in obese individuals after bariatric surgery was in agreement with what was already shown by Szczuko and colleagues 90 . TG levels were different between lipodystrophy and post surgery groups, but not between lipodystrophy and obese groups. Although our results did not establish a direct effect of bariatric surgery on lipodystrophy patients, several previous studies have demonstrated the beneficial effects of bariatric surgery in lipodystrophy patients with BMI < 30 91-95 . Furthermore, our results showed that bariatric surgery had an profound effect on gene expression and epigenetic profiles of macrophages, monocytes, neutrophils, and platelets larger than those observed when comparing obese (or lipodystrophy) with lean individuals ( Tables S8 to S17 ). Although there author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint were genes whose expression levels, after surgery, were indistinguishable from the level observed in lean individuals. There were also many more genes whose expression levels changed after surgery to assume values not observed before in any of the conditions tested. These findings indicated that the reduction in inflammatory signatures observed after bariatric surgery in macrophages, monocytes, neutrophils, and platelets were due to novel gene expression landscapes. Of interest, in these cells we did not observe changes in DNA methylation levels of the same magnitude as those observed at gene expression level. The short life span of the cell types analysed and the overall small number of changes in DNA methylation observed in the different comparison, suggested that the changes from a pro-inflammatory to a healthy bone marrow environment had little effect on the hematopoietic stem cell epigenome. Plasma proteomic allowed us to obtain a whole body snapshot of the changes that occur after surgery. The majority of the changes were in proteins whose level decreased after surgery (532 out of 605; Table S28 ). These showed that the effect of the surgery inflammatory response (including NLRP3), insulin signalling, WNT signalling, VEGF signalling were reduced because of the reduction in fat mass but also vascular integrity was restored, as confirmed by the tissue specificity analysis that identified amongst others artery and blood as the sources of production of the reduced plasma proteins ( Table S29 ).
Interestingly, we also observed that genes previously been implicated in trained immunity 75 , a phenomenon associated with innate cells response to stimuli and also atherosclerosis 96 , were found in the comparison between obese and lean individuals.
However, post surgery we observed that many of the changes involved genes that belonged to the non-challenged set 75 , suggesting that bariatric surgery had a positive impact on innate immune cells or that trained immunity acted downstream the hematopoietic stem cell pool and its effects were eventually diluted and lost. Lastly, the transcriptional events were accompanied by a decrease in adhesion observed in platelets and neutrophils. This could be due to a diminution of proinflammatory signals in the cell environment and would result in a diminished propensity to form a thrombus.
author/funder. All rights reserved. No reuse allowed without permission.

Conflict of interest.
Authors have no CoI to declare. author/funder. All rights reserved. No reuse allowed without permission.

Number of metabolites in each module is indicated in brackets. Cell colour
represents Pearson's correlation as shown by legend. Significance is annotated as follows: * P≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.001, **** P ≤ 0.0001 (Fisher's test). Red stars indicate module-trait associations found to be significant in both groups. Left

Cell types isolation
Whole blood (50ml) in citrate tubes was obtained after informed consent. Platelet rich plasma (PRP) was separated from the cellular fraction by centrifugation (20', 150g and very gentle break). Platelets then isolated from PRP after 2 more spins as above and leukodepleted using anti CD45 Dynabeads (Thermofisher) following the author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint manufacturer's instructions. Purified platelets were stored in TRIzol (Invitrogen) until RNA extraction. The remaining cells were resuspended in buffer 1 and separated on a Percoll gradient. Neutrophils were harvested from the pellet after red cell lysis

RNA extraction
RNA extraction from samples stored in TRIzol was performed following the manufacturer's instructions. Briefly, tubes were retrieved in small batches and author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint thawed on ice. Prior to extraction samples were vortexed for 30" to ensure complete lysis and let for 5' at room temperature. Samples were then transferred to heavy phase lock tubes (5prime), chloroform was added and the tubes spun to separate RNA in the aqueous phase from the organic phase. RNA was precipitated from the former with isopropanol and glycogen. The RNA pellet was washed with 75% ethanol and resuspended in RNase free water. Purified RNA was stored in single use aliquots. Each sample was quality controlled by Bioanalayser (Agilent) and quantified via Qubit (Thermofisher).

Library preparation and sequencing
For cell types isolated from obese and lipodystrophy patients and day controls we WP10 RNA-seq data (extensively described in Chen et al. 75

Sample preparation
Cells were fixed immediately after purification with 1% w/v formaldehyde for 10 min and quenched using 125 mM glycine before washing with PBS. Samples were with --skipZeros --numberOfSamples 50000 options. Peaks were called with MACS2 author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint (v.2.1.1) with --nomodel --shift -100 --extsize 200, a qvalue threshold of 1e-3 options and celltype matching input file scaled to sample read number. We used MACS2 randsample function to downscale inputs. We then computed a score by summing values obtained for each range of these metrics. We applied a threshold of -3 (total) to select the best quality data.

Processing and quantification
All Infinium Human Methylation 450 array data pre-processing steps were carried out using established analytical methods incorporated in the R package RnBeads (v.1.13.4) 112 . First, we performed background correction and dye-bias normalization using NOOB 113 , followed by normalization between Infinium probe types with SWAN 114 . Next, we filtered out probes based on the following criteria: median detection p value 0.01 in one or more samples; bead count of less than three in at least 5% of samples; ambiguous genomic locations 115 ; cross-reactive and SNP-overlapping probes 116 .
The RRBS samples were sequenced on Illumina HiSeq3000 platform in 50bp single-end mode. Base calling was performed by Illumina Real Time Analysis author/funder. All rights reserved. No reuse allowed without permission.
Trimmomatic (v0.32) 117 was used for trimming the adapter sequences. Trimmed short read sequences were aligned onto the GRCh38/hg38 human reference genome with BSMAP(v2.90) 118 aligner in RRBS mode which was optimized for aligning the RRBS data while being aware of the restriction sites and with the following options: -D C-CGG -D T-CGA -w 100 -v 0.08 -r 1 -p 4 -n 0 -s 12 -S 0 -f 5 -q 0 -u -V 2. R package RnBeads was used to filter out low confidence sites: sites overlapping any SNP, having a coverage lower than 5 and high coverage or missing in more than 5% or individuals were filtered out.
Integration analysis required to attenuate technology effect between 450K arrays and RRBS. To this goal, we generated RRBS data for 14 BluePrint donors for which we already have 450K array data in monocytes, and 9 in neutrophils. We first removed non reproducible sites between technologies as follows: for monocytes and neutrophils, 1) liftover 450K sites to Hg38 using UCSC liftover tool 119  author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint Differential analysis For differential analysis, we used methylKit R package (v.1.8.1) 122 and we compared only RRBS data. We first extracted methylation ratios from BSMAP mapping results with methratio.py python script provided with BSMAP. We then removed all sex chromosomes sites and filtered out non-retained sites from RnBeads RRBS processing. Finally, we used the methRead function from methylKit R package in CpGs context at base resolution to read in the input files and calculateDiffMeth function correcting for overdispersion (overdispersion="MN") and applying Chisq-test. Q Values are then computed using SLIM method 122,123 . We applied two thresholds: difference of methylation > 25 and qvalue < 0.05 and retrieved differentially methylated sites (DMS) with getMethylDiff function specifying type="hypo" or type="hyper" option to get down and up methylated CpGs

Plasma metabolites measurement
Metabolites quantification author/funder. All rights reserved. No reuse allowed without permission.

Plasma lipids measurement
Plasma was frozen in dry ice immediately after collection and stored at -80C until analysis. Samples were prepared essentially as previously described 124  The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint TriVersa Nanomate (Advion) interfaced to the Thermo Exactive Orbitrap (Thermo Scientific). Briefy, a mass acquisition window from 200 to 2000 m/z and acquisition in positive and negative modes were used with a voltage of 1.2kV in positive mode and −1.5 kV in negative mode and an acquisition time of 72 s. Raw spectral data were processed as previously described 125 . Raw data were then converted to.mzXML (usingMSconvert 126 with peakpick level 1), parsed with R and 50 spectra per sample (scan from 20 to 70) were averaged using XCMS42, with a signal cutoff at 2000. The files were aligned using the XCMS 127,128 grouping function using "mzClust" with a m/z-window of 22 ppm and a minimum coverage of 60%. Compound annotation was automated using both an exact mass search in compound libraries as well as applying the referenced Kendrick mass defect approach. Signal normalisation was performed by summing the intensities of all detected metabolites to a fixed value to produce a correction factor for the efficiency of ionisation. Exact masses were fitted to the lipid species library and subsequently annotated to the peak as described before 70 .

Sample preparation
Plasma was precleared by centrifugation at 3,000 g for 10 minutes and bound to 100 µL of calcium silicate matrix (CSM, 4 mg/mL) by rotation for 1 hour. The sample was centrifuged at 14,000 g for 1 minute and the supernatant was removed for further analysis. The pellet was washed in ammonium bicarbonate (50 mMoL, 1 mL) 3 times using the same centrifugation settings. The sample was then reduced for 30 minutes at 65°C using 200 µL of DL-dithiothreitol (DTT) premix (ADC 2%: ammonium bicarbonate 50 mMoL: DTT 1 MoL in the ratio of 50:49:1) and alkylated for 30 minutes in the dark with iodoacetamide (IAA) at 20 mMoL. Ammonium bicarbonate was added to dilute the ADC to 0.5%. Trypsin was added in the ratio of 1:25 trypsin to plasma and incubated overnight at 37°C. The ADC was precipitated with 1% formic acid (FA) and centrifuged at 14,000 g for 10 minutes. The peptides were isolated using solid phase EMPORE C18 discs which had been washed with 1 stem of methanol and 3 stem of 0.1% FA. The sample was left to bind to the column for 30 minutes before washing with 0.1% FA and eluting with 60% acetonitrile (ACN) with author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint 0.1% FA and then 80% ACN with 0.1% FA. The ACN was removed by speed vacuum for 1 hour 15 minutes and freeze dried overnight. Peptide suspended in 30 µL of 0.1% FA and a peptide assay was performed to calculate the amount of peptides. 10 µL of peptides were removed from each sample and 0.1% FA added to equalise the volume and spiked with an internal standard protein (yeast alcohol dehydrogenase, ADH), with a known amount of 50 fmol injected for each run.   Figure 2).

Proteomic data processing and analysis
Progenesis QI for Proteomics (Nonlinear Dynamics, Waters Corporation, UK) was employed to identify and quantify proteins. The human database from UniProtKB was downloaded and used in FASTA format. The proteomic raw data was searched using strict trypsin cleavage rules with a maximum of two missed cleavages.
Cysteine (Carbamidomethyl C) was set as a fixed modification. Deamidation N, Oxidation M and Phosphoryl STY were selected as variable modifications. Minimum of 2 fragments per peptide, minimum of 5 fragments per protein and minimum of 2 peptides per protein were set for parameters of identification. The maximum protein mass was set to 1000 kDa. The false rate discovery (FDR) for protein identification was set at a maximum rate of 1%. Then, proteomic data generated from using the Progenesis QI was exported to Microsoft Excel for further data analysis.
For differential analysis, we used LIMMA (v.3.37.4) 129 . Because we compared obese and post surgery patients, we performed a paired analysis. We then applied a threshold of 0.1% on ordinary qvalue.
To define whole blood specific genes, we exported GTEx project 130   Model Assessment for Insulin Resistance (HOMAIR) and adipose tissue insulin resistance (AT-IR) indexes and high-sensitivity C-reactive Protein (hsCRP) and also author/funder. All rights reserved. No reuse allowed without permission.
The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.06.961805 doi: bioRxiv preprint weight (WGT), BMI and age) using cor function from stats R base package (R version 3.5.0) and pearson method (default). P Value of each correlation was computed using corPvalueStudent function from WGCNA R package.
Pathways enrichment analysis were performed with MetaboAnalyst 135 and in particular Pathway analysis module by submitting combined list HMDB identifiers for clusters C1 And C2, hypergeometric test, relative-betweenness centrality topology analysis and KEGG database. In addition, we submitted these lists to the Reactome database.

Training datasets
We identified 16 WP10 donors as controls, according to the following criteria: BMI <

Variable selection: multivariable regression approach
For each of the omics layers considered independently, we used elastic-net penalised logistic regression as implemented in the glmnet R package to identify putative signatures that discriminated between all patients (i.e. lipodystrophy + obese) versus controls. The elastic-net ɑ parameter was fixed at ɑ = 0.01, while the λ parameter was determined using cross-validation. Since different cross-validation splits resulted in different choices for λ, we performed multiple rounds of cross-validation, and used the value of λ that resulted in the maximum number of selections.
author/funder. All rights reserved. No reuse allowed without permission.

Clinical predictive model
We trained a ridge-penalised logistic regression model predictive of the binary response (i.e. patient/control status) using the clinical training dataset.

Multi-omics predictive model
We used the omic variables selected by the multivariable approach described above, together with the clinical covariates, to train a ridge-penalised logistic regression model predictive of the binary response (i.e. patient/control status). We fitted this model using the reduced training dataset. We used this model to make predictions for the 96 individuals for which we had measurements across all omics layers. To allow us to make predictions for those individuals for which we only had measurements on a subset of the omics datasets, we additionally fitted models to each combination of subsets.

Validation of selected lipids
To further investigate the lipidomic signature, we identified selected lipid species that were also measured in two other studies: a subset of 1,507 participants of the Fenland study 69,70 which is a population-based cohort of 12,345 volunteers without diabetes born between 1950 and 1975 and recruited within the Cambridgeshire region between 2005 and 2015, and a biopsy-proven nonalcoholic steatohepatitis (NASH) cohort comprising 42 individuals 70 . We used linear regression analysis to test for association between plasma levels of 8 lipid speciess selected into the lipidomic signature and all relevant CMD parameters quantified in the Fenland cohort, adjusting for age and sex, and using the Bonferroni method to control for multiple testing. We repeated this analysis for a set of 10 lipids that were not selected by either our multivariable or univariate variable selection approaches.

Functional tests
Neutrophils Adhesion Method: Polymorphonuclear granulocytes were isolated via density gradient (1.078g/mL) from 3.2% sodium citrated whole blood within 2hours of venipuncture. Neutrophil purity author/funder. All rights reserved. No reuse allowed without permission.         author/funder. All rights reserved. No reuse allowed without permission.   author/funder. All rights reserved. No reuse allowed without permission.