Abstract

   Systemic metabolic dysregulation in sepsis critically impacts patient
   survival. To better understand its onset, untargeted serum metabolomics
   and lipidomics are analyzed from 152 presymptomatic patients undergoing
   major elective surgery, and identified key metabolites, including
   serine and aminoadipic acid, that differentiate postoperative
   uncomplicated infection from sepsis. Using single‐nucleus RNA
   sequencing data from an in vivo mouse model of sepsis,
   tissue‐independent down‐regulation and tissue‐specific differences of
   serine and energy‐related genes including key module roles for the
   mitochondria‐linked genes, Cox4i1, Cox8a, and Ndufa4 are identified.
   Finally, serine‐dependent metabolic shifts, especially in the liver,
   are revealed by using ^12C/^13C murine data with labeled serine, and
   link altered activity of the serine hydroxymethyltransferase (SHMT)
   cycle with perturbed purine metabolism during sepsis. This study
   demonstrates the close interrelationship between early metabolite
   changes and mitochondrial dysfunction in sepsis, improves the
   understanding of the underlying pathophysiology, and highlights
   metabolic targets to prospectively treat presymptomatic, but at‐risk,
   patients.

   Keywords: cell–cell communication, energy imbalance, lipid metabolism,
   metabolic modeling, sepsis biomarkers, single‐nucleus RNA sequencing
   (snRNA‐Seq)
     __________________________________________________________________

   This study shows that patients at risk of sepsis have a distinct
   metabolite and lipid signature, including serine and aminoadipic acid,
   in their serum before clinical diagnosis. A mouse model of sepsis with
   a compatible serum signature reveals underlying metabolic changes,
   including mitochondrial adaptation, altered serine‐dependent purine
   metabolism, and key gene factors as potential drug targets.

   graphic file with name ADVS-12-e04418-g001.jpg

1. Introduction

   Sepsis, defined as a life‐threatening organ dysfunction caused by a
   dysregulated response to infection,^[ [42]^1 ^] is a critical medical
   condition with a significant global impact, responsible for an
   estimated 48.9 million cases and 11 million deaths annually.^[ [43]^2
   ^] Prompt recognition and treatment are key to improving survival,^[
   [44]^3 ^] however early detection remains a substantial challenge,
   highlighting an urgent need for reliable early biomarkers.

   Despite positive effects on survival by antimicrobial treatments (also
   prophylactic) and supportive care for failing organs,
   patient‐unspecific treatment protocols and the initial sepsis research
   focusing on managing inflammation yielded so far insufficient
   therapeutic success.^[ [45]^3 , [46]^4 ^] Over the last decade it
   became increasingly clear, that the failed metabolic adaptation
   disrupting the homeostatic capacity of the infected host is a central
   phenomenon that characterizes sepsis. Alongside, metabolic profiling
   has proposed biomarkers with the potential for clinical application as
   rapid diagnostic and prognostic tools in patients with sepsis.^[ [47]^5
   , [48]^6 , [49]^7 ^] Improving our understanding of the early metabolic
   disturbances in sepsis appears therefore pivotal in our combat against
   its dire consequences.^[ [50]^8 , [51]^9 , [52]^10 , [53]^11 ^]
   Metabolic and energy‐associated changes including mitochondrial
   dysfunction^[ [54]^12 , [55]^13 , [56]^14 , [57]^15 ^] play a critical
   role in sepsis. These changes profoundly affect both immune and
   non‐immune cells and lead to dysfunction in multiple organs such as the
   heart, lungs, kidneys, liver, and brain. However, the vast majority of
   published clinical studies that have investigated the metabolic
   component of sepsis have been conducted on patients who are already
   critically ill.^[ [58]^16 , [59]^17 ^] Very few have been performed at
   an early stage^[ [60]^18 ^] or been targeted at metabolic disturbances
   such as hypoglycemia,^[ [61]^19 , [62]^20 ^] which may explain the
   limited success of interventions targeting identified metabolic
   failures. Early metabolic alterations, particularly those involving
   energy metabolism, may serve as adaptive processes beneficial to the
   host.^[ [63]^11 , [64]^21 , [65]^22 ^] Recognizing and harnessing these
   adaptive changes in patients at risk of developing sepsis hold the
   potential for novel preventive or therapeutic strategies to improve,
   interrupt or prevent disease courses and fatal patient outcomes.

   Another understudied key aspect, besides the lack of appropriate human
   studies for identification of early metabolic adaptations in patients
   at risk of developing sepsis, is the intra‐ and inter‐tissue
   communication in disease. These are potential causes of the multi‐organ
   failure, of the energy‐associated pathways, and the dissipation of
   involved metabolites across organs.^[ [66]^12 , [67]^16 , [68]^22 ^]
   Creating a detailed multi‐tissue map that highlights cross‐tissue
   interactions from integrated data is essential for advancing our
   understanding and treatment (including prophylactically) of the
   metabolic and energetic maladaptation at the onset or during the course
   of early sepsis.

   In order to fill current gaps in our knowledge, we performed multi‐time
   point, untargeted metabolomic, and lipidomic analyses of a unique
   cohort of patient serum samples taken before the onset of sepsis. We
   included a comparison against several control groups, thereby capturing
   early metabolic changes that may serve adaptive roles (Figure [69]1A).
   Furthermore, we extended our research to a mouse model of sepsis and
   performed untargeted metabolomic and lipidomic analyses in serum and in
   five organs. With this additional data, we elucidate tissue‐tissue
   communication connected to the metabolic and lipidomic signature
   derived from our human cohort. To explore metabolic changes at the
   tissue and cell type level for potential conserved and cell‐specific
   responses, we performed single‐nucleus RNA sequencing (snRNA‐Seq)
   coupled with genome‐scale metabolic modeling across various tissues.
   After examining cell‐cell communication within bioenergetic pathways,
   we finally generated ^13C/^12C labeled serine data of a mouse sepsis
   model to further describe shifted metabolic fluxes and highlight a
   serine‐associated mechanistic metabolic shift across tissues and cells
   in sepsis. This data elucidates the location of systemic failure in
   sepsis, can be linked to metabolic differences present already in the
   serum of presymptomatic patients, and allows to explore new
   prophylactic and precision medicine measures to treat sepsis before its
   confirmed diagnosis.

Figure 1.

   Figure 1
   [70]Open in a new tab

   Infection‐specific metabolomics and lipidomics using presymptomatic
   serum samples from elective surgery patients. A) Study design. Serum
   samples were retrieved from 75 patients preoperatively (Pre) and 3 to 1
   day(s) before clinical diagnosis of postoperative infection (39
   developing Sepsis and 36 with uncomplicated Infect status). The
   patients were matched with 77 patients with no postoperative infection
   (38 Control and 39 SIRS). In total 455 samples were processed for
   untargeted metabolomics (n = 2086) and lipidomics (n = 5489) together
   with 26 tracked clinical parameters. Infect: uncomplicated
   postoperative infection; SIRS: systemic inflammatory response. Created
   with Biorender.com. B,C) Visualization of metabolomics/lipidomics data
   following Partial Least‐Squares Discriminant Analysis (PLS‐DA).
   Clinical parameters are indicated by individual arrows. Inf+/Inf‐:
   Patients with/without postoperative infection. D,E) Volcano plot of
   logistic regression models with covariates age, sex, and organ
   dysfunction status built separately with each metabolite (D) or lipid
   (E) to differentiate the groups infection (Inf+) from no infection
   (Inf‐). Metabolites and lipids with raw P‐value ≤ 0.05 are named and
   colored according to their HMDB and Lipid Map subclass. F) Multiscale
   embedded correlation network analysis illustrating the differential
   correlation of lipids and changes in lipid metabolic pathways in
   presymptomatic samples associated to Inf‐ versus Inf+ postoperative
   outcome. Only lipid pairs with significant differential correlations
   (adjusted empirical p ≤ 0.01) were included. Sign/sign in the legend
   indicates the direction and strength of correlation change in Inf‐/Inf+
   followed by the number of associated significant correlation changes.
   For instance, “++/+ 111″ indicates 111 lipid pairs correlate
   significantly stronger positive (++) in Inf‐ than in Inf+ (+). One
   module of biological interest with Diglyceride (DG) 29:1 as hub was
   circled for better visual clarity (see Results for details). Hub nodes
   of further modules (not shown) are named and with a black border.
   Significant lipids shown in (E) are indicated by a red border. APPT:
   activated partial thromboplastin time; SOFA: Sequential Organ Failure
   Assessment score; BP: blood pressure; GCS: Glasgow Coma Scale; INR:
   international normalized ratio; MAP: mean arterial pressure; Total Resp
   Rate: total respiratory rate; CRP: C‐Reactive Protein. TG:
   triglyceride; LPC: lysophosphatidylcholine; LPE:
   lysophosphatidylethanolamine; PC: phosphatidylcholine; Cerp: ceramide
   phosphate; SM: sphingomyeline; PE: phosphatidylethanolamine; DG:
   diglyceride; PI: phosphatidylinositol; DGDG:
   digalactosyldiacylglycerol; PA: phosphatidic acid; PS:
   phosphatidylserine; MG: monoacylglycerol; FA: fatty acid.

2. Results

2.1. Clinical and Molecular Characterization of the Patient Cohort

   A biobank with blood samples collected daily from 4385 patients
   undergoing elective surgery was established.^[ [71]^23 , [72]^24 ^]
   Demographic and clinical data were recorded commencing preoperatively
   and continuing for up to a week after, to enable identification of
   presymptomatic patients developing an infection, complicated or not by
   new‐onset organ dysfunction (sepsis). To capture the dynamic metabolic
   response of the host to infection and/or sepsis we retrieved serum
   samples from our biobank corresponding to time points Pre
   (pre‐operation), and postoperatively on days ‐3, ‐2, and ‐1 before the
   clinical diagnosis of infection or sepsis (Figure [73]1A). A total of
   455 serum samples were retrieved from 152 patients (Figure [74]1A;
   Figure [75]S1A and Table [76]S1, Supporting Information), including 231
   samples from 77 non‐infected (Inf‐) and 224 from 75 infected patients
   (Inf+). These groups were further discriminated by 114 samples from 38
   non‐infected patients (Inf‐) making an uncomplicated postoperative
   recovery (Control), 117 samples from 39 patients developing a
   non‐infectious systemic inflammatory response (SIRS), 107 samples from
   36 patients diagnosed with uncomplicated infection (Infect), and 117
   samples from 39 patients diagnosed with Sepsis (Table [77]S1,
   Supporting Information). In our subsequent analysis, we refer to
   presymptomatic patient samples by their respective postoperative
   outcome.

   All samples were analyzed by liquid chromatography‐mass spectrometry
   (LC‐MS) for untargeted metabolomic and lipidomic characterization.
   After filtering initially detected 2086 metabolites and 5489 lipids, we
   continued further analysis with robustly identified 173 metabolites and
   359 lipids across the different patient groups (see Experimental
   Section for details).

   There was no significant difference in age (Kruskal‐Wallis rank sum
   test, p = 0.7) and sex (Pearson's Chi‐squared test, p = 0.6) between
   Control, SIRS, Infect, and Sepsis patient groups (Table [78]S1,
   Supporting Information). Significant differences in blood urea, total
   respiratory rate, and Sequential Organ Failure Assessment (SOFA)
   score^[ [79]^25 ^] were found among the four patient groups, with the
   highest values recorded in sepsis patients (Table [80]S1, Supporting
   Information, Kruskal–Wallis rank sum test, p ≤ 0.05). Notably, 8 out of
   26 clinical characteristics of the SIRS group were more similar to the
   infected groups (including both Sepsis and Infect, Table [81]S1,
   Supporting Information) than the Control group.

2.2. Metabolomics and Lipidomic Profiles Distinguish Infected Patients Before
Clinical Identification

   To identify early host metabolic responses to infection, we compared
   Inf‐ and Inf+ presymptomatic metabolomic and lipidomic profiles after
   normalizing postoperative by patient‐resolved preoperative metabolite
   and lipid abundances (see Experimental Section for details). Metabolite
   and lipid profiles distinguished (partial least‐squares discriminant
   analysis (PLS‐DA)) non‐infected from infected patients with an area
   under the curve (AUC) of 0.82 and a 95% confidence interval (CI) of
   0.76–0.88 for metabolites and AUC of 0.79 (CI: 0.71–0.86) for lipids
   (Figure [82]S1B,C, Supporting Information). Several clinical indices
   were significantly associated with observed variation in the
   metabolomic and lipidomic profiles based on partial least‐squares
   discriminant analysis (PLS‐DA, false discovery rate (FDR)‐adjusted p ≤
   0.05). Based on the metabolomic profiles, blood creatinine and
   respiratory rate were significantly associated with infection
   (Figure [83]1B). For both lipidomics and metabolomics, C‐reactive
   protein (CRP) was significantly associated with infection
   (Figure [84]1B,C). The top most influential analytes as identified by
   multivariable analysis included the metabolites 3−ureidopropionate,
   tryptophan, and glutamine, and the lipids lysophosphatidylcholine (LPC)
   20:4, lysophosphatidylethanolamine (LPE) 22:5, and monoacylglycerol
   (MG) 22:5, respectively (Figure [85]S1D,E, Supporting Information).

   Early response biomarkers of infection were then sought by comparing
   presymptomatic Inf‐ and Inf+ patient groups for significantly different
   metabolite and lipid abundances (Figure [86]1D,E). Logistic regression
   models with age, sex, and organ dysfunction (observed in Sepsis and
   SIRS patients) as covariates (Figure [87]1D,E) identified significantly
   differing levels of seven metabolites (logistic regression, p ≤ 0.05,
   Cohen's d: 0.33 – 0.47) (Table [88]S2, Supporting Information) and 30
   lipids (logistic regression, p ≤ 0.05, FDR ≤ 0.4, Cohen's d: 0.33 –
   0.51) (Table [89]S2, Supporting Information). Among these,
   3‐ureidopropionate, which acts as an endogenous neurotoxin via
   inhibition of mitochondrial energy metabolism,^[ [90]^26 ^] was
   elevated in Inf+ (Figure [91]1D). Tryptophan and nicotinic acid both
   had lower values in Inf+, which may relate to insufficient cellular
   nicotinamide adenine dinucleotide (NAD^+) supply and possible energy
   substrate deficiency.^[ [92]^27 ^]

   Of note, 29 of the 30 identified lipids from several classes were
   significantly decreased in Inf+ (Figure [93]1E), while only PA 24:0
   (phosphatidic acid) was increased, hinting at potential impairments in
   lipid and/or energy metabolism. To systematically evaluate perturbed
   lipid co‐regulation at the early presymptomatic stages of infection, a
   multiscale embedded correlation analysis was performed (Figure [94]1F).
   Overall, the direction and/or strength of 367 correlation pairs changed
   between Inf‐ and Inf+, suggesting significant alterations in the host's
   lipid metabolism. 285 positive associations were reduced or lost in the
   early stages of infection compared to the non‐infection group, with 170
   indicating a loss of positive correlation (+/0, blue lines), four
   indicating a switch from positive to negative correlation (±, dark blue
   lines) and 111 indicating a significant reduction in positive
   correlation (++/+, light blue lines, see Experimental Section for
   details) based on the difference in z‐scores between Inf‐ and Inf+
   (Figure [95]1F).

   From the global networks, six lipid modules were identified via
   multiscale clustering analysis (Table [96]S3, Supporting Information).
   One of these (Figure [97]1F, encircled) included a cluster hub that
   connected to several significantly differing abundant lipids upon
   infection. This module comprises DG (diglyceride) 29:1 as a hub
   connected to numerous lipids including lysophosphatidylcholines (LPCs),
   lysophosphatidylethanolamines (LPEs), phosphatidylcholines (PCs),
   ceramide phosphates (Cerps) and sphingomyelines (SMs). Enrichment
   analysis showed that next to lysoglycerophospholipids and
   polyunsaturated fatty acids, this module enriches for “lipid−mediated
   signaling”, “endoplasmic reticulum” (ER), and “mitochondrion”
   (FDR ≤ 0.05), which suggests a regulatory influence of lipids involved
   in ER and mitochondria cell wall integrity (Figure [98]S1F, Supporting
   Information).

   In summary, our untargeted metabolomics and lipidomics analysis
   revealed distinct presymptomatic metabolic signatures that
   differentiate infected from non‐infected patients, with notable
   alterations in metabolites and lipid levels potentially linked to early
   host responses to infection.

2.3. Specific Amino Acids and Lipid Classes Changes Indicate Transition from
Uncomplicated Infection to Sepsis

   We next explored, whether early host metabolite or lipid biomarkers can
   be related to infection outcome. Accordingly, a comparative analysis
   was performed of the presymptomatic untargeted metabolomic and
   lipidomic profiles from the Control, Infect, and Sepsis groups with
   PLS‐DA (Figure [99]2A,B). Patients from Control (metabolite: AUC =
   0.79, CI = 0.71–0.87; lipid: AUC = 0.85, CI = 0.79–0.92; Figure
   [100]S2A, Supporting Information) and Sepsis (metabolite: AUC = 0.88,
   CI = 0.81–0.95; lipid: AUC = 0.83, CI = 0.75–0.91; Figure [101]S2A,
   Supporting Information) groups could be well differentiated. However,
   the Infect group with uncomplicated infection outcome remained
   challenging to clearly separate from the other two groups (metabolite:
   AUC = 0.62, CI = 0.51–0.72; lipid: AUC = 0.6, CI = 0.49–0.71; Figure
   [102]S2A, Supporting Information). The overall difference in data
   variance between groups (projection analysis, FDR‐adjusted p ≤ 0.05,
   see Experimental Section for details) was mainly driven by
   Control‐associated lipid classes LPCs, LPEs and triglycerides (TGs) as
   well as serine, and Sepsis‐associated aminoadipic acid
   (Figure [103]2A,B; Figure [104]S2B, Supporting Information).

Figure 2.

   Figure 2
   [105]Open in a new tab

   Presymptomatic severity trend analysis from Control to septic
   postoperative outcome. A,B) Visualization of PLS‐DA using metabolite
   (A) or lipid (B) data with different postoperative outcome. The top 10
   most influential metabolites and lipids are indicated by individual
   arrows. C,D) Ordinal regression models with covariates age and sex for
   each metabolite and lipid with unadjusted p‐value ≤ 0.05 and false
   discovery rate (FDR) ≤ 0.25 of the disease severity trends from Control
   over Infect (uncomplicated infection) to Sepsis within 3 days and 1 day
   before diagnosis. Displayed abundances are normalized to preoperation
   abundances. (C) Significant metabolites in the Control (blue), Infect
   (orange), and Sepsis (red) group. D) Significant lipids in the Sepsis
   group were grouped by their name and present the change from
   preoperation to 1 day before (postoperative) diagnosis. Solid and
   dashed lines indicate higher and lower abundances (normalized by
   preoperative abundance) in the Control compared to the Sepsis group,
   respectively (compare Figure [106]S2). E) Partial Spearman correlation
   analysis of significantly altered compounds from Figures (C), (D)
   together with clinical parameters in the Sepsis group. Significance of
   correlation is indicated by black circles (un‐adjusted P ≤ 0.05) and
   triangles (FDR ≤ 0.05). Annotation bars indicate increasing (red, Up
   Sepsis) and decreasing (blue, Down Sepsis) severity of infection as
   tested by ordinal regression models (compounds) and based on literature
   review (clinical data). APPT: activated partial thromboplastin time;
   SOFA: Sequential Organ Failure Assessment score; BP: blood pressure;
   GCS: Glasgow Coma Scale; INR: international normalized ratio; MAP: mean
   arterial pressure; Total Resp Rate: total respiratory rate. TG:
   triglyceride; LPC: lysophosphatidylcholine; LPE:
   lysophosphatidylethanolamine; PC: phosphatidylcholine; Cerp: ceramide
   phosphate; SM: sphingomyeline; PE: phosphatidylethanolamine; DG:.

   Ordinal regression models (adjusted for age and sex) were built using
   the presymptomatic profiles to identify early changes in the
   metabolites and lipids associated with increased disease severity
   (Figure [107]2C,D; Table [108]S2, Supporting Information). The
   concentrations of 15 metabolites significantly differed between
   Control, Infect, and Sepsis (p ≤ 0.05, FDR ≤ 0.25). Methionine levels
   were lower in the Control group compared to the infection groups Infect
   and Sepsis, particularly three days before actual diagnosis
   (Figure [109]2C). In contrast, the concentration of the essential amino
   acids serine and glutamine decreased with increasing infection severity
   (Figure [110]2C). These are associated with support of the citric acid
   (TCA) cycle, maintaining redox balance and mitochondrial function.^[
   [111]^28 ^] Serine is also recently described as an important regulator
   of lipid homeostasis with clear molecular links with sphingolipid
   metabolism.^[ [112]^29 ^] Aminoadipic acid was also found elevated in
   Sepsis compared to Infect and Control (Figure [113]2C). Spearman's
   correlation analysis showed strong and significant correlations
   (partial Spearman, adjusted for age and sex, p ≤ 0.05, FDR ≤ 0.05)
   between serine, methionine, and aminoadipic acid with several clinical
   parameters (Figure [114]2E). While serine was negatively correlated,
   methionine and aminoadipic acid were positively correlated with organ
   dysfunction (SOFA score, Figure [115]2E), suggesting an involvement of
   these amino acids in the transition from simple infection to sepsis.

   In contrast to metabolites, more lipids were significantly changed in
   abundance between Control, Infect, and Sepsis groups (N = 106; p ≤
   0.05, FDR ≤ 0.25) (Figure [116]2D). Even though a recent study
   suggested an association between lipid metabolism and mitochondrial
   dysfunction as a predictor of COVID‐19 prognosis,^[ [117]^30 ^] it
   remains unclear whether a disrupted lipid homeostasis is cause or
   consequence of mitochondrial dysfunction in various pathophysiological
   conditions. We grouped significantly altered lipids according to their
   class and their abundance changes from the preoperative level to one
   day before diagnosis of sepsis (Figure [118]2D). All significantly
   altered lipids decreased toward the day of diagnosis in the Sepsis
   group. The abundance decrease was greater for Sepsis compared to
   Control and Infect except for two lipids (PA 24:0, TG (triglycerides)
   61:11) (Figure [119]S2C, Supporting Information).

   The most altered lipid group was triglycerides (TGs), however, the
   prognostic or even therapeutic value of TGs in sepsis is
   controversial.^[ [120]^31 , [121]^32 ^] In our study, 30 out of 31
   significantly altered TGs decreased with increasing infection severity
   in presymptomatic diagnosis of sepsis patients (Figure [122]S2C,
   Supporting Information). However, TG abundances increased in one day
   compared to three days before diagnosis of sepsis with 25 of 30 TGs
   recovering more than half of their initially reduced abundance (Figure
   [123]S2C, Supporting Information), which suggests time‐sensitive TG
   dynamics at the onset of sepsis. Likewise, TG 50:1, TG 50:2, and TG
   53:3, which were significantly negatively correlated with the SOFA
   score (FDR ≤ 0.05) (Figure [124]2E), were found to first decrease and
   then increase before diagnosis. Of the other lipid classes, LPCs (n =
   18), LPEs (n = 18) and PCs (n = 11) showed a significant and stable
   decrease with infection severity (Figure [125]2D). Most of the
   significantly altered lipids showed statistically significant negative
   correlations (partial Spearman, adjusted for age and sex, p ≤ 0.05 and
   FDR ≤ 0.05) with the SOFA score and other clinical parameters (Figure
   [126]S2D, Supporting Information). Selected lipids with the strongest
   correlations were LPC 16:0 (r = −0.42), LPE 23:0 (r = −0.38), and PC
   37:3 (r = −0.33).

   We additionally explored whether Sepsis induces different early
   metabolic and lipidomic changes than SIRS by Spearman correlation
   analysis including also clinical indices (302500 pairwise comparisons
   in total, Table [127]S4, Supporting Information). After detection of
   five mega‐clusters (Figure [128]S3, Supporting Information), we
   superimposed the Sepsis‐based cluster hierarchy on SIRS correlations.
   Next to three clusters with moderate similarity (Spearman's r = 0.52. r
   = 0.57, r = 0.57, respectively), two clusters showed lower similarity
   (r = 0.35, r = 0.44, respectively) between Sepsis and SIRS. These later
   two included metabolites associated to TCA cycle, serine, and threonine
   metabolism, as well as several lipid pathways, including mitochondrion
   and membrane components, and several phospholipids, which are
   associated with mitochondrial integrity and function^[ [129]^33 ^]
   (Figure [130]S3, Supporting Information).

   Overall, we observed pronounced changes in the levels of specific amino
   acids and derivatives, as well as glycerophospholipids (LPCs, LPEs, and
   PCs) that correlate with organ dysfunction (indicated by a change in
   SOFA scores) in the Sepsis group. These all point toward metabolic and
   particularly mitochondrial dysfunction. Notably, among the altered
   metabolites, serine was the only amino acid being significantly
   negatively correlated with SOFA scores. Despite these findings, the
   underlying mechanisms of organ dysfunction remain unexplored due to the
   infeasibility of collecting prospective organ data in patient cohorts
   scheduled for operation and having different postoperative
   complications.

2.4. Differential Tissue Responses and Inter‐Tissue Communication in a CLP
Mouse Model of Sepsis

   After having identified human presymptomatic signatures of sepsis
   involving lipid, mitochondrial, and amino acid metabolism and due to
   the lack of organ data of human patients, we sought to further explore
   the signature's relevance for systemic metabolic dysfunction in a
   murine model of sepsis (Figure [131]S4A, Supporting Information). The
   survival rate 24 h after surgery is similar between septic (cecal
   ligation and puncture, CLP) and control (sham) mice (Figure [132]S4B,
   Supporting Information). However, septic mice show reduced body weight,
   food intake, temperature, and glucose levels, along with an increased
   disease score compared to control mice (Figure [133]S4C, Supporting
   Information). Septic mice also exhibit significantly reduced total
   activity during both day and night (Figure [134]S4D, Supporting
   Information). Most immune cell populations and pathogen loads are
   similar between septic and control mice (Figure [135]S4E,F, Supporting
   Information), with the exception of elevated monocytes in the bone
   marrow and increased neutrophils and myeloid cells in the primary
   lymphoid follicle of septic mice.

   We performed untargeted metabolomic and lipidomic analyses in mouse
   serum and in addition, in five organ tissues, including heart, liver,
   kidney, spleen, and white adipose tissue (WAT) 24 h after operation.
   Several lipid groups (LPCs, LPEs, PCs, DGs, and phosphatidylinositols
   (PIs)), and, with the exception of methionine, amino acids and their
   derivatives (serine, glutamine, tryptophan, and aminoadipic acid) that
   were significantly altered in the serum of human presymptomatic Sepsis
   patients, were also altered in the same direction in the serum of
   septic CLP mice (Figure [136]3A). At the tissue level, different
   tissues showed different abundance changes in these metabolites and
   lipids (Figure [137]3A; Table [138]S2, Supporting Information). Lipid
   groups, including LPCs, LPEs, and PCs, showed a decreasing trend in
   heart, kidney, and liver tissues (Figure [139]3A). Specifically, LPCs
   decreased significantly (p ≤ 0.05 and FDR ≤ 0.25; Cohen's d in [−0.61,
   −1.19]) in all three tissues, PCs decreased significantly (p ≤ 0.05 and
   FDR ≤ 0.25) in liver (Cohen's d = −0.46) and kidney (Cohen's d =
   −0.64), while LPEs decreased significantly (p ≤ 0.05, FDR ≤ 0.25,
   Cohen's d = −1.05) in kidney (Figure [140]3A). In contrast, spleen and
   WAT exhibited an increasing trend for these lipids (Figure [141]3A).
   Notably, LPCs (Cohen's d = 1.24) and PCs (Cohen's d = 0.97) increased
   significantly (p ≤ 0.05, FDR ≤ 0.05) in WAT, while LPEs increased
   significantly (p ≤ 0.05, FDR ≤ 0.05, Cohen's d = 1.01) in spleen
   (Figure [142]3A).

Figure 3.

   Figure 3
   [143]Open in a new tab

   A CLP‐induced septic mouse model recaptures human sepsis signature
   compounds. A) Statistical analysis for CLP versus sham mice samples of
   metabolites and lipids associated to Sepsis signature compounds
   identified in the presymptomatic human cohort. Significance according
   to two‐tailed Student's t‐test (circle: unadjusted p‐value ≤ 0.05,
   triangle: FDR ≤ 0.25). Significance of human serum samples
   (Figure [144]2C,D) is additionally indicated for easier comparison. B)
   Fast expectation‐maximization microbial source tracking (FEAST) results
   show the estimated contribution of five tissues (“source”) on the serum
   (“sink”) based on all metabolites and lipids composition for each mouse
   (two septic, CLP1/2, and three control, sham1/2/3, mice). C) Absolute
   number and fraction of human sepsis signature compounds also detected
   in mice serum and five tissues. D) Multiscale embedded correlation
   network analysis illustrating the differential correlation of
   metabolites and lipids in sham relative to CLP. Only correlation pairs
   with significant differential correlations (adjusted empirical P ≤
   0.05) were included. Sign/sign indicates the direction and strength of
   correlation changes in sham/CLP followed by the number of associated
   lipid pairs with this change. Two identified modules are circled for
   clarity (compare Results). Module hub nodes have a black border, while
   nodes with red labels are compounds that significantly correlated with
   SOFA score in the human cohort (Figure [145]S2B). Abbreviations: LPC:
   lysophosphatidylcholine; LPE: lysophosphatidylethanolamine; PC:
   phosphatidylcholine; DG: diglyceride; PI: phosphatidylinositol; TG:
   triglyceride; WAT: white adipose tissue.

   To further understand the contribution of the five tissues to serum
   metabolites and lipids, we adapted the Fast expectation‐maximization
   microbial source tracking (FEAST) algorithm^[ [146]^34 ^]
   (Figure [147]3B). Across all mice, the largest contribution to serum
   lipids was attributed to the liver (mean 42%), while the kidney was the
   largest contributor to metabolites (mean 24%). On average, 53% of serum
   lipids and 38% of serum metabolites could be attributed to the studied
   tissue sources (Figure [148]3B). Further high percentages of the serum
   lipids (47%) and metabolites (62%) were attributed to unknown sources
   (Figure [149]3B), which may be due to the close relationship between
   serum metabolites and the gut microbiome.^[ [150]^35 ^] By
   cross‐checking against the Microbial Metabolites Database (MiMeDB),^[
   [151]^36 ^] we indeed found that 86% of the metabolites and 47% of the
   lipids could be derived from the gut microbiome.

   Of the 121 metabolites and lipids identified in our human study as
   presymptomatic sepsis signatures, 80% of the metabolites and 43% of the
   lipids were detected in the serum or five other tissues in mice with
   mouse serum showing the highest overlap (metabolite: 67%; lipid: 32%,
   Figure [152]3C). Multiscale embedded correlation analysis revealed that
   the direction or strength of 101 correlation pairs involving human
   presymptomatic sepsis signature metabolites and lipids changed between
   the mouse sepsis and control groups, indicating a substantial
   alteration in inter‐tissue communication (Figure [153]3D). Notably, two
   modules were identified from the global network using multiscale
   clustering analysis (Figure [154]3D, encircled). Module 1 includes PC
   37:3 from the kidney and LPC 18:0 from the heart as module hubs
   (Figure [155]3D). LPC 18:0, linked to serine and tryptophan in the
   spleen, has been shown to have therapeutic effects in experimental
   sepsis.^[ [156]^37 ^] Aminoadipic acid from the liver is the hub of
   module 2, connected to multiple compounds correlated significantly with
   the SOFA score in the human cohort, including LPC 16:0 in heart and LPC
   18:0 in spleen (Figure [157]3D). Three further affected amino acids or
   derivatives in this module across different tissues were serine,
   methionine, and N‐acetyl‐alanine (Figure [158]3D).

   Overall, we also observed the early changes of important metabolic and
   lipid classes, we identified in the serum of the presymptomatic human
   cohort, in mouse serum. However, a distinct metabolic regulation in
   different mouse tissues was also notable, suggesting a potential
   rewiring of energy metabolic pathways in a tissue‐specific manner.
   These findings point to the possibility that transcriptional changes
   before or during early sepsis may underlie the observed metabolic
   shifts.

2.5. Genome‐Scale Modeling Identifies Cellular Drivers of Metabolic Changes
in Sepsis

   We next performed snRNA‐seq on septic (CLP) and control (sham) mice to
   characterize metabolic responses of different cell types in liver,
   kidney, WAT, and brain, resulting in 49349; 43347; 48274, and 50929
   cells, respectively (Figure [159]4A). Insufficient high‐quality snRNA
   data prevented downstream analysis with spleen and heart tissue.
   Differential gene expression analysis was performed to identify marker
   genes for cell type annotation and definition, this determined 8, 10,
   8, and 6 cell types in liver, kidney, brain, and WAT, respectively
   (Figure [160]S5, Supporting Information, 26 different cell types and 32
   cell type‐tissue combinations in total).

Figure 4.

   Figure 4
   [161]Open in a new tab

   Single cell mouse RNA sequencing driven Compass analysis to identify
   tissue‐ and cell type‐specific metabolic activity. A) Single‐nucleus
   RNA sequencing (snRNA‐Seq) design. Triplicates of sham and CLP mice
   were sacrificed 24 h after operation. Tissue samples of brain, liver,
   kidney and WAT were collected and yielded between 43347 and 50929 cells
   in total per tissue. Created with Biorender.com. B) Cohen's d effect
   size distribution for four energy relevant pathways. Effect sizes by
   Cohen's d comparing CLP versus sham per pathway associated reaction as
   defined in the human genome‐scale metabolic model RECON2 across all
   cell types associated to a given tissue. Different grey background
   areas reflect different effect sizes. Only reactions with significant
   changes for CLP versus sham (two‐sided Wilcoxon test, FDR adjusted p ≤
   0.05) were considered. C) Volcano plot for reactions associated to the
   pathways shown in (B). Y axis shows significance following two‐sided
   Wilcoxon test of the change CLP versus sham per reaction (FDR adjusted
   p ≤ 0.05). The top 5 reactions with highest effect change and EC
   classification number per pathway are indicated. D) Contextualized
   metabolic network information for highlighted reactions shown in (C).
   Metabolic network was adapted from the KEGG pathways
   ([162]https://www.genome.jp/kegg/) hsa00310 (Lysine degradation),
   hsa00561 (Glycerolipid metabolism), hsa00564 (Glycerophospholipid
   metabolism) and hsa00260 (Glycine, serine and threonine metabolism).
   Abbreviations: Gly, ser, ala & thr mb: Glycine, serine and threonine
   metabolism; Lys mb: Lysine metabolism, Glycerophospholipid mb:
   Glycerophospholipid metabolism. X‐axis labels sham and CLP reflect
   direction of change of calculated effect sizes. EC numbers: 1.1.1.95:
   Phosphoglycerate dehydrogenase; 1.1.99.1: Choline dehydrogenase;
   1.2.1.31: Aminoadipate semialdehyde dehydrogenase; 1.5.1.10:
   Saccharopine dehydrogenase; 1.5.1.8: Saccharopine dehydrogenase (NADP+,
   L‐lysine‐forming); 1.5.1.9: Saccharopine dehydrogenase (NAD+,
   L‐glutamate‐forming); 1.5.3.7: L‐pipecolate oxidase; 1.5.99.2:
   Dimethylglycine dehydrogenase; 2.1.1.17: Phosphatidylethanolamine
   N‐methyltransferase; 2.1.1.20: Glycine N‐methyltransferase; 2.1.1.5:
   Betaine‐homocysteine S‐methyltransferase; 2.3.1.15:
   Glycerol‐3‐phosphate acyltransferase; 2.3.1.51: Lysophosphatidic
   acid‐acyltransferase; 2.3.1.61: Dihydrolipoamide succinyltransferase;
   2.6.1.39: 2‐aminoadipate transaminase; 2.6.1.51: L‐Serine:pyruvate
   aminotransferase; 2.6.1.52: Phosphoserine transaminase; 2.7.1.107:
   Diacylglycerol phosphate kinase; 2.7.7.15: Choline phosphate
   phosphatase; 2.7.7.41: Phosphatidate cytidylyltransferase; 2.7.8.11:
   Phosphatidylinositol synthase; 3.1.1.23: Acylglycerol lipase; 3.1.1.3:
   Triacylglycerol lipase; 3.1.3.3: Phosphoserine phosphatase; 3.1.3.4:
   Phosphatidic acid phosphatase.

   We explored the metabolic activity of the identified key presymptomatic
   sepsis signatures at the cell type level using the Compass algorithm
   (see Experimental Section for details). To specifically analyze the
   metabolic context in sepsis of our observed relevance of serine and
   aminoadipic acid linked with lipid metabolism, we focused on the
   subsystems “Glycine, serine and threonine metabolism”, “Lysine
   metabolism”, “Glycerophospholipid metabolism” and “Triacylglycerol
   synthesis” as defined in the human genome‐scale metabolic
   reconstruction RECON2 (Figure [163]4B, Experimental Section). In total,
   3566 comparisons of Compass scores were computed for CLP versus sham
   over all 32 different tissue‐cell type combinations. The effect size of
   each Compass score change per tissue and cell type was assessed by
   Cohen's d analysis, of which 3126 were significant (FDR ≤ 0.05) and
   subject to the subsequent analyses.

   A distribution analysis revealed elevated metabolic activity
   probabilities in liver CLP samples in the investigated four subsystems
   with the exception of triacylglycerol synthesis, which showed
   widespread density distribution and thus cell‐type specific effects
   (Figure [164]4B). The highest metabolic induction in septic CLP mice
   was in lysine metabolism in liver, where aminoadipic acid acted as the
   hub in inter‐tissue communication (Figure [165]3D). WAT cells showed
   elevated metabolic activity probabilities in glycerophospholipid
   metabolism and triacylglycerol synthesis corroborating their role in
   lipid metabolism. Kidney cells, on the other hand, were characterized
   in sepsis by decreased predicted metabolic activity in
   glycerophospholipid metabolism and triacylglycerol synthesis, and
   increased metabolic activity in lysine metabolism (Figure [166]4B).
   Brain cells were generally less affected and showed a moderate decrease
   in lysine metabolism in septic CLP mice (Figure [167]4B).

   To identify the most pronounced cell‐type specific drivers of changes
   at the metabolic level, the top five most changed reactions for CLP
   versus sham per pathway for which an enzymatic classification by Enzyme
   Commission (EC) number was available were further investigated
   (Figure [168]4C; Table [169]S5, Supporting Information). In all liver
   cells during sepsis the reactions associated with degradation of lysine
   to aminoadipic acid, with glutamate as the side‐product, were
   profoundly activated (Figure [170]4C). Among all liver cell types, the
   enzymes alpha‐aminoadipic semialdehyde synthase (AASS, EC 1.5.1.8,
   1.5.1.9), saccharopine dehydrogenase (SDH, EC 1.5.1.10), and
   aminoadipate semialdehyde dehydrogenase (ASDH, EC 1.2.1.31) showed the
   highest predicted activity increase in liver‐associated cholangiocytes
   (Cohen's d ≥ 2.09), Kupffer (Cohen's d ≥ 2.23), and stellate cells
   (Cohen's d ≥ 2.02) (Figure [171]4C), suggesting that these liver cell
   types are the main contributors to the metabolic hub function of liver
   aminoadipic acid (Figure [172]3D; Table [173]S5, Supporting
   Information).

   In WAT cells, the metabolic activity shift was generally diverse within
   the investigated pathways (Figure [174]4C). For instance, during sepsis
   WAT lysine degradation was reduced in WAT pericytes (Cohen's d in
   [−0.64, −0.24]), but increased in WAT epithelial cells (Cohen's d in
   [0.28, 0.67], Table [175]S5, Supporting Information), hinting toward at
   a protective role of glutamine^[ [176]^38 ^] to selected WAT cells. In
   support of our prior observation of serine and its involvement in WAT
   metabolism, Compass predicted significantly increased metabolic
   activity of serine‐associated enzymatic reactions (among others, ECs
   2.1.1.5, 1.1.99.1, and 3.1.3.3) in selected WAT cell types
   (Figure [177]4C), including epithelial cells (Cohen's d up to in [0.62,
   0.85]), adipose stem and progenitor cells (Cohen's d in [0.36, 0.63]),
   and in part also in adipocytes (Cohen's d = 0.27 for EC 2.1.1.5 and
   0.33 for 1.1.99.1, respectively; Table [178]S5, Supporting
   Information). Likewise, WAT epithelial, adipose stem and progenitor
   cells also exhibit increased activity in two further key metabolic
   reactions in serine metabolism, phosphoglycerate dehydrogenase (PGDH,
   EC 1.1.1.95, Cohen's d ≥0.4) and serine‐pyruvate aminotransferase (EC
   2.6.1.51, Cohen's d ≥0.4) in sepsis (Table [179]S5, Supporting
   Information). This supports an important role of WAT epithelial and
   adipose stem and progenitor cells with respect to serine‐driven lipid
   metabolism during sepsis. In contrast, WAT pericytes showed decreased
   activity of these enzymes in sepsis, suggesting a complex regulation of
   the associated serine hydroxymethyl transferase (SHMT) cycle in WAT. In
   addition, triacylglycerol Lipase (ATGL, EC 3.1.1.3), which converts TGs
   and DGs to fatty acids and glycerol, showed increased activity of
   different degree over all WAT cells (Figure [180]4C).

   By exploring potential metabolic cross‐talk across the investigated
   pathways by Spearman correlations of the Compass scores, the T
   lymphocytes among all liver cells showed a pronounced correlation
   decrease in sepsis compared to other liver cell types (Figure [181]S6,
   Supporting Information). While kidney cells showed an overall increase
   in metabolic crosstalk (especially in endothelial cells and
   macrophages), brain cells (most pronounced in GABAergic and
   dopaminergic neuronal cells), and especially all WAT cell types, showed
   a loss of correlation across all investigated pathways in sepsis. This
   hints at crosstalk challenges for the liver adaptive immune response
   and WAT in lipid, lysine and serine metabolism (Figure [182]S6,
   Supporting Information).

   We next contextualized our identified most changing metabolic reactions
   with pathway information from the Kyoto Encyclopedia of Genes and
   Genomes (KEGG, Figure [183]4D). Next to the interconnected role of
   lysine degradation and aminoadipic acid toward energy replenishment via
   the TCA cycle, serine and cholines are central within this network.
   This analysis highlights their role, and network perturbation during
   sepsis, and in glycero(phospho)lipid metabolism in the metabolic
   response to sepsis.

   In summary, snRNA‐data‐driven Compass simulations suggest significant
   changes in a number of key metabolic reactions, particularly within
   liver Kupffer cells and WAT epithelial cells. Notably, changes in the
   SHMT cycle and lysine degradation suggest elevated roles for converting
   involved metabolites, including serine or aminoadipic acid, to
   countereffect energy shortages and, potentially, mitochondrial
   dysfunction in sepsis.

2.6. Tissue‐Independent as well as ‐Specific Adaptations of Energy Metabolism

   The evidence of metabolic shifts in presymptomatic patients and the
   predicted rewired activity of specific metabolic sub‐systems in the CLP
   mouse model suggested that energy deficiency in early sepsis is
   different across tissues and cell types. To explore the potential
   regulatory component, we investigated energy‐metabolism‐related gene
   alterations across different tissues and cell types based on the mice
   snRNA‐Seq data. A total of 1473 genes associated with 34
   bioenergetic‐related pathways as defined by the KEGG database (Table
   [184]S6, Supporting Information) were identified (liver: 1354; kidney:
   1357; brain: 1364; WAT: 1384). The overall variation in energy
   metabolism‐related genes, based on tissue‐specific average gene
   expression, clearly separated between CLP and sham‐treated animals for
   all tissues with the least variation in brain tissue (Figure [185]5A).

Figure 5.

   Figure 5
   [186]Open in a new tab

   Mouse tissue‐ and cell type‐Specific gene regulation of energy
   metabolic pathways in CLP‐induced sepsis. A) Principal component
   analysis (PCA) based on the average gene expression profiles of energy
   metabolic genes in the four tissues. Axes reflect the first two
   principal components (PC1, PC2) and their explained variance. B)
   Cell‐cell communication of genes belonging to the 34 energy‐related
   pathways based on the KEGG database in different cell types of the four
   tissues. Numbers in the tissue label indicate the number of
   energy‐related gene‐gene interactions for CLP versus sham. Linkages
   connecting the key metabolic cell types of hepatocytes, proximal
   tubule, neurons, and adipocytes are highlighted by saturated color. Red
   and blue linkage colors indicate more interaction pairs in CLP or sham,
   respectively. Linkage width is associated with the number of observed
   interactions between any two connected cell types. C) Number of
   differentially expressed genes (DEGs) between CLP and sham in the four
   tissues. Red and blue represent up‐ and down‐regulated genes in CLP and
   sham, respectively. D) Average log[2](fold‐change) profile of the 37
   DEGs with consistent variation direction across four tissues at the
   cell type level between CLP and sham. Red and blue colors indicate
   upregulated genes in CLP and sham, respectively.

   We further constructed intercellular communication networks within each
   tissue using CellChat^[ [187]^39 ^] to investigate energy‐metabolism
   relevant cell‐cell signaling variations in sepsis (see Experimental
   Section for details). In general, we found reduced cell‐cell
   communications in CLP compared to sham mice in liver (94 versus 119),
   brain (43 versus 63), and WAT (32 versus 55) tissues, while kidney
   displayed the opposite (65 versus 57) (Figure [188]5B). At the cell
   type level, there were decreased numbers of interactions with CLP in
   liver hepatocytes, brain neurons, and WAT adipocytes (Figure [189]5B).
   These findings, combined with the consistent decrease in metabolic
   cross‐talk of serine‐ and aminoadipic acid‐related reactions in the
   corresponding cell types (Figure [190]S6, Supporting Information),
   highlight the potentially critical contributions of serine and
   aminoadipic acid to the overall decay of energy‐related cell‐cell
   communications across these three cell types.

   To further investigate the responses in mice tissues during sepsis, we
   performed differential expression analysis and identified 741, 840,
   511, and 905 energy‐related differentially expressed genes (DEGs) in
   liver, kidney, brain, and WAT, respectively (Figure [191]5C). We
   observed less up‐regulated DEGs in CLP compared to sham irrespective of
   the assessed tissue. We then analyzed expression patterns by
   hierarchical clustering of energy‐related pathways at the cell type
   level. We found distinct expression patterns for liver together with
   kidney, as well as brain together with WAT cells (Figure [192]S7A,
   Supporting Information). These differences corresponded to the
   tissue‐specific pattern of mitochondrial biogenesis and thermogenesis
   related genes, with up‐regulated expression in all WAT cell types but
   down‐regulated expression in most kidney cell types, except for B
   lymphocytes (Figure [193]S7A, Supporting Information). In addition,
   different energy‐related pathway expression patterns were found across
   cell types in each examined tissue. While liver neutrophils and
   hepatocytes showed down‐regulation of energy‐metabolism‐related
   pathways, potentially accounting for the overall reduction of energy
   metabolism in liver tissue, most other liver cell types showed an
   effect in the opposite direction (Figure [194]S7A, Supporting
   Information). Interestingly, serine metabolism‐associated Shmt1
   expression was up‐regulated in most liver cell types, including immune
   cells such as B lymphocytes and Kupffer cells (Figure [195]S7B,
   Supporting Information), and significantly down‐regulated in
   hepatocytes. Genes associated with lysine degradation were
   up‐regulated, especially in cholangiocytes, Kupffer, and stellate cells
   in liver, and in all kidney cell types, but down‐regulated in all brain
   cell types, and in WAT endothelial cells and macrophages (Figure
   [196]S7A, Supporting Information), adding a consistent regulatory
   component to our metabolism‐focused Compass results (Table [197]S5,
   Supporting Information).

   To explore whether there are also tissue‐independent transcriptional
   changes of genes contributing to energy regulation and metabolism in
   response to sepsis, we investigated DEGs for consistent direction of
   change. Thirty‐seven energy metabolism‐associated DEGs were identified
   with the same transcriptional changes in CLP compared to sham mice
   across all four tissues (Figure [198]S7C, Supporting Information).
   Except for Mt2, a mitochondrial biogenesis/thermogenesis‐related gene
   with antioxidant function,^[ [199]^40 ^] all 22 DEGs, which are
   associated with lipid metabolism, were down‐regulated in sepsis. For
   instance, the glycerolipid metabolic gene Akr1a1, a cancer biomarker
   associated with mitochondrial metabolic activity,^[ [200]^41 ^] was
   down‐regulated across all tissues (Figure [201]S7C, Supporting
   Information), again pointing to overall mitochondrial adaptation and
   dysfunction during sepsis. Apart from up‐regulated Mt2, two further
   mitochondrial biogenesis/thermogenesis‐related DEGs, Adcy7 and Mtarc2,
   were consistently down‐regulated at the tissue level in septic mice
   (Figure [202]S7C, Supporting Information), reflecting the dominating
   effect direction we see at the cell type level per tissue
   (Figure [203]5D). Interestingly, the serine‐related DEGs Gamt and Gatm
   also exhibited consistent down‐regulation in septic mice across cell
   types in all tissues.

   Overall, next to tissue‐independent we also observed tissue‐specific
   transcriptional adaptation of energy metabolic genes across all four
   tissue sites and their cell types during sepsis in mice. Furthermore,
   we identified tissue‐independent differential regulation of specific
   mitochondrial biogenesis and thermogenesis genes, including Mt2, Adcy7,
   and Mtarc2. By using tissue‐specific mouse data, we could add another
   facet to the mounting evidence for systemic alterations in energy
   metabolic responses during sepsis.

2.7. Identification of Hub Genes Linking Mitochondrial Biogenesis with
Sepsis‐Induced Lipid Metabolism Dysregulation

   To provide a more comprehensive understanding of the transcriptional
   regulation of energy‐related metabolic genes, we constructed and
   analyzed gene co‐expression networks for each tissue. We identified 4,
   6, 5, and 4 gene co‐expression modules (by weighted gene co‐expression
   network (WGCNA), see Experimental Section for details) in liver,
   kidney, brain, and WAT, respectively (Figure [204]6A). Of note, we
   observed one module, which we refer to as the common module, in every
   tissue where most of the hub genes (highly connected genes within a
   module) were associated with mitochondrial biogenesis/thermogenesis
   function (Figure [205]6A, cluster with red border). In addition, most
   non‐hub genes in the common module also associated with mitochondrial
   biogenesis/thermogenesis function, displaying a pronounced co‐expressed
   gene interaction across all tissues toward mitochondria‐related energy
   metabolism in the septic mice (Figure [206]S7D, Supporting
   Information). Strikingly, three of the mitochondrial genes, Cox4i1,
   Cox8a, and Ndufa4, were hub genes in the common module in all four
   tissues (Figure [207]6A). Cox4i1 and Cox8a encode for subunits of
   cytochrome c oxidase while Ndufa4 encodes for a subunit of nicotinamide
   adenine dinucleotide hydrogen (NADH) dehydrogenase. These enzymes are
   essential parts of the mitochondrial electron transport chain (ETC),
   contribute to adenosine triphosphate (ATP) generation, and have been
   linked to thermogenesis as alternative output of mitochondrial
   respiration.^[ [208]^42 ^]

Figure 6.

   Figure 6
   [209]Open in a new tab

   Gene co‐expression networks associated with energy metabolism and hub
   genes in mouse CLP‐induced sepsis. A) Average log[2](fold‐change)
   profile of the hub genes in all modules per tissue for CLP versus sham.
   Red and blue colors indicate up‐ and down‐regulated genes in CLP,
   respectively. The structurally common module across all four tissues is
   indicated by a red border. Hub genes belonging to the mitochondrial
   biogenesis/thermogenesis pathway are highlighted in green, within which
   the three common genes across tissues are in bold font. B)
   Co‐expression networks based on energy metabolism associated genes in
   the common module per tissue. Each node represents one gene, and each
   edge refers to the co‐expression relationship between two connected
   nodes. Node color reflects pathway association. Nodes of hub genes
   belonging to mitochondrial biogenesis/thermogenesis are enlarged and
   labeled with gene symbols. Colored linkages indicate the pathway origin
   of genes linked with mitochondrial biogenesis/thermogenesis hub genes.
   The three hub genes Cox4i1, Cox8a and Ndufa4 present in each
   tissue‐specific network are indicated in bold font. Akr1a1 was labeled
   with a black arrow in the liver and WAT.

   While the structure of the common module appeared independent of tissue
   type, a tissue‐specific co‐expression pattern in liver and kidney was
   observed. Here, all hub genes in the common modules were down‐regulated
   in sepsis across all cell types except B lymphocytes within liver
   (Figure [210]6A). In contrast, with the exception of brain
   oligodendrocytes, all brain and WAT cell types up‐regulated the hub
   genes of the common module upon early sepsis. We also identified
   tissue‐independent hub gene co‐expression with multiple genes
   associated with serine metabolism, lysine degradation, and lipid
   metabolism (Figure [211]6B). Among these, lipid metabolism associated
   genes were the most connected to the mitochondrial
   biogenesis/thermogenesis hub genes. Akr1a1, identified as one of the 36
   tissue‐independent down‐regulated genes (Figure [212]S7C, Supporting
   Information), was co‐expressed with all 15 mitochondrial
   biogenesis/thermogenesis hub genes in liver and WAT (Figure [213]6B,
   black arrow). Genes involved in glycerolipid and glycerophosphlipid
   metabolism (Table [214]S7, Supporting Information) were also
   co‐expressed with hub genes from the common modules, including Agpat5
   (lysophosphatidic acid‐acyltransferase, EC 2.3.1.51) and Cdipt
   (phosphatidyltransferase, EC 2.7.8.11). These insights corroborate
   their role in energy metabolic adaptations we predicted with Compass
   particularly in liver and WAT, but not in kidney cells in septic mice
   (Figure [215]4C and [216]4D). Since both processes are vital for
   lysophosphatidic acid (LPA) to PA conversion, mitochondrial membrane
   integrity, and PI synthesis,^[ [217]^43 ^] our findings suggest that
   kidney cells require modulation of PIs during sepsis to ameliorate
   tissue damage protection. This is in support of the kidney tissue
   protective role of PI phosphorylation in a murine sepsis model.^[
   [218]^44 ^]

   Taken together, our co‐expression insights underline the potential
   co‐regulation of mitochondrial ETC‐mediated ATP generation with other
   energy‐related metabolic pathways, particularly including lipids
   associated with membrane integrity.

2.8. Labeled Serine Administration Reveals Drop in Hepatic PCs and Increase
in Purine Metabolism

   Until this point, we accumulated evidence for serine involvement in the
   energy adaptive processes in sepsis. This included the reduction in
   serine levels and its inverse correlation with infection severity in
   our human presymptomatic Sepsis group, as well as altered
   serine‐related metabolic activity and gene expression at the tissue and
   cell level in septic mice. Consequently, we further investigated the
   location and impact of serine‐dependent alterations of metabolic
   mechanisms in sepsis. First, we investigated the impact of serine on
   sepsis at the tissue level by administering unlabeled serine to sepsis
   (CLP) and control (sham) mice (Figure [219] 7A, see Experimental
   Section for details). Most pronounced, we found altered lipid levels
   upon serine addition. Heart tissues showed significant increases in
   LPCs and DGs only in septic mice (FDR ≤ 0.05, Cohen's d in [1.22,
   1.39]), while liver LPCs, PCs, and DGs increased only in control mice
   (FDR ≤ 0.05, Cohen's d in [0.43, 0.97]) (Figure [220]7A). All further
   tissues and serum showed less pronounced differences between control
   and sepsis mice (Figure [221]7A). While spleen and kidney PCs were
   significantly reduced in control mice (FDR ≤ 0.05, Cohen's d in [−0.37,
   −0.46]) upon serine addition, septic mice only showed a trend, whereas
   serum and WAT showed sepsis‐independent alteration of serine induced
   lipid changes (e.g., increased PIs in both sepsis and control serum, or
   reduced LPCs, PCs, PIs, and phosphatidylserines (PSs) in WAT, Table
   [222]S8, Supporting Information).

Figure 7.

   Figure 7
   [223]Open in a new tab

   Serine administration and serine labeling experiments reveal drop in
   hepatic PC and elevated serine integration into purines in early
   sepsis. A) Differential lipid abundance in Serine+ versus Vehicle
   (Serine‐) treated mice in both CLP and sham groups. Mean abundance is
   z‐scaled. B) Metabolites with significantly incorporated ^13C‐labeled
   serine‐derived carbon in septic mice. Only metabolites with significant
   higher incorporation percentage in ^13C CLP than ^13C sham (one‐sided
   Student's t test, p ≤ 0.05, FDR adjusted p ≤ 0.25) and higher
   incorporation percentage in ^13C CLP than ^12C CLP (one‐sided Student's
   t test, FDR adjusted p ≤ 0.05) were considered. C) Schematic of purine
   metabolism in mice liver in context of tissue specific gene expression
   profile based on snRNA data. Metabolic network was adapted from the
   KEGG pathways ([224]https://www.genome.jp/kegg/) mmu00260 (Glycine,
   serine and threonine metabolism) and mmu00230 (Purine metabolism). All
   cell types shown next to genes indicated up‐regulated gene expression
   in these cells in liver of septic compared to sham mice. Created with
   Biorender.com. D) Violin plot of nine genes in purine metabolism shown
   in (C). ^* p ≤ 0.05; ^** p ≤ 0.01; ^*** p ≤ 0.001. LPC:
   lysophosphatidylcholine; LPE: lysophosphatidylethanolamine; PC:
   phosphatidylcholine; DG: diglyceride; PI: phosphatidylinositol; PS:
   phosphatidylserine.

   To investigate more precisely in which metabolic components serine is
   integrated, we orally administered ^13C‐labeled serine to both control
   and sepsis mice (see Experimental Section for details). We first
   quantified the enrichment of downstream lipids in serum and five
   tissues (Table [225]S9, Supporting Information). ^13C‐labeled
   serine‐derived carbon was incorporated significantly into liver LPCs
   (Fisher's exact test, p ≤ 0.05) and PCs (Fisher's exact test, p ≤ 0.05,
   FDR ≤ 0.1) in control but not in sepsis mice (Table [226]S9, Supporting
   Information), demonstrating perturbed serine associated PC biosynthesis
   in liver during sepsis.

   Interestingly, when we investigated further metabolites involved in the
   direct incorporation of serine, we found that ^13C‐labeled
   serine‐derived carbon was significantly incorporated into purine
   metabolites in the liver of septic mice (Figure [227]7B; Table [228]S2,
   Supporting Information). Across serum and five tissues, eight
   metabolites showed significantly higher incorporation of ^13C‐labeled
   serine‐derived carbon in sepsis compared to control mice (Student's
   t‐test: p ≤ 0.05, FDR ≤ 0.25, Cohen's d ≥1.7) (Figure [229]7B). All of
   these metabolites were associated with purine metabolism in the liver.
   In serum, only uric acid showed significantly increased incorporation
   of ^13C‐labeled serine‐derived carbon in septic mice (Figure [230]7B).

   By cross‐checking these insights with our snRNA data, we also found
   up‐regulation of genes involved in purine metabolism in liver,
   particularly those linked to converting metabolites incorporating
   ^13C‐labeled serine‐derived carbon in Kupffer cells, cholangiocytes,
   B lymphocytes, and hepatic stellate cells (Figure [231]7C). Downstream
   of Shmt1, particularly Ak3, Uox, Urah, and Urad were consistently
   up‐regulated across all four liver cell types (Figure [232]7D).

   We conclude that next to a serine‐specific effect on the abundance of
   lipids, purine biosynthesis, and salvage processes are differentially
   regulated in sepsis, potentially as a response to the energetic
   shortages given mitochondrial malfunction.

3. Discussion

   Metabolic adaptation is crucial during sepsis progression to provide
   substrates for biosynthetic pathways, and to regulate signaling and
   organ function with the aim of coping with the allostatic load and
   eventually re‐establishing cellular homeostasis.^[ [233]^11 , [234]^45
   ^] Several studies explored the potential of metabolic markers for
   sepsis,^[ [235]^16 , [236]^17 , [237]^46 , [238]^47 , [239]^48 ^] and
   presented machine learning models that can predict sepsis mortality
   with high accuracy. However, these studies primarily utilize blood
   samples from patients with established sepsis, where adaptive and
   maladaptive responses likely coexist and are complicated further by
   factors associated with organ dysfunction and confined body regions for
   immune responses.^[ [240]^22 , [241]^49 , [242]^50 ^] Here, we utilized
   a longitudinal cohort of serum samples taken from presymptomatic
   patients undergoing elective surgery to capture the earliest metabolic
   responses to sepsis (Figure [243]1). We identified serine, aminoadipic
   acid, and specific lipid classes (LPCs, LPEs, and PCs) as the most
   altered metabolic processes (Figure [244]2).

   Using a CLP sepsis mouse model, we identified a tissue‐ and cell
   type‐specific likelihood of metabolic activity in the serine
   hydroxymethyltransferase (SHMT) cycle connected to choline and lipid
   metabolism based on snRNA data and metabolic model analysis
   (Figures [245]4 and [246]5). While the SHMT cycle is essential for
   providing energy in macrophage polarization,^[ [247]^51 ^] cholines
   have been shown to attenuate sepsis‐associated acute kidney injury.^[
   [248]^52 , [249]^53 ^] Serine is also an important regulator of
   mitochondrial biogenesis and functionality,^[ [250]^54 , [251]^55 ^] in
   part acting via serine‐derived lipids (ceramides),^[ [252]^56 ^] and
   has been associated with the maintenance of central carbon metabolism
   and nucleotide metabolism.^[ [253]^57 ^] Alongside, we observed
   increased concentrations of lysine degradation‐associated aminoadipic
   acid during sepsis, especially in liver cells, which potentially
   affects lipolysis,^[ [254]^58 ^] and mitochondrial biogenesis and
   function.^[ [255]^59 , [256]^60 ^] In our data, liver aminoadipic acid
   is central to inter‐tissue communication (Figure [257]3), supporting
   its critical role in the metabolic response to infection‐induced
   multi‐organ dysfunction in sepsis.^[ [258]^61 ^] By using metabolic
   modeling, we further identified liver Kupffer cells and WAT epithelial
   cells as key drivers of the serine and aminoadipic acid pathways,
   respectively. Both, serine and aminoadipic acid, are hitherto
   unrecognized early tissue‐ and cell‐type specific modulators of lipid
   and mitochondrial metabolic pathways in sepsis and, more importantly,
   can be linked to metabolic alterations in serum of presymptomatic
   patients.

   Tissue‐ and cell‐type dependent as well as independent regulatory and
   metabolic shifts suggest effects on mitochondrial biogenesis,
   associated with altered lipid metabolism. In this context, we
   identified common co‐expression modules with mitochondrial‐ and
   thermogenesis‐related genes serving as hubs and linking genes involved
   again in serine metabolism, lysine degradation, and lipid metabolism
   across the four investigated tissues (Figure [259]6). Specifically, we
   identified Cox4i1, Cox8a, and Ndufa4 as three hub genes in the common
   modules, suggesting an important role for these three genes and their
   protein products within the electron transport chain, the mitochondrial
   ATP synthesis, and thermogenesis.^[ [260]^42 ^] Of note, all three
   genes have already been confirmed as drug targets according by the
   DrugBank database,^[ [261]^62 ^] and appear to be relevant also to
   sepsis according to our data.

   Aligned with the elevated SHMT cycle activity predicted by metabolic
   modeling and the increased transcription of Shmt1, our serine
   intervention model demonstrated that serine shifts toward purine
   metabolism in the liver of septic mice. This shift is likely
   facilitated by the SHMT pathway and associated with malfunctioning
   mitochondrial ETC‐driven energy generation.^[ [262]^63 ^] Specifically,
   the ETC proteins COX4i1 and NDUFA4, whose genes are central in our
   identified common module with differential tissue‐dependent expression
   upon sepsis, have been linked to a heme‐dependent switch of
   mitochondrial ATP synthesis to thermogenesis.^[ [263]^42 ^] This link
   highlights the role, these two proteins potentially play in the
   metabolic adaptations and tolerance observed in sepsis.^[ [264]^8 ^]
   Additionally, our serine intervention model revealed reduced serine
   utilization for PC biosynthesis, along with alterations in
   serine‐associated purine abundances in septic mice (Figure [265]7). It
   appears therefore likely that exogenous serine can contribute to
   maintain purine biosynthesis also in sepsis. This includes an impact of
   serine levels on the pentose phosphate pathway and TCA cycle with the
   ultimate goal of increased energy replenishment via mitochondrial
   biogenesis^[ [266]^57 , [267]^63 , [268]^64 ^] within hepatic,
   particularly Kupffer cells to counteract energy deficiency during
   sepsis.

   Our study is not without limitations. Our human cohort enabled us to
   identify presymptomatic human signatures of sepsis before clinical
   manifestations. The current uniqueness of our dataset, however,
   warrants validation of our results in an independent future cohort,
   ideally including multiple centers with varying ethnicities, and causes
   of sepsis. This would also be important to better evaluate potentially
   varying levels of metabolic markers across different patient
   populations, as well as the influence of other comorbidities,
   including, e.g., diabetes, or liver dysfunction on metabolite levels.
   Nevertheless, we cross‐checked the relevance of our identified markers
   against studies with established sepsis and similar data and sample
   sites. We find 82% concordance in significance and direction with our
   human metabolite and lipid signature results compared to cohorts with
   established sepsis, corroborating the relevance of our identified
   markers also in presymptomatic sepsis (Table [269]S10, Supporting
   Information). In particular, our study aligns with the metabolic (67%
   concordance) and lipidomic signatures (78% concordance considering
   lipid class level) observed at the enrollment time point in cohorts
   with community‐acquired sepsis^[ [270]^17 ^] and pneumonia‐associated
   sepsis.^[ [271]^65 ^] Our study design, using a presymptomatic patient
   cohort, already allowed us to reduce potential confounding effects from
   organ dysfunction or sepsis‐induced catabolism. Furthermore, we
   minimized the impact of confounding effects by adjusting for age and
   sex, and by normalizing postoperative data using patient‐resolved
   preoperative samples. Assessing further physiological processes, such
   as early changes in renal clearance^[ [272]^66 ^] can help rule out
   potential effects on the metabolic shifts we have described and may
   help to further stratify patients at risk of developing sepsis.

   In conclusion, our analysis provides a comprehensive view on dynamic
   metabolic adaptations which can be linked to presymptomatic sepsis in
   human patients. We highlighted key metabolic factors including altered
   serine and aminoadipic acid metabolism, as well as substantial lipid
   depletion, which all pinpoint to a bioenergetic deficiency and
   mitochondrial dysregulation. Targeting our identified
   tissue‐independent module and its hub genes could potentially improve
   mitochondrial dysfunction in early sepsis and even offer completely new
   therapeutic strategies to safeguard patients at risk of developing
   sepsis.

4. Experimental Section

Study Design

   Elective surgery patients were prospectively recruited at seven UK
   hospitals and one German center between November 2007 and February
   2017. Ethical approval was granted through the Southampton and
   South‐West Hampshire Multicentre Research Ethics Committee (Reference
   06/Q1702/152) and registered with ISRCTN (No. 17 375 399).^[ [273]^67
   ^] The protocol achieved US Federal Wide Assurance Independent Review
   Board status (IRB00001756). Patients were enrolled if aged between 18
   and 80 years, due to undergo elective major abdominal, cardiac,
   gynae‐urological, thoracic, or vascular surgery, and provided informed
   consent. All patient information including demographics, type of
   surgery, and perioperative physiological, biochemical, and
   microbiological data were compiled into a comprehensive database (Item
   Tracker Software, Ipswich, UK) by research staff at each center. The
   patient grouping selection process included a Clinical Adjudication
   Panel (CAP) as described in detail in Lukaszewski et al.^[ [274]^23 ^]
   Briefly, a CAP of multidisciplinary specialists was established to
   review and classify postoperative patients based on standardized
   criteria. Five CAPs were held between July 2013 and March 2017. CAP
   members independently assessed clinical data for patients suspected of
   infection, assigning them to categories including Sepsis, Infect, SIRS,
   and Control. Final classifications were reached through group voting
   and discussion, ensuring more robust and consensus‐based diagnoses.

Mice

   All experimental animal procedures in this study were evaluated through
   the Independent Ethical Advisory Committee and approved by the Local
   Government Authority of Thuringia, Thüringer Landesverwaltungsamt, per
   European and German laws and regulations under the license HKI‐23‐002.
   For all experiments, male C57BL/6 mice at 26 – 28 weeks of age
   (obtained from Janvier Labs, France) were housed under specific
   pathogen‐free (SPF) conditions in the Leibniz Institute for Natural
   Product Research and Infection Biology (Hans Knöll Institute). All
   animal experiments follow the European (Directive 2010/63/EU)
   legislations, concerning housing, husbandry, and animal welfare.

Cecal Ligation and Puncture

   Polymicrobial abdominal sepsis was induced using the CLP model as
   previously described.^[ [275]^8 ^] Mice were anesthetized with a
   combination of isoflurane and butorphanol (1 mg kg^−1, s.c.) followed
   by 30% ligation of the cecum and double puncture with a 23 Gauge
   needle. A small amount of faeces was extruded and the cecum was placed
   back into the abdominal cavity. Sham treatment was performed without
   ligation and puncture. All animals received saline (40 mL/kg, s.c.)
   directly after the procedure and once Imipenem/Cilastatin (25 mg kg^−1,
   s.c.) after 6 h. This model is associated with a 14‐day mortality of
   30%, closely mimicking the clinical situation. Animals were orally
   gavaged with either drinking water (vehicle), 500 mg kg^−1 12C‐Serin
   (Sigma) or 13C‐Serin (CLM‐1574‐H‐0.25 Serine‐L (13C3 99%13C),
   Euroisotope, Germany) in 10 mL kg^−1 drinking 6 and 18 h after CLP.^[
   [276]^68 , [277]^69 ^] After 24 h, animals were euthanized with an
   intraperitoneal injection of ketamine/xylazine (500 mg kg^−1 and 25 mg
   kg^−1, respectively, i.p.). The animals were then perfused with
   ice‐cold Hanks’ Balanced Salt Solution and tissues were harvested,
   snap‐frozen, and stored at −80° until further analysis.

Human Serum Sample Preparation

   Sample analysis was carried out by MS‐Omics as follows. For semi‐polar
   metabolite analysis, serum samples (50 µL) were diluted with a mixture
   of acetonitrile, methanol, formic acid, and stable isotope labeled
   internal standards (200 µL 1.00:0.99:0.01 v/v/v). The extracts were
   then passed through a phosphor lipid removal cartridge (Phree,
   Phenomenex) using centrifugation (1 400 rpm, 4 °C, 10 min). An aliquot
   (100 µL) was transferred into a high recovery high‐performance liquid
   chromatography (HPLC) vial and the solvent was removed under a gentle
   flow of nitrogen. Extracts were reconstituted in a reverse phase mobile
   phase mixture (100 µL, 10% B in 90% A). For Lipidomics analysis, serum
   samples (10 µL) were transferred to a Spin‐X filter (0,22 µm, Costar)
   and diluted in a mixture of isopropanol, SPLASH Lipidomix (Avanti Polar
   Lipids) and butylated hydroxytoluene (90 µl, 96:4 vol/vol + 10 µg). The
   extract was then left at room temperature for 10 min before storage at
   −20 °C overnight. Samples were brought to room temperature and filtered
   by centrifugation (14 000 rpm, 5 °C, 2 min). An aliquot (25 µL) was
   transferred into a high recovery HPLC vial and an equal parts mixture
   of mobile phase eluent A and B (75 µL) was added. For quality control,
   a mixed pooled sample (QC sample) is created by taking a small aliquot
   from each sample. This sample is analyzed with regular intervals
   throughout the sequence.

Mice Tissue and Serum Sample Preparation

   To prepare samples for semi‐polar metabolites extraction, the different
   tissue samples were weighted and transferred to Eppendorf tubes.
   Ceramic beads and precooled methanol/water (1:2) were added. All tubes
   were placed in a pre‐cooled (−20 °C) bead beater and homogenized (4 ×
   30 sec., 30 Hz) followed by ultrasonication (5 min). After
   centrifugation (18000 RCF, 5 min, 4 °C), the supernatant was collected.
   The sample pellet was reextracted as described above. The two extract
   supernatants were pooled and passed through a phosphor removal
   cartridge (Phree filter plates). A precise aliquot of the extract was
   evaporated to dryness under a gentle stream of nitrogen, before
   reconstitution with 10% Eluent B in Eluent A. All Samples were diluted
   5 times in mobile phase eluent A and stable isotope‐labeled standards
   were added before analysis. Mouse serum (50 µL) was diluted with a
   mixture of acetonitrile and methanol (200 µL 1:1 v/v). The extract was
   then passed through a phosphor removal cartridge (Phree, Phenomenex).
   An aliquot (100 µL) was transferred into a high‐recovery HPLC vial and
   the solvent was removed under a gentle flow of nitrogen. Extracts were
   reconstituted in a mobile phase mixture (100 µL, 10%B in 90%A). To
   ensure high‐quality sample preparation, a quality control sample (QC
   sample) was prepared by pooling small equal aliquots from each sample,
   to create a representative average of the entire set. This sample was
   treated and analyzed at regular intervals throughout the sequence.

   In the different tissue samples, the pellet from the semipolar
   extraction was used to extract lipids as follows. DCM:MeOH (3:1) was
   added to the pellets, and the samples were homogenized using a
   pre‐cooled Bead‐beater (4 × 60 sec at 30 Hz), with colling of the
   blocks between the beading cycles. Subsequently, the samples were
   ultrasonicated for 5 mins and centrifuged (5 mins, 1800 RCF, 4 °C). The
   supernatant was collected and dried under nitrogen gas. Samples were
   then reconstituted with 6x (spleen and WAT) or 3x (heart, liver, and
   kidney) the sample weight with the reconstitution solvent eluent mix
   and vortex mixed for 30 sec. The reconstituted samples were filtrated
   using SpinX filters by centrifugation (5 mins, 1800 RCF, 4 °C). All
   samples were mixed 1: 50 in eluent mix. For serum samples, an aliquot
   of the serum samples is mixed with extraction solvent (0.1 M BHT in
   isopropanol with stable isotope labeled internal standard) in a Spin‐X
   filter. The filters were vortex‐mixed for 60 sec and left at room
   temperature for 10 mins before placing in −20 °C Freezer overnight. The
   following day samples were left at room temperature for 30 mins before
   centrifuging (1800 g /4 °C/2 min). The filtrate was mixed with eluent
   mix 1:3.

Human Semi‐Polar Metabolites and Lipidomics LC‐MS Method

   Sample analysis was carried out by MS‐Omics as follows. For semi‐polar
   metabolites, the analysis was carried out using a Thermo Scientific
   Vanquish LC coupled to a Q Exactive HF Hybrid Quadrupole‐Orbitrap,
   Thermo Fisher Scientific. An electrospray ionization interface was used
   as an ionization source. Analysis was performed in positive and
   negative ionization mode. The UHPLC was performed using an adapted
   version of the protocol described by Doneanu et al.^[ [278]^70 ^]
   Metabolomics processing was performed untargeted using Compound
   Discoverer 3.0 (Thermo Scientific) for peak picking and feature
   grouping, followed by an in‐house annotation and curation pipeline
   written in MatLab (2021b, MathWorks). Identification of compounds was
   performed at four levels; Level 1: identification by retention times
   (compared against in‐house authentic standards), accurate mass (with an
   accepted deviation of 3 ppm), and MS/MS spectra, Level 2a:
   identification by retention times (compared against in‐house authentic
   standards), accurate mass (with an accepted deviation of 3 ppm). Level
   2b: identification by accurate mass (with an accepted deviation of
   3 ppm), and MS/MS spectra, Level 3: identification by accurate mass
   alone (with an accepted deviation of 3 ppm). For lipid analyses,
   lipidomics high‐performance liquid chromatography (HPLC), coupled with
   high‐resolution mass spectrometry (HRMS) analysis was performed using a
   Thermo Scientific Vanquish LC coupled to Thermo Q Exactive HF Hybrid
   Quadrupole‐Orbitrap mass spectrometer (Thermo Scientific, RRID:
   SCR_020545). The chromatographic separation of lipids was carried out
   on a Waters ACQUITY Charged Surface Hybrid (CSH) C18 column (2.1 ×
   100 mm, 1.7 µm). The column was thermostated at 55 °C. The mobile
   phases consisted of (A) Acetonitrile/water (60:40) and (B)
   Isopropanol/acetonitrile (90:10), both with 10 mM ammonium formate and
   0.1% formic acid. Lipids were eluted in a two steps gradient by
   increasing B in A from 40 to 99% over 18 min. The flow rate was kept at
   0.4 mL/min. Ionization was performed in positive and negative
   ionization mode using a heated electrospray ionization interface. The
   mass spectrometer operated at a resolution of 120000 in a scan range of
   200 to 1500 m/z. Iterative data‐dependent MS/MS (dd‐MS^2) acquisition
   was achieved on a TopN of 10 with stepped collision energy (NCE) of 20,
   40, and 60. Peak areas were extracted using Compound Discoverer 3.0
   (Thermo Scientific). Compound annotations were performed against the
   mzCloud (Thermo Scientific) MSMS library, the Human Metabolome Database
   4.0,^[ [279]^71 ^] and the MS‐Omics lipid library covering 17 classes
   (CerP, DG, digalactosyldiacylglycerol (DGDG), Fatty acids, LPC, LPE,
   monoacylglycerol (MG), PA, PC, phosphatidylethanolamine (PE),
   phosphatidylglycerol (PG), PI, Plasmenyl‐PC, Plasmenyl‐PE, PS, SM and
   TG). Identification of compounds was performed at four levels; Level 1:
   identification by retention times (compared against in‐house authentic
   standards), accurate mass (with an accepted deviation of 3 ppm), and
   MS/MS spectra, Level 2a: identification by retention times (compared
   against in‐house authentic standards and lipid class behavior),
   accurate mass (with an accepted deviation of 3 ppm). Level 2b:
   identification by accurate mass (with an accepted deviation of 3 ppm),
   and MS/MS spectra, Level 3: identification by accurate mass alone (with
   an accepted deviation of 3 ppm). Lipid annotations were curated
   manually by comparing retention time behavior over the number of tail
   chain carbons and double bonds as annotated.

Mice Semi‐Polar Metabolites and Lipidomics LC‐MS Method

   Sample analysis was carried out by MS‐Omics as follows. Semi‐polar
   metabolite profiling was carried out using a Vanquish LC (Thermo
   Scientific) coupled to an Orbitrap Exploris 240 MS (Thermo Scientific).
   The UHPLC was performed using an adapted version of the protocol
   described by Doneanu et al.^[ [280]^70 ^] An electrospray ionization
   interface was used as ionization source. Analysis was performed in
   positive and negative ionization mode under polarity switching.
   Metabolomics processing was performed untargeted using Compound
   Discoverer 3.3 (Thermo Scientific) and Skyline 22.2^[ [281]^72 ^]
   (MacCoss Lab Software) for peak picking and feature grouping, followed
   by an in‐house annotation and curation pipeline written in MatLab
   (2022b, MathWorks). Identification of compounds was performed at four
   levels; Level 1: identification by retention times (compared against
   in‐house authentic standards), accurate mass (with an accepted
   deviation of 3 ppm), and MS/MS spectra, Level 2a: identification by
   retention times (compared against in‐house authentic standards),
   accurate mass (with an accepted deviation of 3 ppm). Level 2b:
   identification by accurate mass (with an accepted deviation of 3 ppm),
   and MS/MS spectra, Level 3: identification by accurate mass alone (with
   an accepted deviation of 3 ppm). For lipid analysis, lipidomics
   high‐performance liquid chromatography (HPLC), coupled with
   high‐resolution mass spectrometry (HRMS) analysis was performed using a
   Thermo Scientific Vanquish LC coupled to Thermo Q Exactive HF Hybrid
   Quadrupole‐Orbitrap mass spectrometer (Thermo Scientific, RRID:
   SCR_02 0545). The chromatographic separation of lipids was carried out
   on a Waters ACQUITY Charged Surface Hybrid (CSH) C18 column (2.1 ×
   100 mm, 1.7 µm). The column was thermostated at 55 °C. The mobile
   phases consisted of (A) Acetonitrile/water (60:40) and (B)
   Isopropanol/acetonitrile (90:10), both with 10 mM ammonium formate and
   0.1% formic acid. Lipids were eluted in a two‐step gradient by
   increasing B in A from 40 to 99% over 18 min. The flow rate was kept at
   0.4 mL min^−1. Ionization was performed in positive and negative
   ionization mode using a heated electrospray ionization interface. The
   mass spectrometer operated at a resolution of 120 000 in a scan range
   of 200–1500 m z^−1. Iterative data‐dependent MS/MS (dd‐MS^2)
   acquisition was achieved on a TopN of 10 with stepped collision energy
   (NCE) of 20, 40, and 60. Peak areas were extracted using Compound
   Discoverer 3.2 (Thermo Scientific) and Skyline 21 (MacCoss Lab
   Software). Compound annotations were performed against the mzCloud
   (Thermo Scientific) MSMS library, the Human Metabolome Database 4.0,
   and the MS‐Omics lipid library covering 17 classes (CerP, DG, DGDG,
   Fatty acids, LPC, LPE, MG, PA, PC, PE, PG, PI, Plasmenyl‐PC,
   Plasmenyl‐PE, PS, SM, and TG). Identification of compounds was
   performed at four levels; Level 1: identification by retention times
   (compared against in‐house authentic standards), accurate mass (with an
   accepted deviation of 3 ppm), and MS/MS spectra, Level 2a:
   identification by retention times (compared against in‐house authentic
   standards and lipid class behavior), accurate mass (with an accepted
   deviation of 3 ppm). Level 2b: identification by accurate mass (with an
   accepted deviation of 3 ppm), and MS/MS spectra, Level 3:
   identification by accurate mass alone (with an accepted deviation of
   3 ppm). Lipid annotations were curated manually by comparing retention
   time behavior over the number of tail chain carbons and double bonds as
   annotated. Stable isotope incorporations (^13C) were determined against
   unlabeled reference samples (containing no ^13C‐substrate).
   Corresponding compound pairs (labeled versus unlabeled) were determined
   based on their retention time, accurate mass, isotope pattern, and
   predicted elemental composition. The molecular relative exchange rate
   (in %) was calculated by averaging the peak area response for compounds
   containing ^13C isotopes and normalizing it against the signal for
   present unlabeled analogs. The Relative Isotope Response was determined
   for annotated compounds. Based on the annotated elemental composition,
   accurate masses for all possible incorporations were predicted and
   chromatographic peak areas for extracted isotope ions were determined
   using Skyline. Peak areas were then normalized for each compound and
   sample, resulting in relative isotope responses (%) for each isotope
   incorporation state. The comparison to natural isotope abundances
   measured in unlabeled samples allows for the determination of ^13C
   incorporations.

Nucleus Isolation from Frozen Samples for snRNA Sequencing

   Nucleus isolation of tissue samples was done as previously described.^[
   [282]^73 ^] Briefly, tissue samples were cut into pieces <0.5 cm and
   homogenized using a glass Dounce tissue grinder (Sigma, cat. No.
   D8938). The tissue was homogenized 25 times with pestle A and 25 times
   with pestle B in 2 ml of ice‐cold nuclei EZ lysis buffer. The sample
   was then incubated on ice for 5 min, with an additional 3 mL of cold EZ
   lysis buffer. Nuclei were centrifuged at 500 g for 5 min at 4 °C,
   washed with 5 mL ice‐cold EZ lysis buffer, and incubated on ice for
   5 min. After centrifugation, the nucleus pellet was washed with a 5 mL
   nuclei suspension buffer (NSB; consisting of 1 × PBS, 0.01% BSA, and
   0.1% Rnase inhibitor (Clontech, cat. No. 2313A)). Isolated nuclei were
   resuspended in 2 ml NSB, filtered through a 35 µm cell strainer
   (Corning‐Falcon, cat. No. 352 235), and counted. A final concentration
   of 1000 nuclei per µL was used for loading on a 10x channel.

snRNA‐seq Library Preparation and Sequencing

   snRNA‐Seq libraries with Chromium Next GEM Single Cell 3ʹ Reagent Kits
   v3.1 on the Chromium Controller (10 × Genomics) were prepared. Single
   nucleus suspensions were prepared from cultured cell lines and nuclei
   were suspended in PBS containing 0.04% BSA. The nuclei suspension was
   loaded onto the Chromium Next GEM Chip G and ran the Chromium
   Controller to generate single‐cell gel beads in the emulsion (GEMs)
   according to the manufacturer's recommendation. Captured nuclei were
   lysed and the released RNA was barcoded through reverse transcription
   in individual GEMs. Barcoded, full‐length cDNA was generated and
   libraries were constructed according to the performer's protocol. The
   quality of libraries was assessed by Qubit 4.0 and the Agilent 2100.
   Sequencing was performed on the Illumina NovaSeq 6000 with a sequencing
   depth of at least 50 000 reads per nucleus and 150 bp (PE150)
   paired‐end reads (performed by Biomarker Technologies Corporation,
   Beijing, China).

Pathogen Load

   Peritoneal fluid was obtained by peritoneal lavage (7–8 mL in sterile
   PBS). Whole organs (i.e., liver, kidney, and spleen) were harvested and
   homogenized under sterile conditions by using a tissue homogenizer
   (Ultra Turrax T25 basic, IKA Labortechnik) at 15 000 min^−1. Serial
   dilutions were plated onto Columbia Blood Agar with 5% Sheep Blood
   plates (ThermoFisher) and incubated (24 hr at 37 °C) in 5% CO2
   (aerobes) or airtight container equipped with the OxoidTM AnaeroGenTM
   sachet (ThermoFisher) (anaerobes).

Promethion Behavioral and Phenotyping System

   Promethion Core (Sable Systems, USA) was used for indirect calorimetry
   experiments. Animals were singled cage and acclimatized for one day
   before experimental procedures were performed. The recording continued
   for one day after sham or CLP. The system consists of a standard GM‐500
   cage with a food hopper and a water bottle connected to load cells
   (2 mg precision) with 1 Hz rate data collection. Additionally, the cage
   contains a red house enrichment. Total (ambulatory and fine) activity
   was monitored at 1 Hz rate using an XY beam break array (1 cm spacing).
   Oxygen, carbon dioxide, and water vapor were measured using a CGF unit
   (Sable Systems). This multiplexed system operated in pull‐mode. Airflow
   was measured and controlled by the CGF (Sable Systems) with a set flow
   rate of 2 L/min. Oxygen consumption and carbon dioxide production were
   reported in milliliters per minute (mL/min). Energy expenditure was
   calculated using the Weir equation^[ [283]^74 ^] and Respiratory
   Exchange Ratio (RER) was calculated as the ratio of VCO2/VO2. Raw data
   was processed using Macro Interpreter v2.38 (Sable Systems).

Flow Cytometry

   Single‐cell suspension from peritoneal lavage and bone marrow were
   obtained and immune cell population was quantified by flow cytometry.
   For bone marrow isolation, the femur was removed from mice after
   euthanasia, and ligament and muscles were mechanically cleared. The
   bone was separated in two pieces and bone marrow was collected by
   centrifugation in 1.5 mL microcentrifuge tubes. Peritoneal lavage was
   collected as indicated in Pathogen Load and washed with FACS buffer (2%
   BSA + 2 mM EDTA + 2 mM NaN3). Single‐cell suspension was stained in
   5 mL polystyrene tubes with the following anti‐mouse antibodies to
   identify distinct immune cell populations: CD45 (clone REA737), CD3
   (clone REA641), NK1.1 (clone REA1162), CD11b (clone REA592), Ly‐6G
   (clone REA526), F4/80 (clone REA126), and Ly‐6C (clone REA796)
   (Miltenyi Biotec) for 20 min at 4 °C. For absolute quantification, the
   CountBright Absolute Counting Beads (Invitrogen) kit was used according
   to the manufacturer´s instructions. Fluorescence minus one (FMO)
   control was used to identify positive populations in the analysis.
   Samples were acquired on an LSR II flow cytometer (BD Biosciences) and
   analyzed with FlowJo 10.7.1 software (TreeStar).

Untargeted Metabolomics/Lipidomics Data Normalization

   In both human and mice cohorts, metabolite data within annotation level
   1, level 2a, and level 2b, and lipid data within annotation level 1,
   and level 2a were used. Metabolites and lipids were filtered out with
   Descriptive Power (DP) ≤ 2.5, missing percentage > 30% in all samples
   and replaced value < LOD (The Limit of Detection) with NA. Missing
   values in metabolome and lipidome profiles were imputed by LOD/2. In
   the human cohort, postoperative abundances of metabolites and lipids
   were normalized using the preoperative abundances to remove variability
   in the patient profiles that were unrelated to infection. Specifically,
   the post‐surgery data at three‐time points were log2 transformed,
   subtracted from the log2 transformed pre‐surgery data, and the mean of
   these differences was used. In the mice cohort, the data were
   logarithmically transformed before further analysis. Missing values in
   clinical profiles were imputed by the random forest method from R
   package missForest.^[ [284]^75 ^]

Partial Least‐Squares Discriminant Analysis (PLS‐DA)

   Partial least‐squares discriminant analysis (PLS‐DA) was calculated by
   R package mixOmics.^[ [285]^76 ^] The ROC is calculated based on the
   first 2 predicted components to predict the class of each sample with
   auroc function from R package mixOmics.^[ [286]^76 ^] 95% confidence
   intervals of are under the curve (AUC) were calculated using the
   function roc from R package pROC (v1.18.5). Envfit function in the
   vegan package^[ [287]^77 ^] is used for vector fitting to fit the
   factors onto partial least‐squares discriminant analysis (PLS‐DA)
   ordination. Clinical indices were incorporated into the envfit function
   in the vegan package in R to determine those that were significant. The
   function envfit evaluates how each clinical variable individually
   correlates with axis1 and axis2. Metabolites and lipids were also
   analyzed against the community composition data using the envfit
   function in the vegan package in R. Statistical significance were
   assessed based on 999 permutations.

Metabolic Features Identification in Human and Mice Cohort

   In the human cohort, logistic regression models with sex, age, and
   organ dysfunction as covariates were built with R package stats.^[
   [288]^78 ^] Metabolites and lipids were classified into subclasses by
   The Human Metabolome Database (HMDB)^[ [289]^79 ^] and Lipid Maps,^[
   [290]^80 ^] respectively. Significant metabolites and lipids with raw
   P‐value ≤ 0.05 were shown in the color of the volcano plot. Ordinal
   logistic regression models with a cumulative logit link function were
   employed using the R package ordinal.^[ [291]^81 ^] Models were tested
   to meet the assumption of proportional odds. Benjamini‐Hochberg (BH)
   method was used to correct for multiple hypothesis testing. A list of
   106 lipids and 15 metabolites were found to be significant (P ≤ 0.05 &
   BH P ≤ 0.25) in their respective ordinal regression models. The partial
   Spearman correlation was conducted by R package ppcor^[ [292]^82 ^]
   using significant metabolites and lipids (P‐value ≤ 0.05 and
   Benjamini–Hochberg P‐value ≤ 0.25) from ordinal regression models
   related to infection severity and clinical indices. Benjamini–Hochberg
   method was used to correct for multiple hypothesis testing.

   In mice cohort, Student's t‐test was used to assess differences between
   the two groups for continuous variables. T statistics for lipids within
   the same class were compared against zero to identify significant up‐
   or down‐regulated lipid classes, including only those with at least six
   lipids. One‐sided Fisher's exact test was used to identify lipid
   classes that significantly incorporated ^13C‐labeled serine‐derived
   carbon. The Benjamini‐Hochberg procedure was applied to calculate the
   false discovery rate (FDR) and adjust p‐values for multiple hypothesis
   testing.

Differential Correlation Analysis and Pathway Analysis

   Differential correlations between the non‐infected and the infected
   patients were calculated using Spearman correlation with the R package
   DGCA.^[ [293]^83 ^] The significance of differential correlation was
   calculated for both same and different directions of the correlation
   coefficients between the two compared groups. The difference in
   Z‐scores, based on the correlation coefficients, was used to compute a
   two‐sided p‐value using the standard normal distribution. Empirical
   p‐values were calculated by comparing the Z‐scores from the original
   and permuted datasets. Only pairs with differential correlation
   (empirical p < 0.01) were included to build correlation networks from
   differentially correlated pairs. For sham and CLP mice, Pearson
   correlation was used to calculate differential correlations across five
   tissues. Only pairs with differential correlation (empirical p < 0.05)
   and at least one compound significantly correlated with the SOFA score
   in the human cohort were included. Multiscale embedded correlation
   networks were used to identify modules from the differential
   correlation network by R package MEGENA.^[ [294]^84 ^] Lipid pathway
   topology analysis was conducted by LION^[ [295]^85 ^] with “target‐list
   mode” by a one‐tailed Fisher's exact test. Lipids from the circled
   model in differential correlation analysis (Figure [296]1F) were used
   to form the list of compounds of interest. All lipids in our study were
   used as the background set.

Integrative Analysis

   Three data types (metabolomics, lipidomics, clinical indices) were
   obtained from experiments. The data consisted of 173 metabolites, 359
   lipids, and 18 clinical indices. All data were Z‐scaled and
   concatenated into one matrix. Spearman correlations between all
   compounds were calculated by R package ppcor.^[ [297]^82 ^] Complete
   hierarchical clustering was used to reorder molecules based on 1 –
   Pearson correlation between all molecules by R package
   ComplexHeatmap.^[ [298]^86 ^] The number of clusters was determined
   based on the 5 clusters observed in the heatmap and cutree from R
   package dendextend.^[ [299]^87 ^] For each mega‐cluster, the over
   representation of the KEGG pathways was determined from the included
   metabolite by MetaboAnalyst 6.0^[ [300]^88 ^] and pathway enrichment
   analysis of the lipidomics data using the Lipid Ontology (LION)
   enrichment analysis web application.^[ [301]^85 ^] The SIRS+ group was
   constrained to the hierarchy that was established for the Sepsis group
   and yielded the same mega‐clusters for comparison.

FEAST Analysis

   To estimate the contribution of tissue source compounds to serum
   metabolite and lipid composition (“sink” communities), the fast
   expectation‐maximization (EM) microbial source tracking algorithm
   (FEAST R‐package v0.1.0) was used.^[ [302]^34 ^] The analysis was
   performed on scaled peak area tables, where peak areas were normalized
   by dividing each count matrix value by the maximum peak area and
   multiplying by 1 million, ensuring suitable scaling for integer
   conversion. The EM algorithm was run with 10 000 iterations.

snRNA‐Seq Data Processing

   Alignment was performed to this amended reference using 10x Cell Ranger
   v7.2, which employs the STAR sequence aligner. The reference genome was
   the human genome GRCm39 or the mouse genome mm39. Gene expression
   counts were determined using unique molecular identifiers (UMIs) for
   each cell barcode‐gene combination. Following alignment, cell barcodes
   were filtered to identify those that contain nuclei using the approach
   implemented in Cell Ranger, and only these barcodes were considered for
   downstream analysis. Cell Ranger count function was used to generate
   filtered matrices with “–force‐cell” was set as 10 000 for all liver,
   kidney, brain, and WAT tissues. To remove the nuclei with low quality,
   seurat (v5.1.0)^[ [303]^89 ^] R package was applied to construct 4
   objects according to the four tissues and filter nuclei with gene
   number over 200 and below 5000, as well as the ratio of mitochondria
   lower than 10% were maintained, and genes with at least one feature
   count in more than ten nuclei were used for the following analysis.
   Doublets were identified performing the DoubletFinder (v2.0.4)^[
   [304]^90 ^] R package with principal components (PCs) 1–20. nExp was
   set to 0.08 ×  nCells2/10000 and pN to 0.25. pK was determined through
   paramSweep_v3, and cells classified as doublets were removed before
   downstream analysis. In total, 43347, 48274, 49349, and 50929 cells
   were retained for liver, kidney, brain, and WAT, respectively. Next,
   Seurat's SCTransform function was applied to normalize and scale each
   sample, respectively. Replicates from the same tissue were integrated
   and removed batch effects using CCAIntegration method, and the top 50
   PCs were used to conduct unsupervised clustering analysis, and
   subsequently visualized by t‐distributed stochastic neighbor embedding
   (tSNE) dimensionality reduction (resolution 0.8, 0.5, 0.5, and 0.3 for
   liver, kidney, brain, and WAT, respectively).

Cell Type Annotation and Differentially Expressed Gene (DEG) Identification

   Automatic and manual marker gene identification were conducted to
   annotate the cell types in different tissues. scTyper^[ [305]^91 ^] was
   first used to automatically annotate cell type of each cluster based on
   a combined reference database consisting of scType default database and
   cell type markers in CellSTAR database.^[ [306]^92 ^] Clusters with
   unknown labels were subsequently used for manual annotation. Seurat's
   PrepSCTFindMarkers function was applied at first, and FindAllMarkers
   function (avg_log2FC ≥ 0.5 and p_val_adj < 0.05) was used to find
   expressed markers of each cell cluster. Each unknown cluster was
   annotated on the basis of the expression level of typical gene markers
   for specific cell types. The detailed information of marker genes for
   each cell type across the four tissues is listed in Table [307]S11
   (Supporting Information). Seurat's FindMarkers function (avg_log2FC ≥
   0.5 and p_val_adj < 0.05) was used to identify the significantly
   different expressed genes between CLP and sham conditions in both
   tissue and cell type levels.

Compass Analysis

   For metabolic analysis, COMPASS was run with default parameters and the
   RECON2 model using the integrated SCTransform snRNA data as provided by
   Seurat's Cellranger pipeline ([308]https://yoseflab.github.io/Compass/,
   –species parameter set to mus_musculus).^[ [309]^93 ^] The generated
   penalty reaction scores were subsequently log1p‐normalized for each
   single cell to receive probabilities of metabolic reaction activities
   per cell as described before.^[ [310]^93 ^] Reactions from “Glycine,
   serine and threonine metabolism”, “Lysine metabolism”,
   “Glycerophospholipid metabolism” and “Triacylglycerol synthesis”
   pathways were selected and further analyzed. Statistical significance
   (two‐sided wilcoxon test, FDR‐adjusted p ≤ 0.05) and effect size
   (Cohen's d measure of effect size) of change between CLP and sham mice
   were determined using the functions “wilcox_test” and “cohens_d” of the
   R package rstatix (v0.7.2, R version 4.3.0). To investigate identified
   metabolic reactions with most changes in CLP mice, these reactions were
   contextualized with information from the KEGG pathway maps hsa00310
   (Lysine degradation), hsa00561 (Glycerolipid metabolism), hsa00564
   (Glycerophospholipid metabolism) and hsa00260 (Glycine, serine and
   threonine metabolism, cf. Figure [311]4D).

Energy‐Metabolism Related Cell–Cell Communication Identification and Gene
Co‐Expression Network Construction

   To investigate the variation of energy metabolism profiles in the four
   tissues, 34 energy pathways were selected belonging to mus musculus
   KEGG Orthology (KO) from the KEGG database Table [312]S6 (Supporting
   Information), within which 13 genes specifically encoded by
   mitochondria were removed, leading to 1473 energy metabolic genes for
   downstream analysis. Principal component analysis (PCA) was performed
   based on average gene expression profiles of energy genes from the four
   different tissues. The four tissue‐specific Seurat objects were
   screened to four sub‐objects based on the 1473 energy metabolic genes.
   Cell‐cell signaling pathways in each cell type between CLP and sham
   groups across the four tissues were identified using the CellChat
   package (v2.1.2)^[ [313]^39 ^] in R based on the four sub‐objects. The
   meticulously curated database of ligand‐receptor (L‐R) interactions for
   mice was used for cell‐cell communication analysis, and the truncated
   mean was set to 5% to calculate the average energy metabolic gene
   expression by discarding 5% from each end of the data. CellChat takes
   snRNA‐Seq expression data and cell type annotations as input, and
   computes the number of each combination of sender and receiver cell
   types, as well as for each L‐R interaction.

   Energy metabolic gene co‐expression networks of the four tissues were
   constructed using hdWGCNA (v0.3.3)^[ [314]^94 ^] R package. After
   filtering energy‐related metabolic genes with 5% prevalence across
   cells in each tissue, 1044, 1029, 957, and 973 genes were used for
   network construction for liver, kidney, brain, and WAT, respectively.
   Pseudobulk metacells aggregating nuclei were created that belonged to
   the same sample and cell type setting a nearest‐neighbour parameter k
   of 30 and using 10 as the maximum number of shared nuclei between
   metacells. Modules were identified through the ConstructNetwork
   function. To identify highly connected hub genes, ModuleConnectivity
   function was used to calculate the eigengene‐based connectivity (kME)
   of each energy metabolic gene. Ten genes were presented in each module
   with the highest kME value as hub genes, and used Cytoscape (v3.10.1)^[
   [315]^95 ^] with edge‐weighted spring‐embedded layout to visualize the
   networks based on the topoligcal overlap matrices (TOMs) generated by
   GetTOM function in hdWGCNA.

Statistical Analysis

   R (version 4.3.0) and GraphPad Prism (version 10.3.1) were utilized for
   data analysis. Untargeted metabolomics/lipidomics data and snRNA‐Seq
   data were normalized by log2 transform and SCTransform function from R
   package Seurat respectively before downstream analyses. The values were
   presented as the means ± SDs. The data from the two unpaired groups
   were compared via a two‐tailed unpaired Student's t‐test. Statistical
   analyses requiring adjustment for confounding variables were performed
   using logistic regression and ordinal logistic regression models.
   Values of ^* p < 0.05, ^** p < 0.01, and ^*** p < 0.001 were considered
   as statistically significant. The Benjamini‐Hochberg procedure was
   applied to calculate the FDR to adjust P values for multiple hypothesis
   testing.

Conflict of Interest

   M.S. received grants from NewB, Apollo Therapeutics, and UCL Technology
   Fund and others from Abbott, Amormed, bioMérieux, Biotest, Deltex
   Medical, Fresenius, Mindray, NewB, Pfizer, Radiometer, Roche
   Diagnostics, Safeguard Biosystems, Shionogi, and Spiden outside of this
   project. M.S. is an unpaid advisor to Presymptom Health Ltd, hemotune,
   deepUll, and Santersus, and R.L. is a part‐time advisor to Presymptom
   Health Ltd.

Author Contributions

   L.X. and Z.Z. contributed equally to this work. Conceptualization,
   G.P.; methodology, L.X., S.S., Z.Z.; Investigation, K.D., L.X., S.S.,
   W.V., Z.Z.; writing—original draft, G.P., L.X., S.S., Z.Z.;
   writing—review & editing, all co‐authors; funding acquisition, G.P.;
   resources, G.P., M.B., M.S., R.L., S.W.; supervision, G.P.

Supporting information

   Supporting Information
   [316]ADVS-12-e04418-s002.docx^ (10.2MB, docx)

   Supporting Information
   [317]ADVS-12-e04418-s001.xlsx^ (250KB, xlsx)

Acknowledgements