Dynamic Lipidome Alterations Associated With Human Health, Disease and Ageing

nature metabolism
Article https://doi.org/10.1038/s42255-023-00880-1
Dynamic lipidome alterations associated

with human health, disease and ageing
Received: 28 December 2022 Daniel Hornburg 1,6, Si Wu 1,6, Mahdi Moqri 1, Xin Zhou1,
Kevin Contrepois 1, Nasim Bararpour 1, Gavin M. Traber 1, Baolong Su2,
Accepted: 28 July 2023
Ahmed A. Metwally 1, Monica Avina1, Wenyu Zhou1, Jessalyn M. Ubellacker1,3,
Published online: 11 September 2023 Tejaswini Mishra 1, Sophia Miryam Schüssler-Fiorenza Rose 1,
Paula B. Kavathas 4, Kevin J. Williams2,5 & Michael P. Snyder 1
Check for updates
Lipids can be of endogenous or exogenous origin and affect diverse

biological functions, including cell membrane maintenance, energy
management and cellular signalling. Here, we report >800 lipid species,
many of which are associated with health-to-disease transitions in diabetes,
ageing and inflammation, as well as cytokine–lipidome networks. We
performed comprehensive longitudinal lipidomic profiling and analysed
>1,500 plasma samples from 112 participants followed for up to 9 years
(average 3.2 years) to define the distinct physiological roles of complex
lipid subclasses, including large and small triacylglycerols, ester- and
ether-linked phosphatidylethanolamines, lysophosphatidylcholines,
lysophosphatidylethanolamines, cholesterol esters and ceramides. Our
findings reveal dynamic changes in the plasma lipidome during respiratory
viral infection, insulin resistance and ageing, suggesting that lipids may have
roles in immune homoeostasis and inflammation regulation. Individuals
with insulin resistance exhibit disturbed immune homoeostasis, altered
associations between lipids and clinical markers, and accelerated changes
in specific lipid subclasses during ageing. Our dataset based on longitudinal
deep lipidome profiling offers insights into personalized ageing, metabolic
health and inflammation, potentially guiding future monitoring and
intervention strategies.
Lipids are an important and highly diverse class of molecules that High-throughput omics technologies provide new avenues to
have critical roles in cell structure, cell signalling and bioenergetics. understand the molecular landscape of human physiology and its
Despite their critical roles in many biological processes, there is much dynamic changes during health and disease. To date, many studies
to be learned about the diversity of lipids in humans, how their com- have used next-generation sequencing owing to its accessibility and
position differs across people, and how they change over time at an cost-effectiveness1. Recently, mass spectrometry (MS) strategies
individual level and during disease. Such information is expected to have provided quantitative insights into the proteome2,3 at scale and
provide insights into biological processes such as ageing as well as the depth. Metabolites, which can also be investigated using MS, have been
possible roles of lipids in health and disease. studied to a lesser extent given their complex chemical diversity4,5.
Department of Genetics, Stanford University, Stanford, CA, USA. 2Department of Biological Chemistry, David Geffen School of Medicine, University
1
of California, Los Angeles, Los Angeles, CA, USA. 3Department of Molecular Metabolism, Harvard T.H. Chan School of Public Health, Boston, MA,
USA. 4Departments of Laboratory Medicine and Immunobiology, Yale School of Medicine, New Haven, CT, USA. 5Lipidomics Laboratory, University of
California, Los Angeles, Los Angeles, CA, USA. 6These authors contributed equally: Daniel Hornburg, Si Wu. e-mail: [email protected]
Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1578

Lipids comprise a major, heterogeneous family of biomolecules within and with an increased frequency of three to seven collections over
the metabolome and remain challenging to characterize owing to their 3 weeks during periods of illness (for example, RVI) or notable stress,
wide range of physicochemical properties6 and the relatively small as previously reported17,18. In addition to lipid profiling, we collected
number of lipidomics studies. 50 clinical laboratory measurements at each sampling timepoint along
Complex lipids can be divided into several classes and subclasses with medical records (Supplementary Data 2). Finally, because samples
that are distinguished by lipid head groups and linkages to different were collected during periods of stress and illness, we also profiled
aliphatic chains7. Lipids such as triacylglycerols (TAGs), diacylglycer- 62 cytokines, chemokines and growth factors in plasma at the same
ols (DAGs), phosphatidylcholines (PCs), phosphatidylethanolamines timepoints.
(PEs), ceramides (CERs), sphingomyelins (SMs) and cholesterol esters The human lipidome was characterized using a high-throughput
(CEs) each consist of a specific backbone architecture conjugated to quantitative lipidomics pipeline (Lipidyzer) consisting of a
various fatty acids (FAs). The attached FAs can vary in the number of triple-quadrupole mass spectrometer (Sciex QTRAP 5500) in con-
unsaturated bonds and their positions within acyl chains; together junction with a differential mobility separation (DMS) device15,16. This
with the backbone, FAs confer distinct physicochemical properties setup allows the identification and robust quantification (estimated
and physiological roles. Lipids carry out and regulate many key func- concentrations) of >1,000 lipid species across 16 subclasses (free FA
tions, including redox homoeostasis, energy storage, intracellular and (FFA), TAG, DAG, CE, PC, lysophosphatidylcholine (LPC), PE, alkyl ether
extracellular signalling, induction and resolution of acute and chronic substituent containing PE (PE-O), alkenyl ether (Plasmalogen) substitu-
inflammation8–10, and maintenance of electrochemical gradients across ent containing PE (PE-P), lysophosphatidylethanolamine (LPE), SM,
subcellular compartments. Abnormal lipid profiles (dyslipidaemia) PI, CER, hexosylceramide (HCER), lactosylceramide (LCER) and dihy-
have been associated with a range of diseases, including metabolic syn- droceramide (DCER); Fig. 1b). In addition, we observed the differential
drome, type 2 diabetes (T2D), cancer, nephropathy and cardiovascular behaviour of smaller and larger TAGs, which comprise ≤48 and ≥49
and neurodegenerative diseases, and may result from a combination of carbons across all FAs, respectively, and evaluated these separately in
factors such as genetic heterogeneity, lifestyle and, as recently shown, most analyses. For accurate quantification and to control for variance
inflammation related to coronavirus disease 2019 infection11–13. introduced during lipid extraction, we included a mix of 54 deuterated
One of the key roles of lipids in maintaining metabolic homoeo- spike-in standards for nine lipid subclasses at known concentrations.
stasis is to mediate the induction and attenuation of inflammatory Lipid species that were not present as labelled spike-in standards were
processes (for example, leukotriene, prostanoid and endocannabinoid normalized against the spike-in standards based on structural similarity
signalling)8,9,14. Because of the various roles lipids have in maintaining and signal correlation (described in Methods).
homoeostasis in humans, different lipid species or classes may influ- We randomized the samples separately for lipid extraction and
ence perturbations that induce acute inflammation (for example, MS data acquisition. After filtering (described in Methods), we quanti-
respiratory viral infections (RVIs)), as well as the resolution of inflam- fied, on average, 778 lipids in each sample and 846 lipid species across
mation, metabolic diseases (for example, T2D) and physiological pro- >1,600 samples (including quality control (QC) samples). We found
cesses (for example, ageing) that have been associated with changes the highest number (373) of lipid species in the large TAG subclass and
in the regulation of chronic inflammation. In light of the diverse roles the smallest number (4) in the DCER subclass (Fig. 1c). Lipids comprise
of lipids, it is important to understand their quantitative differences chemically heterogeneous molecules that exert a broad spectrum of
among individuals and their dynamics across phenotypes to character- biological functions ranging from bioenergetics to cellular signalling.
ize their potential roles in health and disease. This is partially visible in lipid subclass-specific abundance distribu-
Here, we characterize the lipidome dynamics in >100 human par- tions. Figure 1d shows the abundance distributions across more than
ticipants followed for up to 9 years, covering periods of health and four orders of magnitude and for each lipid subclass, and depicts two
disease, using an MS-based approach that allows a broad array of lipid distinct properties: (1) the median abundance of that subclass and (2)
types to be measured rapidly, quantitatively and rigorously15,16. We the abundance range across all interrogated plasma samples (includ-
identified distinct longitudinal lipid signatures that link lipid profiles ing healthy and disease timepoints). SMs and FFAs were observed, on
to the microbiome, ageing and different clinical pathophysiologies, average, as the most abundant subclasses, but they spanned a relatively
including insulin resistance (IR) and chronic and acute inflammation. small dynamic range. Other lipid subclasses, including LPCs, CEs and
Our results provide valuable insights into the associations of key lipids TAGs, had a lower median abundance but a much wider dynamic range.
and lipid subclasses with distinct metabolic health states in humans, Our study demonstrated high technical reproducibility. As antici-
and serve as a unique resource to the scientific community. pated, the 104 QC samples clustered distinctly (Extended Data Fig. 1);
the median coefficient of variation (CV) for the QC samples was low,
Results with values between 6.5% (small TAGs) and 20.7% (DAGs). In contrast,
Comprehensive lipid profiling of a longitudinal cohort CVs calculated across participants and sampling timepoints ranged
From a cohort of >100 participants with IR or insulin sensitivity (IS), we from 19.9% (SMs) to 91.4% (small TAGs), indicating sufficient assay
previously collected longitudinal molecular data comprising genome, reproducibility to discern biological differences. To ensure the highest
transcriptome, proteome, metabolome and 16S microbiome data robustness in our analysis, we focused on 736 lipid species for which (1)
across different timepoints (~1,000 in total17). Within this cohort, we QC CVs were <20% and (2) CVs in biosamples were larger than CVs in QC
explored various molecular signatures in health and disease and identi- samples. Except for FFAs, intraparticipant variance was consistently
fied hundreds of molecular pathways associated with metabolic, cardio lower than interparticipant variance, suggesting that individual lipid
vascular and oncologic pathophysiologies17,18. Here, we investigate signatures are distinct and stable over time (Fig. 1e). Interestingly, both
the dynamics of a largely unexplored molecular layer—the ‘plasma small and large TAGs and ester- and ether-linked PEs (PE versus PE-O and
lipidome’—and extend the longitudinal duration by 2 years to obtain PE-P) exhibited significant differences within their respective subclasses
a total of 1,539 samples. in terms of variance (Fig. 1e) and abundance distribution (Fig. 1d). This
To investigate lipidome alterations associated with health, dis- implies the existence of unique physiological and participant-specific
ease and lifestyle changes, plasma samples from 112 participants were differences, which may provide new insights into biological processes.
profiled at a median of ten timepoints across 2–9 years (average 3.2
years; one participant was sampled 163 times across 9 years; Fig. 1a, Lipid signatures are highly individualized
Supplementary Data 1 and 2 and Supplementary Figs. 1 and 2). Samples We first sought to investigate lipid abundance differences across indi-
were collected every 3 months when the participants were healthy viduals by characterizing the lipidome in ‘healthy’ baseline samples,

a c 400
373
Health record
No. of lipids
IR/IS 117
Immunome Laboratory measures
100 79
47
38 32
26 26 26
16 17 9 12 10 7 7 4
112 participants
Lipidome 0
9 years (plasma)
al AG
AG
PC
C
G
PE
-P
-O
E
PI
E
A
SM
H R
ER
DC R
ER
(average ~3.2 years)
LP
E
FF
LP
DA
PE
PE
C
C
LC
Sm e T
lT
rg
b
La
Complex lipids DAGs and TAGs
P R2–PO4–
d
+
FFAs R–NH3
FA1 FA2
O
FA1 R–OH R2–O
FA3 R–(CH3)3N
+
OH
FA2
DAG(FA1/FA2) TAG(FA1/FA2/FA3) CEs

FA(18:3) 2
log10(concentration)
No. of FAs
No. of carbons across
(nmol ml−1)
Carbon chain length all FAs in TAGs
No. of unsaturations Small TAGs: <49 C FA1
Position of unsaturations Large TAGs: >48 C CE(FA1)
0
PEs CERs
FA1am
P P
CER(FA1)
−2
FA1 FA2 FA1mod FA2
al AG
AG
PC
C
G
PE
-P
-O
E
PI
E
A
SM
H R
ER
DC R
ER
LP
E
FF
LP
DA
PE
PE(FA1/FA2) PE-O(FA1mod/FA2)
PE
C
C
LC
Sm e T
lT
rg
LCER(FA1)
La
FA1mod FA1am
P
P e 150
FA2
FA1
QC samples
LPE(FA1) PE-P(FA1mod/FA2) DCER(FA1) FA1am
HCER(FA1) Intraparticipant samples
FA1am
Functionalization: hydrolysed Interparticipant samples
(lyso) and non-hydrolysed 100
Alkyl ether substituent (–O) and Functionalization: LCERs and HCERs,
alkenyl ether (plasmalogen) amide bond
CV (%)
PCs PIs SMs

50
FA2
FA1 FA1am
FA1
P FA2 P P FA1 25
P
SM(FA1)
PI(FA1/FA2) 0
PC(FA1/FA2) LPC(FA1)
16 subclasses
al AG
AG
PC
C
G
PE
-P
-O
E
PI
E
A
SM
H R
ER
DC R
ER
LP
E
FF
LP
DA
Functionalization: hydrolysed
PE
PE
C
C
LC
Sm e T
lT
(lyso) and non-hydrolysed >1,100 lipid species

rg
La
Fig. 1 | Longitudinal lipidomics profiling. a, Profiling, using >1,500 biosamples, participants (median estimated concentration across all participant-specific
across 112 participants followed for up to 9 years. Dynamic changes in the samples), lipid species (846) spanned a dynamic range of more than four orders
lipidome were characterized in the context of health status and medication of magnitude, with distinct estimated concentration ranges for each lipid species
history and in comparison with the participants’ cytokine, chemokine and and subclass. e, Comparison of the CVs of QC (n = 104), intraparticipant and
metabolic profiles, as well as microbiome. b, Lipid subclasses investigated interparticipant samples. All boxplots report the 25% (lower hinge), 50% (centre
in this study. Lipid species, defined by a specific combination of backbone line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or
architecture and FAs, can be grouped based on their physicochemical properties. outside the hinge ± 1.5× the interquartile range (IQR). Outliers (beyond 1.5× the
c, We analysed 846 lipids (y axis) across multiple subclasses. d, Across all 112 IQR) are not plotted.
defined as samples from participants in the absence of any self-reported we examined which lipid subclasses show the largest interindividual
acute disease. This does not preclude latent, asymptomatic chronic differences and quantified how much of the variance observed for each
conditions such as prediabetes or potential undiagnosed conditions. lipid species can be attributed to interparticipant differences (Fig. 2a).
Overall, we analysed 802 healthy baseline samples derived from 96 Many lipids, in particular among TAGs, SMs, HCERs and CEs, showed
participants from whom we collected samples at two or more time- a high degree of participant-specific variance, in some cases >50%. In
points. The number of baseline samples per participant is shown in contrast, FFAs were found to have relatively low participant-specific
Supplementary Fig. 3; most participants had approximately ten healthy variation. To further illustrate participant specificity, we performed
visits, except one outlier with 52 healthy baseline samples. t-distributed stochastic neighbour embedding (t-SNE) on data from
In comparison with the transcriptome, proteome and general participants with >12 healthy visits, based on 100 lipids that we deter-
metabolome, lipid signatures can be highly personalized when assessed mined to be the most personalized (Fig. 2b). Most, but not all, samples
longitudinally19. To investigate the participant specificity of lipid pro- clustered by individual participants (Fig. 2b,c), showing that some
files for healthy sampling timepoints at timescales of months to years, lipids can comprise personalized signatures even across years.

a d
350
150 M1
WGCNA modules
10
No. of
–log10(FDR)
lipids
100 M2
20
M3
30
0
−log10(concentration) (nmol ml )
M4
M5
Lipid mean estimated

M6
Variance explained by
1.0
60 M7 0.5
Correlation
participant (%)
Lipid panel T2D Immune system Haematologic Electrolytes Renal Hepatic

0.5 0
TGL
CHOL
HDL
LDL
NHDL
CHOLHDL
LDLHDL
GLU
A1C
INSF
HSCRP
WBC
NEUT
NEUTAB
LYM
LYMAB
MONO
MONOAB
EOS
EOSAB
BASO
BASOAB
IGM
RBC
HGB
HCT
MCV
MCH
MCHC
RDW
PLT
Na
K
Cl
CO2
AG
Ca
CR
BUN
EGFR
UALB
ALCRU
UALBCR
TP
ALB
TBIL
ALKP
ALT
AST
GLOB
40
−0.5
0
20
−0.5
e f
−1
0 91 lipids 88 lipids 54 lipids M1 M2 M3 M4 M5 M6 M7
CE
HCER
SM
Large TAG
PE
CER
DAG
PE-P
LCER
PC
PI
Small TAG
LPE
LPC
DCER
PE-O
FFA
Large TAG
CE: 1 CE: 7 SM: 1
CER: 8 Small TAG: 26 Small TAG
FFA: 3 PI: 26 CE: 4 SM: 7 PC
DAG: 9
LPC: 7 HCER: 6
LPC
DCER: 2
b c 20 P < 2.2 × 10
−16 LCER: 1
M1
LPE: 5
M2
LCER: 4
M3
PE
PE: 2 PE-P
PE: 18
Distance
PE-O
PE-O: 8
10 PE-P: 23 LPE
PC: 39
Large TAG: 26 DAG
PI
t-SNE2
0
50 lipids 45 lipids 22 lipids 10 lipids FFA
a
r
te
tr
In
In
CE: 1 Large TAG: 12 CE

CER: 1
Participant DAG: 7 DAG: 8 SM
ZOZOW1T ZYHHR4Z CE: 4 M7
FFA: 2 LPC: 4 CER
ZN9YTFN ZLTUJTN
LPC: 2 LPE: 1 15
ZN0JE53 ZL9BTWF M4 FFA: 6 M5 M6 HCER
–log10(FDR)
ZM7JY3G ZL63I8R 10
PE-P: 6 LCER
ZLZQMEV ZKVR426 FFA: 6
PE-O: 2 5
ZLZNCLZ PC: 17 DCER
t-SNE1 PC: 16 PE: 2 Large TAG: 30 0
Fig. 2 | Interindividual differences in healthy baseline. a, Top, bar plot CHOL, total cholesterol; NHDL, non-HDL; CHOLHDL, cholesterol to HDL ratio;
showing the number of lipid species per class ordered by the variance explained LDLHDL, LDL to HDL ratio; GLU, glucose; INSF, fasting insulin; HSCRP, high-
by the participant factor; bottom, boxplot showing the variance explained by sensitivity CRP; WBC, white blood cell count; NEUT, neutrophil percent; NEUTAB,
participants in each lipid class (left y axis) and line graph showing the mean neutrophil absolute count; LYM, lymphocyte percent; LYMAB, lymphocyte
log10(estimated concentration) (red line, right y axis) of each lipid class. Variance absolute count; MONO, monocyte percent; MONOAB, monocyte absolute count;
decomposition analysis was conducted using n = 802 healthy samples. b, t-SNE EOS, eosinophil percent; EOSAB, eosinophil absolute count; BASO, basophil
clustering of 11 participants who contributed ≥12 healthy samples (n = 191), percent; BASOAB, basophil absolute count; IGM, immunoglobulin M; RBC, red
based on the 100 most personalized lipids. c, Intraparticipant distance, which blood cell count; HGB, haemoglobin; HCT, haematocrit; MCV, mean corpuscular
refers to the Euclidean distance between each pair of samples belonging to the volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular
same participant, and interparticipant distance, which refers to the distance haemoglobin concentration; RDW, red cell distribution width; PLT, platelet;
between the centroids from each pair of participants, for the t-SNE results. AG, albumin to globulin ratio; CR, creatinine; BUN, blood urea nitrogen; EGFR,
Boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) estimated glomerular filtration rate; UALB, urine albumin; ALCRU, aluminium
quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× to creatinine ratio, urine; UALBCR, urine albumin to creatinine ratio; TP, total
the IQR. Outliers (beyond 1.5× the IQR) are not plotted. The intraparticipant and protein; ALB, albumin; TBIL, total bilirubin; ALKP, alkaline phosphatase;
interparticipant distances were compared using a two-sided t test. d, WGCNA ALT, alanine aminotransferase; AST, aspartate aminotransferase; GLOB, globulin.
modules and their correlation (BH-adjusted FDR cut-off of 5%) with clinical e, Module composition for the WGCNA analysis shown in c, coloured by lipid
measures. Dot size depicts the BH-adjusted −log10(FDR). The colour scale subclass. f, Enrichment analysis results based on Fisher’s exact test, depicting
indicates the degree and direction of the correlation. TGL, total triglyceride; the BH-adjusted −log10(FDR) for enriched subclasses for each WGCNA module.
Key lipids are associated with important clinical measures in M1 and M2 have negative health associations based on conventional
Given the high variation in specific lipid classes among individuals, we clinical measures. In contrast, M7, which contained some FFAs and
next investigated the degree to which global lipidome profiles from LPCs, correlated with lower CRP and A1C levels. M3, which was enriched
healthy baseline samples are associated with clinical measures. We first for PE-P and PE-O, showed an association with higher levels of HDL
grouped lipids into seven modules by applying weighted gene correla- and lower levels of fasting insulin, and, compared with the dominant
tion network analysis (WGCNA) based on the similarity of lipid profiles T2D patterns in M1 and M5, demonstrated a signature that is generally
and then associated these seven modules with 50 clinical measures considered healthier.
while controlling for the covariates sex, age, ethnicity and body mass In addition, we investigated lipid–microbiome associations and
index (BMI; Fig. 2d–f and Supplementary Fig. 4). Controlling for these observed mostly negatively correlated lipids, including several TAG
covariates allows investigation of the direct associations between the species for the bacterial family Oscillospiraceae and (L)PE, PC and CE
lipid modules and clinical measures by ruling out potentially confound- for Clostridiaceae (Supplementary Fig. 5 and Supplementary Note 1).
ing effects from sex, age and BMI. Modules M1 and M5, which were These microorganisms are known to be abundant in the gut of par-
enriched for CER and PE, as well as small TAG (mainly M1) and large ticipants with IS in this cohort20, suggesting a potentially beneficial
TAG (mainly M5), showed the strongest positive association with T2D role of Clostridia in host lipid metabolism. Finally, an outlier analysis
measures, including glycated haemoglobin (A1C), fasting blood glucose identified participants with abnormally high or low lipid signatures,
and fasting insulin. Moreover, they showed a positive association with some of which we could correlate with underlying medical conditions
inflammatory markers, including high-sensitivity C-reactive protein such as hepatic steatosis (Supplementary Fig. 6 and Supplementary
(CRP) level and white blood cell count, and a negative association with Note 2). Overall, this global analysis suggests that many lipid sub-
high-density lipoprotein (HDL; ‘good cholesterol’) levels. Hence, lipids classes are associated with and potentially have a role (for example,

proinflammatory, anti-inflammatory or metabolic role) in clinical in energy regulation, cell signalling and immune homoeostasis exhibit
conditions, or may serve as biomarkers to stratify health states. broad dysregulation in IR.
Global lipidome disruption in IR Dynamic lipidome alterations during viral infections

As many clinical measures were associated with specific lipid subclasses, In addition to their role in chronic inflammatory and metabolic condi-
we next determined how the lipidome is influenced by the chronic tions such as IR, complex lipids are key mediators of acute inflamma-
metabolic disorder IR. IR commonly occurs in T2D and is a condition tory responses, for example, by releasing arachidonic acid (FA(20:4)).
in which cells, mainly muscle cells and adipocytes, are unresponsive Hence, complex lipids may be modified, released and activated during
to insulin, leading to high glucose levels in the blood. IR is often associ- RVIs and possibly vaccinations while also having important roles
ated with chronic inflammation as well as metabolic syndrome, includ- in these processes in an IR-dependent manner.
ing dyslipidaemia, and can lead to non-alcoholic fatty liver disease. Participants in this cohort were densely sampled during periods
Elucidating how the lipid network is perturbed in individuals with of RVI (72 distinct RVI episodes in 36 participants for a total of 390
IR is important to better understand the molecular mechanisms and samples) and vaccination (44 episodes in 24 participants for a total
prognosis of metabolic disorders. of 275 visits; Supplementary Fig. 9). For both RVI and vaccination,
IR can be diagnosed by measuring the steady-state plasma glu- we classified longitudinally collected samples as early-phase (days
cose (SSPG) level after endogenous insulin secretion is suppressed 1–6), later-phase (days 7–14) and recovery-phase (weeks 3–5) samples
and insulin and glucose are infused at fixed concentrations21. IR or IS (Fig. 4a). Using linear models, we identified 210 lipids that were sig-
(IR/IS) status was measured using SSPG assays in 69 participants, of nificantly changed during RVI (false discovery rate (FDR) < 10%) across
whom 36 and 33 were classified as having IR (SSPG >150 mg dl−1) and most subclasses (Fig. 4b), some of which have previously been impli-
IS (SSPG ≤150 mg dl−1), respectively. At the global level, we observed cated in acute inflammation. For instance, PEs have been reported
some capacity of lipid signatures to distinguish IR and IS (Fig. 3a). to have a critical role in apoptotic cell clearance and the aetiology of
Using regression analyses that controlled for age, sex, ethnicity and various viruses26. Another example is PIs, which bind to the respira-
baseline BMI, we resolved comprehensive differences between IR and tory syncytial virus with high affinity, preventing virus attachment to
IS across most lipid subclasses (Fig. 3b–d), such that more than half of epithelial cells27. LPCs, which we observed in increased abundance
the lipids (424) were significantly associated with SSPG levels. Lipids during inflammation, have been demonstrated to have therapeutic
and lipid subclasses that had a significant positive correlation with effects (after intraperitoneal administration in mice) in severe infec-
SSPG included TAGs and DAGs, which is consistent with our observations through immune cell recruitment and modulation28.
tions (Fig. 2d) and previous reports of higher levels of these lipids in To further investigate the lipid-associated processes that are
individuals with dyslipidaemia and metabolic syndrome22,23. We also involved in acute infection, we examined enriched lipids during infec-
observed subsets of CERs to have increased abundance, contribut- tion (Fig. 4c). We observed significant changes in specific lipid sub-
ing to the development of obesity-induced IR in mice and humans24 classes, including ether-linked PEs and TAGs containing saturated
(Fig. 3b,c), and making possible the lipid-based differentiation of IR FAs (SFAs) such as dodecanoic acid (FA(12:0)), following RVI. Dode-
and IS (Fig. 3a and Supplementary Fig. 7). canoic acid and palmitic acid (FA(16:0)) are proinflammatory com-
To investigate over- and under-representations of specific sub- pounds that upregulate cyclooxygenase 2 (ref. 29) and have key roles in
groups of lipids, we performed an enrichment analysis on positive and the activation of inflammatory responses. Overall, this suggests that
negative model coefficients. As TAGs comprise the largest subclass different key lipid subclasses may be important for various aspects of
of lipids in our data and could dominate the results, we performed viral biology as well as the immune response, and undergo significant
enrichment analyses separately for each lipid subclass and across all changes during RVI.
lipids (Fig. 3e). Enrichments were evaluated at the subclass level (Fig. 1b) To explore the choreography of lipid dynamics over time, we
and for FA composition (global saturation level and specific FAs). examined their trajectory during the different phases of RVI. The 210
Importantly, to our knowledge, new associations were found, including significantly changed lipids were mapped to four major clusters, using
an association of ether-linked PE (PE-P)—in contrast to PE in general— a hierarchical clustering approach based on the Euclidean distance
with lower SSPG levels. Ether-linked PEs are involved in cell signalling between lipid species as the similarity measure (Fig. 4d), and these
and can act as antioxidants25. Together with increased levels of TAGs main clusters were linked with clinical measures (Fig. 4e). Except for
with higher SSPG levels, reduced PE-P levels suggest IR-associated the green cluster, which was significantly enriched for PC, all profiles
inflammation and may indicate a PE-mediated link between oxidative showed decreased levels during infection. The blue cluster was sig-
stress, inflammation and IR. nificantly enriched for small TAG and showed sharply decreased lipid
In Fig. 2d, we demonstrated lipid modules that correlate with a levels in early RVI, with a rapid recovery that correlated with clinical
variety of clinical measures. As it is well documented that IR affects measures of total lipids, including cholesterol and LDL. This indicates
both lipid regulation (dyslipidaemia) and clinical phenotypes, we a metabolic shift in early infection, potentially to support increased
investigated whether associations between lipids and clinical measures energy metabolism. The orange cluster, enriched for LPC, large TAG
are affected by the IR/IS status, which would have important health and ether-linked PE, showed a similar profile to the blue cluster but
implications for these participants (Fig. 3f and Supplementary Fig. 8). a delayed recovery to baseline levels. Lipids in this cluster were posi-
Intriguingly, we found many significant differences in both the effect tively correlated with the clinical lipid panel and blood glucose lev-
sizes and the direction of the correlation of lipid signatures and clinical els but negatively correlated with CRP level and neutrophils. This
measures between participants with IR and those with IS. For instance, suggests that early changes in energy metabolism (reduction in lipid
in IR unlike in IS, the low-density lipoprotein (LDL) to HDL ratio was and blood glucose levels) are coupled with increased inflammation
positively associated with the ether-linked PE-P and PE-O, and nega- (reduction in LPC and ether-linked PE levels, as well as an increase in
tively associated with LPE (Fig. 3f). Moreover, in participants with IR the CRP level and neutrophils) followed by a slow attenuation of inflam-
and IS, we observed opposite correlations of immune and blood cell mation at later stages of RVI. The purple cluster, which was enriched
measurements with lipid subclasses, including A1C–SM, SSPG–CER for FFA, represents slowly decreasing lipid levels and reached the
and SSPG–PI, as well as immunoglobulin M–PE-P/PE-O, monocyte– lowest point during the RVI recovery phase before reverting to the
PE-P/PE-O, eosinophil–TAG and white blood cell–PI (Fig. 3f). Overall, baseline levels. In particular, the late-stage correlation with immune-
these data indicate that, depending on the IR/IS status, lipid–clinical related parameters (CRP level, lymphocyte count) suggests that
measure associations can vary significantly and the key lipids involved reduction in the levels of some lipids in this cluster may relate to a

a e f IR/IS
10 TAG (all)
20 TGL
log2(odds)
5 CHOL
Diabetes Lipid panel

Aggregated
0 PE (all) HDL
subclasses LDL
−5 NHDL
−10 CER (all) CHOLHDL
LDLHDL
PC2 (12.468%)
0
Small TAG GLU
A1C
INSF
Large TAG SSPG
HSCRP
P = 2.2 × 10
SM WBC
−20 NEUT
Immune system
NEUTAB
PE-P LYM
IS Detailed LYMAB
subclasses MONO
−16
PE
IR MONOAB
EOS
−40 LPC EOSAB
–20 0 20 Density BASO
BASOAB
LCER IGM
PC1 (29.062% )
RBC
Electrolytes Haematologic
HCER HGB
HCT
b 80
82 80
79 FA(24:1)
MCV
MCH
FA MCHC
composition RDW
60 FA(20:3)
Significant
54 PLT
lipids (%)
50
Na
40 35 FA(18:1) K
31
22 Cl
20 18 17 CO2
12 12 10
PC
PE
All
6 AG
c
0 0 0 Ca
0 0.6 CR
200 IR IS BUN
0.4 EGFR
β coefficient
Renal
100 IS Correlation 0.2 UALB
ALCRU
IR/IS
0 IR 0 UALBCR
All TP
−100 −0.2 ALB
Hepatic
TBIL
−200 −0.4 ALKP
ALT
−0.6 AST
Small TAG
LCER
Large TAG
DAG
HCER
FFA
LPC
CER
PE
PE-P
SM
PI
PE-O
PC
CE
DCER
LPE
GLOB
Small TAG
Large TAG
DAG
PE
PE-O
PE-P
LPE
CER
DCER
HCER
LCER
PC
LPC
PI
SM
CE
FFA
d 68
4 13
4
17
72
9 LCER 4 5
CDP-Cho CPT 1
CDP-Eth PC LPC CE
GPAT LPAAT PAP DGAT SPT CerS1–6 DES

G-3-P LPA PA DAG TAG FFA Acetyl-CoA DCER CER HCER
13 374 6 3
EPT CDS 2
6 24 470 17 2 9 6
11
96
11 TCA cycle 7
3
LPE CDP-DAG Energy metabolism

SMase SM synthase
PSS PIS
PGS
PSD Lipid higher in SM
PE PS PG PI
IR IS Unchanged 1
6 8
61 4 26 3 7
51
23
Fig. 3 | IR- and IS-associated lipid signatures. a, Principal component analysis diphosphoethanolamine; CPT, choline phosphotransferase; EPT, ethanolamine
comparing IR and IS. The density plot on the right indicates the distribution of phosphotransferase; GPAT, glycerol-3-phosphate acyltransferase; LPAAT,
eigenvectors for each data point along the second principal component (PC2). lysophosphatidic acid acyltransferases; PAP, phosphatidate phosphatase; DGAT,
Eigenvector comparison between IR and IS was conducted using a two-sided DAG acyltransferase; G-3-P, glyceraldehyde-3-phosphate; CDS, CDP-diacylglycerol
t test. b, Regression analysis (n = 69): 424 of 736 lipids had significant correlations synthase; PSD, PS decarboxylase; PSS, PS synthase; PGS, PG synthase; PIS, PI
with SSPG (BH FDR < 5%; corrected for age, sex, ethnicity and baseline BMI). synthase; SPT, serine palmitoyl transferase; CerS, ceramide synthase; SMase,
c, Boxplot depicting regression coefficients for the respective lipid classes by sphingomyelinase; DES, dihydroceramide desaturase; Acetyl-CoA, acetyl coenzyme
using 69 samples for which the SSPG level was measured at the visit. Larger A; TCA, tricarboxylic acid. e, Enrichment analysis (Fisher’s exact test) performed
coefficients indicate stronger associations with higher SSPG levels. Colour on the coefficients from SSPG regression. Enriched annotations were calculated
indicates distributions for which the 25th or 75th percentile is positive or for positive coefficients with BH FDR < 10% (positive log2(odds)) and negative
negative. Boxplots report the 25% (lower hinge), 50% (centre line) and 75% coefficients with BH FDR < 10% (negative log2(odds)). For enriched annotations,
(upper hinge) quantiles. Whiskers indicate observations equal to or outside a BH FDR cut-off of 5% was applied. f, Correlations between clinical measures and
the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. lipid profiles for IR and IS. Correlations are shown when the correlations in IR and
d, Proportional differences between IR and IS detected in participants. Centre IS were significantly different and the absolute Δ correlations in IR and IS were >0.2.
numbers indicate the total number of lipids in each class. Enzyme names In addition, the overall correlations between lipids and clinical measures across IR
are shown in red. CDP-Cho, cytidine diphosphocholine; CDP-Eth, cytidine and IS are depicted when the aforementioned two criteria were met.
temporary strong immunosuppression to attenuate early- to mid-phase We next investigated whether individuals with IR and IS respond
inflammation and promote a return to homoeostasis. Overall, our differently to infections and vaccination (Fig. 4f and Supplementary
data suggest links between differential responses of lipids and Fig. 10). Through a longitudinal differential analysis, we found distinct
specific biological roles, with rapid shifts in energy metabolism to longitudinal profiles for IR and IS. We observed a higher abundance of
support inflammation early in infection and possible attenuation several FFAs during the early stage of RVI and greater elevation of PC
in later stages. Reflecting important global shifts in cell signalling, levels in the middle to late stage in participants with IR than in par-
metabolism and inflammation during RVI, these lipids may allow the ticipants with IS. In contrast, TAGs and some PEs were differentially
assessment of disease severity and prognosis or offer an opportunity elevated in IS compared with IR throughout the middle to later stages
for therapeutic intervention. of infection. The IR/IS-specific FFA and TAG responses may reflect the

a c d f
Purple z score
5 CER(16:0)
Infection PI (all)
log2(odds)
4 1.5
Aggregated 1.0 CER(20:0)
Before (healthy) Early Late Recovery After (healthy) 3 PE (all)
2
subclasses 0.5 FFA(14:0)
PC (all) 0 FFA(14:1)
1
−0.5
–186 d1 d7 d15 d40 d186 Small TAG −1.0 FFA(16:0)
Days Green −1.5 FFA(16:1)
PI
FFA(22:6)
PEP Large TAG HCER(20:0)
b Significantly associated lipids PEO Detailed Small TAG
PC
LPC(22:6)
LPE(18:1)
PC subclasses
300 LPC PC(15:0/20:4)
LPE Blue PE PC(18:0/22:4)
No. of lipids
LPC
PE-P PC(18:1/20:3)
a PE-O PC(18:1/22:4)
CE LPE PC(18:1/22:5)
100 DAG
SFA Saturation PC(18:2/18:3)
a PI
PC(18:2/20:5)
39 42 FA(20:5) FFA
34 a Orange PE(18:2/18:2)
a FA(20:0) CE
23 17 a a a FA SM
PE-P(16:0/22:6)
13 8 a
8 12 5 5 FA(18:0) composition CER
PI(18:1/20:3)
1 1 0 1 0 1
0
FA(16:0) HCER TAG 36:0(FA(12:0))
LCER TAG 38:0(FA(12:0))
FA(12:0) TAG 40:0(FA(14:0))
DCER
Large TAG
Smal TAG
PC
PE-P
PI
DAG
PE
CE
FFA
LPC
PE-O
CER
SM
HCER
LPE
LCER
DCER
TAG TAG 44:0(FA(18:0))

All
PC
Subclasses
Before
Early
Late
Recovery
After
Before
Early
Late
Recovery
After
TAG 46:1(FA(18:0))
TAG 48:4(FA(20:4))
TAG 52:1(FA(16:1))
TAG 52:1(FA(20:0))
e 0
TAG 54:0(FA(18:0))
TAG 54:1(FA(18:0))
Orange
−log10(FDR)
2.5
TAG 54:1(FA(18:1))
5.0
Green TAG 54:1(FA(20:0))
7.5
TAG 54:2(FA(20:0))
≥10
Purple TAG 56:1(FA(18:1))
0.50 TAG 56:3(FA(20:2))
Correlation
Blue
Subclass
0.25 Days in infection episode
0 OmicsLonDA test statistics

Lipid panel Diabetes Immune system Haematologic Electrolytes Renal Hepatic
−0.25
TGL
CHOL
HDL
LDL
NHDL
CHOLHDL
LDLHDL
GLU
A1C
INSF
HSCRP
WBC
NEUT
NEUTAB
LYM
LYMAB
MONO
MONOAB
EOS
EOSAB
BASO
BASOAB
IGM
RBC
HGB
HCT
MCV
MCH
MCHC
RDW
PLT
Na
K
Cl
CO2
AG
Ca
CR
BUN
EGFR
UALB
ALCRU
UALBCR
TP
ALB
TBIL
ALKP
ALT
AST
GLOB
−3 −2 −1 0 1 2 3
Fig. 4 | RVIs and vaccination. a, Longitudinal sampling at five timepoints their corresponding profiles in each cluster. e, Associations of lipid profiles in
during RVIs: before infection (healthy), early event, late event, recovery and RVI and clinical measures. Depicted are correlations between the identified
after infection (healthy). b, Lipid class breakdown for all detected lipids. Dark lipid clusters (d) and 50 clinical laboratory measures (BH FDR cut-off of 5%). Dot
green depicts 210 significantly changed lipids throughout RVI. aEnriched size depicts −log10(FDR); colour scale represents the correlation direction and
subclass. Fisher’s exact test was used for the lipid class enrichment analysis of the degree. f, Differential profile of lipids that were significantly changed during RVI,
significant lipids (BH FDR for each lipid subclass: CE, 3.35 × 10−4; CER, 0.95; DCER, comparing IR and IS. For each lipid feature, the shaded blocks demonstrate the
0.49; HCER, 0.87; LCER, 1; DAG, 1; FFA, 0.56; LPC, 6.32 × 10−8; LPE, 8.40 × 10−3; time intervals during which the corresponding lipid was significantly different
PC, 3.01 × 10−4; PE, 0.27; PE-O, 1.01 × 10−3; PE-P, 1.00 × 10−8; PI, 7.65 × 10−5; between IR and IS. The orange shaded blocks representing the lipid profiles at
SM, 1; large TAG, 1; small TAG, 3.66 × 10−2). c, Lipid enrichment analysis for this time interval are dominant (with higher lipid levels) in IR, and the blue shaded
significantly changed lipids during RVI, across (left column) and within classes. blocks representing the lipid profiles at this time interval are dominant in IS.
d, Trajectory analysis of the 210 significantly changed lipids following RVI and
altered energy metabolism in IR, whereas differences in PCs and other owing to biological ageing or the period during which the cohort aged,
lipid classes may indicate changes in immune-associated signalling or other cohort effects. Periods and cohorts are social contexts affect-
pathways. Importantly, we found that the patterns after vaccination ing individuals and are inherently and mathematically confounded by
were distinct from those during infection (Supplementary Fig. 10). the individuals’ age34. These comprise environmental factors differ-
For example, fewer TAG species showed elevated levels in IS, whereas a ently affecting young and old participants, due to them being born in
distinct population of LCERs were upregulated in IR after vaccination. As different generations, and include generation-dependent exposures
individuals with T2D associated with IR often exert a more compromised that may also affect lipidome compositions (for example, diet, lifestyle
immune response to RVI17, such changes may be biologically significant. and/or diseases) rather than actual age. However, we note that the
longitudinal nature of our data better enabled us to eliminate some
Altered ageing of participants with IR biases and focus on the same individual across time34. Furthermore,
Ageing increases the risk of cardiovascular diseases and is accompanied we previously did not observe major dietary changes in the cohort18.
by a variety of diseases including T2D30,31 and chronic inflammation32. To identify lipid changes that occur with ageing in our longitudinal
In our study, the participants spanned an age range of 20–79 years cohort, we used a linear model that estimates relative lipid changes
(healthy timepoints, median 57 years) and were longitudinally sampled as a function of the change in age (Δage model) while also controlling
on average over 3 years (Fig. 5a). Across the cohort, we observed an for sample storage length and BMI. With this model, we determined
increase in BMI with higher age (Fig. 5b). We previously identified the ‘ageing’ effect (β coefficient) for each lipid subclass (Fig. 5c) and
age-associated molecular signatures in a subset of this cohort, including across lipid species (Fig. 5d).
inflammation (acute-phase proteins), blood glucose and lipid metab- We found that the levels of most lipid subclasses increased with
olism (A1C, apolipoprotein A-IV protein), but had not investigated ageing, most prominently CERs (LCER, HCER, DCER), SMs, LPCs and
age-associated lipidome changes33. To identify lipids and pathways CEs, with some of the observed variance suggesting more complex
that change with ageing and may be associated with the development lipid–ageing dependencies (Fig. 5c). A general increase in the levels
of age-related pathologies, such as chronic low-grade inflammation, of multiple lipid species and subclasses is consistent with previous
we investigated longitudinal changes in the lipidome. In cross-sectional observations35,36. Intriguingly, the levels of TAGs generally increased
studies, lipid content can differ across participants with different ages, over time (Supplementary Fig. 11), but this trend disappeared when

a 50
d e
TAG (all) Female
10
Male
log2(odds)
5 PE (all) Aggregated 0.02
0 subclasses
No. of visits
PC (all)
−5
∆age model coefficients

10 −10 CER (all) 0.01
Large TAG
SM 0
0
PE-P
PE −0.01
20 40 60 80 LPE
Detailed
Age (years) LPC subclasses −0.02
b 50
LCER
HCER −0.03
CER
LPC
HCER
CER
DCER
CE
LPE
LCER
SM
PE-P
FFA
PE-O
PC
PE
PI
DAG
Large TAG
Small TAG
CE
SFA
40
R = 0.37 PUFA
BMI (kg m−2)
P = 0.00038 Saturation
Omega-3
MUFA f IR
0.02
IS
30 FA(26:1)
FA(24:1)
∆age model coefficients

FA(24:0) 0
FA(22:6)
20 FA(22:4)
20 40 60 80 FA(22:0) −0.02
FA
Age (years) FA(20:5) composition
c 4
FA(20:0)
−0.04
LPC FA(18:3)
HCER
Change (%)
3 FA(18:2)
CER
FA(18:1)
2 CE
SM FA(16:1)
HCER
LCER
CE
LPC
SM
DCER
CER
FFA
PC
PE-P
LPE
PE-O
PI
PE
DAG
Large TAG
Small TAG
1
LCER
PC
PE
TAG
All
0
0 1 2 3 4 5
Year
Fig. 5 | Age-associated changes in the lipidome. a, Median ages, age range Δage model coefficients at the lipid species level with a BH FDR of <10%) and
(horizontal lines) and number of visits (y axis) of 90 healthy participants. Violin those associated with lower age (negative log2(odds), blue, determined for
plot shows the distribution of age within the cohort. Inner boxplot reports the all negative Δage model coefficients at the lipid species level with a BH FDR of
25% (left hinge), 50% (centre line) and 75% (right hinge) quantiles. Whiskers <10%). Enrichments were calculated independently within lipid subclasses, as
indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers well as across all lipid species (‘all’). log2(odds) values are depicted for significant
(beyond 1.5× the IQR) are not plotted. b, Correlation of median BMI and median associations with lower or higher age (BH FDR < 5%). Infinite log2(odds) values are
age across healthy participants. Vertical lines depict the BMI range for each imputed with 0.5× the mean value of positive/negative log2(odds) determined
participant across all collected healthy samples. Regression line (red) from across all data. MUFA, monounsaturated FA. e, Δage coefficients (ageing–sex)
a linear model is shown with the 95% confidence band (grey). c, Significantly of individual lipid subclasses for male and female participants, controlling for
(BH FDR < 5%) changed lipid subclasses (percentage change for the summed sample storage length and BMI. f, Δage coefficients (ageing–IR/IS) of individual
untransformed concentration of respective lipid species) with ageing across lipid subclasses for IR and IS, controlling for storage length, BMI and sex. For
5 years based on the Δage model controlling for BMI and sample storage length. e and f, data are presented as the mean of estimated coefficients ± s.d.,
d, Fisher’s exact test enrichment analysis comparing physicochemical properties determined using an ordinary least-squares regression test.
associated with higher age (positive log2(odds), red, determined for all positive
controlling for BMI. We performed an enrichment analysis on the significant ageing biomarker in blood plasma, suggesting a more sys-
Δage model coefficients at the species level and observed a shift in the temic decrease. Through desaturation and elongation, linoleic acid is
physicochemical properties of lipids associated with ageing, including metabolized to arachidonic acid (FA(20:4)), which we found to increase
increased levels of SFAs and monounsaturated FAs, whereas the levels in abundance with increasing age when we applied less stringent filter
of polyunsaturated FAs (PUFAs) were reduced (Fig. 5d). This pattern ing (Supplementary Fig. 12), further substantiating a general shift
has been previously associated with dyslipidaemia and inflammation37, towards inflammation with ageing. Furthermore, large and small TAGs
underlining progressive deterioration of metabolic health during showed distinct patterns, underlining the different functional roles
ageing. We also observed depleted levels of (beneficial) omega-3 FAs. along the TAG spectrum. Interestingly, the levels of LPCs, which have
In particular, the levels of docosahexaenoic acid (FA(22:6), TAG) and been implicated in cardiovascular diseases and neurodegeneration41
eicosapentaenoic acid (FA(20:5), PE) decreased with ageing. These and some of which are anti-correlated with CRP (Fig. 2), increased with
omega-3 FAs have been indicated to have beneficial health effects ageing, further underlining their pleiotropic role in human health.
by lowering plasma cholesterol levels and serving as precursors for We also observed a strong sex dimorphism for multiple subclasses
mediators that resolve inflammation, such as resolvins, protectins and (Fig. 5e). Beyene et al. reported sex-associated differences in lyso- and
maresins38,39. In addition, decreased levels of linoleic acid (FA(18:2)) ether-phospholipid metabolism42, which we confirmed in our study.
have been reported in aged skin40; our data show that this is also a In addition, we observed sex-associated differences for small TAGs as

a prominent signature in ageing, with higher levels in men and lower and small TAGs. In contrast, large TAGs were negatively associated with
levels in women. IL-6 and IL-10. Moreover, we observed a hub of negative associations
We next investigated the extent to which IR alters molecular age- between TAGs containing FA(22:5) and multiple cytokines, including
ing signatures and observed that participants with IR had larger coef- the anti-inflammatory IL-10 and the proinflammatory IL-23, as well as
ficients for multiple subclasses, including HCER, LCER, SM and CE, than IL-6. Enrichment of TAG subclasses for positive and negative associa-
participants with IS. Larger coefficients indicate that ageing-related tions within both proinflammatory and immunoregulatory cytokines
changes may be accelerated in IR versus IS (controlling for storage suggests that TAG subclasses (in terms of both the length and satura-
length, sex and BMI; Fig. 5f). In contrast to previous reports that did tion of the acyl chain) have distinct roles in immunoregulation and
not distinguish IR status35, our study identified a negative association signalling.
between DAGs and ageing in participants with IR. Intriguingly, higher In Fig. 2, we found that some LPCs were associated with anti-
DAG levels are commonly linked to dyslipidaemia and IR37,38; however, inflammatory, hence healthier, signatures. Here, LPCs were positively
similar to TAGs (see above), DAGs may have a stronger association with associated with several growth factors, such as epidermal growth fac-
BMI, which was controlled for in the model. Moreover, PI and PE showed tor (EGF), vascular endothelial growth factor (VEGF) and brain-derived
opposite ageing effects in participants with IR and IS, which suggests neurotrophic factor (BDNF), and resistin. VEGF is involved in promoting
IR-specific changes in phospholipid metabolism with ageing. In sum, angiogenesis, whereas BDNF and EGF promote cell proliferation, with
the composition of many lipid subclasses (that is, degree of unsatura- BDNF having a cardinal role in neurogenesis and plasticity52. In addition,
tion, omega-3 FAs, large TAGs, ether-linked PEs) changes significantly LPCs were positively associated with the soluble CD40 ligand (sCD40L),
with ageing, a process that—for some lipid subclasses—differs between which is secreted by activated T cells and platelets during inflamma-
the sexes and is distinctly accelerated in the presence of IR. tion, as well as with the inflammatory cytokine IL-1⍺ and the adipose
tissue-specific secretory factor resistin, which induces other cytokines
Specific associations of lipids with cytokines and chemokines and has been suggested to contribute to a chronic proinflammatory
Given the importance of cytokines, chemokines and growth factors cascade in T2D53. Together, LPCs demonstrated contrasting associa-
in diverse biological processes, we characterized their relationship tions, including some anti-inflammatory and tissue repair as well as pro-
to lipids across homoeostasis and various pathophysiological disease inflammatory signatures. These associations may highlight a difference
processes in our longitudinal cohort. We investigated the degree to between chronic inflammation (for example, mediated by factors such
which the abundance of a particular lipid predicts the level of cytokines, as resistin during T2D) and acute inflammation (for example, during
chemokines or growth factors, controlling for BMI, sex, ethnicity and an infection), which is strongly associated with high CRP levels. It may
multiple measurements per participant as random effects across also reflect that both inflammatory and anti-inflammatory mediators
all samples and timepoints for which both measures were available are present in amounts that regulate a response so that it is effective
(1,180 samples). Overall, we found 1,245 significant (FDR < 5%) positive but not excessive. Moreover, PCs containing linoleic acid (FA(18:2))
and negative associations between a majority of lipids (580) and were negatively associated with the chemokines CXC motif ligand 9
40 cytokines, chemokines and growth factors (Fig. 6a). (CXCL9; also known as MIG (monokine induced by interferon-γ (IFNγ)))
The largest numbers of positive associations were between granu- and CXCL10 (also known as IP-10 (IFNγ-induced protein 10 kDa); Fig.
locyte–macrophage colony-stimulating factor (GM-CSF) and TAGs 6b,f). CXCL9 and CXCL10 are induced by IFNγ to recruit cells to sites of
and between leptin and TAGs (Fig. 6a and Supplementary Fig. 13). The inflammation; they bind to the same chemokine receptor, CXCR3. This
adipokine leptin regulates caloric intake and is commonly present in association suggests that these lipids may affect immune cell migra-
elevated levels in obesity, contributing to the associated inflammatory tion during inflammation, in addition to their immune modulation
state43. Its amount in the blood correlates with the amount of adipose role that we observed during RVI (Fig. 4). Overall, our multiomics data
tissue. Its receptor is expressed in the hypothalamus, hippocampus outline complex associations between cytokines and lipid subclasses
and many immune cells; thus, it also acts as a neuroregulator and an as well as differential associations of lipids with specific FA composi-
immunoregulator43,44. The cytokine GM-CSF, originally defined as a tions, suggesting distinct roles ranging from immune activation to
haemopoietic growth factor, has other biological roles, including exert- immunosuppression.
ing proinflammatory effects45–47. These signatures are consistent with
the inflammatory effect of the high TAG levels that we observed and Discussion
are also found as a consequence of a high-fat diet, obesity and hepatic Until recently, most omics studies have focused on transcriptomics,
adiposity48,49. The pleiotropic cytokine interleukin-6 (IL-6), whose proteomics and, more recently, metabolomics, which is more closely
inflammatory and anti-inflammatory effects are context dependent50, associated with many phenotypes54. However, lipidomics remains
together with the anti-inflammatory cytokine IL-10 (ref. 51), showed largely underexplored despite lipids’ important roles in cell signalling,
negative associations with some TAGs and clustered distinctly from the cell structure and energy management. The lipidome has been difficult
positive TAG–leptin and TAG–GM-CSF associations, suggesting func- to study owing to its complexity and the fact that it is derived from
tional differences among different TAG species in immunoregulatory both endogenous and exogenous factors, such as the microbiome and
networks (Fig. 6a). TAGs showed the overall highest number of associa- lifestyle (diet and physical activity). Investigating the dynamic range
tions with leptin and GM-CSF, whereas lipids from other subclasses, of lipid changes, including which lipids change with acute or chronic
such as PE, PC and DAG, were also positively associated (Fig. 6b,c and conditions over what period, may reveal markers of early disease onset
Supplementary Fig. 13). In contrast, lyso species of PE and PC (Fig. 6d,e) and progression as well as mechanistic insights that can be used to
showed fewer associations with and less central roles for GM-CSF and develop better and personalized treatments.
leptin. Overall, these results suggest regulatory commonalities across By following participants for up to 9 years, we identified highly
lipid classes (for example, positive associations of TAGs, DAGs, PCs and participant-specific lipids and lipid subclasses (Figs. 1e and 2a–c),
PEs with leptin) and differences within subclasses for proinflammatory functional modules of lipids that map to clinical measures at base-
and immunoregulatory pathways. line and throughout diseases (Figs. 2d,f and 3f), and lipid outlier
To elucidate the extent to which specific subsets of lipid species are signatures that may be predictive of diseases such as hepatic steato-
associated with cytokines and chemokines, we performed an enrich- sis (Supplementary Fig. 6). Across perturbations such as IR (Fig. 3),
ment analysis (Fig. 6f). Overall, we observed strong associations of viral infections (Fig. 4), ageing (Fig. 5) and cytokine–lipid associations
FAs with distinct cytokines. For instance, positive leptin–TAG associa- (Fig. 6), we observed distinct behaviours among many lipid subclasses,
tions were significantly enriched for SFA, the polyunsaturated FA(18:3) such as ether- and ester-linked PEs, small and large TAGs, and lipids

a b HGF c IL-2
IL-1Ra
IL-17A
HGF
IP-10
Leptin IL-12p70
Eotaxin EGF
GM-CSF
IL-21 MIG
IP-10 MIG
VEGF ICAM1 CD40L
MIP1β IL-23
IL-10
GM-CSF EGF
Eotaxin
MIP-1α Leptin
IL-1α GROα IL-18
TNFβ TRAIL GM-CSF
Leptin IL-6
HGF IL-12p40 SDF1α
IL-7 FGFb
MIG PC PE
IL-2 IL-21 IL-15
d e
Eotaxin IL-18 SCF IL-13
IL-1Ra PDGF-BB BDNF
IL-8 IL-8 TNFβ
SDF1α
IL-17A IL-23 Resistin
IP-10 RANTES MIP-1α MIG
IP-10 IL-4
EGF IL-13 CD40L IP-10
CD40L Leptin
Resistin Leptin
IL-1α IL-6 IL-21
ICAM1 SCF IL-1α GM-CSF
IL-7 IL-23 IL-1α
SDF1α FGFb
IL-12p70 IL-9 IL-8 Eotaxin EGF VEGF EGF
Eotaxin
VEGF TNFβ TRAIL IL-15
BDNF TNFβ IL-23
IL-4 FasL FasL
IL-10 GROα IL-12p40
GM-CSF
MIP1β CD40L
IL-7
Positive association IL-9
Negative association All LPC LPE
GROα
f TAG CER PE PC log2(odds) (iii) Across all

TAG (all)
Large TAG HCER PE-P LPC PI (all) Aggregated
Small TAG LCER PE-O PE (all) subclasses
CE
−10
−5
0
5
10
PC (all)
DAG DCER LPE (ii) Within non-TAGs CER (all)
PI
FFA SM PI PE-P
PE-O
PC
LPE Detailed
LPC subclasses
LCER
(i) Within TAGs HCER
FFA
Small TAG
Large TAG
SFA
PUFA Saturation
Omega-6
Omega-3
FA(24:1)
FA(22:6)
FA(22:5)
FA(22:4)
FA(20:5)
FA(20:4) FA
FA(20:3) composition
FA(20:2)
FA(18:3)
FA(18:2)
FA(17:0)
FA(16:0)
FA(14:0)
Lipid class
SDF1α
RANTES
PDGF-BB
MIG
Leptin
IL-23
IL-10
IL-6
GROα
GM-CSF
FGFb
FasL
Eotaxin
PI_IP-10
PE_IL-2
PE_HGF
PE_GM-CSF
PE_EGF
PC_VEGF
PC_SDF1α
PC_SCF
PC_Resistin
PC_MIG
PC_IP-10
PC_IL-1α
PC_HGF
PC_EGF
PC_CD40L
PC_BDNF
VEGF
TRAIL
TNFβ
SDF1α
SCF
Resistin
RANTES
PDGF-BB
MIP-1α
MIG
Leptin
IP-10
IL-23
IL-13
IL-12p40
IL-10
IL-8
IL-6
IL-2
IL-1Ra
IL-1α
HGF
GROα
GM-CSF
FGFb
FasL
Eotaxin
EGF
CD40L
BDNF
Fig. 6 | Lipid–cytokine associations. a–e, Network of 1,245 significant β coefficients (BH FDR < 10%) are indicated in red; enrichments (log2(odds))
(BH FDR < 5%) lipid–cytokine associations, indicating positive (red) and negative among lipids with negative β coefficients (BH FDR < 10%) are indicated in
(blue) associations calculated across 1,180 samples, across all lipids (a) and blue; black denotes cases for which a certain property was enriched in both
for PCs (b), PEs (c), LPCs (d) and LPEs (e). Networks were pruned based on populations (positive and negative associations). log2(odds) values are depicted
a BH FDR of 5% for coefficients determined in linear mixed-effects models. when the respective annotation was significantly associated with a BH FDR
Colour indicates lipid class; edge width represents coefficients; and node of <5%. Infinite log2(odds) values are imputed with 0.5× the positive/negative
size represents node connectivity (popularity). The network was assembled log2(odds) values determined across all data. IL-1Ra, IL-1 receptor antagonist;
using the ‘graphopt’ layout algorithm. f, Fisher’s exact test enrichment analysis ICAM1, intercellular adhesion molecule 1; SDF1⍺, stromal cell-derived factor 1⍺;
comparing the physicochemical properties of lipids (y axis), at the subclass, RANTES, regulated on activation, normal T cell expressed and secreted;
global FA and individual FA level, that are associated with a particular cytokine PDGF-BB, platelet-derived growth factor-BB; GRO⍺, growth-regulated ⍺ protein;
(x axis). The analysis was performed for TAGs only (i), for all non-TAG lipids (ii) FasL, Fas ligand; TRAIL, tumour necrosis factor-related apoptosis-inducing
and across all lipids (iii). Enrichments (log2(odds)) among lipids with positive ligand.
with specific FA configurations (for example, omega-3/omega-6 FAs, across all FAs)). Ether-linked PEs have been implicated in cell signalling
PUFAs and SFAs). Overall, our results point to the distinct biological and as antioxidants25, and we found them to be significantly associated
roles of lipid subclasses and demonstrate that conventional clinical with healthy phenotypes including low SSPG levels and high HDL levels.
lipid profiles (that is, overall TAG levels) do not resolve many changes In addition, ether-linked PEs are depleted early during infection, puta-
relevant to metabolic health. tively increasing the inflammatory state, and/or depleted by scavenging
Throughout our analyses, we consistently observed distinct radical oxygen species resulting from inflammation. In addition, PE,
behaviours between ester-linked and ether-linked PEs (PE versus PE-O/ PE-P and PE-O show sex- and IR/IS-specific signatures during ageing
PE-P), as well as between two functionally distinct subgroups of TAGs (see below). Together, these observations suggest that ether-linked
(small TAGs (≤48 carbons across all FAs) and large TAGs (≥49 carbons PEs are associated with health and chronically low levels of these PEs

may have detrimental effects in humans. In this context, it will also be ageing) inflammatory conditions. The role of lipids in immunity is
interesting to investigate other ether-linked lipid subclasses known to only partially understood given the complexity of both the immune
be detectable in blood plasma, such as PCs with alkyl ether substituent system and lipid metabolism. Our analysis covered myriad positive and
(PC-O) and PCs with alkenyl ether (Plasmalogen) substituent (PC-P), negative cytokine–lipid associations, providing a valuable reference
in future studies. and resource for future studies. An important finding in this study is
Recently, we reported the differential regulation of small and large the strong positive associations of GM-CSF, leptin and the chemokine
TAGs within a 60-min recovery phase after exercise19. Here, we confirm eotaxin (CCL11) with small TAG species (Fig. 6). As GM-CSF and leptin
and expand our previous observations suggesting new clinically rele- are involved in regulating and promoting inflammation, their strong
vant physiological roles of these TAGs within a much larger longitudinal associations with small TAGs underscore the central and distinct role
cohort. Our data demonstrate that (1) many biological variations are these lipids may have in immunoregulation in comparison with groups
captured in TAGs’ abundance profiles (Figs. 1e and 2a), which (2) are of large TAGs.
distinct for TAG subgroups including large and small TAGs, and those Moreover, our data provide strong evidence for the pleiotropic role
containing specific FAs (Figs. 5d and 6f). For instance, small TAGs show of LPCs, which are categorized as proinflammatory41. We found a strong
distinct associations with certain cytokines and chemokines and are rap- positive association with proinflammatory signalling molecules, includ-
idly depleted during early RVI, followed by a rapid recovery to baseline ing IL-1⍺ and sCD40L, in addition to positive associations with growth
levels. Depletion of small TAGs during infection suggests an important factors such as BDNF and negative correlations with the proinflamma-
role in energy metabolism and signalling to support early inflammatory markers CRP and SSPG. There may be feedback loops such that a
tion. In addition, during ageing, large and small TAGs differ markedly, response is balanced, leading to the production of both inflammatory
suggesting distinct roles in ageing-related energy metabolism and and anti-inflammatory cytokines with different kinetics. Inflamma-
lipid-mediated signalling. TAGs also showed the highest technical repro- tory responses must be proportionate to the stimulus so that they are
ducibility in this study, making them an ideal target for new biomarkers effective but not excessive (leading to damage, for example, a cytokine
at the subclass and species levels. We therefore propose that small and storm). After the initial acute inflammation, the response needs to dimin-
large TAGs as well as ether-linked PEs could be further explored as health ish and anti-inflammatory signals need to increase. This process needs
biomarkers. Moreover, dietary supplements affecting plasma levels to be tuned in both magnitude and kinetics, and associated lipids may
may provide a therapeutic avenue to reduce chronic inflammation have a role in these responses. Although this study was not designed
and the detrimental effects of ageing and other conditions. to mechanistically resolve the role of LPCs and other lipid subclasses
In addition to identifying lipids that decrease in abundance with that we observed to be longitudinally associated with, for instance,
ageing, such as PUFAs, omega-3 FAs, FA(18:3) and FA(18:2), we identified viral infections, our data provide a resource for future studies and high-
many lipid subclasses and properties that are enhanced with ageing, light putatively competing roles that may depend on physiological and
including the proinflammatory CEs, CERs, SMs, arachidonic acid and pathological contexts, including comorbidities, age, BMI or IR/IS status.
SFAs, as well as LPCs whose role could be more ambiguous. Intriguingly, Although the study cohort is ethnically diverse and sex balanced
some of these effects were stronger (for example, HCERs, CEs, SMs) or similar to the US population, there are some limitations. (1) There is
directionally different (for example, PEs, DAGs) in participants with IR, a bias towards middle-aged and highly educated participants with
which can be interpreted as accelerated or differential ageing. IR is asso- a higher proportion of individuals living in northern California.
ciated with chronic inflammation and can lead to dyslipidaemia and (2) Some of the insights were generated based on small sample numbers
metabolic syndrome, which, in turn, increases the risk of age-related (for example, outlier analysis; Supplementary Fig. 6). Therefore, not
morbidities such as diabetes or cardiovascular diseases. Together with all observations and findings may be generalizable to a wider popula-
a distinct regulation between IR and IS, these observations indicate tion that is subject to different lifestyles. Although we identified many
a large-scale realignment of chronic inflammatory processes during signatures that have been observed in previous studies, supporting the
ageing, which may be accelerated in individuals with IR. Moreover, validity of our findings, lifestyle differences can have an effect19,55–57 and
we observed strong sex-specific ageing signatures for lipid subclasses future studies should be designed to confirm and extend our observa-
including CE. As cholesterol is a precursor of steroid hormones, many tions. (3) Our lipidomics pipeline targets >1,100 lipid species but does
of these sex differences in CE levels may relate to or even cause differ- not always resolve the exact molecule identity (for example, position
ences in sex-specific hormone levels. of a double bond in FAs). To expand the set of lipids investigated in
In addition to ageing-related differences, we found significant this study, we included phosphatidylinositol (PI) as a new lipid class
differences in both the effect sizes and the direction of the correlation in multiple-reaction-monitoring transitions (MRMs) and observed
of lipid subclasses and clinical measures between participants with IR multiple PI species to be significantly altered, for example, during
and those with IS. Importantly, our models controlled for—among other RVIs. For this lipid class, however, no internal spike-in standards were
variables—BMI, which is commonly associated with IR. Our analysis included in the sample preparation, which limits the accuracy of the
therefore highlights significant IR/IS differences separate from BMI abundance information for all PIs. Moreover, not all putatively detect-
effects. Key IR/IS-related differences include clinical markers related able blood plasma lipids, including PC-O and PC-P, were monitored.
to T2D and the immune response; haematological, hepatic and renal (4) The biphasic extraction method used here is an established and
measures; as well as electrolytes. For instance, PE-P/PE-O, LPE and LPC efficient procedure that is compatible with high-quality non-glass labo-
showed distinct directions of associations with multiple clinical meas- ratory equipment. However, we have observed that FFAs have higher
ures, such as the LDL to HDL ratio and various cell populations, for IR baseline levels when extracted with non-glass material, which can lead
and IS, expanding the previously suggested interplay between clinical to an overestimation of the levels of certain FFAs. This ratio compres-
measures and complex lipids41. As many lipids have both bioenergetic sion can reduce the sensitivity for detecting subtle changes between
and signalling functions, our observation indicates important differ- participants. (5) We used descriptive models (that is, we did not evalu-
ences in cellular signalling in participants with IR. ate the predictive power by using cross-validation and hold-out data).
Overall, these results underscore that the holistic assessment of Although we confirmed many previous findings and observed new
metabolic health state is highly useful, and in some cases necessary, lipid–phenotype associations, correlations are not proof of causation.
to improve the interpretation of conventional clinical measures such In addition, our models controlled for multiple covariates (for example,
as the LDL to HDL ratio. sex and BMI), allowing for separation of effects, such as lipid–age-
We identified distinct lipid subclasses and species that change ing versus lipid–BMI associations. Ageing is a complex process, and
in various disease states such as acute (RVI) and chronic (diabetes, increasing BMI may be a default process for industrialized societies and

lifestyles. Moreover, high BMI has been linked to inflammation. Thus, The tuning solution contained 40 μl of the SPLASH internal standard
controlling for BMI may not always be desirable for understanding mix, 100 μl of the Sciex tuning mix, 100 μl of Lyso-tune mix (1 mg ml−1
age- and inflammation-related changes in these cohorts. (6) The roles 17:1 lysophosphatidylglycerol (LPG), 1 mg ml−1 17:1 lysophosphatidyl-
of complex lipids are diverse, and the physiological impacts of lipids serine (LPS), 0.1 mg ml−1 17:1 lysophosphatidylinositol (LPI), 10 μg ml−1
with specific properties (for example, the fraction of complex lipids lysophosphatidic acid (LPA)) and 760 μl toluene–methanol (1:9) with
containing arachidonic acid) may not be due to direct effects. 10 mM ammonium acetate.
Our study provides a longitudinal, in-depth analysis of intricate For lipid extracts from 40 μl of plasma, three acquisition methods
lipid–health relationships during acute and chronic inflammation, were used. The injection volumes were 42, 50 and 39 μl for methods
metabolic diseases and ageing. The multitude of lipid–phenotype 1, 2 and 3, respectively. The source temperature was set to 150 °C for
connections we have revealed here serve as a valuable resource for all methods. Methods 1 and 3 were operated with DMS enabled and at
biomarker discovery, a starting point to investigate disease mecha- a separation voltage of 3,700 V. Lipid classes were monitored as fol-
nisms, and a basis to conceive therapeutic and preventive strategies. lows: method 1—PC (140), PE (119), PE-O (36), PE-P (61), LPC (26), LPE
For instance, although many lipids undergo intestinal lipolysis driven (26); method 2—CE (26), CER (12), DCER (12), HCER (12), LCER (12), FFA
by endogenous enzymes and microbial processes, some dietary lipids (26), TAG (519), DAG (59); method 3—SM (12), phosphatidic acid (PA;
are absorbed directly. Our research suggests a variety of potential 77), LPA (12), phosphatidylglycerol (PG; 78), LPG (16), PI (77), LPI (16),
dietary interventions that could improve human health. Supplements phosphatidylserine (PS; 78), LPS (16). Each transition was acquired 20
such as ether-linked PEs may alleviate inflammation, providing effec- times (see Supplementary Data 2 for the compensation voltage, Q1 and
tive treatment for chronic inflammation and facilitating recovery Q3 masses, and dwell times). Method 1, method 2 and the positive mode
after RVI by minimizing damage related to reactive oxidative agents. of method 3 (SM) contained transitions of the Lipidyzer original setup.
Alongside adjusting the small TAG to large TAG ratio, these lipids Method 3 negative mode targets additional lipids.
could also be instrumental in addressing ageing- and IR-associated
dyslipidaemia. Moreover, some lipid species and subclasses, includ- Raw data extraction and processing
ing FFAs, ether-linked PEs, CEs or CERs, demonstrate sex-specific and Data acquisition was performed similarly to the processing of the
ageing-related patterns. This prominent sex dimorphism suggests that Lipidyzer Workflow Manager. First, *.wiff files were converted to
sex- and age-specific interventions should be considered and might *.mzML files with MSConvert (v.3.0), setting ‘write index’ and ‘TPP
enhance therapeutic effects. It would also be worthwhile to assess the compatibility’ to true. For each raw file, data extraction was performed
impact of genetic polymorphisms, particularly in relation to lipid spe- in R. In brief, *.mzML files were imported with
cies and subclasses exhibiting high variance, such as certain TAG, DAG
and PE species, which may be relevant for personalized interventions. penMSfile(FileAndPath, backend = “pwiz”)
o
Together, future studies should explore how altering the exogenous chromatograms(‘openMSfile_output)
lipid intake (for example, through diet) or targeting lipid conversion
enzymes can affect both plasma lipid signatures and clinical pheno- using mzR (v.2.6.2). Next, all transitions with more than two zero
types such as IR, acute and chronic inflammation, and molecular ageing. intensities throughout the 20 repeated measurements were excluded
(reported as ‘not available’). For all remaining transitions, the mean
Methods intensity was calculated (excluding zero-intensity recordings). Lipid
Study design species identities were matched based on the Q1 and Q3 masses and the
Participants were enrolled as ‘healthy volunteers’ in the framework of corresponding scan index (the order in which MRMs were scheduled)
the National Institutes of Health integrated Human Microbiome Project in methods 1 (negative mode: PC, PE, LPC, LPE), 2 (negative mode: FFA;
2 (ref. 17). Inclusion and exclusion criteria were previously described positive mode: TAG, CE, DAG, CER, DCER, HCER, LCER) and 3 (negative
in detail18. Participants provided informed written consent for the mode: LPG, PG, LPI, PI, LPS, PS, LPA, PA; positive mode: SM). Note that
study under research study protocol 23602 approved by the Stanford the transitions PG, PS, PA and their respective lyso forms were not
University Institutional Review Board. considered for analysis. For lipids monitored in methods 1 and 2, as well
as SM (method 3), internal standards from the Sciex Lipidyzer internal
Lipid extraction spike (LPISTDKIT-102b) were matched according to the Sciex Lipidyzer
Plasma samples were prepared and analysed in a randomized order. protocol. Individual concentrations were estimated based on the known
Plasma was thawed on ice, and lipids were extracted using a biphasic abundance of the corresponding spike-ins. In brief, concentrations
separation technique (ice-cold methanol, methyl tert-butyl ether and (‘actual concentration’) of all spike-in standards were retrieved from the
water). A 260-μl volume of methanol and 40 μl of a spike-in standard ‘certificate of analysis’ of the internal spikes and converted to ‘nmol ml−1’.
(cat. no. 5040156, Sciex) were added to 40 μl of plasma, and the mix- Lipidyzer assumes that, in plasma, nmol g−1 = nmol ml−1. Internal spike
ture was vortexed for 20 s. Lipids were extracted by adding 1,000 μl stocks of individual lipid classes were mixed, dried down and resus-
of methyl tert-butyl ether and incubating the samples under agitation pended in a volume to adjust their respective stock concentration to
for 30 min at 4 °C. Phase separation was induced by adding 250 μl of the expected plasma levels (using the Lipidyzer Workflow Manager as a
ice-cold water, followed by vortexing for 1 min and centrifugation at reference). The internal spike area measured by MS was compared with
14,000g for 15 min at 4 °C. The upper phase containing the lipids was that of the respective endogenous lipids to approximate the absolute
collected, dried down under nitrogen and stored at −20 °C in 200 μl of concentrations of the endogenous lipids. For complex lipids with two
methanol. On the day of MS acquisition, lipids were dried down under identical FAs, it was assumed that the measured signal from the frag-
nitrogen and reconstituted with 300 μl of 10 mM ammonium acetate ment ions was at 2× intensity. Lipids belonging to the additional classes
in a 9:1 mixture of methanol and toluene. in method 3 had no corresponding standards and were normalized
based on one of the other spiked-in lipids, as detailed in the next section.
Lipidomics data acquisition The Lipidyzer resolves an individual FA as part of a TAG within each
The QTRAP 5500 system (Sciex) equipped with a DMS device (Lipidyzer) transition, a layer of information that we use to evaluate changes in FA
was operated with a Shimadzu SIL30AC autosampler for targeted lipi- compositions. For analyses that do not rely on the specific FA compo-
domics, with a modified strategy to include additional lipid species sition depicted in Fig. 3c, we aggregated TAGs to groups defined by
in the acquisition. To ensure robustness of results, the Lipidyzer was summed FA carbons and the number of unsaturations, summing the
cleaned and tuned after each batch (every 48 h; Supplementary Fig. 14). untransformed concentrations of the corresponding TAGs.

Normalization of lipid intensities in method 3 Data normalization

Spike-in standards (internal spikes) were not available at the time of Lipids were normalized based on the internal spike-in standards (see
lipid extraction for the lipids monitored here. Although this allows a above), similar to the standard Lipidyzer workflow that has been vali-
relative comparison across samples given a reproducible workflow, dated previously. Within an expanded method published by Su et al.,
we desired to leverage the information of the other internal spikes to additional MRMs can be used for isotope correction58. Here, we did not
further normalize for variances introduced across samples. To that acquire all of the MRMs needed for isotope correction. Although the
end, we performed a correlation analysis by using QC samples derived extent of correction depends on the abundance of interfering species
from the same stocks that were measured across all batches. The lipid and can significantly mask the signal of a targeted lipid species, Su et al.
intensities in these samples are expected to be the same, and the ratios reported that no species was corrected by >6% and, outside of TAGs,
of internal lipids compared with internal spikes will be the same if the no species was corrected by >3% (ref. 58). Cytokines were obtained in
variation introduced by lipid extraction or MS analysis affects them in three separate batches. Data were log2 transformed and corrected for
a similar manner. This setup allowed us to identify spike-in standards the effect of the batches using the ‘dbnorm’ (v.0.2.2) package59. The
to normalize the additional lipids monitored in method 3 (PA, LPA, ComBat model (sva (v.3.38.0))60 showed the best performance and
PG, LPG, PS, LPS, PI, LPI). For this normalization, we only considered thus was considered in this study.
internal spikes of the classes PC, PE, LPC and LPE, as those were also
acquired in positive mode with DMS enabled. Data imputation
We calculated the Pearson correlation coefficients for Lipidomics data. Missing values were imputed using a
log10(intensities) comparing internal spikes and the new lipid spe- K-nearest-neighbour strategy that accounts for a truncated distribu-
cies, and selected pairs according to the following hierarchy: (1) the tion (Extended Data Fig. 2)61. This approach involves drawing from
highest correlating internal spike with at least 50% complete obser- intensities at the detection limit defined for each lipid class separately.
vations across QC samples was selected; (2) if a match could not be This was a reasonable yet conservative assumption that allowed for
determined for a lipid species–internal spike pair, we selected the the imputation of missing values without inflating fold changes by
internal spike that showed the highest correlation with any other lipid considering the sensitivity of MS. Missing weight measures in the age-
of the same class; and (3) if both (1) and (2) did not select an internal ing analysis were imputed by taking the mean between the two closest
spike, the highest correlating spike-in standard across all additional adjacent timepoints (Supplementary Data 2). Missing levels of several
lipid classes was selected. cytokines in batch 1 (that is, hepatocyte growth factor (HGF), basic
fibroblast growth factor (FGFb), IL-8, IL-9, MIP-1⍺, stem cell factor (SCF)
Abundance estimation of lipids in method 3 and tumour necrosis factor-β (TNFβ)) and batch 2 (that is, IFN⍺2 and
For all original lipid classes, internal spikes of known concentrations FGFb) were imputed using the K-nearest-neighbour method with the
allow the approximation of absolute abundances. As described above, number of neighbours being 10.
the missing internal spikes for new lipids in method 3 do not allow
direct inference of absolute abundances. Using a linear regression Estimation of precision
model based on all the known concentrations of lipids in the samples, CV was calculated for non-imputed, untransformed data. If, for a
we predicted the concentrations of the new lipids. As the normal- participant, multiple samples were collected on the same day, those
ized abundances for these lipids (method 3) are not based on labelled were excluded. Only participants with at least three sampling time-
spike-in of the same molecular class and thus do not account for ioni- points were considered. Lipids with fewer than three quantifications
zation efficiency differences, they provide an estimate of the absolute were excluded. Note that, although the average intraparticipant CV
abundance range of the new classes. Importantly, this normalization was larger than the average QC CV, a small subset of low-abundance
does not affect the comparison of the relative abundance of the same lipid species for a subset of participants showed lower CVs. These CVs
lipid species across samples. emerged from the low signal that results in discrete quantifications
from the mass analyser.
Lipidomics data filtering
To ensure the accuracy and reliability of our analysis, we implemented Dimensionality reduction
several data filtering criteria. First, we excluded from the analysis t-SNE scatterplots were generated after log2 transformation and z-score
biosamples with >25% missing data. Additionally, lipids with <10% valid scaling of the data using the R package ‘Rtsne’ (v.0.15) with the following
values, as determined by the Lipidyzer reporting requirements, were parameters: perplexity = 5, θ = 0.5.
also excluded. To further ensure high-quality quantitative results for
the results presented in Figs. 2–6, we removed any lipid with a CV of WGCNA
>20% in QC samples and the few lipids for which the CV in QC samples Network analysis using self-reported healthy samples (Fig. 2d) was
was higher than the CV across the remaining biosamples. Furthermore, performed using the WGCNA R package (v.1.70-3). The soft-threshold
owing to limitations in separation by DMS, we did not include PAs in power was optimized to achieve approximate scale-free topology
our analysis. We also excluded PS/LPS and PG/LPG from the analysis (R2 > 0.8). Networks were constructed using the ‘blockwiseModules’
as they showed a significant number of missing values. PI(16:0/18:3) function. The network dendrogram was created using average linkage
was removed from the dataset owing to its association with incorrect hierarchical clustering of the topological overlap dissimilarity matrix
masses. Finally, QC_73 from batch 21 was removed owing to separate (1 − TOM). Modules were defined as branches of the dendrogram by
clustering compared with all other QC samples. using the hybrid dynamic tree-cutting method62, selecting a minimum
module size of 5. Modules were presented by their first principal com-
Internal spike-in reassignment ponent (module eigengene) of the standardized expression profiles.
Four internal spikes were not consistently quantified across the sam- Modules with eigengene correlations of >0.8 were merged together,
ples (missingness rate >5%) and were substituted with similar deu- generating seven lipid modules. Next, the Pearson correlation coef-
terated (d) standards belonging to the same class: dDAG(16:0/18:3) ficients between the module eigengene and clinical measures were
missing in 21% of the samples was substituted with dDAG(16:0/18:2); calculated using the ‘cor.test’ function in R (stats (v.3.6.2)), and all the
dDAG(16:0/20:5) missing in 12% was substituted with dDAG(16:0/20:4); obtained P values for the correlations were corrected for multiple
dPE(18:0/22:5) missing in 44% was substituted with dPE(18:0/20:4); hypotheses through the Benjamini–Hochberg (BH) procedure (stats
and dPE(18:0/20:5) missing in 8% was substituted with dPE(18:0/20:4). (v.3.6.2)).

Annotation enrichment analysis Linear mixed models

To determine over- and under-representation of functional subgroups SSPG association. To identify lipids that were associated with SSPG,
of lipids, we classified all lipid species based on their physicochemical linear mixed models were applied using log-transformed lipid measure-
properties (Supplementary Data 2) as reported in the LION database63. ments, controlling for participants, sex, ethnicity, age and BMI (Fig. 3).
Under- or over-representation was evaluated using a hypergeometric The R package ‘lme4’ (v.1.1-27.1) was used to construct the linear mixed
test (Fisher’s exact test) or one-dimensional annotation enrichment64,65. models, as well as output estimates and nominal P values. The obtained
Our dataset is dominated by TAG species, which could bias the enrich- raw P values were corrected for multiple hypotheses through the BH
ment analysis results. For that reason, we performed the enrichment procedure by using the ‘p.adjust’ function in R.
analysis across all lipids and within subclasses. For each figure, the
following statistics were used to calculate enrichments for categories: Infection. To identify lipids that were significantly changed dur-
‘OddEven_All’, ‘Omega_All’, ‘Saturation_All’, ‘Lipid_Class_Detailed_Special’ ing infection episodes, linear mixed models were applied using
and ‘FA_All’, applying a BH FDR set to 0.1 by using the ‘p.adjust()’ function. log-transformed lipid measurements, controlling for participants,
Figure 3e shows the Fisher’s exact test comparing positive SSPG sex, ethnicity, age and BMI (Fig. 4). The R package ‘lme4’ (v.1.1-27.1) was
coefficients with negative SSPG coefficients. Figure 4c shows the used to construct the linear mixed models, as well as output estimates
Fisher’s exact test determining whether the significantly changed and nominal P values. The obtained raw P values were corrected for
lipids during infection (RVI) were enriched. Figure 5f shows the Fisher’s multiple hypotheses through the BH procedure by using the ‘p.adjust’
exact test comparing positive Δage coefficients with negative Δage function in R.
coefficients. Figure 6f shows the Fisher’s exact test calculating enrich-
ments in positive coefficients as well as negative coefficients. Nega- Ageing. For each individual, ageing-associated lipid changes were
tive infinite log2(odds) and positive infinite log2(odds) were imputed calculated by subtracting measurements obtained at each visit from
with 0.5× the minimum log2(odds) and 0.5× the maximum log2(odds), the baseline values (Fig. 5). Accordingly, the number of years since
respectively. If a category was enriched among negative and positive onset was calculated as the number of years from the first recorded
coefficients for a lipid (that is, an over-representation was observed in measurement. To estimate the fractional changes in lipid measure-
the FDR-significant positive and negative coefficients), the enrichment ments, we used a linear regression model with log-transformed lipid
was set to 0 and highlighted in black in the heat map. measurements, controlling for BMI and storage length (and sex if
indicated in the figure). To control for potential biases related to the
Correlation between lipids and clinical measures in IR/IS number of samples per individual, we excluded measurements from
Correlations between lipid levels and clinical measures were calculated one participant with a uniquely large number of samples. To control
using the ‘cor.test’ function in R. Lipidomics data and laboratory meas- for potential biases related to a few participants with measurements
ures were both standardized before correlation calculation. To inves- spread across a longer enrolment time, we excluded a few samples
tigate differences between the correlations in IR and IS, we calculated collected >5 years since onset. All coefficients and s.d. values were
the correlations by using only the healthy samples from participants estimated using the ordinary least-squares implementation of the
with IR and IS, and we highlight only the correlation contrasts that were linear regression method in the ‘statsmodels’ package with the default
significantly different between IR and IS. The correlations between parameters in Python (v.3.7). Linear models were run either at the lipid
lipid levels and clinical measures from using all healthy samples from species level (Fig. 5d) or the lipid class level (sum of raw concentrations)
participants with IR and IS were also calculated and are presented as for all participants (Fig. 5c), as well as for sex and IR/IS (Fig. 5e,f).
reference values.
Cytokine–lipid association. We used linear mixed-effects models
RVI clustering and correlation with clinical measures (lmer {lme4}) controlling for BMI, sex, ethnicity and participants (ran-
K-means clustering was performed to investigate lipid similarities fol- dom effects) to estimate cytokine levels as a function of estimated lipid
lowing infection by using the lipidomics data of infection events after concentrations (Fig. 6). Both cytokine levels and lipid signals were
log2 transformation and z-score scaling. We calculated the minimum scaled and centred (scale()). Restricted maximum likelihood was set to
centroid distance for a range of cluster numbers, and the optimal false, and P values were estimated using summ {jtools} and corrected
number was chosen using the ‘elbow’ method. The median values of for multiple-hypothesis testing with p.adjust() applying a BH FDR of
the lipid profiles belonging to each cluster were correlated with the <5% for network generation.
clinical measures to indicate medical implications. The correlations
were calculated using the ‘cor.test’ function in R, and all the obtained Cytokine network
P values for the correlations were corrected for multiple hypotheses The cytokine–lipid network was constructed based on model coeffi-
using the BH procedure. cients by using the ‘graphopt’ layout algorithm in graphlayout{ggraph}
(v.2.1.0), igraph (v.1.5.0) and tidygraph (v.1.2). The network was pruned
RVI longitudinal IR and IS analysis to exclude all coefficients with a BH FDR of >0.05.
To detect the time intervals of differentially abundant lipids between IR
and IS during infection events, we used a longitudinal analysis method, Reporting summary
OmicsLonDA66. For each lipid in each group (IR or IS), we used a gene Further information on research design is available in the Nature Port-
ralized additive mixed model for modelling nonlinear time-series folio Reporting Summary linked to this article.
abundance during the inflammation episodes. OmicsLonDA is an
extension of MetaLonDA67 to account for correlated data, repeated Data availability
measurements and multiple covariates. We accounted for sex, age, Processed lipid data are provided as Supplementary Data 1. Raw mass
ethnicity and BMI as covariates, whereas participant identifiers were spectrometry data are hosted on our portal at http://hmp2-data.stan-
used as random effects. The P values for each lipid at each time interval ford.edu/index.php under the substudy iPOP lipidomics as well as
(the time interval unit was set to 1 day) were obtained and then adjusted at https://www.metabolomicsworkbench.org/ under the direct link
for multiple testing by using the BH procedure. We implemented this https://doi.org/10.21228/M8ZM5P. Cytokine and microbiome data are
process on both infection and immunization events, identified the hosted at http://hmp2-data.stanford.edu/index.php. Lipids were clas-
significantly different time intervals between IR and IS in both kinds sified partially based on their physicochemical properties reported in
of events, and compared these significant time intervals. the LION database63. Source data are provided with this paper.

Code availability 23. Erion, D. M. & Shulman, G. I. Diacylglycerol-mediated insulin

All software and algorithms are mentioned in the Methods section. resistance. Nat. Med. 16, 400–402 (2010).
Custom analysis scripts are hosted on the Stanford iPOP site (http:// 24. Chaurasia, B. et al. Targeting a ceramide double bond improves
med.stanford.edu/ipop.html). Enrichment analysis was conducted insulin resistance and hepatic steatosis. Science 365, 386–392
using a modified version of the R AnnoCrawler64 package, with the (2019).
corresponding lipid properties provided as Supplementary Data 2. 25. Dean, J. M. & Lodhi, I. J. Structural and functional roles of ether
lipids. Protein Cell 9, 196–206 (2018).
References 26. Richard, A. S. et al. Virion-associated phosphatidylethanolamine
1. Adewale, B. A. Will long-read sequencing technologies replace promotes TIM1-mediated infection by Ebola, dengue, and West
short-read sequencing technologies in the next 10 years? Afr. J. Nile viruses. Proc. Natl Acad. Sci. USA 112, 14682–14687 (2015).
Lab. Med. 9, 1340 (2020). 27. Numata, M. et al. Phosphatidylinositol inhibits respiratory
2. Aebersold, R. & Mann, M. Mass-spectrometric exploration of syncytial virus infection. J. Lipid Res. 56, 578–587 (2015).
proteome structure and function. Nature 537, 347–355 (2016). 28. Smani, Y., Domínguez-Herrera, J., Ibáñez-Martínez, J. & Pachón, J.
3. Ferdosi, S. et al. Engineered nanoparticles enable deep Therapeutic efficacy of lysophosphatidylcholine in severe
proteomics studies at scale by leveraging tunable nano–bio infections caused by Acinetobacter baumannii. Antimicrob. Agents
interactions. Proc. Natl Acad. Sci. USA 119, e2106053119 (2022). Chemother. 59, 3920–3924 (2015).
4. Pinu, F. R., Goldansaz, S. A. & Jaine, J. Translational metabolomics: 29. Lee, J. Y., Sohn, K. H., Rhee, S. H. & Hwang, D. Saturated fatty
current challenges and future opportunities. Metabolites 9, 108 acids, but not unsaturated fatty acids, induce the expression of
(2019). cyclooxygenase-2 mediated through Toll-like receptor 4. J. Biol.
5. Newgard, C. B. Metabolomics and metabolic diseases: where do Chem. 276, 16683–16689 (2001).
we stand? Cell Metab. 25, 43–56 (2017). 30. Jaul, E. & Barron, J. Age-related diseases and clinical and public
6. Murphy, R. C. Challenges in mass spectrometry-based lipidomics health implications for the 85 years old and over population.
of neutral lipids. Trends Analyt. Chem. 107, 91–98 (2018). Front. Public Health 5, 335 (2017).
7. Fahy, E. et al. Update of the LIPID MAPS comprehensive 31. Franceschi, C. et al. The continuum of aging and age-related
classification system for lipids. J. Lipid Res. 50, S9–S14 (2009). diseases: common mechanisms but different rates. Front. Med. 5,
8. Levy, B. D., Clish, C. B., Schmidt, B., Gronert, K. & Serhan, C. N. 61 (2018).
Lipid mediator class switching during acute inflammation: signals 32. Franceschi, C. & Campisi, J. Chronic inflammation (inflammaging)
in resolution. Nat. Immunol. 2, 612–619 (2001). and its potential contribution to age-associated diseases.
9. Serhan, C. N. Pro-resolving lipid mediators are leads for resolution J. Gerontol. A Biol. Sci. Med. Sci. 69, S4–S9 (2014).
physiology. Nature 510, 92–101 (2014). 33. Ahadi, S. et al. Personal aging markers and ageotypes revealed by
10. van Meer, G., Voelker, D. R. & Feigenson, G. W. Membrane lipids: deep longitudinal profiling. Nat. Med. 26, 83–90 (2020).
where they are and how they behave. Nat. Rev. Mol. Cell Biol. 9, 34. Bell, A. & Jones, K. Age, period and cohort processes in
112–124 (2008). longitudinal and life course analysis: a multilevel perspective. in
11. Shevchenko, A. & Simons, K. Lipidomics: coming to grips with A Life Course Perspective on Health Trajectories and Transitions
lipid diversity. Nat. Rev. Mol. Cell Biol. 11, 593–598 (2010). (eds. Burton-Jeangros, C. et al.) (Springer, 2015); https://doi.org/
12. Savelieff, M. G., Callaghan, B. C. & Feldman, E. L. The emerging 10.1007/978-3-319-20484-0_10
role of dyslipidemia in diabetic microvascular complications. 35. Slade, E. et al. Age and sex are associated with the plasma
Curr. Opin. Endocrinol. Diabetes Obes. 27, 115–123 (2020). lipidome: findings from the GOLDN study. Lipids Health Dis. 20,
13. Caterino, M. et al. Dysregulation of lipid metabolism and 30 (2021).
pathological inflammation in patients with COVID-19. Sci. Rep. 11, 36. Carrard, J. et al. Metabolic view on human healthspan: a
2941 (2021). lipidome-wide association study. Metabolites 11, 287 (2021).
14. Dennis, E. A. & Norris, P. C. Eicosanoid storm in infection and 37. Klop, B., Elte, J. W. F. & Cabezas, M. C. Dyslipidemia in obesity:
inflammation. Nat. Rev. Immunol. 15, 511–523 (2015). mechanisms and potential targets. Nutrients 5, 1218–1240 (2013).
15. Contrepois, K. et al. Cross-platform comparison of untargeted 38. Sears, B. & Perry, M. The role of fatty acids in insulin resistance.
and targeted lipidomics approaches on aging mouse plasma. Lipids Health Dis. 14, 121 (2015).
Sci. Rep. 8, 17747 (2018). 39. Sinclair, H. M. Essential fatty acids in perspective. Hum. Nutr. Clin.
16. Ghorasaini, M. et al. Cross-laboratory standardization of Nutr. 38, 245–260 (1984).
preclinical lipidomics using differential mobility spectrometry 40. Kim, E. J. et al. Skin aging and photoaging alter fatty acids
and multiple reaction monitoring. Anal. Chem. 93, 16369–16378 composition, including 11,14,17-eicosatrienoic acid, in the
(2021). epidermis of human skin. J. Korean Med. Sci. 25, 980–983 (2010).
17. Zhou, W. et al. Longitudinal multi-omics of host–microbe 41. Law, S.-H. et al. An updated review of lysophosphatidylcholine
dynamics in prediabetes. Nature 569, 663–671 (2019). metabolism in human diseases. Int. J. Mol. Sci. 20, 1149 (2019).
18. Schüssler-Fiorenza Rose, S. M. et al. A longitudinal big data 42. Beyene, H. B. et al. High-coverage plasma lipidomics reveals
approach for precision health. Nat. Med. 25, 792–804 (2019). novel sex-specific lipidomic fingerprints of age and BMI:
19. Contrepois, K. et al. Molecular choreography of acute exercise. evidence from two large population cohort studies. PLoS Biol. 18,
Cell 181, 1112–1130 (2020). e3000870 (2020).
20. Zhou, X. et al. Longitudinal analysis of serum cytokine levels 43. Pérez-Pérez, A. et al. Role of leptin as a link between metabolism
and gut microbial abundance links IL-17/IL-22 with Clostridia and and the immune system. Cytokine Growth Factor Rev. 35, 71–84
insulin sensitivity in humans. Diabetes 69, 1833–1842 (2020). (2017).
21. Hwu, C. M. et al. Surrogate estimates of insulin sensitivity in 44. Abella, V. et al. Leptin in the interplay of inflammation,
subjects with hypertension. J. Hum. Hypertens. 21, 246–252 metabolism and immune system disorders. Nat. Rev. Rheumatol.
(2007). 13, 100–109 (2017).
22. Ginsberg, H. N., Zhang, Y.-L. & Hernandez-Ono, A. Regulation of 45. Bhattacharya, P. et al. Dual role of GM-CSF as a pro-inflammatory
plasma triglycerides in insulin resistance and diabetes. Arch. Med. and a regulatory cytokine: implications for immune therapy.
Res. 36, 232–240 (2005). J. Interferon Cytokine Res. 35, 585–599 (2015).

46. Reed, J. A. et al. GM-CSF action in the CNS decreases food intake Acknowledgements
and body weight. J. Clin. Invest. 115, 3035–3044 (2005). We thank the entire Snyder laboratory for critical discussion and the
47. Lee, Y. et al. Anti-obesity effects of granulocyte-colony great research environment, in particular O. Dagan-Rosenfeld and
stimulating factor in Otsuka–Long–Evans–Tokushima fatty rats. E. Higgs for helping with the institutional review board process, as
PLoS ONE 9, e105603 (2014). well as A. Honkala for editing the manuscript. We thank M. Pearson
48. Lee, K. M. C., Achuthan, A. A. & Hamilton, J. A. GM-CSF: a promising and E.M. Elgierari for her help with the Lipidyzer as well as Y.R. Hasson
target in inflammation and autoimmunity. Immunotargets Ther. 9, for her help with cytokine assay processing. X.Z. received support
225–240 (2020). from the Stanford Aging and Ethnogeriatrics (SAGE) Research Center
49. Caro, J. F., Sinha, M. K., Kolaczynski, J. W., Zhang, P. L. & under National Institutes of Health (NIH)/National Institute on Aging
Considine, R. V. Leptin: the tale of an obesity gene. Diabetes 45, (NIA) grant P30AG059307. The SAGE Center is part of the Resource
1455–1462 (1996). Centers for Minority Aging Research program led by the NIA at the
50. Tanaka, T., Narazaki, M. & Kishimoto, T. Interleukin (IL-6) NIH. The contents are solely the responsibility of the authors and
immunotherapy. Cold Spring Harb. Perspect. Biol. 10, a028456 do not necessarily represent the official views of the NIA or the NIH.
(2018). S.M.S.-F.R. was supported by an NIH Career Development Award
51. Pedersen, B. K. Muscle as a secretory organ. Compr. Physiol. 3, (K08ES028825). We are deeply grateful to the families of our team
1337–1362 (2013). for their patience and support, as our work often required us to spend
52. Numakawa, T., Odaka, H. & Adachi, N. Actions of brain-derived weekends and nights extracting, measuring and analysing the human
neurotrophic factor and glucocorticoid stress in neurogenesis. lipidome.
Int. J. Mol. Sci. 18, 2312 (2017).
53. Siddiqui, K., Joy, S. S. & George, T. P. Circulating resistin levels Author contributions
in relation with insulin resistance, inflammatory and endothelial D.H., S.W. and M.P.S. planned the project. D.H. and G.M.T. performed
dysfunction markers in patients with type 2 diabetes and impaired experimental bench work and mass spectrometry acquisition. D.H.,
fasting glucose. Endocr. Metab. Sci. 1, 100059 (2020). K.J.W. and B.S. conceived and conducted mass spectrometry raw
54. Fiehn, O. Metabolomics—the link between genotypes and data processing. M.A. and W.Z. managed sample inventory. S.W., D.H.
phenotypes. Plant Mol. Biol. 48, 155–171 (2002). and S.M.S.-F.R. curated clinical data. Cytokine data normalization
55. Lancaster, S. M. et al. Global, distinctive, and personal changes in was done by N.B., T.M., S.W. and D.H. Microbiome data analysis was
molecular and microbial profiles by specific fibers in humans. done by X.Z. and K.C. Modelling and integrative lipid analysis were
Cell Host Microbe 30, 848–862 (2022). performed by S.W., M.M. and D.H. OmicsLonDA longitudinal analysis
56. Shen, X. et al. Multi-omics microsampling for the profiling was performed by S.W. and A.A.M. Results were interpreted by D.H.,
of lifestyle-associated changes in health. Nat. Biomed. Eng. S.W., J.M.U., P.K., S.M.S.-F.R., N.B., T.M., X.Z., K.C., A.A.M. and M.P.S. The
https://doi.org/10.1038/s41551-022-00999-8 (2023). manuscript was written by D.H., S.W., J.M.U., P.K. and M.P.S., with input
57. Zhang, X., Gao, P. & Snyder, M. P. The exposome in the era of the and edits from all authors.
quantified self. Annu. Rev. Biomed. Data Sci. 4, 255–277 (2021).
58. Su, B. et al. A DMS shotgun lipidomics workflow application to Competing interests
facilitate high-throughput, comprehensive lipidomics. J. Am. Soc. M.P.S. is a co-founder and on the advisory board of Personalis,
Mass Spectrom. 32, 2655–2663 (2021). SensOmics, January AI, Filtricine, Qbio, Protos, iollo, RTHM and
59. Bararpour, N. et al. DBnorm as an R package for the comparison Mirive. M.P.S. is on the advisory board of Jupiter, Abbratech, Neuvivo
and selection of appropriate statistical methods for batch effect and Mitrix. D.H. has financial interests in Seer and PrognomiQ. K.C. is
correction in metabolomic studies. Sci. Rep. 11, 5657 (2021). currently an AstraZeneca employee. A.A.M. is currently an employee
60. Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in of Google. K.J.W. is a consultant for Verso Biosciences, Inc. The
microarray expression data using empirical Bayes methods. remaining authors declare no competing interests.
Biostatistics 8, 118–127 (2007).
61. Shah, J. S. et al. Distribution based nearest neighbor imputation Additional information
for truncated high dimensional data with applications Extended data is available for this paper at
to pre-clinical and clinical metabolomics studies. BMC https://doi.org/10.1038/s42255-023-00880-1.
Bioinformatics 18, 114 (2017).
62. Langfelder, P., Zhang, B. & Horvath, S. Defining clusters from a Supplementary information The online version
hierarchical cluster tree: the Dynamic Tree Cut package for R. contains supplementary material available at
Bioinformatics 24, 719–720 (2008). https://doi.org/10.1038/s42255-023-00880-1.
63. Molenaar, M. R. et al. LION/web: a web-based ontology
enrichment tool for lipidomic data analysis. Gigascience 8, giz061 Correspondence and requests for materials should be addressed to
(2019). Michael P. Snyder.
64. DansenCode. DansenCode/AnnoCrawler: AnnoCrawler. Zenodo
https://doi.org/10.5281/ZENODO.3939260 (2020). Peer review information Nature Metabolism thanks the anonymous
65. Cox, J. & Mann, M. 1D and 2D annotation enrichment: a reviewers for their contribution to the peer review of this work.
statistical method integrating quantitative proteomics with Primary Handling Editor: Christoph Schmitt, in collaboration with the
complementary high-throughput data. BMC Bioinformatics 13, Nature Metabolism team.
S12 (2012).
66. Metwally, A. A. et al. Robust identification of temporal biomarkers Reprints and permissions information is available at
in longitudinal omics studies. Bioinformatics 38, 3802–3811 www.nature.com/reprints.
(2022).
67. Metwally, A. A. et al. MetaLonDA: a flexible R package for Publisher’s note Springer Nature remains neutral with regard
identifying time intervals of differentially abundant features in to jurisdictional claims in published maps and institutional
metagenomic longitudinal studies. Microbiome 6, 32 (2018). affiliations.

Open Access This article is licensed under a Creative Commons indicated otherwise in a credit line to the material. If material is not
Attribution 4.0 International License, which permits use, sharing, included in the article’s Creative Commons license and your intended use
adaptation, distribution and reproduction in any medium or format, is not permitted by statutory regulation or exceeds the permitted use, you
as long as you give appropriate credit to the original author(s) and the will need to obtain permission directly from the copyright holder. To view
source, provide a link to the Creative Commons license, and indicate a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons license, unless © The Author(s) 2023

Extended Data Fig. 1 | Principal component analysis for biosamples and quality controls. Principal component analysis of imputed log10 estimated nmol/ml
concentrations calculated using prcomp (scaled and centered data). Quality control (red) compared to biosamples (light blue) show distinct clustering.
Nature Metabolism
Extended Data Fig. 2 | KNN-TN class-wise imputation. KNN-TN imputation of missing values (blue) for different lipid classes. Y-axis denotes counts, x axis denotes
log10 estimated concentrations (nmol/ml).
Nature Metabolism
≥

Dynamic Lipidome Alterations Associated With Human Health, Disease and Ageing

Uploaded by

Copyright:

Available Formats

Dynamic Lipidome Alterations Associated With Human Health, Disease and Ageing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dynamic Lipidome Alterations Associated With Human Health, Disease and Ageing

Uploaded by

Copyright:

Available Formats

nature metabolism

Dynamic lipidome alterations associated

Lipids can be of endogenous or exogenous origin and affect diverse

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1578

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1579

DAG(FA1/FA2) TAG(FA1/FA2/FA3) CEs

PCs PIs SMs

(lyso) and non-hydrolysed >1,100 lipid species

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1580

Lipid mean estimated

Lipid panel T2D Immune system Haematologic Electrolytes Renal Hepatic

CE: 1 Large TAG: 12 CE

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1581

Global lipidome disruption in IR Dynamic lipidome alterations during viral infections

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1582

Diabetes Lipid panel

GPAT LPAAT PAP DGAT SPT CerS1–6 DES

LPE CDP-DAG Energy metabolism

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1583

TAG TAG 44:0(FA(18:0))

0 OmicsLonDA test statistics

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1584

∆age model coefficients

∆age model coefficients

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1585

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1586

f TAG CER PE PC log2(odds) (iii) Across all

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1587

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1588

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1589

Normalization of lipid intensities in method 3 Data normalization

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1590

Annotation enrichment analysis Linear mixed models

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1591

Code availability 23. Erion, D. M. & Shulman, G. I. Diacylglycerol-mediated insulin

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1592

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1593

Nature Metabolism | Volume 5 | September 2023 | 1578–1594 1594

You might also like