Plasma Lipidome and Prediction of Type 2 Diabetes in the Population

Diabetes Care 1

Plasma Lipidome and Prediction Celine´ Fernandez,1 Michal A. Surma,2 Christian Klose,3 Mathias J. Gerl,3 of Type 2 Diabetes in the Filip Ottosson,1 Ulrika Ericson,1 Nikolay Oskolkov,4 Marju Ohro-Melander,1 Population-Based Malmo¨ Diet Kai Simons,3 and Olle Melander1 and Cancer Cohort https://doi.org/10.2337/dc19-1199 PDMOOYHAT EVCSRESEARCH SERVICES EPIDEMIOLOGY/HEALTH

OBJECTIVE Type 2 diabetes mellitus (T2DM) is associated with dyslipidemia, but the detailed alterations in lipid species preceding the disease are largely unknown. We aimed to identify plasma lipids associated with development of T2DM and investigate their associations with lifestyle.

RESEARCH DESIGN AND METHODS 178 lipids were measured at baseline by mass spectrometry in 3,668 participants without diabetes from the Malmo¨ Diet and Cancer Study. The population was randomly split into discovery (n 5 1,868, including 257 incident cases) and replication (n 5 1,800, including 249 incident cases) sets. We used orthogonal projections to latent structures discriminant analyses, extracted a predictive component for T2DM incidence (lipid-PCDM), and assessed its association with T2DM incidence using Cox regression and lifestyle factors using general linear models.

RESULTS

A T2DM-predictive lipid-PCDM derived from the discovery set was independently associated with T2DM incidence in the replication set, with hazard ratio (HR) among 1Department of Clinical Sciences, Malmo,¨ Lund subjects in the fifth versus first quintile of lipid-PCDM of 3.7 (95% CI 2.2–6.5). In University, Lund, Sweden comparison, the HR of T2DM among obese versus normal weight was 1.8 (95% CI 2Łukasiewicz Research Network–PORT Polish 1.2–2.6). Clinical lipids did not improve T2DM risk prediction, but adding the lipid- Center for Technology Development, Wroclaw, Poland PCDM to all conventional T2DM risk factors increased the area under the receiver 3Lipotype GmbH, Dresden, Germany 4 operating characteristics curve by 3%. The lipid-PCDM was also associated with a Department of Biology, National Bioinformatics dietary risk score for T2DM incidence and lower level of physical activity. Infrastructure Sweden, Science for Life Labora- tory, Lund University, Lund, Sweden CONCLUSIONS Corresponding author: Celine´ Fernandez, celine. A lifestyle-related lipidomic profile strongly predicts T2DM development beyond [email protected] current risk factors. Further studies are warranted to test if lifestyle interventions Received 18 June 2019 and accepted 3 November 2019 modifying this lipidomic profile can prevent T2DM. This article contains Supplementary Data online at http://care.diabetesjournals.org/lookup/suppl/ Human plasma contains several hundred lipid molecular species, which collectively doi:10.2337/dc19-1199/-/DC1. make up the plasma lipidome (1). Disturbed lipid metabolism is a hallmark of type 2 © 2019 by the American Diabetes Association. diabetes mellitus (T2DM). T2DM-associated dyslipidemia is traditionally described in Readers may use this article as long as the work is properly cited, the use is educational and not for terms of circulating total triacylglycerides (TAGs) and lipoprotein cholesterol levels profit, and the work is not altered. More infor- (i.e., high total TAG and low HDL cholesterol [HDL-C] levels) (2). However, lipidomics- mation is available at http://www.diabetesjournals based studies of T2DM have shown that the traditional lipid markers are insufficient .org/content/license. Diabetes Care Publish Ahead of Print, published online December 8, 2019 2 Plasma Lipidome and Type 2 Diabetes Risk Diabetes Care

to predict disease risk, and more specif- RESEARCH DESIGN AND METHODS Follow-Up and End Points Retrieval ically molecular lipid species associated Study Participants and Data Collection Subjects were followed for incidence of with T2DM risk factors, such as obesity, The MalmoDietandCancer¨ –Cardiovascular T2DM until 31 December 2014. The hyperglycemia, and insulin resistance, Cohort(MDC-CC)is aprospectivepopulation- median follow-up time was 21.2 years – – have been identified (3,4). In addition, based cohort designed to study the (25 75% interquartile range [IQR] 16.9 fi plasma molecular lipid species associ- epidemiology of carotid artery disease 22.0). T2DM was de ned as a fasting $ ated with pre-T2DM or overt T2DM collected from 1991 to 1994 (15). At plasma glucose 7.0mmol/L, a history of have been discovered (5,6). baseline, all the MDC-CC participants physician diagnosis of T2DM, being on Longitudinal studies assessing asso- underwent a medical history, physical diabetes medication, or having been ciations between the plasma lipidome examination, and laboratory and lifestyle registered in local or national diabetes and risk of developing T2DM (7–10) or assessment. Of the 5,405 overnight- registries, as described previously (19). gestational diabetes mellitus (11) or fasted participants, citrate plasma sam- progressing from gestational diabetes ples were available in 4,067 subjects for Analytical Process Design to T2DM (12) have emerged. However, analysis of the lipidome. Of these, a total Samples were divided into analytical prospective studies of the plasma lipi- of377participantswerefurtherexcluded batches of 84 samples each. Each batch was accompanied by a set of four blank dome in a general population-based from statistical analyses because of the samples and eight identical control setting with many incident cases and presence of T2DM (for definition, see reference samples (human plasma). long follow-up for the prediction of below) at baseline examination. Addi- These control samples in groups of T2DM are lacking. Phenotypically well- tionally, lipidomic results from 22 sam- one blank and two reference samples characterized prospective cohorts are ples were excluded because of too many were distributed evenly across each necessary to study whether lipid molec- missing values (for threshold, see below). batch, extracted, and processed together ular species can be of clinical use for The remaining 3,668 study participants with study samples to control for back- T2DM risk prediction on top of conven- without diabetes were then randomly ground and intrarun reproducibility. All tional risk factors, including level of gly- divided into discovery (n 5 1,868, of batches were measured within 4 weeks. cemia and standard lipids. Identification whom 257 developed incident T2DM of such lipids is essential to provide during follow-up) and replication (n 5 Mass Spectrometry Lipidomics targets for novel lifestyle and pharma- 1,800, of whom 249 developed incident Lipid extraction of 1 mL of overnight- cological intervention studies. T2DM during follow-up) sets using the fasted citrate plasma samples stored Orthogonal projections to latent struc- “ function Select a random sample of at 280°C upon collection, followed by tures discriminant analysis (OPLS-DA) ; ” cases, 50% of all cases in SPSS. quantitative mass spectrometry–based is a supervised and linear multivariate Data on baseline levels of BMI, systolic lipid analysis, was performed at Lipotype statistical method for the unbiased blood pressure (SBP), current smoking, GmbH using a high-throughput shotgun selection of independent features use of antihypertensive treatment, phys- lipidomics technology (20). predicting events, which uses cross- ical activity during leisure time, clinical validation to reduce the likelihood of lipids, and blood glucose were obtained Postprocessing false-positive results. OPLS partitions as previously described (16). Physical Spectra were analyzed with an in house– the systemic variation in the data into activity during leisure time was trans- developed lipid identification software one predictive and one or several or- formed into a score (17). A diet risk score based on LipidXplorer (21). Data postpro- thogonal components (i.e., unrelated to previously shown to predict T2DM cessing and normalization were per- events) (13,14). (DRSDM) based on four foods consistently formed using an in house–developed In the current study, our aim was to associated with T2DM in epidemiological data management system. Only lipid test if plasma lipids, measured by a top- studies (processed meat, sugar-sweet- identifications with a signal-to-noise ratio down shotgun lipidomics method, pre- ened beverages, whole grain, and coffee) .5 and a signal intensity fivefold higher dict T2DM incidence during long-term was constructed as previously described thanincorrespondingblanksampleswere follow-up in a large population-based and expressed as risk points from 0 to considered for further data analysis. prospective cohort study from Malmo,¨ 8. The DRSDM was divided into three Batch correction was applied using Sweden. To avoid false-positive results, groups: low DRSDM (0–2 points), medium eight reference samples per 96 wells. we randomly divided the population DRSDM (3–5 points), and high DRSDM (6–8 Amounts were corrected for analytical into a set for discovery and a set for points) (18). drift if the P value of the slope was ,0.05, replication. Because many molecular T2DM at baseline examination was with an R2 .0.75, and the relative drift lipids are highly correlated, we used defined as a fasting whole-blood glucose was .5%. The median coefficient of lipid OPLS-DA for unbiased selection of in- $6.1 mmol/L (corresponding to plasma molecular species variation as assessed dependent and informative lipid pre- glucose $7.0 mmol/L), self-report of a by reference samples was 10.49%. An dictors for T2DM. In addition, we physician diagnosis, or use of diabetes occupational threshold of 75% was ap- investigated if dietary habits and level medication. plied to the data, resulting in 178 lipid of leisure time physical activity are re- All participants provided written in- molecular species for further analysis. lated to molecular lipid species levels, in formed consent, and the study (LU 51-90) Samples with .20% of the 178 lipids order to pave the way for potential was approved by the Ethics Committee missing were removed (22 samples). strategies to modulate their levels. at Lund University. Missing values were imputed using the care.diabetesjournals.org Fernandez and Associates 3

NIPALS algorithm (R 3.4.3, nipals package), models. Participants with incomplete models. In the first model (i.e., conven- a linear and multivariate statistical-based data on covariates were excluded. Haz- tional nonlipid risk factors), the predic- method that can cope with a moderate ard ratios (HRs) were expressed per SD tive value of age, sex, BMI, SBP, use amount of data missing at random (22). increment or per increasing quintiles of antihypertensive treatment, current (quintiles two through five compared smoking, and fasting plasma levels of Statistical Analyses with quintile one). Model 1 was adjusted glucose were investigated. In a second Statistical analyses were conducted in for age and sex, and model 2 was further model (i.e., all conventional risk factors), SPSS (version 25.0) unless otherwise adjusted for BMI, SBP, use of antihyper- fasting plasma levels of triglycerides, x2 stated. Student t tests and Pearson tensive treatment, current smoking, and LDL-C, and HDL-C were further added. tests were used to compare continuous fasting plasma levels of glucose, trigly- In the last model (i.e., all conventional and dichotomous variables, respectively, cerides, LDL cholesterol (LDL-C), and riskfactorsand PCDM),thelipid-PCDM was at baseline examination between the HDL-C at baseline examination. Model also added. two cohorts. Due to nonnormality, lipid 2 was additionally adjusted for family Second, continuous net reclassifica- species were transformed with the base history of diabetes, lipid-lowering med- tion improvement (cNRI) and integrated 10 log. The lipids data were mean cen- ication, leisure time physical activity discrimination improvement (IDI) were tered and unit variance scaled prior to level, and DRS . In model 3, BMI was calculated using the survIDINRI package statistical analysis. DM included as a variable divided into in R 3.4.3 (23) in order to compare the In the discovery set, OPLS-DA was three clinically defined categories (#25 predictive capacity ofthe lipid-PCDM toall adopted in order to simultaneously ana- kg/m2,normalweight;25–30 kg/m2, conventional risk factors. lyze correlations between the 178 lipid In the whole cohort, general linear overweight; .30 kg/m2,obese)instead molecular species and T2DM incidence models were used to study associations ofasacontinuousvariable.Normalweight (13,14). The systematic variation in lipid between the lipid-PC and lifestyle subjects were used as the reference DM levels that correlated with T2DM inci- factors (i.e., tertiles intake of energy- category. In sensitivity analyses, we dence was captured in one lipid pre- adjusted food groups, DRS , and level only included males, females, or partic- DM dictive component for T2DM incidence of physical activity during leisure time) ipants with overnight-fasted plasma glu- (lipid-PCDM), and the variation that did adjusting for potential confounders, cose concentration ,5.6 mmol/L and not correlate with T2DM was contained which were successively added in three performed adjustments according to in orthogonal components. The model models (age, sex, BMI, smoking, alcohol performance was validated by sevenfold models 1 and 2. intake, and physical activity). cross-validation. In the next step, data In the replication set, Kaplan-Meier Bonferroni correction was used to from the replication set were projected survival curve was used to describe the determine significance for multiple tests on the OPLS-DA model from the discov- rate of T2DM incidence over time in performed in the discovery set. For all ery set, and lipid loadings from the quintiles of baseline levels of lipid-PCDM. other tests, a two-sided P value of ,0.05 discovery set were used to calculate a The added value of the lipid-PCDM for was considered statistically significant. lipid-PCDM in the replication set. Analy- T2DM risk prediction was assessed by ses using OPLS-DA were performed in two methods in the replication set. First, SIMCA, version 14.1 (Sartorius Stedim receiver operating characteristic (ROC) RESULTS Biotech, Gottingen,¨ Germany). curves wereconstructed and Delong test, Baseline Characteristics The associations between lipids and performed in R 3.4.3 using the pROC The baseline characteristics of the study T2DM incidence were investigated using package, was used to compare the area participants in the discovery and repli- multivariable Cox proportional hazards under the curve (AUC) of the different cation sets are listed in Table 1. There

Table 1—Characteristics of the study participants Discovery set (n 5 1,868) Replication set (n 5 1,800) Pa Age (years) 57.50 6 5.95 57.39 6 6.04 NS Sex (% women) 58.7 61.7 NS Incident T2DM (%) 13.8 13.8 NS BMI (kg/m2) 25.38 6 3.81 25.38 6 3.64 NS Fasting glucose (mmol/L) 4.91 6 0.43 4.91 6 0.44 NS Fasting triglycerides (mmol/L) 1.27 6 0.59 1.26 6 0.60 NS Fasting LDL (mmol/L) 4.16 6 0.97 4.16 6 0.99 NS Fasting HDL (mmol/L) 1.41 6 0.37 1.42 6 0.37 NS SBP (mmHg) 141.28 6 18.53 140.68 6 18.96 NS Antihypertensive medication use (%) 14.3 14.8 NS Lipid-lowering therapy (%) 2.6 1.4 0.012 Smoker (%) 27.6 26.8 NS Values are displayed as means 6 SD or frequency in percent. ns, nonstatistical signiﬁcant. aP values were obtained from Student t tests or Pearson x2 tests for the binary variables. 4 Plasma Lipidome and Type 2 Diabetes Risk Diabetes Care

was no statistical difference for the in- of the OPLS-DA model were 12% and 8%, activity level, and DRSDM (Supplementary vestigated parameters, including the respectively. A total of 77 lipids had Table 4). There was a linear relationship clinical lipids, between the two study significant loadings (P , 0.05) and con- between the lipid-PCDM and the risk of sets except for the intake of lipid-low- tributed to the lipid-PCDM, explaining incidentT2DM in both model 1and 2 (P for ering medication. 5.3% of the 178 lipids proportion of trend respectively equal to 4.48E227 or variance (R2X 5 0.053). Results from 5.00E214), and the top versus bot- the multivariate analysis were concor- tom quintile of the lipid-PC in the fully Lipid Species and Risk of T2DM in the DM dant with the Cox regression analyses, adjusted model was associated with an MDC-CC Discovery Set HR of 5.12 (95% CI 2.92–8.96) (Table 2). In the MDC-CC discovery set with 1,868 with 49 lipids overlapping between the participants, 257 developed new-onset two analyses. Furthermore, as in the Cox regression models, lipids belonging to T2DM during a median follow-up time Validation of Results in the MDC-CC the TAG and DAG lipid classes had pos- of 21.3 years (IQR 17.2–22.0). Out of Replication Set 178 individual lipid species analyzed itive loadings whereas lipids belonging In the MDC-CC replication set with 1,800 (Supplementary Table 1), 58 associated to the PC O- and C18:2-containing lipid individuals, 249 developed new-onset with T2DM incidence at Bonferroni sig- species had generally negative loadings T2DM during a median follow-up time fi nificance (P , 2.8E24) in Cox propor- (Supplementary Table 3). More speci - of 21.2 years (IQR 16.8–22.0). As com- tional hazards models adjusted for age cally, TAG 48:0;0 and LysoPC 18:2;0 (LPC pared with the discovery set, all 58 and sex (Supplementary Fig. 1A and 18:2;0) displayed the strongest positive individual lipids were consistently signifi- Supplementary Table 2). Increased base- and negative associations, respectively, cantly related to T2DM incidence in line levels of several lipid species belong- with T2DM incidence in terms of HR the replication set at P , 4.0E23 (Sup- ing to the TAG and diacylglyceride (DAG) effect size and loadings. plementary Fig. 1B and Supplementary lipidclasses (HR/SD change of lipidlevels) Next, we tested if the generated lipid- Table 2). associated with increased risk of incident PCDM associated with T2DM incidence in In the next step, data from the rep- T2DM (1.25# HR #1.61), whereas in- Cox regression models. A higher value of lication set were projected on the creased levels of ether phospatidyl- the lipid-PCDM (HR/SD increment of the OPLS-DA model from the discovery set choline (PC O-) and of lipid species PCDM) was associated with increased risk (Supplementary Fig. 2B). Loadings from containing a fatty acyl chain with C18:2 of T2DM after age and sex adjustment the discovery set were used to calculate – (except one, DAG 16:0;0_18:2;0) associa- (model 1) (HR 1.93 [95% CI 1.77 2.10], a lipid-PCDM in the replication set. The ted with decreased risk of future T2DM P 5 4.50E249) (Table 2). The association cumulative incidence rate of T2DM ac- (0.64# HR #0.80). remained significant after further adjust- cording to quintiles of the calculated Because individual lipid species are ment for BMI, SBP, use of antihyperten- lipid-PCDM was higher for subjects in highly correlated, in a next step, we sive treatment, current smoking, and theupper quintiles(SupplementaryFig. 3). used OPLS-DA to simultaneously study fasting plasma levels of glucose, trigly- A higher value of the lipid-PCDM (HR/SD the association between all 178 lipid cerides, LDL-C, and HDL-C at baseline increment of PCDM) was associated species and T2DM incidence, which re- (model 2), with a multivariable-adjusted with increased risk of T2DM after age sulted in one lipid predictive compo- HR of 1.74 (95% CI 1.58–1.93, P 5 and sex adjustment (model 1) (HR 1.89 2 nent for T2DM incidence (lipid-PCDM) 4.07E 27) (Table 2). The HR was un- [95% CI 1.72–2.07], P 5 4.07E242) as well as two orthogonal components changed after additional adjustment for (Table 2). The association remained (Supplementary Fig.2A).Thegoodnessof family history of diabetes, intake of lipid- significant after full adjustment for fit (R2Y) and the predictive ability (Q2Y) lowering therapy, leisure time physical the traditional risk factors for T2DM

Table 2—Incidence of T2DM in relation to baseline levels of the lipid predictive component in the discovery (n 5 1,868) and the replication (n 5 1,800) set Discovery set Replication set Model 1 (n 5 257/1,610) Model 2 (n 5 246/1,571) Model 1 (n 5 249/1,550) Model 2 (n 5 243/1,517) Overall HRa (95%CI) 1.93 (1.77–2.10) 1.74 (1.58–1.93) 1.89 (1.72–2.07) 1.57 (1.40–1.76) P value ,0.001 ,0.001 ,0.001 ,0.001 Quintiles HRb Q1 1.0 (referent) 1.0 (referent) 1.0 (referent) 1.0 (referent) HRb Q2 1.30 (0.68–2.45) 1.11 (0.57–2.13) 1.45 (0.79–2.66) 1.32 (0.72–2.43) HRb Q3 2.72 (1.55–4.76) 2.02 (1.13–3.63) 2.04 (1.16–3.61) 1.23 (0.68–2.21) HRb Q4 3.64 (2.11–6.26) 2.41 (1.35–4.29) 3.41 (2.00–5.82) 2.05 (1.18–3.56) HRb Q5 8.30 (5.00–13.81) 5.12 (2.92–8.96) 7.90 (4.80–13.00) 3.49 (2.01–6.05) Pc for trend ,0.001 ,0.001 ,0.001 ,0.001 Model 1 is adjusted for age and sex. Model 2 is adjusted for age, sex, BMI, SBP, use of antihypertensive treatment, current smoking, and fasting plasma levels of glucose, triglycerides, LDL-C, and HDL-C at baseline. Q, quintile. aHRs are expressed per one-SD increment of the lipid predictive component. 95% CIs of the HRs are reported in parentheses. bHRs are expressed per quintile of the lipid predictive component, using quintile one as reference. 95% CIs of the HRs are reported in parentheses. cP for trend across quintiles of the lipid predictive component in a linear regression model. care.diabetesjournals.org Fernandez and Associates 5

(model 2), with a multivariable-adjusted clinical lipids did not improve risk dis- with total dairy intake (b 5 0.05, HR of 1.57 (95% CI 1.40–1.76, P 5 crimination on top of conventional non- P trend ,0.01) but was not found to 3.76E215) (Table 2). The HR was un- lipid risk factors in ROC curves (AUC 5 associate with intake of fish and shellfish, changed after additional adjustment for 0.775 [95% CI 0.744–0.806] vs. 0.770 nor with intake of fruits and vegetables. family history of diabetes, lipid-lowering [95% CI 0.738–0.801], P 5 0.176), adding Level of leisure time physical activity was medication, leisure time physical ac- the lipid-PCDM to the conventional risk negatively correlated with the lipid-PCDM tivity level, and DRSDM (Supplementary factors for T2DM (including clinical lipids) (b 5 0.09, P trend ,0.001), and the Table 4). There was also a linear rela- significantly improved the AUC by 3% association remained significant when tionship in the replication set between (AUC 5 0.807 [95% CI 0.778–0.837] vs. further adjusting for BMI and lifestyle the lipid-PCDM and the risk of incident 0.775 [95% CI 0.744–0.806], P , 0.001) factors (Supplementary Table 7). T2DM in both model 1 and 2 (P for (Fig. 1). Furthermore, estimates of IDI trend respectively equal to 5.88E226 or and cNRI for 18-year T2DM event rates CONCLUSIONS 2.95E208), and the top versus bottom increased when adding the lipid-PCDM on quintile of PCDM in the fully adjusted top of all the traditional risk factors for Here, we identified individual molecular model was associated with an HR of 3.49 T2DM (IDI 5 0.037 [95% CI 0.02–0.07], lipid species as well as a pattern of lipids (95% CI 2.01–6.05) (Table 2). The stan- P , 0.001; cNRI 5 0.17 [95% CI 0.06– (i.e., lipid-PCDM) prospectively associa- dardized lipid-PCDM was more strongly 0.27], P 5 0.013). ted with T2DM risk during long-term associated with incident T2DM than the follow-up in a general population. Our standardized BMI in model 2 (HR 1.57 Association of Diet Pattern and results are robust as they were very [95% CI 1.40–1.76], P 5 3.76E215, vs. HR Physical Activity with the Plasma similar in the replication set as compared 1.21 [95% CI 1.08–1.36], P 5 0.001). Lipidome with the discovery set. The lipid-PCDM Furthermore, being in the fourth or fifth In the whole study cohort of 3,668 par- was more strongly related to T2DM quintile of lipid-PCDM was associated ticipants, we investigated if the lipid- than BMI was, both in terms of its effect with a higher T2DM risk than being obese PCDM associated with the DRSDM, which size and statistical significance. As BMI in all models (Table 3 and Supplementary is based upon intake of whole grain, is considered a key clinical risk factor Table 5). coffee, sugar-sweetened beverages, and for T2DM, this result demonstrates the In sensitivity analyses, the lipid-PCDM processed meat (18). The DRSDM (high- clinical importance of the lipid-PCDM. significantly associated with T2DM in- vs. low-risk diet) was also associated with Moreover, clinical lipids used in many cidence in males, females, and partici- increased risk of T2DM in our cohort after T2DM-predictive algorithms were outper- pants with a normal fasted plasma adjustment for age and sex (HR 1.73 formed by the detailed lipid measure- glucose concentration (Supplementary [95% CI 1.29–2.30], P 5 2.12E204). ment contained within the lipid-PCDM. Table 6). The effect size of the lipid- There was a positive association between Last, we identified associations between PCDM association with T2DM incidence the lipid-PCDM and the DRSDM when the lipid-PCDM and dietary intake, in- was even strengthened when remov- adjusting for age and sex (b 5 0.12, cluding a previously T2DM-associated diet- ing participants with impaired fasting P trend ,0.001), which remained un- arypattern,aswellasphysicalactivity. glucose and adjusting for all traditional changed when further adjusting for We identified several individual lipids risk factors (HR 2.01 [95% CI 1.63–2.47], BMI and lifestyle factors. All components prospectively associated with T2DM and P 5 5.15E211). of the DRSDM, except for whole grain replicated all associations in the replica- Next, we tested whether the lipidomic- intake, contributed significantly to this tion set. Among the lipids of interest, based lipid measurements contributed association (Supplementary Table 7). several ether lipids displayed a signifi- to T2DM risk discrimination. Whereas The lipid-PCDM was positively associated cant inverse association with T2DM incidence, which is in accordance with previous results from the Prevencion´ Table 3—Incidence of T2DM in relation to baseline levels of the lipid predictive con Dieta Mediterranea´ trial case-cohort component in the replication set (n 5 1,800), adjusting for traditional risk factors, study of T2DM incidence (8). Previous including BMI categories studies also found that ether lipid levels 5 Model 3 (n 243 of 1,517) werelower in individuals with obesity (4), BMI (kg/m2) categoriesa prediabetes, and T2DM (6). This is of #25 1.0 (referent) interest because of the suggested role for – – 25 30 1.03 (0.76 1.41) plasmalogens, the most common form of .30 1.79 (1.23–2.62) ether lipids, as sacrificial antioxidants, Quintiles of lipid predictive component sparing the oxidation of other lipids, and Q1 1.0 (referent) Q2 1.33 (0.72–2.44) as scavengers of reactive oxygen species Q3 1.27 (0.71–2.29) (24). Plasmalogens might also modulate Q4 2.18 (1.25–3.80) inflammation by acting as storage depots Q5 3.72 (2.15–6.45) for arachidonic acid and docosahexae- Data are displayed as HR (95% CI). In model 3, the following independent variables were noic acid (25). simultaneously entered: age, sex, categories of BMI, SBP, use of antihypertensive treatment, We also found that several lysophos- current smoking, and fasting plasma levels of glucose, triglycerides, LDL-C, and HDL-C at baseline. pholipids (i.e., LPC and lysophosphatidy- Q, quintile. aNumber of individuals per BMI category is 904, 717, and 179, repectively. lethanolamine [LPE]) were inversely 6 Plasma Lipidome and Type 2 Diabetes Risk Diabetes Care

Figure 1—ROCcurves showingthe different AUCs forthe predictionof T2DMcalculatedfor differentmodels in the replicationset (n 5 1,800).In the ﬁrst model (i.e., conventional nonlipid risk factors), the predictive value of age, sex, BMI, SBP, use of antihypertensive treatment, current smoking, and fasting plasma levels of glucose was investigated. In a second model (i.e., all conventional risk factors), fasting plasma levels of triglycerides, LDL-C, and HDL-C were further added. In the third and last model (i.e., all conventional risk factors and PCDM), the lipid PCDM was also added.

related with T2DM risk, including LPC primary preventive treatment targeting prediction over all traditional risk factors 18:2;0, which is in accordance with pre- the lipid/lipid metabolic pathway enough was confirmed using IDI and cNRI. This vious studies (8,10). Furthermore, lower time to interfere with the T2DM path- highlights the need of a refinement of the LPC levels have been previously shown to ophysiological process. concept of “dyslipidemia” (referring to associate with an increased risk of car- Currently, BMIand presence of obesity low HDL-C and high TAG) related to the diovascular disease (26) and appear to is the key risk factor used for T2DM metabolic syndrome (28), as a key role of play a role in inflammation and athero- prediction in the clinic. In our study, the dyslipidemia component within the genesis (27). We found that a decreased the lipid-PCDM was more strongly related metabolic syndrome is to signal in- plasma level of C18:2 fatty acid (linoleic to T2DM than BMI both in terms of its creased risk of T2DM and thus leads acid) containing complex lipids is asso- effect size and statistical significance to initiation of preventive actions. Col- ciated with an increased risk of T2DM, even when entered into the same model, lectively, the superior role of lipid-PCDM which we have also previously shown to suggesting it identifies individuals at high inT2DMpredictionoverbothobesity and occur in overt T2DM in the MDC-CC T2DM risk independently of BMI (regard- clinical lipids suggests it would bring reexamination cohort (5). less of whether lipid-PCDM is causally clinical value if introduced on top of Our results also confirm previous stud- related toT2DM or not). Thus, individuals currently used T2DM risk factors. ies that have established that DAG and with high lipid-PCDM, irrespective of their Plasma lipid levels could be influenced TAG, including TAG 48:0;0, are associated BMI, might be candidates for intensive by dietary intake (29,30) and physical with a higher risk of T2DM (6,8,9). Over- lifestyle preventive actions in order to activity (31). The finding that the lipid- all, our study with .20 years of follow-up prevent T2DM in a similar way as indi- PCDM associated with a diet score pre- repeats several findings obtained in case- viduals with high BMI are treated today. viously shown to be linked to elevated control studies with shorter follow-up Furthermore, while clinical lipids used in T2DM risk (18), although the effect size of times (i.e., 4 years for the Prevencion´ T2DM prediction models, i.e., total TAG the association was relatively weak, sug- con Dieta Mediterranea´ trial [8] and and HDL-C, did not improve T2DM risk gests that diet modification targeted at 12 years for the Framingham Heart Study prediction on top of nonlipid traditional components of that score, i.e., increased [9]). This indicates that changes in the risk factors, the lipid-PCDM improved risk coffee intake and decreased sugar- plasma lipidome occur several years be- prediction by 3% beyond all risk factors. sweetened beverages and processed fore T2DM onset, which could give any The lipid-PCDM added value for T2DM risk meat, might be interesting candidates care.diabetesjournals.org Fernandez and Associates 7

for future controlled lifestyle interven- at testing if modifying the plasma lipidome 6. Meikle PJ, Wong G, Barlow CK, et al. Plasma tion trials. Moreover, based on our may reduce the risk of T2DM. lipid profiling shows similar associations with results, one can speculate that the prediabetes and type 2 diabetes. PLoS One 2013; 8:e74341 plasma level of the lipid-PCDM could 7. Mamtani M, Kulkarni H, Wong G, et al. Lip- be a useful tool for monitoring adher- Funding. C.F. was supported by the Albert idomic risk score independently and cost- ence to lifestyle therapy in high-risk Pahlsson˚ Foundation, the Crafoord Research effectively predicts risk of future type 2 diabetes: subjects for T2DM. Foundation, the Ernhold Lundstrom¨ Research results from diverse cohorts. Lipids Health Dis Foundation, the Royal Physiographic Society 2016;15:67 The main strength of this study is its ˚ of Lund, and the Ake Wiberg Foundation. C.F., 8. Razquin C, Toledo E, Clish CB, et al. Plasma large size in terms of number of subjects M.O.-M., and O.M. additionally acknowledge lipidomic profiling and risk of type 2 diabetes in and prospective end points during long- support from the Swedish Foundation for the PREDIMED trial. Diabetes Care 2018;41: term follow-up, which permitted us to Strategic Research (IRC15-0067). M.O.-M. was 2617–2624 supported by the European Research Coun- perform a robust replication of the find- 9. Rhee EP, Cheng S, Larson MG, et al. Lipid cil (Consolidator grant 649021), the Swedish fi fi ings. The very similar and significant pro ling identi es a triacylglycerol signature of Research Council, the Swedish Heart and Lung insulin resistance and improves diabetes predic- fi ˚ ndings in both the discovery and rep- Foundation, Region Skane, the Swedish Diabetes tion in humans. J Clin Invest 2011;121:1402–1411 lication sets greatly decrease the risk of Foundation, the Novo Nordisk Foundation, and the 10. Suvitaival T, Bondia-Pons I, Yetukuri L, et al. Albert Pahlsson˚ Foundation. O.M. was supported Lipidome as a predictive tool in progression to false-positive results, which are very by the Knut and Alice Wallenberg Foundation, the type 2 diabetes in Finnish men. Metabolism common in smaller “omics” studies. Goran¨ Gustafsson Foundation, the Swedish Heart 2018;78:1–12 and Lung Foundation, the Swedish Research Moreover, we based our conclusions 11. Lu L, Koulman A, Petry CJ, et al. An unbiased Council, the Novo Nordisk Foundation, Region solely on the replication set, thus min- lipidomics approach identifies early second tri- Skane,˚ and Skane˚ University Hospital. imizing the risk of, for example, inflation mester lipids predictive of maternal glycemic Duality of Interest. M.A.S., C.K., and K.S. are traits and gestational diabetes mellitus. Diabetes of effect sizes and significance levels. shareholders of Lipotype GmbH. M.J.G. is an – On the other hand, both the discovery employee of Lipotype GmbH. K.S. is the chief Care 2016;39:2232 2239 12. Lappas M, Mundra PA, Wong G, et al. The and replication sets were derived from executive officer of Lipotype GmbH. No other fl prediction of type 2 diabetes in women with the same South Swedish Caucasian pop- potential con icts of interest relevant to this article were reported. previous gestational diabetes mellitus using lip- – ulation, and it would be important to Author Contributions. C.F. drafted the man- idomics. Diabetologia 2015;58:1436 1442 test our findings in other populations uscript. C.F., F.O., and N.O. contributed to the 13. Bylesjo M, Rantalainen M, Cloarec O, with, for example, different lifestyle data analyses. C.F., K.S., and O.M. contributed to Nicholson JK, Holmes E, Trygg J. OPLS discrim- inantanalysis:combiningthe strengthsof PLS-DA habits and race for external validation. study concept and design. M.A.S., C.K., M.J.G., U.E., and M.O.-M. contributed to acquisition of and SIMCA classification. J Chemometr 2006;20: Another strength of this study comes data. C.F., M.A.S., C.K., M.J.G., F.O., U.E., N.O., 341–351 from the detailed clinical and lifestyle M.O.-M., K.S., and O.M. provided intellectual 14. Trygg J, Wold S. Orthogonal projections to information available in the cohort, al- contributions to drafting and/or revising the latent structures (O-PLS). J Chemometr 2002;16: – though gestational diabetes history was manuscript and approved the final version. 119 128 C.F. is the guarantor of this work and, as such, 15. Rosvall M, Janzon L, Berglund G, Engstrom¨ G, missing. Last, the broad coverage of the had full access to all the data in the study and Hedblad B. Incident coronary events and case lipidome provided a detailed insight into takes responsibility for the integrity of the data fatality in relation to common carotid intima- the metabolic changes preceding T2DM, and the accuracy of the data analysis. media thickness. J Intern Med 2005;257:430–437 even though we cannot exclude that Data and Resource Availability. The MDC data 16. Fernandez C, Rysa¨ J, Almgren P, et al. Plasma because of the long sample storage discussed in this article will be made available to levels of the proprotein convertase furin and appropriate academic parties based on a written incidence of diabetes and mortality. J Intern Med time, some of the lipids have been par- application to the MDC steering committee; 2018;284:377–387 tially degraded. However, if this would please contact Anders Dahlin (anders.dahlin@ 17. Hellstrand S, Sonestedt E, Ericson U, et al. be the case, it should bias our results med.lu.se). Intake levels of dietary long-chain PUFAs modify toward no findings. the association between genetic variation in – We do acknowledge that despite ex- References FADS and LDL-C. J Lipid Res 2012;53:1183 1189 18. Ericson U, Hindy G, Drake I, et al. Dietary and tensive adjustments and the prospective 1. Quehenberger O, Armando AM, Brown AH, et al. Lipidomics reveals a remarkable diversity of genetic risk scores and incidence of type 2 di- study design, we cannot prove causality lipids in human plasma. J Lipid Res 2010;51: abetes. Genes Nutr 2018;13:13 19. Enhorning¨ S, Hedblad B, Nilsson PM, between the plasma lipidome and T2DM, 3299–3305 ¨ given the observational nature of the 2. Taskinen MR, Boren´ J. New insights into the Engstrom G, Melander O. Copeptin is an independent predictor of diabetic heart disease study. This may not necessarily limit its pathophysiology of dyslipidemia in type 2 di- and death. Am Heart J 2015;169:549–556.e1 abetes. Atherosclerosis 2015;239:483–495 role as a risk biomarker. However, cau- 20. Surma MA, Herzog R, Vasilj A, et al. An 3. Huynh K, Barlow CK, Jayawardana KS, et al. sality needs to be proven if T2DM risk– automated shotgun lipidomics platform for high High-throughput plasma lipidomics: detailed throughput, comprehensive, and quantitative associated lipids should be targets for mapping of the associations with cardiometa- analysis of blood plasma intact lipids. Eur J Lipid novel preventive drug therapies and – bolic risk factors. Cell Chem Biol 2019;26:71 Sci Technol 2015;117:1540–1549 lifestyle interventions. 84.e4 21. Herzog R, Schuhmann K, Schwudke D, et al. Our results indicate that the plasma 4. Weir JM, Wong G, Barlow CK, et al. Plasma LipidXplorer: a software for consensual cross- fi lipidome contains information beyond lipid pro ling in a large population-based cohort. platform lipidomics. PLoS One 2012;7:e29851 known predictors of T2DM, such as J Lipid Res 2013;54:2898–2908 22. Wold S, Sjostrom M, Eriksson L. PLS-regression: 5. Kjellqvist S, Klose C, Surma MA, et al. Iden- obesity and clinical lipids, which could a basic tool of chemometrics. Chemometr Intell tification of shared and unique serum lipid – ’ fi Lab 2001;58:109 130 be used for patients risk strati cation. profiles in diabetes mellitus and myocardial 23. Uno H, Tian L, Cai T, Kohane IS, Wei LJ. A Furthermore, our results encourage con- infarction. J Am Heart Assoc 2016;5(12). pii: unified inference procedure for a class of mea- trolled lifestyle intervention trials aimed e004503 sures to assess improvement in risk prediction 8 Plasma Lipidome and Type 2 Diabetes Risk Diabetes Care

systems with survival data. Stat Med 2013;32: 27. Schmitz G, Ruebsaamen K. Metabolism and plasma lipidomics in a twin study. Sci Rep 2017;7: 2430–2442 atherogenic disease association of lysophospha- 3750 24. Dean JM, Lodhi IJ. Structural and functional tidylcholine. Atherosclerosis 2010;208:10–18 30. Toledo E, Wang DD, Ruiz-Canela M, et al. ﬁ roles of ether lipids. Protein Cell 2018;9:196–206 28. Fedorowski A, Burri P, Hulthen´ L, Melander Plasmalipidomicpro lesandcardiovascularevents in a randomized intervention trial with the Med- 25. Braverman NE, Moser AB. Functions of plas- O. The metabolic syndrome and risk of myocar- iterranean diet. Am J Clin Nutr 2017;106:973–983 malogen lipids in health and disease. Biochim dial infarction in familial hypertension (hyper- 31. Grace MS, Dempsey PC, Sethi P, et al. Break- – Biophys Acta 2012;1822:1442 1452 tension heredity in Malmo¨ evaluation study). J ing up prolonged sitting alters the postprandial 26. Fernandez C, Sandin M, Sampaio JL, et al. Hypertens 2009;27:109–117 plasma lipidomic proﬁle of adults with type 2 Plasma lipid composition and risk of developing 29. Frahnow T, Osterhoff MA, Hornemann S, diabetes. J Clin Endocrinol Metab 2017;102: cardiovascular disease. PLoS One 2013;8:e71846 et al. Heritability and responses to high fat diet of 1991–1999