Changes in the lipidome in type 1 diabetes following low carbohydrate diet: Post‐hoc analysis of a randomized crossover trial

Abstract Aims Lipid metabolism might be compromised in type 1 diabetes, and the understanding of lipid physiology is critically important. This study aimed to compare the change in plasma lipid concentrations during carbohydrate dietary changes in individuals with type 1 diabetes and identify links to early‐stage dyslipidaemia. We hypothesized that (1) the lipidomic profiles after ingesting low or high carbohydrate diet for 12 weeks would be different; and (2) specific annotated lipid species could have significant associations with metabolic outcomes. Methods Ten adults with type 1 diabetes (mean ± SD: age 43.6 ± 13.8 years, diabetes duration 24.5 ± 13.4 years, BMI 24.9 ± 2.1 kg/m2, HbA1c 57.6 ± 2.6 mmol/mol) using insulin pumps participated in a randomized 2‐period crossover study with a 12‐week intervention period of low carbohydrate diet (< 100 g carbohydrates/day) or high carbohydrate diet (> 250 g carbohydrates/day), respectively, separated by a 12‐week washout period. A large‐scale non‐targeted lipidomics was performed with mass spectrometry in fasting plasma samples obtained before and after each diet intervention. Longitudinal lipid levels were analysed using linear mixed‐effects models. Results In total, 289 lipid species were identified from 14 major lipid classes. Comparing the two diets, 11 lipid species belonging to sphingomyelins, phosphatidylcholines and LPC(O‐16:0) were changed. All the 11 lipid species were significantly elevated during low carbohydrate diet. Two lipid species were most differentiated between diets, namely SM(d36:1) (β ± SE: 1.44 ± 0.28, FDR = 0.010) and PC(P‐36:4)/PC(O‐36:5) (β ± SE: 1.34 ± 0.25, FDR = 0.009) species. Polyunsaturated PC(35:4) was inversely associated with BMI and positively associated with HDL cholesterol (p < .001). Conclusion Lipidome‐wide outcome analysis of a randomized crossover trial of individuals with type 1 diabetes following a low carbohydrate diet showed an increase in sphingomyelins and phosphatidylcholines which are thought to reduce dyslipidaemia. The polyunsaturated phosphatidylcholine 35:4 was inversely associated with BMI and positively associated with HDL cholesterol (p < .001). Results from this study warrant for more investigation on the long‐term effect of single lipid species in type 1 diabetes.


| INTRODUC TI ON
It is now recognized that type 1 diabetes is a major risk factor for developing cardiovascular events. [1][2][3] Dyslipidaemia is a modifiable risk factor for cardiovascular disease and highly prevalent in type 1 diabetes. [4][5][6] The present approach to diagnosing dyslipidaemia is based on the clinical measurement of the three main types of serum lipids, namely HDL cholesterol, LDL cholesterol and triglycerides. 7 However, since HDL-cholesterol metabolism might be compromised in type 1 diabetes, the identification of alternative biomarkers is critically important. [3][4][5][6] Lipidomics studies suggest lipids as indicators or predictors for dyslipidaemia and cardiovascular disease. [8][9][10][11][12][13][14] However, to date, the plasma lipidome from individuals with type 1 diabetes has not been interrogated to understand diet-dependent changes in a clinical setting.
Lipidomics expands the lipid information of the three main traditional mentioned lipids, since there are many more lipid species that can be measured with modern lipidomics methods. [15][16][17] Lipidomics provides a profile of hundreds of individual lipid species and on low-abundance lipid species from a biological sample. Therefore, there is the potential to uncover novel insights into the pathophysiology in individuals with type 1 diabetes. The composition of the lipidome can reveal a fingerprint in relation to dyslipidaemia and cardiovascular disease in type 1 diabetes; thus, it can be an important tool for early screening.
In this lipidomics analysis of a randomized crossover trial, we compared the plasma lipids concentration from individuals with type 1 diabetes before and after ingesting isocaloric low carbohydrate diet (LCD) and high carbohydrate diet (HCD) for 12 weeks (<100 g vs. >250 g) separated by a 12-week washout period obtained from a previous published study. 18 We hypothesized that (1) the lipidomics profiles before and after ingesting low or high carbohydrate diet for 12 weeks will be different; and (2) specific annotated lipid species have significant associations to metabolic characteristics.

| Clinical trial
Fasting plasma samples for lipidomics analysis were collected from a previously published study. 18 Participants were adults with insulin pump-treated type 1 diabetes (mean ± SD: age 43.6 ± 13.8 years, diabetes duration 24.5 ± 13.4 years, BMI 24.9 ± 2.1 kg/m 2 , HbA 1c 57.6 ± 2.6 mmol/mol). They underwent two 12-week diet interventions separated by a 12-week washout. Fasting plasma samples were collected before and after each intervention.
Main study outcomes from the previous article for the 14 study participants are given in Table 1. In brief, the study showed that glycaemic variability and time spent in hypoglycaemia were lower during LCD than during HCD. Mean glucose, however, was the same.
Total daily bolus insulin doses were lower during LCD reflecting the lower carbohydrate intake. Finally, weight decreased during LCD, whereas it increased during HCD with a significant difference in change between groups. Systolic and diastolic BP increased during HCD (not significant); however, the between-group differences were insignificant. The between-group difference in HDL-cholesterol levels was significant (p = .005) with the LCD showing an increase (LCD: 0.06 mmol/L; HCD: −0.03 mmol/L). No changes in fasting LDL cholesterol and triglyceride were detected. 18 In the present study, we therefore investigated whether the diet-induced changes in relevant clinical variables (BMI, BP and HDL cholesterol) were associated with specific lipid species.
All participants were given personalized advice by a dietitian focused on eating strategies for meeting the carbohydrate criteria.
LCDs contained less than 100 g carbohydrates per day, and HCDs contained a minimum of 250 g carbohydrates per day. Suggestions for a healthy composition of carbohydrates, fat and protein sources were provided; however, only the amount of carbohydrates was fixed. Carbohydrates intake was recorded in grams in the participants' insulin pumps on a meal-by-meal basis, but there was no registration of actual food intake. 18 In this study, only samples from the 10 participants who completed both intervention periods are included.

| Lipidomics analysis
A modified Folch lipid extraction method was used to analyse the total lipids from plasma samples. 19 Briefly, plasma samples were randomized and lipids were extracted from 10 µL plasma using chloroform:methanol (2:1 v/v) method following addition of nine different internal standards (stable isotope labelled and non-physiological lipid species). A detailed name list of the used internal standards sphingomyelins and phosphatidylcholines which are thought to reduce dyslipidaemia.
The polyunsaturated phosphatidylcholine 35:4 was inversely associated with BMI and positively associated with HDL cholesterol (p < .001). Results from this study warrant for more investigation on the long-term effect of single lipid species in type 1 diabetes.

K E Y W O R D S
cardiovascular disease, dyslipidaemia, lipidomics, low carbohydrate diet, randomized trial, type 1 diabetes is available in the Methods S1. Samples were analysed in random order in positive and negative electrospray ionization modes using ultra-high-performance liquid chromatography-quadrupole time-offlight mass spectrometry (UHPLC-Q-TOF-MS). The UHPLC system was from Agilent Technologies (Santa Clara, CA, USA) and was used as previously described. 20,21

| Data pre-processing
The lipidomics mass spectrometry data were pre-processed in MZmine 2.18.2. 22 The workflow includes raw data import, filtering, peaks detection, chromatogram building, chromatogram deconvolution, peak list de-isotoping, peak list alignment, gap filling and finally, peak annotation was carried out by combining MS and retention time information with in-house lipid library with an m/z tolerance of 0.006 m/z and RT tolerance of 0.2 min. Lipidomics data post-processing and analysis were performed in R programming language for statistical computing (https://www.r-proje ct.org/). 23 The pre-processed data were normalized to internal standards. The missing values in the lipidomics data set were imputed with the k-nearest neighbour algorithm. 24 To achieve normal distribution, data were log-transformed.
Coefficient of variation (Relative Standard Deviation; %RSD) for peak areas and retention times of lipid-class specific internal standards were calculated. The measurement of 14 lipid classes was calculated by summing the individual lipid species within each class.

| Statistical analysis
Data were analysed and visualized in R. Lipid species-wise mixed-effect models were used to consider the four repeated measures from each participant in the crossover trial. All statistical tests were corrected for multiple testing using the Benjamini-Hochberg method. 25 Tests were corrected for multiple comparisons with false discovery rate (FDR). First, differences between the two diets were modelled with lipid species-wise mixed-effect models, with time, diet and time-diet interaction as fixed effects and the participant ID as random effects.
We further conducted the lipid concentration changes found from the mentioned analysis within groups.
Second, associations between key lipid levels and the clinical covariates of interest were assessed: BMI, HDL cholesterol, systolic BP and diastolic BP. In this analysis, only lipid species with significant difference between groups from the first analyses were included.
Association was tested by adding one clinical covariate at a time as an independent variable to the lipid-wise mixed-effect model detailed in the first step. A detailed description about the data analysis plan is available in the Methods S2. Finally, lipids of interest with significant association to clinical variables were semi-quantified (relative quantified) and visualized in box plots grouped by diets and scatterplots. To support further research on the topic, the power to replicating the main finding in an identical study was estimated.
This was done with a post hoc calculation of the power to finding an identical-sized effect in an identical-sized crossover study at an alpha level of 0.05.   Figure S1. The coefficient of variation (%RSD) of peak areas was on average 17.41% for internal standards (Table S1).

| Difference in the outcome between the diets
In total, 11 lipid species from phosphatidylcholine and sphingomy- All the aforementioned lipid species were present in significantly higher amounts after LCD than after HCD (see the "Slope-LCD" and "Slope-HCD" columns in Table 2). (β ± SE: −0.21 ± 0.06, p = .003). The same trend of inverse association was also observed in diastolic BP, although not to statistical significance after correcting for multiple testing. PC(35:4) and PC(31:0) lipid species were elevated during the LCD intervention (see column "Slope-LCD" and "Slope-HCD" in Table 2), inferring a link between diet, circulating lipid concentrations, BMI and BP. There were no associations with the systolic BP that could be detected with statistical significance before and after correcting for multiple testing in this small study.

| Association between lipid species and clinical variables
There was a positive association with the HDL cholesterol and  Table S2.
One of the 11 significantly changed lipid species was associated both with BMI, diastolic BP and HDL, namely the polyunsaturated phosphatidylcholine 35:4. Figure 1A shows box plot of semi-quan-

TA B L E 2 Results of the mixed-effects model
The main result of this study was that when both LDC and HDC were compared, 11 lipid species belonging to sphingomyelin and phosphatidylcholine were significantly changed in outcome between the diets (Table 2). Additionally, and most importantly, when we studied each trial arm separately, we observed that all the 11 significantly changed lipid species in the outcome were in greater concentration in the LCD arm, but none were altered significantly in the HCD arm. This was an unexpected result since total cholesterol, LDL and triglycerides had not changed significantly in this trial and only HDL cholesterol was significantly elevated in the LCD arm, suggesting a moderate lipid improvement with LCD. Also this suggests that a prolonged LCD could help with adjusting lipid classes since there is existing evidence that individuals with type 1 diabetes have decreased levels of blood sphingolipids and phosphatidylcholines, abundant in HDL. 26 Phosphatidylcholines and sphingomyelins are two important classes of phospholipids essential for cell membrane function and major components of HDL cholesterol specific those containing polyunsaturated fatty acids. 14,27,28 Interestingly, a noteworthy result from our study is that the LCD arm showed that HDL cholesterol was increased and positively associated with polyunsaturated sphingomyelins and phosphatidylcholines (Table S2). In rodents, sphingomyelin supplementation has been shown to help reduce the intestinal absorption of serum lipids such as cholesterol and triglycerides. 12 Sphingomyelin levels have also been associated with mild-to-moderate hypertension. 29 Based on these findings, supplementation could be an avenue to explore in the context of dyslipidaemia in type 1 diabetes. The findings for this study were in line with previous work on LCD in which similarly HDL cholesterol and diastolic BP were increased and decreased, respectively, with less carbohydrate intake, in a cross-sectional carbohydrate intake study (carbohydrate intake of < 130 g/day and > 253 g/day) in type 1 diabetes. 2 Analogous results have been reported in individuals with type 2 diabetes, in which LCD resulted in a significant increase in HDL cholesterol and decrease in high BP. 30,31 Previous studies have demonstrated that abnormal sphingolipids levels are found at the onset of type 1 diabetes in regulating beta cell biology and inflammation. 26 Fenofibrate, a lipid lowering medication, is known to regulate sphingolipid metabolism and has been suggested as an important treatment in the management of dyslipidaemia. 32 It has been shown that very-long-chain sphingomyelin were increased in three weeks fenofibrate-treated NOD mice and this had a beneficial effect on blood glucose homeostasis. 32 Interestingly, the main study outcome from Schmidt et al 18 showed that glycaemic variability and time spent in hypoglycaemia were lower during LCD than during HCD.
As the total amounts of cholesterol did not change with LCD, this was a positive result since hypercholesterolaemia is a major risk factor for developing cardiovascular disease. [33][34][35] The presence of CVD is related to disturbances in lipoprotein metabolism and dyslipidaemia. 36 Dyslipidaemia in individuals with diabetes not only fuels the reduction of HDL-cholesterol concentration, but also it modifies the composition of lipoprotein. [3][4][5][6] Increasing evidence suggests that in patients with chronic inflammatory disorders, HDL cholesterol may lose important antiatherosclerosis properties and become dysfunctional. 36 So far, no therapeutic strategy to raise HDL cholesterol has been successful in reducing CVD. 36 In this study, we found and PC (35:4)) and two sphingomyelin lipid species (SM(d36:2) and SM(d34:2)) associated with HDL cholesterol (Table S2). Interestingly, F I G U R E 1 Jittered box plots of prepost lipid levels comparisons of PC (35:4) in each diet intervention and association with BMI and HDL cholesterol. PC (35:4): (β ± SE: 1.04 ± 0.28, p = .043) was elevated during LCD (a) and inversely associated with BMI (b) and positively associated with HDL cholesterol (c). HB: High carbohydrate diet at baseline, HF: High carbohydrate diet at follow-up, LB: Low carbohydrate diet at baseline, LF: Low carbohydrate diet at followup all these lipid species were increased during the LCD arm suggesting that LCD led to better functionality and composition of HDL-cholesterol particles as they are enriched in plasma phosphatidylcholine (see the "Slope-LCD" column in Table 2).
Remarkably, a previous lipidomics study in 3779 type 2 diabetes-based cohort researchers found that PC(35:4) lipid species out of the 7 novel lipid species were associated with CVD. 37 PC(35:4) was suggested to improve prediction of CVD mortality. Remarkable result to emerge from our data is that PC(35:4) was the strongest positively associated lipid species with HDL cholesterol (Table S2) and increased during the LCD arm. Importantly, PC(35:4) was also inversely associated with BMI ( Figure 1B) and diastolic BP. This and our investigation demonstrate the potential of lipid species as biomarkers for CVD risk in type 1 diabetes. in a larger cohort to fully understand the therapeutic opportunities in type 1 diabetes. Based on our estimate, the 94% probability of replicating the finding already in an identical-sized study is considerably higher than the power of 80% that is typically used as a starting point at planning a study. For a deeper understanding of the mechanisms behind these lipidomic responses, however, a larger study is warranted.

ACK N OWLED G EM ENTS
The authors thank the study participants.

CO N FLI C T O F I NTE R E S T
None of the investigators has personal interests in the conduct or the outcomes of the study.