In this study, the samples were obtained from the Finnish Type 1 Diabetes Prevention and Prediction Study (DIPP) (15), an observational birth cohort study following the participants from the age of 3 months to 15 years of age or to the diagnosis of type 1 diabetes. The DIPP study was initiated in 1994 and is currently running in three university hospitals in Finland (Oulu, Tampere, and Turku). The participants involved in the current study were chosen from the subset of DIPP children who were born in Turku University Hospital since sequential unthawed serum samples, collected between 1998 and 2012 and stored at -80°C were available from these participants. The study protocol was approved by the ethics committee of the Hospital District of Southwest Finland. The study was conducted according to the guidelines in the Declaration of Helsinki. All families provided written, informed consent for participation in the study. For each child, longitudinal samples for metabolomics analysis were obtained at the age of 3, 6, 12, 18, 24 or ≥ 36 months ( Figure 1 ).

This study included samples (n=239 for lipidomics and n=213 for polar metabolites) from 66 children, divided into two groups, being 25 RPs and 41 SPs. Selected characteristics of the participants in this study are listed in Table 1 . In this study, non-fasting blood samples were collected, serum was prepared within 3 hours of sample collection and stored at -80°C until analyzed.

Screening for HLA-conferred susceptibility to type 1 diabetes was carried out using cord blood samples. The HLA-genotyping was performed with a time-resolved, fluorometry-based assay for four alleles using lanthanide chelate-labelled, sequence-specific oligonucleotide probes detecting DQB1*02, DQB1*03:01, DQB1*03:02, and DQB1*06:02/3 alleles (16). The carriers of DQB1*02/DQB1*03:02 or DQB1*03:02/x genotypes (here x≠ DQB1*03:01, DQB1*06:02, or DQB1*06:03 alleles) were categorized as children with increased HLA susceptibility to type 1 diabetes and accordingly invited to the DIPP follow-up program.

The study participants underwent more extensive HLA genotyping. This genotyping defined all common European HLA-DR-DQ haplotypes at low resolution and at higher resolution haplotypes where this was relevant for the assessment of the risk for type 1 diabetes, e.g. HLA-DR4 subtypes in DR4-DQ8 haplotypes. In a series of 2,991 family trios from the Finnish Pediatric Diabetes Register, the genotype risks were defined and genotypes were combined into six groups from O (strongly protective) to 5 (high risk) which did not overlap for 95% confidence intervals of their OR values for type 1 diabetes (17).

The children with HLA-conferred genetic susceptibility were prospectively observed for levels of diabetes-associated autoantibodies [islet cell antibodies (ICA), insulin autoantibodies (IAA), and glutamic acid decarboxylase antibodies (GADA) and insulinoma-like antigen 2 antibodies (IA-2A)]. These autoantibodies were analyzed from serum samples taken at each follow-up visit as previously described (18). ICA levels were determined using an approved, immunofluorescence assay with a detection limit of 2.5 Juvenile Diabetes Foundation Units (JDFU) (19). GADA and IAA levels were quantified using specific radiobinding assays, the threshold of positivity being 5.36 and 3.48 relative units (RU) respectively (20, 21). Similarly, IA-2A levels were measured with a radio-binding assay with a threshold of 0.43 RU (22).

The samples were randomized and extracted using a modified version of the previously-published Folch procedure (23). 10 µl of serum was mixed with 10 µl 0.9% NaCl and extracted with 120 µl of CHCl₃: MeOH (2:1, v/v) solvent mixture containing internal standard mixture (c = 2.5 µg/ml).; 1,2-diheptadecanoyl-sn-glycero-3-phosphoethanolamine [PE(17:0/17:0)], N-heptadecanoyl-D-erythro-sphingosylphosphorylcholine [SM(d18:1/17:0)], N-heptadecanoyl-D-erythro-sphingosine [Cer(d18:1/17:0)], 1,2-diheptadecanoyl-sn-glycero-3-phosphocholine [PC(17:0/17:0)], 1-heptadecanoyl-2-hydroxy-sn-glycero-3-phosphocholine [LPC(17:0)] and 1-palmitoyl-d31-2-oleoyl-sn-glycero-3-phosphocholine [PC(16:0/d31/18:1)], cholest-5-en-3ß-yl heptadecanoate [CE(17:0)] and, triheptadecanoylglycerol [TG(17:0/17:0/17:0)].

The samples were vortex mixed and incubated on ice for 30 min after which they were centrifuged at 7800 × g for 5 min. Finally, 60 µL from the lower layer of each sample was collected and mixed with 60 µL of ice-cold CHCl3: MeOH (2:1, v/v) in an LC vial.

The Ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry (UHPLC-QTOFMS) analyses were carried out in a similar manner to that described earlier, with some modifications (24, 25). The UHPLC-QTOFMS system was from Agilent Technologies (Santa Clara, CA, USA) combining a 1290 Infinity LC system and 6545 QTOFMS, interfaced with a dual jet stream electrospray (dual ESI) ion source. MassHunter B.06.01 software (Agilent Technologies, Santa Clara, CA, USA) was used for all data acquisition and MZmine 2.53 was used for data processing as described below (26). Identification of lipids was based on in house laboratory LC-MS/MS data on retention time and mass spectra.

Chromatographic separation was performed using an Acquity UPLC BEH C18 column (100 mm × 2.1 mm i.d., 1.7 µm particle size) and a C18 precolumn, both from Waters Corporation (Wexford, Ireland). The mobile phases were water (phase A) and acetonitrile:2-propanol (1:1, v/v) (phase B), both containing 1% 1M ammonium acetate and 0.1% (v/v) formic acid ammonium acetate as ionization agents. The LC pump was programmed at a flow rate of 0.4 mL min^–1 and the elution gradient was as follows: from min 0–2, the percentage of phase B was modified from 35% to 80%, from min 2-7, the percentage of phase B was modified from 80% to 100% and then, the final percentage was held for 7 min. A post-time of 7 min was used to regain the initial conditions for the next analysis. Thus, the total analysis time per sample was 21 min (including post processing). The settings of the dual ESI ionization source were as follows: capillary voltage 3.6 kV, nozzle voltage 1500 V, N₂ pressure in the nebulizer 21 psi, N₂ flow rate and temperature as sheath gas 11 L min^–1 and 379°C, respectively. Accurate mass spectra in the MS scan were acquired in the m/z range 100–1700 in positive ion mode.

The peak area obtained for each lipid was normalized with lipid-class-specific internal standards. A (semi)quantitation was performed using lipid-class-specific calibration curves. Pooled samples were used for quality control, in addition to NIST CR-1950 serum and in-house plasma QC serum. The RSD of the concentrations of the identified lipids in QC samples and pooled extracts was on average 6.3% and 16.2%, respectively.

To a 10 µL aliquot of plasma, 225 µL of ice-cold MeOH (LC-grade, Honeywell) containing the following internal standards (all from Sigma Aldrich): Heptadecanoic acid (5 ppm), DL-valine-d8 (1 ppm), and succinic acid-d4 (1 ppm) was added. The samples were then sonicated in an ice bath for 30 s prior to centrifugation (5500 g, 5 min). 250 µL of the supernatant was transferred to a 2 mL glass auto sampler vial. The pellet was stored at –20°C for protein analysis. The protein content was measured by the Bradford method. The supernatant was dried under a stream of nitrogen at 45°C. Prior to the mass spectrometry measurements, the samples were derivatized using a two-step procedure. Initially the samples were methoximated by incubating the samples with methoxyamine hydrochloride (25 µL, 20mg/mL in pyridine, Sigma Aldrich) at 45°C for 1 h. N-Methyl-N-trimethylsilyltrifluoroacetamide (25 µL, Sigma Aldrich) was then added and the samples were incubated for a further 60 min. A retention index standard containing straight chain, even alkanes (n 10-40, 10 µL, Sigma Aldrich) was added. All this sample prep was performed on a Gerstel robot and timed to the injection into the GC-MS. The derivatized samples were analyzed using gas chromatography (Agilent 7890B) coupled to a time of flight mass spectrometer (Pegasus BT LECO). The metabolites were separated using a 30 m × 0.25 mm (ID) with a film thickness of 0.25 µm HP-5 (Agilent). A guard column (10 m) with an ID of 0.25 mm was used. 1 µL of the sample was injected in split less mode with an inert glass liner (Agilent) held at a temperature of 240°C. The GC was set to constant flow mode (1.2 mL/min) using helium (Aga) as the carrier gas. The GC oven was programed as follows: 50°C (isothermal for 0.2 min), then 7°C/min until 240°C, then 20°C/min until 300°C (isothermal for 5 min). The transfer line was held at 230°C for the whole run. The ion source was set to electron ionization mode and held at 250°C. The MS was scanning from 50 – 650 Da with an extraction rate of 30 kHz and a scan rate of 16 per second.

Lipidomics data processing was performed using open source software MZmine 2.53 (26). The following steps were applied in the processing: 1) Crop filtering with a m/z range of 350 – 1700 m/z and a RT range of 2.0 to 12 min, 2) Mass detection with a noise level of 1200, 3) Chromatogram builder with a minimum time span of 0.08 min, minimum height of 1000 and a m/z tolerance of 0.006 m/z or 10.0 ppm, 4) Chromatogram deconvolution using the local minimum search algorithm with a 70% chromatographic threshold, 0.05 min minimum retention time (RT) range, 5% minimum relative height, 1200 minimum absolute height, a minimum ratio of peak top/edge of 1 and a peak duration range of 0.08 - 5.0, 5) Isotopic peak grouper with a m/z tolerance of 5.0 ppm, RT tolerance of 0.05 min, maximum charge of 2 and with the most intense isotope set as the representative isotope, 6) Join aligner with a m/z tolerance of 0.008 or 10.0 ppm and a weight for of 2, a RT tolerance of 0.1 min and a weight of 1 and with no requirement of charge state or ID and no comparison of isotope pattern, 7) Peak list row filter with a minimum of 12 peaks in a row (= 10% of the samples), 8) Gap filling using the same RT and m/z range gap filler algorithm with an m/z tolerance of 0.006 m/z or 10.0 ppm, 9) Identification of lipids using a custom database search with an m/z tolerance of 0.006 m/z or 10.0 ppm and a RT tolerance of 0.1 min, 10) Normalization using lipid-class-specific internal standards and (semi) quantitation with lipid-class-specific calibration curves, 11) Normalization with total protein amount 12) Data imputation of missing values were done with half of the row’s minimum.

The GC-QMS data was processed in ChromaTOF (v5.51, LECO) using the peak find algorithm. The peak intensities were normalized to heptadecanoic acid. The peaks were manually checked and corrected if needed for correct integration. Metabolites which had a coefficient of variation greater than 30% in the pooled quality control sample or fell below the limit of quantification were excluded from subsequent analysis.

The metabolites and lipids data values were log transformed prior to analysis. The difference in the lipidome and metabolome between the studies groups were compared using a multivariate linear model. For longitudinal samples, linear mixed effects model were regressed with fixed effect (~ sex +case + age) and random effect ~ (1 | Subject). For age wise comparisons the lipids/metabolites were regressed with various factors such as sex, and disease conditions (e.g. RP vs. SP) using MaAsLin2 package in R (lipids ~ sex +case). To subsequently visualize metabolite level, violin plots from the ggplot2 R package were used.

Pathway analysis of the significant metabolites (p-value < 0.05) was performed in MetaboAnalyst 5.0 (27). The compounds unmatched during compound name matching were excluded from the subsequent pathway analysis. We implemented global test hypergeometric testing for the functional enrichment analysis. The pathway topological analysis was based on the relative betweenness measures of a metabolite in a given metabolic network and for calculating the pathway impact score. Based on the impact values from the pathway topology analysis, the impact value threshold was set to >0.10.

We performed untargeted metabolomics and lipidomics analysis in a longitudinal setting among two study groups: 25 rapid progressors who progressed to type 1 diabetes within 18 months after the appearance of the first autoantibody (RP) and 41 slow progressors, who presented with clinical disease after a period of at least 60 month from the initial seroconversion to clinical disease (SP). For each child, samples were analyzed from up to six time points, corresponding to the ages of 3, 6, 12, 18, 24, or ≥36 months ( Figure 1 ). The lipidomics dataset (n = 239) included the identified lipids from the following lipid classes: cholesterol esters (CE), ceramides (Cer), diacylglycerols (DG), lysophosphatidylcholines (LPC), phosphatidylcholines (PC), phosphatidylethanolamines (PE), sphingomyelins (SM), and triacylglycerols (TG). Metabolomic dataset (n = 213) included the metabolites from chemical classes, including amino acids, carboxylic acids, phenolic compounds, sugars, and sugar derivatives.

To identify the sources of variation in the lipidomics and metabolomics datasets, multivariable linear model was performed. The concentrations of lipids/metabolites were regressed on factors including age, sex, and case-related phenotypes (e.g., SP vs. RP). Among these factors, age showed the strongest effect on the circulating lipidomics and metabolomics profiles. A total of 44 polar metabolites and 106 lipids were related to age (p-value < 0.05, Supplementary Material SM Tables 1, 2 ). A total of 35 out of 175 lipids differed between male and female in the longitudinal setting (p-value < 0.05, SM Table 3 ). However, no sex differences were observed in the metabolomics dataset.

A total of 21 polar metabolites differed between the study groups (SP vs. RP; p-value < 0.05, Figure 2 , SM Table 4 ). The tryptophan-derived microbial catabolite 3-indole acetic acid was downregulated in the SP group ( Figures 2A, B , p-value <0.05). While most of the sugar derivatives were upregulated in the RP compared to the SP group, quinic acid remained upregulated in the SP group ( Figure 2C ). Figures 2A-C shows the local polynomial regression fitting (LOESS) plot of selected metabolite class with time for the case-related phenotypes.

Considering that age is a confounding factor in the metabolomics analysis, we also performed age-matched comparisons between SP and RP ( Figure 2D ). We found that the main metabolic differences between these two groups were after the median age of seroconversion (i.e., around 24-36 months of age). The result showed that at the age of 3, 6, 12, 18, 24 and ≥36 months, a total of 12, 3, 2, 14, 17 and 23, metabolic features were altered, respectively, between the study groups ( Figure 2D ).

Next, the changes in plasma lipid concentration between the study groups in the longitudinal series sample collection were examined. We found that lipid-related differences were less pronounced as compared to polar metabolites. Levels of two lipids, PE (34:0) and TG (18:2/18:1/18:1), were different between SP and RP (p-value < 0.05). With the exception of lipid PE (34:0), there was no consistent lipid related trend between the study groups. The age-wise comparison showed that at the age of 3, 6, 12, 18, 24 or ≥36 months, a total of 1, 1, 1, 13, 3 and 10 lipids were different between the SP and RP groups, respectively ( ESM Figure S1 ).

As progression of type 1 diabetes is associated with distinct islet autoantibody trajectories, we compared differences in metabolite changes in crelation to the development of specific islet autoantibodies (IAA vs. others). With the exceptions of a carboxylic acid derivative, no other differences (p-value <0.05) between the IAA vs. other autoantibodies were observed, which may be due to the small sample size per autoantibody grouping.

We then examined the differences between the SP and RP groups at the metabolic pathway level. The selected metabolites/lipids that differed between the case-related phenotypes (SP vs. RP) were subjected to metabolic pathway analysis in MetaboAnalyst 5, integrating enrichment-based analysis and pathway topology analysis. Using pathway impact estimates, we found that mainly amino acid -related metabolic processes including phenylalanine, tyrosine and tryptophan biosynthesis, ascorbate and aldarate metabolism, and tyrosine metabolism, pentose and glucuronate interconversions were different between SP and RP ( SM Figure S2 , impact value threshold was set to >0.10).

We also studied the metabolic changes between the SP and RP groups (1) prior to the appearance of islet autoantibodies and (2) after the first appearance of islet autoantibodies. We identified differences in polar metabolite profiles after the appearance of the first islet autoantibody when comparing rapid and slow progressors to type 1 diabetes. Altogether, 13 metabolites remained upregulated in rapid progressors as compared to slow progressors (p-value <0.05, SM Figure S3 ). These metabolites can be divided into major chemical classes including sugars, amino acids, and a microbiota-derived metabolite indole-3-acetic acid (i.e., tryptophan catabolite). Among these 13 metabolites, comparative profiles of indole-3-acetic, threonine, beta-tocopherol and hexose are shown in Figure 3 (p-value <0.05, ESM Table 5 ). We also analyzed metabolite concentration differences between rapid progressors and slow progressors prior to the appearance of islet autoantibodies. Quinic acid and an unknown metabolite were downregulated in rapid progressors when compared to slow progressors before seroconversion (p-value <0.05).

Next, we studied whether lipidomic patterns were related to the development of islet autoantibodies. For that, we analyzed the lipid concentration differences between rapid and slow progressors before and after the emergence of the first islet autoantibody. The results revealed weak differences in lipid levels between the slow and rapid progressors, after or before seroconversion to islet autoantibody positivity. We found two lipid species [PE (O-16:0/22:6) and Cer (d18:1/24:0)] downregulated in slow progressors after the appearance of first islet autoantibodies, when compared to rapid progressors (p-value <0.05). Likewise, there were two lipids [PC (O-38:6) and PE (34:0)] that differed between rapid and slow progressors before the appearance of the first islet autoantibodies (p-value <0.05).