2.1 Participant characteristics and study design

Three hundred fourteen participants were enrolled in this study, dubbed the Diachron cohort, 274 of which met inclusion–exclusion criteria listed in Tables S1 and S2. Age- and sex-matched study participants belonged to four categories: normoglycemic non-diabetic non-obese (referred to as control subjects in this study), normoglycemic non-diabetic obese (referred to as obese non-T2D subjects), obese with T2D (referred to as obese T2D subjects) and non-obese with T2D (referred to as non-obese T2D subjects). Obesity was defined by body mass index (BMI) > 30; T2D—by glycated hemoglobin (HbA_1c) > 6.5% (equivalent to 48 mmol/mol in IFCC unit). A list of the baseline characteristics of the participants in each group is presented in Table 1. All participants gave informed consent, and the study had ethics committee approval (CER11-015) and was registered at ClinicalTrials.gov (registration no. NCT02384148). All study participants filled out the Munich Chronotype Questionnaire (MCTQ), allowing calculation of MSF_sc (sleep-corrected local time of mid-sleep on work-free days) values that characterize an individual's chronotype. The participants were asked to follow a moderate diet without excess fat or alcohol intake, 24 h prior to the testing day.

2.2 Harvesting of sera

Blood samples for all study participants were collected between 08:00 am and 10:00 am, following overnight fasting from 10 pm. Blood samples were collected in clot-activator vacutainers and immediately analysed by the Geneva University Hospital laboratory for blood glucose, HbA_1c, hormones, lipids, liver and kidney functions (detailed list of the measured blood clinical parameters is reported in Table S3). Serum was immediately prepared from blood samples by centrifugation (10 min, 1650 × g, 4°C) and stored at −80°C for further analyses.

2.3 Primary dermal fibroblast culture and DNA extraction

Cutaneous biopsies were taken from each participant's shoulder between 8:00 am and 10:00 am and processed as described previously (Du & Brown, 2021). DNA was extracted from fibroblast cultures using QIAamp DNA Mini Kit (Qiagen AG, Cat# 51304) and eluted in a final volume of 15 μL.

2.4 Lentivector production

Bmal1-luciferase lentiviral particles (Brown et al., 2005) were produced at the Viral Vector Facility of the University of Zurich. Transient transfection in 293T cells was performed using the polyethylenimine method (Toledo et al., 2009). Lentiviral particles were harvested at 48 h post-transfection, PEG precipitated, titred and used for the transduction of the U2OS cells with multiplicity of infection (MOI) of 3.

2.5 U2OS cell culture, in vitro synchronization, real-time bioluminescence recording and period length analysis

U2OS cells (ATCC, Cat# HTB-96, RRID:CVCL_0042) were cultured in DMEM low glucose (GIBCO) supplemented with 1% Penicillin/Streptomycin (GIBCO, Cat# 15140122), .5% Amphotericine B (Sigma, Cat# A2942), .5% Gentamycin (Sigma, Cat# G1397) and 10% FCS (GIBCO, Cat# A5256701). Cells were transduced with lentivirus expressing Bmal1-luciferase and selected with Blasticidin S (Gibco, Cat# R21001) at 25 μg/mL final concentration. The same batch of transduced U2OS cells was used for all the circadian measurements. After synchronization of the cells with a 30-min pulse of 100 nmol/L dexamethasone, the circadian bioluminescence recording was performed in DMEM low glucose without phenol red (GIBCO, Cat# 31053028) supplemented with 1% Penicillin/Streptomycin (GIBCO, Cat# 15140122), .5% Amphotericine B (Sigma, Cat# A2942), .5% Gentamycin (Sigma, Cat# G1397), 1 mmol/L of luciferin and in the presence of 10% of the individual's sera. Bioluminescence was monitored by a home-made robotic device equipped with photomultiplier tube detector assemblies, allowing the recording of technical triplicates in 24-well plates for 1 week. Photon counts were integrated over intervals of 1 min. After removing the first oscillation cycle (to avoid a potential bias stemming from the immediate early response to synchronization), the remaining cycles were detrended by a moving average with a window of 24 h. Period length was subsequently analysed using the Actimetrics LumiCycle program. The recordings were done in triplicates wells of a 24-well plate (triplicate wells of U2OS cells, with the serum aliquots from the same patient added to the culture medium), and average of the replicates was reported for each experiment. The standard deviation of the replicates for each serum was .566 h ± .410 h for the whole cohort.

2.6 Metabolomics by ultra-performance liquid chromatography-mass spectrometry (UPLC–MS)

2.7 Statistical analysis

All data analyses were conducted in R (v3.6.1). In order to assess the correlation between the circadian period length measured in U2OS cells cultured in the presence of patient's serum and clinical parameters or metabolite levels in serum, we performed Kolmogorow–Smirnow (KS) tests between the first and the fourth data quartile. These comparisons were performed within the different patient groups (non-obese T2D, obese T2D, obese non-T2D, or control, or grouped as stated). P values were reported without or with correction for multiple testing using the Benjamini–Hochberg method in the stats R package. Enrichment analysis of genome wide association study (GWAS) p values against those of other traits was performed using gset package in R. LocusCompare package in R (Liu et al., 2019) was used to compare GWAS and expression quantitative trait loci (eQTL) signal at the March1 locus. Coloc package in R was used to visualize GWAS and eQTL signals at the March1 locus.

2.8 Genotyping

Fibroblasts were genotyped using the Illumina CoreExome 24 v1.3 array. Only samples with variant calling rate > 98% were considered. Population stratification was done by principal component analysis using the phase 31,000 genome variants to select for European subjects, resulting in 269 subjects qualified for GWAS. Variants were then filtered to only choose those from the European panel. Next, variants were filtered using vcftools with the following parameters: --mac 2, --max-missing .95, --hwe .000001, yielding 290′867 genotyped variants. Genotyped variants were imputed using the Michigan Imputation Server with the phase 31,000 genome genotypes as reference. Imputed variants were filtered out according to these criteria: imputation quality > .5, minor allele frequency (MAF) > .05, Hardy–Weinberg probability < 1e-6. A total of 5′630′127 variants were left after filtering these steps.

2.9 GWAS

GWAS was performed on circadian period length (inverse transformed using the following command line in R: period length = qnorm((rank(x)-.5)/length(x)) using PLINK 1.90. Sex, age, disease (control, obese non-T2D, non-obese T2D and obese T2D), date of circadian measurement, experimenter and the first 10 MDS dimensions of the genotypes were included as covariates.

2.10 External database

GWAS summary statistics for chronotypes or diabetes-related traits were downloaded from http://www.nealelab.is/uk-biobank, GWAS round 2. GWAS summary statistics for metabolites were taken from Shin et al. (2014).

2.11 Identification of genes flagged by GWAS signals and analysis of their period length phenotypes in published siRNA screen

Genes flagged by GWAS signals, dubbed here GWAS genes, were identified as genes within 100 kb of the GWAS variants with p < 5 × 10⁻⁴. Detrended bioluminescence data of the GWAS genes and scramble controls were taken from Table S5 from published siRNA screen (Zhang et al., 2009). Period length and rhythmicity were computed by submitting the bioluminescence data to Biodare2 website using the default settings (https://biodare2.ed.ac.uk/, n.d.; Zielinski et al., 2014). Only genes whose rhythmicity were significant (p < .05, not corrected for multiple testing) were considered for period length comparison with scramble controls.

3.1 Study overview

To study the effects of circulating factors on circadian traits in obesity and T2D, circadian rhythms of U2OS cells were continuously recorded in the presence of 10% serum from non-obese T2D, obese T2D, obese non-T2D or non-obese non-diabetic control subjects in the medium. Patients were genotyped to investigate genetic origin of individual differences in sera that affect cellular oscillations. For deep metabolic phenotyping, patients' sera were subjected to metabolomics analysis by LC–MS (Figure 1a).

3.2 Circadian period length of U2OS cells assessed in the presence of obese patients' serum increases concomitant with severity of obesity

We observed huge inter-individual effects of sera on circadian period length measured in U2OS cells that varied between 20.69 and 25.62 h across the entire cohort, with mean and standard deviation that were 22.9 and .8 h, respectively (Figure 1b). The mean period length and standard deviation for each patient group were the following: control, 23.2 ± .7 h; obese non-T2D, 22.8 ± .8 h; non-obese T2D, 22.9 ± .7 h; and obese T2D, 23.0 ± .8 h. A one-way analysis of variance (ANOVA) revealed that there was a statistically significant difference in period length by patient group (F(3) = 4.308, p < .01). A Tukey post hoc test found that the mean value of period length was significantly different between obese non-T2D and control (p < .01, 95% C.I. = [−.68, −.1]) but not for any other pair-wise comparisons. Although statistical significance was observed, the huge inter-individual differences within each patient group and across the entire cohort suggest that such significance is of low biological relevance. It is more likely that inter-individual differences represent the main driver of the variation in circadian modifying effects of subjects' sera.

We next investigated the sources of inter-individual differences in serum factors that could explain the observed effects on circadian period length within each patient group. To this end, sera were divided into quartiles based on various clinical parameters listed in Table 1. We then compared the cumulative distribution of cellular period length in the presence of sera from the first quartile with the one from the fourth quartile. For serum from T2D patients (non-obese T2D and obese T2D combined), such comparison yielded no statistically significant differences in the distribution of period length, except for triglycerides and ASAT (Figure S1A, Table S7, p < .05 for triglycerides and ASAT, KS test, not corrected for multiple testing). However, in the presence of sera from obese subjects (obese non-T2D and obese T2D combined), a period lengthening in U2OS cells was observed between quartiles separated based on increased Homeostatic Model Assessment for Insulin Resistance (HOMA-IR), insulin, HbA_1c, fasting blood glucose, triglycerides and decreased high density lipoprotein (HDL) (Figure 2, Table S8, all p < .05, KS test, not corrected for multiple testing). Overall, deteriorated metabolic health in obese patients correlated with cellular period lengthening. For parameters that reflect liver and kidney function (ASAT and creatinine), a similar relationship was observed: worse liver and kidney function correlate with longer period length. Importantly, the same observation was not present for non-obese subjects (control and non-obese T2D combined) for these parameters (Figure S1B, Table S9, all p > .05, KS test, not corrected for multiple testing), nor for control subjects (Table S10, all p > .05, KS test, not corrected for multiple testing). Overall, these observations suggest that severity of obesity is the most important aspect that correlates with differential effects of patients' sera on circadian period length.

3.3 Genome wide association analysis identified March1 as the gene most associated with period lengthening effects

Reasoning that individual differences in period lengthening by serum have genetic origin, we sought for genetic variants associated with period length measured in the presence of patient serum. GWAS identified 613 variants that belong to 128 loci at suggestive genome-wide significance (p < 5 × 10⁻⁴, Tables S11 and S12) that were associated with period length across the entire cohort. Interestingly, while the three top identified variants were intergenic, the fourth most associated variant (rs7654787) mapped to an intron of March1 gene, known for its functions in antigen-presenting cells (Liu et al., 2019) (Figure 3a, Figure S2). Noteworthy, March1 knockout animals exacerbate obesity-induced insulin resistance stemming from its effects on CD8 T cell fate (Majdoubi et al., 2020). In addition, March1 knockout animals also exhibit enhanced insulin sensitivity, and knockdown experiments show that MARCH1 degrades surface insulin receptor in the basal state (Nagarajan et al., 2016). The involvement of March1 in insulin regulation prompted us to investigate the relationship between the genotype of March1 variant, metabolic status and insulin resistance in our cohort. A two-way ANOVA was performed to analyse the effect of March1 variant and obese status on insulin and HOMA-IR. Simple main effects analysis showed that genotype of March1 variant did not have a statistically significant effect on both insulin and HOMA-IR (p = .257 and .238, respectively). Obese or T2D status had a statistically significant effect on insulin and HOMA-IR (p < .005 for all) as expected. Interestingly, the interaction effect between genotype and obese status on insulin showed a modest significance, F(1, 2) = 2.579, p = .078, indicating that the effect of obese status on insulin levels differs among genotypes. Indeed, we observed that obese subjects carrying the homozygous AA allele had similar insulin levels compared with non-obese subjects (Figure 3b, p = .13), in contrast to the expected higher insulin for obese subjects bearing other genotypes (both p < .0001). Similarly, the interaction effect between genotype and obese status on HOMA-IR showed a modest significance, F(1, 2) = 2.935, p = .055. We also observed that obese subjects carrying the homozygous AA allele had similar HOMA-IR compared with non-obese subjects (Figure 3c, p = .45), in contrast to the expected higher HOMA-IR when being obese with the other genotypes (both p < .01). The above observations are in line with allele A being associated with shorter period length (Figure 3a). In addition, the interaction effect of genotype and metabolic status was specific for obese status, as the interaction effect of genotype and T2D on insulin and HOMA-IR did not yield statistical significance (p = .94 and .52, respectively). Accordingly, the differences in insulin and HOMA-IR by AA genotype versus other genotypes of this March1 variant were not observed when patients were stratified by T2D status (Figure S3). This is consistent with our former observation regarding the correlation of the metabolic status in obese patients with the period length in U2OS cells.

We next sought for evidence of activity at rs7654787 in publicly available data. We found that this variant is located within 2 kb of cis regulatory elements (CRE, H3K4me3, H3K4me1 and H3K27Ac signatures) active in 7 cell lines in the ENCODE dataset (Figure S5A). In addition, eQTL analysis shows the correlation between genotypes and gene expression. We sought for published eQTLs in the pancreas (Viñuela et al., 2020) at rs7654787 and found suggestive evidence that rs6536810, which is in linkage disequilibrium with rs7654787 (R² = .9956), is associated with lower expression of March1 (p = 4e-07, slope = −.21, T statistic = −5.15, Figure S5B). In other words, combined with our data, allele A of March1 variant is associated with lower expression of the gene, shorter period length and lower insulin and lower HOMA-IR. This observation is in line with the reported finding that MARCH1 expression increased in insulin-resistant versus insulin-sensitive subjects (Nagarajan et al., 2016). This is also consistent with the results that March1 knockdown and knockout mice showed improved glucose tolerance without an increase in insulin levels, suggesting enhanced insulin sensitivity in mice (Nagarajan et al., 2016). Noteworthy, although different analyses suggested that subjects carrying allele A of March1 would have lower gene expression, leading to lower blood insulin levels that are associated with a shorter period length when added to U2OS cells, these analyses only show association and not causality. Therefore, the relevance of the identified variants for the circadian clockwork was further validated below.

3.4 Genes flagged by GWAS signals showed period phenotype in siRNA screen

To investigate the biological relevance of the GWAS variants, we identified 270 genes within 100 kb of the significant variants (Table S13) and searched for their period length phenotype in the published siRNA screen in U2OS cells (Zhang et al., 2009). Strikingly, knockdown of March1 led to shortening of cellular oscillations measured in U2OS cells (Figure S4A, Welch two sample t test, p = .013, not corrected for multiple testing), consistently with our findings. The shorter period length phenotype was expected from the above-presented correlation analyses. For the 270 genes flagged by GWAS signals, 178 of them were included in the siRNA screen. Among the 178 genes, 87 showed statistically significant difference (not corrected for multiple testing) in period length compared with scramble controls (Figure S4B). When p values were corrected for multiple testing, 60 of them remained significant, including March1 (Figure S4C). Although the knockdown efficiency in an siRNA screen could not be confirmed, the prevalence of multiple period length phenotypes among the GWAS genes suggested their relevance in circadian properties.

3.5 Identified GWAS variants are enriched in previously reported variants associated with extreme chronotype and T2D diabetes

Because the here detected phenotype links circadian properties to diabetic status, we reason that there would be overlaps of our GWAS variants with those for chronotypes and diabetes-related traits. Indeed, our GWAS identified variants were enriched in those associated with ‘self-reported chronotype’ and related traits, for instance variants linked to ‘sleep duration’ and ‘nap during the day’ in the UK Biobank database (Figure 4a). There was also enrichment for ‘diagnosed type 2 diabetes’ variants but interestingly not for ‘self-reported type 2 diabetes’ variants. In addition, GWAS identified variants were enriched in ‘job involving night shift’ variants, as well as ‘depression’, two traits that are often associated with T2D (Holt et al., 2014; Strohmaier et al., 2018). Moreover, the beta coefficient of the association, which denotes the directions of association, for example if a variant is associated with longer or shorter period length, called shortly here beta coefficient directions, were largely consistent between GWAS variants and trait variants (Figure 4b–d). Variants that were associated with shorter period length were also associated with morningness, whereas variants that were associated with diabetic status were also associated with longer period. This is in agreement with our previous finding that in vitro circadian period length correlates with human chronotype (Brown et al., 2005) and published work showing that late chronotype is associated with worse glycemic control (Iwasaki et al., 2013; Reutrakul et al., 2013).

3.6 Metabolomics reveals that circadian period lengthening is associated with insulin resistance in obese individuals

We observed that within the obese group, circadian period length did not correlate with BMI (p = .43, KS test), indicating that obesity index alone cannot explain the period lengthening effects of sera from these subjects. It has been recognized that risk for cardiometabolic abnormalities varies among obese patients (Iacobini et al., 2019). In our Diachron cohort, we indeed observe such variation, with obese patients whose sera were classified in the first quartile in terms of triglycerides and fasting blood glucose and the fourth quartile of HDL cholesterol fitting the criteria for low cardiometabolic risk based on these parameters (Tsatsoulis & Paschou, 2020) (Table S14, fasting serum triglycerides ≤ 1.7 mmol/L, fasting blood glucose ≤ 6.1 mmol/L and HDL cholesterol serum concentrations > 1.0 mmol/L in men or >1.3 mmol/L in women). This suggests that metabolic signature in the serum of obese patients affects circadian period length.

In order to obtain more detailed metabolic signature, we performed targeted and untargeted metabolomics of the patient sera by UPLC-MS. We detected 371 features in total, including 40 from targeted, and 331 from untargeted ones (Tables S15 and S16). Annotation of untargeted features by MS-DIAL (Tsugawa et al., 2015) identified 78 of them (Table S17). We next analysed the metabolic features in obese subjects (non-diabetic and diabetic combined) by quartiles based on each of detected metabolites. The cellular period length distribution measured in serum from the first and the fourth quartile group was compared. We observed a statistically significant shift in the distribution of period length between the first and fourth quartile for a range of compounds, including targeted and annotated untargeted ones (Figure 5a,b). For most of the compounds, the distribution of period length shifted towards longer values with higher levels of compounds, with few compounds only exhibiting the opposite tendency (one of lysophosphatidylcholine species, succinic acid and serotonin, see Section 4). In addition to annotation of untargeted feature, we also employed the mummichog algorithm to infer pathway activity without feature identification to gain information from the whole untargeted metabolites (Li et al., 2013). Metabolic pathway enrichment analysis on all untargeted features suggested that BCAA degradation and biosynthesis were as among the most involved metabolic pathways (Table S18). It has been reported that circulating BCAAs are elevated in subjects with insulin resistance and T2D (Lynch & Adams, 2014). Concordantly, we observe that many of the metabolites associated with period length were also associated with insulin resistance score (HOMA-IR) in our cohort (Figure S6).

In addition, we explored the association between serum lipid landscape and cellular period length, based on the serum lipidomics analyses on the sub-set of this cohort that we have recently reported (see Table S19) (Sinturel et al., 2023). We detected statistically significant association between longer cellular period and lower levels of phospholipids, specifically LysoPE, phosphatidylcholines (PC) and LysoPC species (Figure 5c). This finding is consistent with the earlier observation that lower levels of phospholipids are associated with insulin resistance (Tonks et al., 2016; Yin et al., 2020). Overall, we report that serum metabolic status associated with insulin resistance in obese patients may account for the period lengthening effect in U2OS cells.

3.7 GWAS variants are enriched in BCAA variants

If serum metabolite levels are associated with serum's effects on period length that are partially explained by genetic variants, there should be overlaps between our GWAS variants and those explaining metabolite levels. Indeed, we found that GWAS variants are enriched in BCAA and branched-chain keto acid (BCKA) variants (Figure S7A). Comparison of beta coefficient direction between GWAS variants and the metabolite variants confirmed that longer period measured in U2OS cells is associated with higher levels for all BCAAs and BCKAs in the patients' sera (Figure S7B).