Determinants of plasma levels of proglucagon and the metabolic impact of glucagon receptor signalling: a UK Biobank study

The UK Biobank is a large prospective research resource including half a million participants aged 40 to 69 at the time of inclusion from the UK. The biobank encompasses genetic, lifestyle and health data derived from various sources such as questionnaires, physical assessments, biological specimens, imaging and the continual monitoring of health-related outcomes, as described in detail previously [9]. Participants who withdrew from the biobank were excluded from all analyses (n=179, updated 14 November 2023). We excluded individuals who were not categorised as white and individuals with sex chromosome aneuploidy. A list of field names used in the study is available in electronic supplementary material (ESM) Table 1.

Proteomics data for 1464 proteins and 54,219 individuals were generated by Olink (Uppsala, Sweden) using the Olink Explore 1536 PEA (proximity extension assay). The proteomics measurements are available as Normalised Protein eXpression (NPX) units (an arbitrary and normalised unit on a log₂ scale). The data was accessed through the DNAnexus platform (DNAnexus, Mountain View, CA, USA) and processed locally. Processing of the raw data and normalisation has been described elsewhere [10, 11]. Overall, there was 2.9% missing data (4.5% missing data for proglucagon). In preliminary analyses, we found no association between the proportion of missing data and sex or having a glucagon receptor variant compared with the wild-type (see Methods: Exome sequencing analysis and variant annotation). We filtered out individuals and proteins with >10% missing data (4 proteins and 3298 individuals were excluded). On the basis of the assumption that missing values are missing because they are below the limit of detection, we used the MinProb method (ImputeLCMD package for R, v. 2.0; R Foundation, Vienna, Austria) to impute the remaining 1.1% missing data (1.7% missing data for glucagon). This approach aims to preserve the overall structure of the dataset by conservatively estimating the missing values; however, imputation may lead to a loss of information, particularly in terms of variability. The resulting proteomics sub-cohort consisted of 40,164 individuals, and differential expression analysis was conducted on 1460 proteins to uncover plasma proteins potentially regulated by glucagon receptor signalling. In a prospective analysis the association between concentrations of proglucagon and the risk of incident diabetes was evaluated.

We analysed the whole-exome sequencing data of 469,914 individuals from the UK Biobank [9]. The UK Biobank whole-exome sequencing data was reference-aligned with the Original Quality Functional Equivalent protocol previously described [12]. This protocol uses Burrows–Wheeler Alignment Maximal Exact Matches (BWA-MEM) [13] to map all the reads to the human reference genome GRCh38 [14]. Variant calling was performed using DeepVariant (v.1.5.0) [15]. We filtered the genomic Variant Call Format (gVCF) files for each sample, restricted to the location of the glucagon receptor at chr17: 81,804,150 to 81,814,008 forward strand. The analyses were conducted on the Research Analysis Platform (https://​ukbiobank.​dnanexus.​com, accessed August 2023). A computing instance equipped with 36 CPU threads was chosen to run 36 bcftools (Swiss army knife v.4.9.1; DNAnexus, Mountain View, CA, USA) parallel jobs and automatically launch new jobs when the previous jobs had finished. The output csv files were subsequently merged using Python. The scripts used to filter the gVCF files from UK Biobank and merge the output csv files are available at https://​github.​com/​nicwin98/​UK-Biobank-GCG.

We filtered for genotype quality (GQ) <20, depth (DP) <10, and allele balance (AB, for the minor allele) <0.2. The genetic variants were annotated for their sequence effect with opencravat.org (accessed August 2023) [16]. We created a variant group ‘Frameshift’ for predicted loss-of-function (LoF) alleles. This group included the sequence ontologies frameshift elongation, frameshift truncation, in-frame deletion, in-frame insertion, start lost, and stop gained (see ESM Table 2). Missense variants were denoted as the reference amino acid (1 letter code), the amino acid position, followed by the alternative amino acid. Missense variants were subsequently categorised as G40S heterozygotes and G40S homozygotes (pooled in the proteomics sub-cohort), or as ‘cAMP LoF’ based on a previous study reporting the molecular phenotype of 38 missense variants [5].

We defined type 2 diabetes based on hospital diagnoses encoded as E11 or E14 in the ICD-10 classification system (https://​icd.​who.​int/​browse10/​2019/​en). We excluded individuals with a diagnosis of type 1 diabetes (E10). Prevalent cases were defined as: (1) probable and possible type 2 diabetes based on the Eastwood algorithm [17], (2) with a baseline HbA_1c greater than 48 mmol/mol (6.5%) (a recommended cutoff point for diagnosing type 2 diabetes), or (3) a diagnosis before or within 6 months after the enrolment visit. A list of fields used for the definition of diseases is available in ESM Table 3. After exclusion of 1833 prevalent cases, a total of 1562 developed incident type 2 diabetes during follow-up (median follow-up time: 14.75 years).

Risk time (in months) was defined from date of baseline examination (between 2006 and 2010) where the blood sample for proteomics analysis was obtained, to date of type 2 diabetes diagnosis, death, end of follow-up (defined as the last updated version of the hospital register [3 October 2021]), or loss to follow-up, whichever occurred first. Data on loss to follow-up was last updated in May 2017, and censored individuals were included in the analysis up to the point of censoring.

For the Kaplan–Meier survival analysis, the cohort was stratified into tertiles based on their plasma proglucagon levels at the baseline visit. Pairwise comparison using the Logrank test with Bonferroni correction for multiple testing was used to compare survival curves between the three subgroups. Using Cox proportional hazard regression with adjustment for age and sex, we explored BMI (continuously) as a potential intermediate. However, due to lacking model fit, our final analysis instead stratified on BMI categories (BMI <25 kg/m²; 25 kg/m² ≤ BMI <30 kg/m²; BMI ≥30 kg/m²) and addressed whether the association between proglucagon and incident type 2 diabetes differed across BMI categories. Proportional hazards and linearity assumptions for all covariates were assessed with Schoenfeld and Martingale residual plots, respectively. Survival analyses were done using the Survival package for R (v3.5-4) [18].

For association analyses, we used the following variables: age when attending the assessment centre; sex; BMI; baseline type 2 diabetes (defined as probable and possible type 2 diabetes based on the Eastwood algorithm) [17]; liver fat (quantified by MRI–proton density fat-fraction [PDFF]) at the second repeat visit) [19]; weekly alcohol consumption; glucose; HbA_1c; amino acids. BMI was used continuously or as a binary variable (non-obese, BMI <25 kg/m²; obese, BMI ≥30 kg/m²). Liver fat was used as a continuous variable or a binary variable to define MASLD (non-MASLD, <5.5%; MASLD, ≥5.5%), with individuals having a weekly alcohol intake above 17.5 units for women and 26.25 units for men being excluded in both cases. A list of type 2 diabetes medications tested for confounding is provided in ESM Table 4.

Differences in plasma levels of proglucagon by disease groups and respective control groups were assessed by unpaired t tests. For linear models, increments of independent variables were set to 5% for liver fat, 5 years for age, and 5 mmol/l for serum creatinine. BMI, glucose, HbA_1c and amino acids were normalised to the SD of each variable. Glucagon receptor variant groups were tested for association of binary traits with logistic regression with Firth correction.

Differential expression analysis of the wild-type reference group vs the glucagon receptor variant groups (i.e. frameshift, cAMP loss-of-function, and G40S) respectively, was conducted using Limma software (v.3.56.2) [20]. Protein NPX values from the Olink proteomics analysis were the outcome and variant group, age and sex were predictors. Multiple comparisons were adjusted using the Benjamini–Hochberg method. R version 4.3.0 was used for all analyses.

Plasma proglucagon was available as a part of the proteomics analysis in a sub-cohort that matched the UK Biobank cohort on age, sex, and recruitment centre [10]. BMI, type 2 diabetes prevalence, liver fat, and HbA_1c were also similar between the two cohorts (ESM Table 5). We investigated whether plasma proglucagon (available as NPX units [an arbitrary and normalised unit on a log₂ scale]) was associated with BMI, type 2 diabetes and MASLD. Plasma proglucagon was increased in individuals with obesity (fold change [FC] 1.43, p=1×10⁻⁷²) (Fig. 1a), type 2 diabetes (FC 2.36, p=3×10⁻⁷⁵) (Fig. 1b) and MASLD (FC 1.23, p=1×10⁻⁶) (Fig. 1c).

We performed multiple linear regression models to assess the association between plasma proglucagon as independent variable and type 2 diabetes, BMI and liver fat as dependent variables. Model 1 was adjusted for age, sex, fasting time and plasma creatinine. All variables associated with higher plasma proglucagon (Fig. 1d). To investigate whether these metabolic diseases were independently associated with higher proglucagon, we adjusted the linear models for each of the other variables. Plasma proglucagon remained positively associated with BMI (p=0.0004), type 2 diabetes (p=1.4×10⁻⁶) and liver fat (p=0.0027) (Fig. 1d), suggesting that each of these metabolic disorders independently associate with elevations in plasma proglucagon. This was confirmed in an additional analysis with a tenfold increase in the sample size for type 2 diabetes and BMI specifically (Fig. 1e). Individuals with type 2 diabetes on metformin treatment had higher proglucagon levels, whereas insulin treatment lowered proglucagon to levels below the reference levels of individuals without diabetes (ESM Fig. 1).

To investigate if high plasma levels of proglucagon associate with an increased risk of developing type 2 diabetes, we first performed a Kaplan–Meier survival analysis stratified on tertiles of plasma proglucagon at baseline (Tertile 1: mean proglucagon (NPX): −1.50, n=13,385; Tertile 2: mean proglucagon: −0.005, n=13,386; Tertile 3: mean proglucagon: 1.62, n=13,385). The median follow-up time was 14.75 years, and the number of incident type 2 diabetes cases included in the models was 1551 (893 men and 658 women). The risk of incident type 2 diabetes increased stepwise from low to medium levels (p=1.5×10⁻⁵) and from medium to high levels (p=2.4×10⁻⁵) (Fig. 2a). Second, we applied a Cox proportional hazard regression analysis to evaluate the impact of baseline proglucagon levels of incident type 2 diabetes with adjustment for age and sex. Proglucagon was positively associated with the risk of type 2 diabetes development (HR 1.13; 95% CI 1.09, 1.17, p=1.3×10⁻¹²). We also stratified our model on BMI, and the association between proglucagon and incident type 2 diabetes was of similar magnitude across BMI categories (Fig. 2b).

We tested the impact of selected confounders on plasma levels of proglucagon using linear models. Proglucagon levels were higher in male than female participants and increased with age (Fig. 3a). Plasma creatinine served as a renal function estimate, as circulating products of proglucagon are cleared in the kidneys. Increased plasma creatinine, suggestive of reduced renal clearance, was associated with increased proglucagon (Fig. 3a).

Next, we used multiple linear regression to evaluate the association between plasma proglucagon and amino acids and glucose, adjusting for BMI and the confounders in Fig. 3a. Plasma proglucagon was positively associated with tyrosine, phenylalanine, alanine and histidine as well as the branched-chain amino acids (BCAA) valine, leucine and isoleucine. Interestingly, glutamine and glycine were not associated with proglucagon levels (Fig. 3b). Plasma glucose and HbA_1c were also positively associated with proglucagon levels (Fig. 3b).

To investigate the effect of genetic variation in the glucagon receptor on the same metabolic outcomes as above, we identified individuals with genetic variants of the glucagon receptor from whole-exome sequencing data. We grouped the glucagon receptor variants the following way: (1) the SNP G40S, previously associated with non-insulin-dependent diabetes, hypertension and adiposity [21‐23], but normal cAMP signalling and reduced β-arrestin signalling [5], (2) the missense variants V368M, R378C, R225H, R308W and D63N were grouped as ‘cAMP LoF’ based on previous research [5], and (3) variants annotated as frameshift or stop-codon gained were grouped as ‘Frameshift’ variants (ESM Table 2). Figure 4a–c outlines the number of individuals in each group in the UK Biobank cohort and the sub-cohort included in the proteomics analysis.

Although only 11 individuals heterozygous for a cAMP LoF variant were included, we observed an elevation in proglucagon compared with the wild-type reference group (β 0.85; 95% CI 0.04, 1.66; p=0.04) (Fig. 4d). G40S or frameshift variants were not associated with proglucagon levels, implying that cAMP rather than β-arrestin signalling may be involved in metabolic processes directly or indirectly regulating proglucagon levels.

In the frameshift variant group including 17 individuals, liver fat was increased compared with the wild-type reference group (β 0.49; 95% CI 0.03, 0.95; p=0.04) (Fig. 4e). None of the glucagon receptor variant groups were associated with BMI as a continuous or binary trait. We observed no difference in the prevalence of type 2 diabetes between individuals carrying G40S and the wild-type reference group (Fig. 4f). The sample size of type 2 diabetes was inadequate for cAMP LoF and Frameshift variant groups.

To test if the association between plasma proglucagon and type 2 diabetes and BMI, respectively, was dependent on the glucagon receptor genotype, we performed logistic and linear models stratified on the variant groups with inclusion of the interaction term proglucagon×genotype. Interestingly, although plasma proglucagon was associated with type 2 diabetes in the whole cohort (Fig. 1d,e), this was not the case in individuals with G40S (p=0.1) (Table 1). The interaction term between G40S and proglucagon was p=0.073. The effect of plasma proglucagon on BMI was larger in carriers of the G40S variants compared with carriers of the wild-variant (β 0.68; 95% CI 0.43, 0.93 vs β 0.33; 95% CI 0.29, 0.36; p=0.013) (Table 1). The effect of plasma proglucagon on BMI in carriers of cAMP LoF and frameshift variants were not different from individuals with wild-type receptors; however, the results are limited by the low sample size (11 and 24, respectively) (Table 1).

Plasma levels of the individual and the sum of amino acids did not show notable variations between individuals with the glucagon receptor variant groups and those with the wild-type variant (Table 2).

We next performed differential expression analysis on the proteomics dataset (~1500 proteins measured in each sample) to identify plasma proteins potentially regulated by glucagon receptor signalling. After correction for multiple testing, no proteins reached statistical significance. A list of the top ten up- and downregulated proteins for each variant group is provided in Table 3.