Multiomic integration analysis identifies atherogenic metabolites mediating between novel immune genes and cardiovascular risk

Abstract

Background

Understanding genetic-metabolite associations has translational implications for informing cardiovascular risk assessment. Interrogating functional genetic variants enhances our understanding of disease pathogenesis and the development and optimization of targeted interventions.

Methods

In this study, a total of 187 plasma metabolite levels were profiled in 4974 individuals of European ancestry of the GCAT| Genomes for Life cohort. Results of genetic analyses were meta-analysed with additional datasets, resulting in up to approximately 40,000 European individuals. Results of meta-analyses were integrated with reference gene expression panels from 58 tissues and cell types to identify predicted gene expression associated with metabolite levels. This approach was also performed for cardiovascular outcomes in three independent large European studies (N = 700,000) to identify predicted gene expression additionally associated with cardiovascular risk. Finally, genetically informed mediation analysis was performed to infer causal mediation in the relationship between gene expression, metabolite levels and cardiovascular risk.

Results

A total of 44 genetic loci were associated with 124 metabolites. Lead genetic variants included 11 non-synonymous variants. Predicted expression of 53 fine-mapped genes was associated with 108 metabolite levels; while predicted expression of 6 of these genes was also associated with cardiovascular outcomes, highlighting a new role for regulatory gene HCG27. Additionally, we found that atherogenic metabolite levels mediate the associations between gene expression and cardiovascular risk. Some of these genes showed stronger associations in immune tissues, providing further evidence of the role of immune cells in increasing cardiovascular risk.

Conclusions

These findings propose new gene targets that could be potential candidates for drug development aimed at lowering the risk of cardiovascular events through the modulation of blood atherogenic metabolite levels.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-024-01397-2.

Background

The global burden of dyslipidaemias in adult population has increased over the past 30 years [1]. This progressive increase is an alarming scenario because altered levels of blood lipids and fatty acids are a major risk factor for atherosclerotic cardiovascular disease (CVD) [2, 3]. CVDs are a leading cause of disease worldwide, with near a 30% of overall deaths, and 40% of premature deaths [4]. In this context, blood biomarkers would help to detect and manage the disease and prevent their development.

Blood metabolite levels are influenced by both germline genetic variations [5] and environmental factors, such as lifestyle habits, temperature, or long-term air pollution [6, 7]. To understand and estimate the contribution of single-nucleotide polymorphisms (SNPs), many metabolite genome-wide association studies (GWAS) have been carried out during the last decade [8–15]. These studies identified several genes that can be explored as pharmacological targets to modulate metabolite levels. However, most SNP-metabolite associations lay in intronic or intergenic regions and do not provide either a clear target gene or a functional contribution of the genetic region to the associated outcome. Therefore, GWAS findings are difficult to translate into clinical interventions. Over the last decade, these challenges have been addressed integrating GWAS results with functional genomics datasets through genetic instrumental analyses, such as colocalization analyses, transcriptome‐wide association studies (TWAS) and Mendelian randomization (MR) approaches [16]. A major advantage of instrumental analyses is that germline genetic associations cannot be explained by reverse causation and are less susceptible to confounding. Instrumental analyses can therefore circumvent many of the inherent limitations of traditional observational studies.

The transcriptome-wide association studies (TWAS) allow the identification of associations between predicted gene expression and metabolite levels. This instrumental approach interrogates SNPs tagging functional elements that alter expression levels to identify nearby genes causally associated with an outcome. TWAS capitalize and integrate expression quantitative trait loci (eQTL) studies and large GWAS datasets to provide associations that are not products of reverse causation [17]. In addition, TWAS results from individual tissues can be jointly analysed to leverage multi-tissue eQTLs and enlarge power to detect gene expression variation potentially altering metabolite levels [18]. Tissue transcriptome profiles reflect the average gene expression across heterogeneous cells within the analysed tissue. However, it has been observed that most cells follow a few broad transcriptional programs which can be classified into major cell types: epithelial, endothelial, mesenchymal (i.e. fibroblasts, stem cells, muscle cells), neural and immune cells [19]. Therefore, the analysis of tissue type-specific eQTLs can be very informative to identify genes altering blood metabolite levels expressed within tissues categorized by major cell type composition. Previous TWAS studies associated gene expression with lipid levels and different cardiovascular outcomes [20–24]; however, the interpretation of the results is not straightforward, as within a single locus more genes than the causal gene can be identified by pleiotropic effects (bystander genes) [25, 26]. Therefore, additional lines of evidence are needed to rule out pleiotropic effects.

In this study, we utilize genetic instrumental approaches to prioritize candidate genes across tissue categories. These genes are responsible for variations in plasma metabolites that contribute to an elevated risk of cardiovascular disease. To do this, as summarized in Fig. 1, we profiled a total of 187 unique plasma metabolite levels and 3 metabolite-ratios, including lipids particles, amino acids and fatty acids, in 4974 individuals of European ancestry of the GCAT| Genomes for Life cohort, and generated the corresponding GWAS results. Then, we generated a GWAS meta-analysis, including our study and publicly available datasets, and integrated these results with gene expression reference panels. The identified fine-mapped genes were also assessed for cardiovascular outcomes in three large European studies. Finally, for genes whose expression was associated with both metabolite levels and cardiovascular risk, we conducted MR mediation analyses. These analyses aimed to estimate the effect of gene expression on cardiovascular risk, considering both the direct effect and the effect mediated by metabolite levels. This multiomic integration approach enabled us to identify candidate genes that influence variations in blood metabolite levels and contribute to cardiovascular risk.

Fig. 1.

Graphical abstract of the performed analyses and main results. GWAS: Genome-wide association study. TWAS: Transcriptome-wide association study. MR: Mendelian randomization

Methods

A detailed workflow diagram of the methods is depicted in the Additional file 1.

Study participants

Participants included in this study belong to the GCAT|Genomes for life cohort, a population-based cohort from Southern Europe (Catalonia, NE Spain). The GCAT cohort comprises 20,000 volunteers, recruited between 2014 and 2018, from Catalan general population, aged between 40 and 65 years old at the time of recruitment, and 59.16% women. All the participants completed a detailed, self-reported, baseline questionnaire. All participants gave their consent, and all procedures were carried out in accordance with ethical standards [27]. About 5000 participants were randomly selected from the whole cohort based on overall demographic distribution (i.e. gender, age, residence) to perform genome-wide analyses [28]. This cohort subset, named GCATcore, included 55.6% women, with a mean age of 51.0 ± 6.9 years and an average body mass index (BMI) of 27.1 ± 4.7 (Additional file 2: Table S1).

Genome data

A total of 5459 GCAT participants were genotyped using the Infinium Expanded Multi-Ethnic Genotyping Array (MEGAEx) (ILLUMINA, San Diego, California, USA). A final dataset of 4974 GCAT participants of European ancestry that passed strict quality control were considered for genome-wide analyses, the GCATcore [28]. Genetic variants of these individuals were phased using SHAPEIT2 [29] and imputed using IMPUTE2 [30] and four reference panels: 1000 Genomes [31], UK10K [32], Genomes of the Netherlands [33] and Haplotype Reference Consortium [34]. Following the recommendations of the developers of GUIDANCE [35], the best imputation quality score from each panel was retained. After merging the imputation results of each reference panel, a total of ~ 20 M unique autosomal variants with minor allele frequency (MAF) > 0.001 and imputation quality score (R²) > 0.3 were retained for subsequent analysis.

Metabolome data

A total of 188 metabolite levels were assessed in plasma samples of 4974 GCATcore participants by the Centre for Omic Sciences (COS) Joint Unit of the Universitat Rovira i Virgili-Eurecat. Details on sample processing and metabolite profiling are included in Additional file 1. In brief, each metabolite was profiled using one of these three platforms: Gas Chromatography-mass spectrometry (GCms, four batches), Liquid Chromatography-mass spectrometry (LCms, five batches) and Nuclear Magnetic Resonance Chromatography-mass spectrometry (NMRms, two batches). Redundant metabolite measures were discarded when measured using different profiling platforms. This was the case for total cholesterol levels, discarding GCms measures and keeping NMRms measures, providing a final set of 187 metabolites (Additional file 2: Table S2). Metabolites were grouped according to chemical classification provided by the Human Metabolome DataBase (HMDB) 4.0 [36]. We classified the metabolites into 4 super-classes: lipids and lipid-like molecules, organic acids, organic oxygen compounds and organoheterocyclic compounds; and 11 classes: fatty acyls, glycerophospholipids (phosphatidylcholines, lysophosphatidylcholines and phosphatidylethanolamines), glycerolipids (triglycerides), lipoprotein lipids (characteristics of lipoprotein lipids), prenols (tocopherols), sphingolipids (sphingomyelins), steroids (cholesterols), carboxylic acids (amino acids), hydroxy acids, organooxygen compounds (carbohydrates) and indoles. Additionally, we computed 3 ratio parameters: the ratio of the concentration of low-density lipoprotein (LDL) particles to high-density lipoprotein (HDL) particles [LDL/HDL], the ratio of the concentration of total lipoprotein particles to HDL particles [Total/HDL] and the ratio of leucine to isoleucine [leucine/isoleucine] (Additional file 2: Table S2). Since plasma was not always obtained in fasted conditions, analyses were adjusted correcting by chylomicrons as a closely related measure of the post-prandial phase.

Genome-wide association analysis

A linear regression was fitted for each metabolite controlling by age, sex, chylomicron levels and number of batch and the first four principal components (PC) to account for the genetic population structure of the GCAT participants. Residuals were rank-based inverse normalized and considered as the outcome for GWAS. Allele dosage for each imputed genetic variant was considered in the GWAS analysis by using PLINK2 [37].

Meta-analysis

Meta-analysis was performed using METAL software [38] and combining GWAS summary statistics from the GCAT and six additional published studies of similar ancestry, age, BMI distribution and metabolomic platforms [8–13]. Sample size, sex ratios, ancestry and age and BMI distributions are included in the Additional file 2: Table S1. A full description of the enrolling process and metabolite profiling of the additional datasets is included in the Additional file 1. The combination of the studies included up to 26,533 independent individuals and each plasma metabolite was meta-analysed by studies comprising only independent individuals using similar metabolomics platforms (GCms, LCms or NMRms). In case a metabolite was analysed by two or more studies including non-independent individuals, we considered the study with the largest sample size (Additional file 2: Table S2). Since each study performed the GWAS analysis using different transformations, meta-analyses were based on Z-scores (p-value direction) weighted by sample size using a random effects model implemented in the METAL software [38]. Assuming a standardized trait with a mean 0 and a standard deviation 1, the SNP effect (b) and standard error (se) can be estimated as suggested by Zhu et al. [39];

$$b=\frac{z}{\sqrt{2p\left(1,-,p\right)\left(n,+,z^2\right)}}se=\frac{1}{\sqrt{2p\left(1,-,p\right)\left(n,+,z^2\right)}}$$

where p is the minor allele frequency (MAF) of the single-nucleotide polymorphism (SNP), n represents the sample size and Z the Metal Z-score. The threshold used to determine significant genetic variants was calculated by dividing the standard genome-wide significant threshold of P < 5 × 10⁻⁸ by the number of effective tests identified through a PC analysis for all analysed metabolites [13]. We found that 48 PCs explained 95% of the total variance and were considered as effective tests. Then, the significance threshold was P < 1.04 × 10⁻⁹.

Genetic heritability and correlation

Phenotypic variance explained by all genome-wide SNPs (heritability) was estimated through the genomic-relatedness-based restricted maximum-likelihood (GREML) approach implemented in the Genome-wide Complex Trait Analysis (GCTA) software [40]. Genetic correlations between traits were estimated using the bivariate GREML approach in GCTA [41]. Analyses were performed using the individual-level data of GCAT samples with SNPs restricted to those with MAF > 0.01. To obtain correlation results with enough statistical power, we included metabolites with nominally significant genetic heritability (P < 0.05). To adjust for multiple testing, we considered genetic correlations with a false discovery rate (FDR) < 0.05 as robust evidence of shared genetic contribution.

Mash analysis

To validate top SNP associations, multivariate adaptive shrinkage (mash) analysis of the meta-analysed summary statistics was performed [42]. In the presence of structured effects, mash analysis shares information across metabolites and shrinks estimates towards zero allowing the identification of associated loci. We performed this analysis using the R package mashr. As a condition to nominate significant genetic variants in each GWAS, we considered a genome-wide significance threshold of 1.04 × 10⁻⁹ in the meta-analysis and a standard local false sign rate (lfsr) < 0.05 in the mash approach.

SNP annotation

Top associated SNPs were functionally annotated and gene-mapped using the integrative web-based Functional Mapping and Annotation (FUMA) platform (https://fuma.ctglab.nl/) [43]. Overlapping top SNP regions among metabolites were considered as a single locus. Previous associations with complex phenotypes were as well assessed using FUMA. Regional plots for top associated SNPs were generated, which enables inspection of the strength of association, the extent of association signal and linkage disequilibrium (LD), and the position of findings relative to genes in the region. Plots were generated using the software LocusZoom v1.3 [44]. Measures of LD were estimated using European populations of the 1000 Genomes Project.

Transcriptome-wide association analysis for metabolites

We integrated results of GWAS meta-analyses with reference panels for gene expression under the summary-based PredXcan approach [45] to identify predicted gene expression associated with metabolite blood level. Elastic net prediction models for genetically regulated gene expression in multiple tissues from GTEx v8 were obtained from PredictDB (http://predictdb.org/) [46, 47]. In addition, we retrieved the Correlated Expression and Disease Association Research (CEDAR) dataset that included transcriptome data of six relevant blood cell types (CD4, CD8, CVD19, CD14, CD15 and platelets) [48]. CEDAR data quality control and elastic net prediction models were performed by our group and was described in a previous work [49]. For each metabolite, TWAS results were meta-analysed among tissues, using the summary-based MultiXcan approach [18], for all tissues and for different tissue groups regarding the abundance of major cell type [19]: epithelial, mesenchymal, immune or neural (Additional file 2: Table S3). In order to correct the results for multiple testing, we defined as robust evidence of candidate causal gene those results with a P-value lower than the significance threshold corrected by Bonferroni using the number of effective tests and the total number of tested genes (P < 1.21 × 10⁻⁸).

Fine-mapping in extended LD regions for metabolites

To tease out the causal gene in genetic regions with more than one associated gene due to extended LD, probabilistic fine-mapping was performed using the fine-mapping of causal gene sets (FOCUS) software [50]. FOCUS models the correlation structure across predictive models and computes posterior probabilities (PIP) for a gene to explain observed association signals at a specific locus. As a condition to nominate significant genes whose gene expression is probably causally associated with metabolite levels, we considered a significance threshold of 1.21 × 10⁻⁸ in the TWAS approach, and a credible set with a nominal confidence > 90% and with a PIP > 0.5 in the FOCUS approach.

Finally, fine-mapped genes were linked to diseases using the DisGeNET platform that integrates data from curated repositories, GWAS catalogues and scientific literature [51]. Diseases were classified into 23 categories according to the Medical Subject Headings 2022 (MeSH) database (https://meshb.nlm.nih.gov/treeView), and we tested whether these categories were overrepresented in our gene set using a fisher test. Significance threshold was corrected by Bonferroni using the number of tested categories (P < 2.2 × 10⁻³).

Transcriptome-wide association analysis of fine-mapped genes for cardiovascular outcomes

For genes which predicted gene expression was found associated with metabolite levels, we tested whether predicted gene expression was additionally associated with cardiovascular risk. To do so, we used cardiovascular-related GWAS results from large studies of European individuals (CARDIoGRAMplusC4D [52], FinnGen [53] and UK Biobank [54, 55]) (Additional file 2: Table S4). A full description of the enrolling process of the cardiovascular datasets is included in the Additional file 1. The analysed cardiovascular outcomes included coronary artery disease (CAD, 3 datasets), myocardial infarction (MI, 3 datasets), angina pectoris (ANG, 2 datasets), atherosclerosis (ATH, 2 datasets) and atrial fibrillation (AF, 2 datasets). First, we integrated GWAS summary statistics with reference panels for gene expression under the summary-based PredXcan approach [45], and TWAS results were meta-analysed among tissues, using the summary-based MultiXcan approach [18], for all tissues and for different tissue groups regarding the abundance of major cell type [19]: epithelial, mesenchymal, immune or neural (Additional file 2: Table S3). In order to correct the results for multiple testing, we defined as robust evidence of candidate causal gene those results with a P-value lower than the significance threshold corrected by Bonferroni using the number of tested cardiovascular datasets (12 datasets) and the total number of tested genes (53 genes) (P < 7.9 × 10⁻⁵).

Mendelian randomization approach

Genetically informed mediation inference analysis was performed to make causal mediation inference about the relationship between gene expression, metabolite levels and cardiovascular risk. To do so, we apply a Mendelian randomization (MR) approach in several steps. MR is a method that uses genetic variants as instrumental variables to test for and estimate causal effects between risk factors and outcomes. The underlying principle is that genetic variants are randomly assigned at conception, akin to a natural randomized control trial, and germline genetic associations cannot be explained by reverse causation and are less susceptible to confounding [56]. The most used MR approach is the “two-sample” MR, where the genetic instrument of the putative risk factor is identified in one genetic study (the largest for the risk factor), but is subsequently evaluated for association to the outcome in a second genetic study (the largest for the outcome) [57].

As genetic instruments for gene expression, it was used the top associated expression quantitative trait loci (eQTL), not showing pleiotropic effects for nearby genes also identified in this study, from the most relevant tissue in the GTEx Consortium [46]. In the case of CELSR2 and PSRC1 genes, the selected eQTLs were not either associated to expression of SORT1 or in LD (R² = 0.08; European samples) (https://ldlink.nih.gov/). However, the eQTL for PSRC1 showed residual association with CELSR2 expression in the liver. The genetic instruments for metabolite levels were selected from the top independent SNPs identified in this study. The strength of associations between a genetic instrument and an exposure is reflected in the F-statistic, which is inversely related to weak instrument bias, being 10 the minimum estimation for a F-statistic to avoid bias of this nature [58]. The F-statistic was estimated as $F=(\frac{n-k-1}{k})(\frac{R^2}{1-R^2})$, where R² is the proportion of phenotypic variance explained by the genetic instrument, n is the sample size, and k the number of genetic variants [58]. Explained phenotypic variance for a single SNP was estimated as $R^2={2b}^2(\text{p})(1-\text{p})$, a function of effect size for the risk factor in standard deviation units (b) and minor allele frequency (p) [59]. The parameters of association of genetic instruments with CAD risk will be obtained from the cardiovascular-related GWAS results from large studies of European individuals (CARDIoGRAMplusC4D [52], FinnGen [53] and UK Biobank [54, 55]) (Additional file 2: Table S4).

Each genetic instrument provides an estimation of the exposure levels effect on outcome risk (Wald ratio: genetic effect on the exposure/ genetic effect on the outcome). In the case of eQTLs on metabolites and CAD risk, this was the main MR estimate. In the case of metabolite levels on CAD risk, the main MR estimate was the combination of SNP Wald ratios in a single causal estimation through an inverse-variance weighted MR estimator [57]. Because of the presence of pleiotropic variants can lead to biased causal effect estimates, several MR sensitivity analyses for data with potentially invalid instruments were applied. Initially, to evaluate the extent to which directional pleiotropy (non-balanced horizontal pleiotropy) may affect the effect estimate, we used the intercept test within an MR-Egger weighted linear regression approach [60]. Furthermore, the weighted median method relaying on the distribution on SNP effects were applied, which is less sensitive to SNPs with biased effect [61]. These MR estimates were obtained using the “TwoSampleMR” R package (R software).

Finally, we investigated complex networks of relationships between variables, in particular where some of the effect of gene expression on CAD risk may operate through an intermediate variable (metabolite levels). Under an instrumental approach, mediated and non-mediated effects of the exposure on the outcome can be estimated using a regression-based method. If all effects are linear without interaction terms, the non-mediated effect can be obtained, under the assumption of homogeneity of causal effects across individuals in the population, as the difference between the total effect of gene expression on CAD risk and the product of the effects of gene expression on metabolite levels and metabolite levels on CAD risk. The standard error and confidence intervals for these quantities can be estimated by bootstrapping [62].

Results

Metabolomic profile

We profiled metabolite blood levels of 4974 individuals of the GCATcore. The quality control procedure provided a final dataset of 187 metabolites and 3 ratios for subsequent analyses. We classified the metabolites into 11 classes according to the Human Metabolite Database (HMDB) 4.0 [36]; 10 fatty acyls, 74 glycerophospholipids (phosphatidylcholines, lysophosphatidylcholines and phosphatidylethanolamines), 26 glycerolipids (triglycerides), 23 lipoprotein lipids (characteristics of lipoprotein lipids), one prenol (tocopherol), 25 sphingolipids (sphingomyelins), 3 steroids (esterified, free and total cholesterol), 16 carboxylic acids (amino acids), 3 hydroxy acids, 5 organooxygen compounds (carbohydrates) and one indol (Additional file 2: Table S2). The distribution of metabolite levels is depicted in Additional file 1.

Genome-wide association analysis

The GCAT GWAS summary statistics were meta-analysed with six additional published studies (Additional file 2: Table S1), generating summary statistics comprising up to 26,533 independent individuals (Additional file 2: Table S2). Manhattan plots and quantile–quantile (QQ) plots of meta-analyses results for each metabolite are included in Additional file 1. As result, a total of 44 genetic loci were associated with 124 metabolite parameters, comprising 350 independent locus-metabolite associations. LocusZoom plots for the 350 locus-metabolite associations were included in Additional file 1. We identified 165 lead SNPs tagging 76 genes (Fig. 2A, B, Additional file 2: Table S5). We did not identify any significant genetic association for the other 66 analysed metabolites. The 44 genetic loci had been previously related to metabolite levels, and half of them were specific to metabolite classes; 8 of them were only associated with carboxylic acids, 6 with glycerophospholipids, 5 with lipoprotein lipids, 2 specific loci for sphingolipids and 1 for indoles. The other 22 loci were shared among metabolite classes (Fig. 2A, C). Two metabolite classes, prenols and organooxygen compounds, showed no genome-wide significant SNPs.

Fig. 2.

Genome-wide associated variants for 190 metabolites. A Summary of the GWAS results. B 3D Manhattan plot of significant SNPs according to thresholds in meta-analyses and mashr approaches. C Significant loci per metabolite class

The functional annotation of the 165 identified lead SNPs revealed that there were 10 non-synonymous SNPs and 1 non-frameshift deletion variant (Table 1 and Additional file 2: Table S5). Among them, we identified common variants such as a non-synonymous SNPs and a non-frameshifting deletion in APOB gene (rs1367117-Thr98Ile in exon 4 and rs878853971-p.Leu12_Ala16delinsProAlaLeu in exon 1) related to cholesterol and lipoprotein lipid levels; one SNP in GCKR gene (rs1260326-Leu446Pro in exon 14) associated with fatty acyls, lipoprotein lipids and glycerolipids; and the two SNPs in APOE gene responsible of the APOE protein isoforms (rs429358- p.Cys130Arg and rs7412-Arg176Cys in exon 4) related to fatty acyls, cholesterol and lipoprotein lipids. We also identified less frequent variants (MAF < 0.1 in the GCAT samples) such as the rs11591147 in exon 1 of PCSK9 gene (p.Arg46Leu), the rs6756629 in exon 2 of ABCG5 (p.Arg50Cys), the rs268 (exon 6 of LPL; p.Asn318Ser), the rs2228603 (exon 3 of NCAN; p.Pro92Ser), the rs58542926 (exon 6 of TM6SF2; p.Glu167Lys) and the rs1800961 (exon 4 of HNF4A; p.Thr139Ile) associated with fatty acyls, cholesterol, lipoprotein lipids, phosphatidylcholine and sphingomyelin levels (Table 1).

Table 1.

Non-synonymous and non-frameshifting deletion variants associated with metabolite levels in GWAS meta-analyses

Lead SNP						Metabolite		Meta-analysis				Mashr	Gene
rs number	Chr	Bp	GCAT Freq	Eff Al	Ref Al	Class	Name	NTotal	Beta	SE	P-value	lfdr	Encode number	Symbol	Exon	Codon mutation
rs11591147	1	55,505,647	0.02	T	G	Fatty acyls	Linoleic acid	18,499	− 0.298	0.039	4.E − 14	5.E − 47	ENSG00000169174	PCSK9	exon1	p.Arg46Leu
						Lipoprotein lipids	Large LDL	20,959	− 0.553	0.037	8.E − 51	7.E − 48
							Medium LDL	20,959	− 0.540	0.037	2.E − 48	4.E − 48
							Small LDL	20,959	− 0.484	0.037	3.E − 39	9.E − 48
							Cholesterol in LDL	20,959	− 0.567	0.037	2.E − 53	4.E − 48
						Sphingolipids	Sphingomyelin	18,447	− 0.328	0.039	9.E − 17	1.E − 43
						Steroids	Esterified cholesterol	18,466	− 0.452	0.039	2.E − 30	6.E − 47
							Free cholesterol	18,466	− 0.466	0.039	2.E − 32	1.E − 47
							Total cholesterol	20,959	− 0.476	0.037	6.E − 38	5.E − 48
rs1367117	2	21,263,900	0.29	A	G	Lipoprotein lipids	Large LDL	24,245	0.099	0.010	4.E − 23	3.E − 28	ENSG00000084674	APOB	exon4	p.Thr98Ile
							Cholesterol in LDL	26,531	0.105	0.010	3.E − 28	6.E − 29
rs878853971	2	21,266,774	0.29	G	GGCAGCGCCA	Lipoprotein lipids	Medium LDL	24,243	0.097	0.010	2.E − 22	7.E − 28	ENSG00000084674	APOB	exon1	p.Leu12_Ala16delinsProAlaLeu
							Small LDL	24,243	0.088	0.010	2.E − 18	8.E − 28
rs1260326	2	27,730,940	0.47	T	C	Fatty acyls	MUFA	18,509	0.078	0.010	5.E − 14	1.E − 07	ENSG00000084734	GCKR	exon15	p.Leu446Pro
							Omega 3	18,518	0.066	0.010	2.E − 10	5.E − 07
						Glycerolipids	Triglycerides	26,519	0.078	0.009	2.E − 19	2.E − 13
						Lipoprotein lipids	Small HDL	24,247	0.069	0.009	2.E − 14	3.E − 04
rs6756629	2	44,065,090	0.06	A	G	Lipoprotein lipids	Large LDL	24,247	− 0.120	0.019	3.E − 10	8.E − 13	ENSG00000138075	ABCG5	exon2	p.Arg50Cys
							Cholesterol in LDL	26,533	− 0.132	0.018	4.E − 13	7.E − 13
rs268	8	19,813,529	0.01	A	G	Lipoprotein lipids	Large HDL	24,243	0.264	0.039	9.E − 12	2.E − 21	ENSG00000175445	LPL	exon6	p.Asn318Ser
rs2228603	19	19,329,924	0.05	T	C	Glycerophospholipids	Phosphatidylcholine diacyl C34:4	12,450	− 0.183	0.029	5.E − 10	1.E − 12	ENSG00000130287	NCAN	exon3	p.Pro92Ser
rs58542926		19,379,549	0.06	T	C	Glycerolipids	Triglycerides	26,514	− 0.165	0.019	1.E − 18	2.E − 32	ENSG00000213996	TM6SF2	exon6	p.Glu167Lys
						Lipoprotein lipids	Triglycerides in IDL	24,242	− 0.146	0.020	8.E − 14	1.E − 33
						Steroids	Total cholesterol	26,460	− 0.127	0.019	1.E − 11	4.E − 29
rs429358	19	45,411,941	0.10	T	C	Glycerolipids	Triglycerides	26,513	− 0.128	0.014	7.E − 19	3.E − 40	ENSG00000130203	APOE	exon4	p.Cys130Arg
						Lipoprotein lipids	Small VLDL	24,241	− 0.113	0.015	5.E − 14	2.E − 35
rs7412	19	45,412,079	0.06	T	C	Fatty acyls	Linoleic acid	18,497	− 0.191	0.022	4.E − 18	1.E − 32	ENSG00000130203	APOE	exon4	p.Arg176Cys
						Lipoprotein lipids	Triglycerides in HDL	4974	0.284	0.042	2.E − 11	8.E − 28
							Cholesterol in IDL	20,955	− 0.350	0.021	2.E − 64	3.E − 72
							LDL	4974	− 0.420	0.042	4.E − 23	1.E − 71
							Large LDL	20,955	− 0.439	0.021	8.E − 101	5.E − 114
							Medium LDL	20,955	− 0.454	0.021	5.E − 108	3.E − 122
							Small LDL	20,955	− 0.394	0.021	1.E − 81	2.E − 107
							Cholesterol in LDL	23,241	− 0.495	0.019	2.E − 142	6.E − 148
							Non HDL	4974	− 0.398	0.042	7.E − 21	6.E − 65
							LDL/HDL	4974	− 0.368	0.042	5.E − 18	1.E − 74
							Total/HDL	4974	− 0.340	0.042	1.E − 15	1.E − 64
						Steroids	Esterified cholesterol	18,466	− 0.323	0.022	8.E − 49	1.E − 61
							Free cholesterol	18,466	− 0.226	0.022	2.E − 24	1.E − 37
							Total cholesterol	23,173	− 0.304	0.020	1.E − 53	7.E − 66
rs1800961	20	43,042,364	0.03	T	C	Lipoprotein lipids	Cholesterol in HDL	26,529	− 0.166	0.026	3.E − 10	3.E − 07	ENSG00000101076	HNF4A	exon4	p.Thr139Ile

Genetic heritability and correlations

SNP-based heritability estimates observed in this study (Additional file 2: Table S2) were similar than estimates observed in previous studies [63]. A total of 47 out of 190 metabolites levels and ratios showed a significant SNP-based heritability (P < 0.05) [h² = 0.59 − 0.12] (Additional file 2: Table S2). The metabolites showing significant genetic heritability belonged to the metabolite classes of glycerophospholipids (15 phosphatidylcholines, 2 lysophosphatidylcholines and 2 phosphatidylethanolamines), and glycerolipids (16) and lipoprotein particles mainly related with very low-density lipoproteins (VLDL) (5) (Additional file 2: Table S2). Genetic correlations were performed among the 47 metabolites with significant heritability, and a total of 40 genetic correlations were observed (FDR < 0.05), involving 15 glycerolipids, 9 glycerophospholipids and the 5 lipoprotein particles. The 85% of correlations were among metabolites of the same class (Additional file 2: Table S6).

Transcriptome-wide association analyses

We integrated GWAS meta-analyses results with reference panels for gene expression from 58 tissues and cell types. A total of 2537 gene expression-metabolite associations were identified in the TWAS meta-analyses results including all tissues (Fig. 3A), while additional 1807 gene expression-metabolite associations were identified in the meta-analyses of tissue categories (data not shown). The fine-mapping FOCUS approach validated 53 genes as associated with 108 metabolites, comprising 196 gene expression-metabolite associations (Table 2, Fig. 3B). Manhattan plots of meta-analyses results including all tissues, and highlighting fine-mapped genes, for each metabolite were included in Additional file 1.

Fig. 3.

Transcription-wide associated genes for 190 metabolites. A 3D Manhattan plot of the 2537 gene-metabolite associations in the overall multi-tissue meta-analysis. B Summary of the fine-mapped 53 genes for metabolites. C Upset plot of the 53 fine-mapped gene-metabolite associations per metabolite class. D Distribution of the 53 fine-mapped genes among metabolite classes and specific cell type categories

Table 2.

Fine-mapped associations of predicted gene expression for metabolites and cardiovascular outcomes

Genes				Metabolites		Cardiovascular phenotypes
Name	Chr	Bp start	Bp end	Class	Name
TRIT1	1	40,306,723	40,349,183	Glycerophospholipids	LPC16:0n
PCSK9	1	55,505,221	55,530,525	Lipoprotein lipids	Cholesterol in IDLai; Triglycerides in IDLai; Large LDLai; Medium LDLai; Small LDLai; Cholesterol in LDLai	CGC4D CADai; UKBB CADai; FinnGenn CADai; CGC4D MIi; FinnGenn MIai; FinnGenn ANGai; UKBB ATHai; FinnGenn ATHai
				Steroids	Esterified Cholesterola; Free Cholesterolai; Total Cholesterolai
DOCK7	1	62,920,399	63,153,969	Glycerolipids	Triglyceridesaeimn
				Lipoprotein lipids	Large VLDLeim
CELSR2	1	109,792,641	109,818,372	Lipoprotein lipids	Cholesterol in IDLaei; Large LDLaei; Medium LDLaei; Small LDLaei; Cholesterol in LDLaei	CGC4D CADaeim; UKBB CADaeimn; FinnGenn CADaei; CGC4D MIaei; UKBB MIaei; UKBB ANGi; FinnGenn ANGaei; UKBB ATHaeimn; FinnGenn ATHaei
				Steroids	Esterified Cholesterolei; Total Cholesterolaei
PSRC1	1	109,822,178	109,825,808	Lipoprotein lipids	LDLim	CGC4D CADaeimn; UKBB CADaeimn; FinnGenn CADaeimn; CGC4D MIaeimn; UKBB MIaeimn; UKBB ANGeim; FinnGenn ANGaeimn; UKBB ATHaeimn; FinnGenn ATHaeimn
ZNF697	1	120,162,045	120,190,396	Carboxylic acids	Serinen
SEC22B	1	120,698,575	120,719,144	Carboxylic acids	Serineae
APOB	2	21,224,301	21,266,945	Glycerolipids	Triglyceridesan	CGC4D CADam; UKBB CADam
				Lipoprotein lipids	Cholesterol in IDLaemn; Triglycerides in IDLaemn; Large LDLaemn; Medium LDLaemn; Small LDLamn; Cholesterol in LDLaemn; Small VLDLamn
				Steroids	Esterified Cholesterolaemn; Free Cholesterolaemn; Total Cholesterolaemn
PPM1G	2	27,604,061	27,632,554	Fatty acyls	MUFAe
				Lipoprotein lipids	Mean diameter for VLDLe
NRBP1	2	27,650,657	27,665,126	Glycerophospholipids	LPC16:1e
C2orf16	2	27,799,389	27,805,588	Glycerolipids	Triglyceridesae
				Lipoprotein lipids	Large VLDLae; Medium VLDLe; Small VLDLae
CPS1	2	211,342,406	211,543,831	Carboxylic acids	Glycineaeimn
ELOVL2	6	10,980,992	11,044,547	Glycerophospholipids	PC38:5aei
MOG	6	29,624,758	29,640,149	Carboxylic acids	Tyrosinea
RPP21	6	30,312,908	30,314,661	Glycerolipids	TG54:6ae
PSORS1C2	6	31,105,313	31,107,127	Glycerolipids	TG50:0aeim
HCG27	6	31,165,537	31,171,745	Glycerolipids	TG52:4aem	CGC4D CADm; UKBB CADai; FinnGenn CADi; CGC4D MIm; FinnGenn ATHi
HLA-B	6	31,321,649	31,324,965	Glycerolipids	TG48:0aimn	UKBB CADan
MICA	6	31,371,356	31,383,092	Glycerophospholipids	PCe42:5i
BAG6	6	31,606,805	31,620,482	Glycerolipids	TG51:2aemn
MSH5	6	31,707,725	31,732,622	Glycerolipids	TG46:0ae; TG46:1aen; TG46:2ae	UKBB CADi
SKIV2L	6	31,926,857	31,937,532	Glycerolipids	TG48:2a
HLA-DRB1	6	32,546,546	32,557,625	Glycerophospholipids	PCe38:5aemn	UKBB CADm; FinnGenn ATHn
MFSD4B	6	111,580,551	111,592,370	Carboxylic acids	Tyrosineaen
MLXIPL	7	73,007,524	73,038,873	Glycerolipids	Triglyceridesaeimn
				Lipoprotein lipids	Large VLDLaeimn; Medium VLDLaeimn; Small VLDLaeimn; Mean diameter for VLDLaeimn
LPL	8	19,759,228	19,824,769	Glycerolipids	Triglyceridesai; TG52:2a; TG52:3a	CGC4D CADaim; UKBB CADaim; CGC4D MIi; UKBB MIam; UKBB ATHai
				Lipoprotein lipids	Large HDLaim; Mean diameter for HDLaim; Cholesterol in HDLaim; Large VLDLai; Medium VLDLai; Small VLDLaim; Mean diameter for VLDLai; Cholesterol in VLDLai
SCD	10	102,106,881	102,124,591	Fatty acyls	Stearic acidaim	UKBB CADi
				Glycerophospholipids	LPC16:1aim
TMEM258	11	61,535,973	61,560,274	Glycerophospholipids	LPC20:5aem
FADS2	11	61,560,452	61,634,826	Glycerophospholipids	PC38:3aeimn; LPC20:3aeimn
FADS1	11	61,567,099	61,596,790	Fatty acyls	Omega 3aeimn; PUFAaeimn; Docosahexaenoic acidaeimn; Linoleic acidaeimn
				Glycerophospholipids	PC32:0aeimn; PC32:2aeimn; PC34:2aeimn; PCe34:2aeimn; PC34:3aeimn; PC34:4aeimn; PC36:2aeimn; PCe36:2aeimn; PC36:3aeimn; PCe36:3aeimn; PC36:4aeimn; PCe36:4aeimn; PC36:5aeimn; PCe36:5aeimn; PC37:4aeimn; PC38:2aeimn; PCe38:3aeimn; PC38:4aeimn; PCe38:4aeimn; PC38:5aeimn; PCe38:5aeimn; PC38:6aeimn; PCe38:6aeimn; PC40:4aeimn; PCe40:4aeimn; PC40:5aeimn; PCe40:5aeimn; PC40:6aeimn; PCe42:4aeimn; PCe42:5aeimn; LPC18:2aeimn; LPC20:4aeimn; PEe38:5aeimn;
				Glycerolipids	TG52:6aeimn
				Lipoprotein lipids	Mean diameter for HDLaeimn; Cholesterol in HDLaeimn; Mean diameter for VLDLm
BEST1	11	61,717,293	61,732,987	Glycerolipids	TG54:4 m
ZPR1	11	116,648,436	116,658,766	Glycerolipids	Triglyceridesaemn	FinnGenn CADaem; FinnGenn ATHaem
				Lipoprotein lipids	Large VLDLaemn; Medium VLDLaemn; Small VLDLaemn; Mean diameter for VLDLaem
SYNE2	14	64,319,683	64,693,165	Sphingolipids	SM32:1am; SM33:1a
TMEM229B	14	67,913,801	68,000,456	Glycerophospholipids	PCe32:1aeimn; PCe36:5aeimn; PCe38:6aeim
MYZAP	15	57,884,139	57,977,562	Glycerophospholipids	PE36:4ae
ALDH1A2	15	58,245,622	58,790,065	Lipoprotein lipids	Large HDLai; Triglycerides in IDLai; Mean diameter for VLDLi
LIPC	15	58,702,768	58,861,151	Fatty acyls	MUFAa; PUFAa
				Lipoprotein lipids	Small HDLa; Mean diameter for HDLaeimn; Triglycerides in HDLaeimn; Mean diameter for LDLaeimn; Triglycerides in LDLam
FAM81A	15	59,664,892	59,815,748	Glycerophospholipids	Phosphatidyl Cholinean
				Steroids	Free Cholesteroln
NTAN1	16	15,131,710	15,149,921	Glycerophospholipids	PC36:3 m; PC38:3aeimn; PCe38:3 m; LPC20:3aeimn; LPC20:4 m
CETP	16	56,995,762	57,017,757	Glycerophospholipids	Phosphatidyl Cholineaeimn; PCe32:1aeimn; PCe34:1aeim; PCe34:2aem; PCe34:3aeim	UKBB CADaemn
				Lipoprotein lipids	HDLaeimn; Large HDLaeimn; Medium HDLaeimn; Small HDLaei; Mean diameter for HDLaeimn; Cholesterol in HDLaeimn; Triglycerides in IDLm; Small VLDLaeimn; LDL/HDLaeimn; Total/HDLaeimn
SLC7A6	16	68,298,433	68,335,722	Carboxylic acids	Lysinee
MARVELD3	16	71,660,064	71,676,017	Carboxylic acids	Tyrosinemn
HPR	16	72,088,522	72,111,145	Lipoprotein lipids	Triglycerides in IDLm	UKBB CADam
GCSH	16	81,115,566	81,130,008	Carboxylic acids	Glycinem
SNRPD1	18	19,192,228	19,210,417	Glycerophospholipids	PC32:2an
CERS4	19	8,271,620	8,327,305	Sphingolipids	SM36:1aeimn; SM36:2aeimn; SM38:1aen; SM38:2aeimn
KRI1	19	10,663,761	10,676,713	Lipoprotein lipids	Large LDLe
				Steroids	Total Cholesterole
ATP13A1	19	19,756,007	19,774,502	Glycerolipids	Triglyceridesaeim
				Lipoprotein lipids	Small VLDLe
PVR	19	45,147,098	45,166,850	Lipoprotein lipids	Cholesterol in IDLaen; Large LDLan; Medium LDLaen; Small LDLaen; Cholesterol in LDLaen	UKBB CADan; FinnGenn CADa; UKBB ATHan
				Steroids	Esterified Cholesterolan; Total Cholesterolan
APOC2	19	45,445,495	45,452,822	Sphingolipids	SM42:1e	UKBB CADai
SPTLC3	20	12,989,627	13,147,411	Sphingolipids	SM32:0ai; SM32:1a; SM33:1ai; SM35:1ai; SM36:0ai; SM39:1ai; SM40:0ai; SM43:1ai; SM43:2ai
PLTP	20	44,527,399	44,540,794	Lipoprotein lipids	Large HDLaeim; Medium HDLem; Small HDLaeimn; Mean diameter for HDLaeimn
DGCR9	22	19,061,632	19,061,632	Carboxylic acids	Prolineaen

Genes were significant in the meta-analyses including ^aall, ^eepithelial, ^mmesenchymal, ⁱimmune and/or ⁿneural tissues. Metabolite description in Table S1

PC Phosphatidylcholine diacyl C, PCe Phosphatidylcholine acyl-alkyl C, LPC Lysophosphatidylcholine acyl C, PE Phosphatidylethanolamine acyl C, Pee Phosphatidylethanolamine alkyl C, TG Triglycerol C, SM Sphingomyeline C, CGC4D CardiogramplusC4D, UKBB UK Biobank, CAD Coronary artery disease, MI Myocardial infraction, ANG Angina pectoris, ATH Atherosclerosis

The 95% of fine-mapped gene expression-metabolite associations (187) identified genes potentially explaining previously observed locus-metabolite associations, from which the 74% (139) comprised the closest gene of the associated SNP (Fig. 3B, Additional file 2: Table S6). The other 5% of gene-metabolite associations (9) identified associations not observed in GWAS results (Fig. 3B), most of them within the 44 genetic loci and increasing the number of associated metabolites, except for SEC22B gene (1p12) related to serine levels and GCSH gene (16q23.2) to glycine levels (Additional file 2: Table S7). Among the 53 fine-mapped genes in Table 2, it is worth highlighting some results. The fine-mapped TWAS results suggest that variations in gene expression levels might be responsible for some of the associations observed in GWAS. For instance, PCSK9, CELSR2, APOB, LPL, FADS1, LIPC, CETP and PLTP on lipoprotein lipids; PCSK9, CELSR2 and APOB on steroids; APOB, MSH5 and LPL on glycerolipids; FADS1 and LIPC on fatty acyls; FADS1 and CETP on glycerophospholipids; and CERS4 and SPTLC3 on sphingomyelins (Table 2). Beside these observations, we identified PSRC1 gene (1p13.3) associated with total LDL concentration, despite that GWAS results pointed to the SNP rs12740374, in the UTR3 region of CELSR2 gene (Additional file 2: Table S5). We observed another similar case in the 2p23.3 region. The SNP rs6547692, located in an intron of GCKR gene, was found associated with concentration of small, medium and large VLDL particles (Additional file 2: Table S5). However, the gene whose predicted expression was found associated to these lipid traits was the C2orf16 gene (Table 2). Other results to highlight are found in 7q11.23 region. This genetic region was found associated to small VLDL concentration in the GWAS results; but TWAS results identified the gene MLXIPL, within the region, associated with concentration of small but also with medium, large and total VLDL, and triglycerides (Table 2). Also, in 10q24.31 region, the expression of SCD gene appeared related to levels of stearic acid and lysophosphatidylcholine acyl C16:1 (LPC16:1), while GWAS results pointed to PKD2L1 gene. Finally, in the APOE loci, GWAS results only identified the two non-synonymous APOE SNPs; however, TWAS results also pointed to PVR and APOC2 genes as potentially associated with lipoprotein lipids, steroids and sphingomyelin levels.

Most of the 53 fine-mapped genes were specific to a metabolite class (Fig. 3C), being the carboxylic acids and sphingolipids the metabolite classes with all the genes specific to the class (9 and 4 genes, respectively). These were followed by glycerophospholipids (11 genes out of 15 total genes associated in this class: 11/15), glycerolipids (8/16), and lipoprotein lipids (4/18)). Steroids and fatty acyls were the classes without specific genes, 6 and 4 genes, respectively, that were mostly shared with lipoprotein lipids (Fig. 3C). On the other hand, 15 of the 53 fine-mapped genes were specific to a single tissue type category (Fig. 3D; pale colour portion of bars). For instance, the genetically predicted gene expression of SLC7A6 was associated with lysine levels only in epithelial tissues meta-analysis, while expression of GCSH was related to glycine levels only in the mesenchymal tissues meta-analysis (Table 2). Finally, it is also worth to mention that some genes showed stronger effects in meta-analysis of specific tissue types compared with meta-analysis including all or other tissue types. This was the case for PCSK9 and SPTLC3 genes which associations with metabolite levels were stronger in immune tissue meta-analysis, and for CELSR2 and C2orf16 genes in epithelial tissues meta-analysis (Additional file 2: Table S7).

Cardiovascular diseases

Gene set enrichment of the 53 fine-mapped genes was tested for association with disease categories using the DisGeNET platform [51]. Results revealed that several fine-mapped genes were associated with the following DisGeNET terms: coronary heart disease (21 genes, FDR = 3 × 10⁻⁹), coronary artery disease (20 genes, FDR = 9 × 10⁻⁸) and cardiovascular diseases (19 genes, FDR = 6 × 10⁻⁷), while enrichment was also observed for diabetes (20 genes, FDR = 9 × 10⁻⁶), and diabetes mellitus (20 genes, FDR = 1 × 10⁻⁴) (Fig. 3B; Additional file 2: Table S8).

To identify potential causal genes whose expression contributes to cardiovascular risk, we tested for association the predicted expression of the 53 fine-mapped genes with a set of cardiovascular outcomes. The tested outcomes were coronary artery disease (CAD), myocardial infraction (MI), angina pectoris (ANG), atherosclerosis (ATH) and atrial fibrillation (AF) from three large studies of European population (CARDIoGRAMplusC4D [52], FinnGen [53] and UK Biobank [54, 55]) (Additional file 2: Table S4). As result, the gene expression of 15 genes were identified as associated with cardiovascular outcomes except for AF, comprising 51 gene expression-cardiovascular outcome associations (Table 2 and Additional file 2: Table S9). Most of these associations were identified in the overall multi-tissue meta-analyses results, but additional associations were identified in the meta-analyses results including only tissues within immune (7 out of 36 total associations in this category: 7/36), mesenchymal (3/24), neural (1/15) and epithelial (0/20) tissue categories (Fig. 4A). The genes associated with an outcome in at least two datasets, thus, replicating results were PCSK9, CELSR2, PSRC1, APOB, HCG27 and LPL (Fig. 4A and Table 2). Interestingly, the PCSK9 gene showed stronger effects in the association with cardiovascular outcomes in immune tissues meta-analysis, while CELSR2 gene in epithelial tissues meta-analysis. Additionally, the genes APOB, HCG27 and LPL appear to play roles in immune and mesenchymal tissues. This suggests that these genes might influence immune functions and the behaviour of mesenchymal cells, which are critical for the structure but also function of connective tissues (Table 2 and Additional file 2: Table S9).

Fig. 4.

Transcription-wide associated genes for cardiovascular outcomes. A Upset plot of the 15 gene-cardiovascular phenotypes associations per tissue category and summary of the 6 replicated genes. B Summary of the genetically informed mediation inference analysis among the 6 genes, metabolite levels and cardiovascular risk

Mendelian randomization approach

Genetically informed mediation analysis was performed to infer causal mediation in the relationship between gene expression of the 6 identified genes (PCSK9, CELSR2, PSRC1, APOB, HCG27 and LPL), metabolite levels and cardiovascular risk. To do so, we apply a Mendelian randomization approach in several steps. First of all, we identified the metabolites, related to the 6 genes of interest, that were associated with CAD risk under a MR approach using the SNPs identified in this study as instruments [57]. The metabolites associated with CAD risk were mainly atherogenic metabolites, including the steroid parameters (esterified, free and total cholesterol), levels of small, medium, large and total LDL particles, cholesterol in IDL, LDL and VLDL particles, triglycerides in IDL and levels of triglycerol C52:2 and C52:4 (Additional file 2: Table S10). Second, for each gene, we identified the main eQTL in a relevant tissue (blood for PCSK9, CELSR2, PSRC1, HCG27 and LPL, and adipose subcutaneous tissue for APOB) of GTEx Consortium [46]. Then, we estimated the MR effect of gene expression on the metabolite levels identified in the previous step and on the CAD risk in the three datasets (CGC4D, UK Biobank and FinnGen study) [57]. As expected, all results were significant, except for the genes APOB and HCG27 on the FinnGen CAD risk (Additional file 2: Table S11). Finally, we decomposed the total effects of gene expression on CAD risk into mediated effects and non-mediated effects by metabolites [62]. We observed that for the genes APOB, HCG27 and LPL, the effects mediated by metabolites were significant and explaining the total effects, letting null contribution to non-mediated effects of gene expression of these genes on CAD risk (Fig. 4B, Additional file 2: Table S12). For the genes PCSK9 and CELSR2, the effects mediated by metabolites were significant and partially explaining the total effects, thus, some effect non-mediated by the tested metabolites remained (Fig. 4B, Additional file 2: Table S12). Finally, the gene PSRC1 showed opposite effects of mediated and non-mediated effects resulting in a total negative contribution of gene expression on CAD risk, i.e. being protective. This was the result of an increasing risk effect mediated by concentrations of total LDL, i.e. expression of PSRC1 increases concentrations of total LDL and CAD risk, but a protective effect non-mediated by the tested metabolite, i.e. expression of PSRC1 decreases CAD risk by independent pathways than LDL levels (Fig. 4B, Additional file 2: Table S12).

Discussion

Over the past decade, numerous GWAS have investigated metabolite levels and cardiovascular outcomes, and provided hundreds of genetic regions associated with these traits [64]. However, the interpretation of GWAS findings, the identification of the causal gene underlying susceptibility and the evaluation of relationships between genetic regions, metabolite levels and cardiovascular risk have remained challenging. In this study, to link metabolite-mediated cardiovascular genetic risk loci to key cell types and tissues, we performed integrative analyses spanning multiple GWAS on metabolite levels and cardiovascular diseases, and gene expression reference panels with different broad transcriptional programs under a genetically informed inference framework. We performed genome-wide and transcription-wide association analyses on levels of 190 plasma metabolites and metabolite ratios, from large European datasets including novel 4974 individuals of the Spanish GCAT cohort, and on risk for cardiovascular diseases, from three large European studies. In addition, because the interpretation of the TWAS results is not straightforward by the potential identification of bystander genes, we infer mediation in the relationship between gene expression of six top associated genes (PCSK9, CELSR2, PSRC1, APOB, HCG27 and LPL), metabolite levels and cardiovascular risk. Interestingly, our analysis revealed that the following atherogenic metabolites were the mediators of cardiovascular risk: steroid parameters (esterified, free and total cholesterol), levels of small, medium, large and total LDL particles, cholesterol in IDL, LDL and VLDL particles, triglycerides in IDL and levels of triglycerol C52:2 and C52:4.

Some of the observed genetic associations are predicted to have functional impact, such as the rare non-synonymous SNP in PCSK9 gene [52], the common non-synonymous SNPs in APOB [65], and the non-frameshifting deletion in APOB gene [66], inducing LDL cholesterol and consequently affecting CAD risk. However, the SNPs identified in APOB in our analysis were different from the ones observed in a Japanese study [65], pointing to population-specific functional associations.

Results from the multiomic integration approach identified 53 fine-mapped genes, most of them associated with a single metabolite class; while specificity for tissue type category was not widely observed. In contrast, in the analyses for cardiovascular outcomes, we identified 15 genes, out of the 53 genes, replicating the results for six of them, which gene expression was associated with cardiovascular outcomes showing stronger association in immune tissues. The identification of HCG27 gene as the causal gene for CAD risk in 6p21 locus is a new finding of this study. PCSK9 on 1p32 and CELSR2 and PSRC1 on 1q13 were already associated with cardiovascular risk in previous GWAS [52, 67–70]. As well, APOB (2p24) and LPL (8p21) genes have been previously associated with LDL cholesterol and triglycerides metabolism, and CAD risk [52, 65, 71, 72].

Regarding the MHC region, previous studies did not identify a clear target gene for the observed cardiovascular association [70]. HCG27 is a long noncoding RNA (lncRNA) whose locus was associated with monocyte/leukocyte ratio [73], and circulating HCG27 lncRNA has been associated with acute ischemic stroke and intracranial haemorrhage via inflammation-related signalling pathway [74, 75]. Inflammation in tissues can disrupt normal lipid storage and mobilization processes, which would explain the previous associations between MHC region and circulating triglycerides levels [76, 77]. Our results showed that HCG27 lncRNA expression in immune cells is associated with cardiovascular risk meditated by triglycerol levels. Therefore, HCG27 is a functional candidate gene to explain the GWAS association with CAD risk.

In the 1p32 region, the pointed gene which gene expression would be related to CAD risk was PCSK9 in tissues enriched with immune cells, which included the liver tissue. As observed in the Human Protein Atlas browser (https://www.proteinatlas.org/), the protein encoded by PCSK9 is mainly enriched in the liver, but also expressed in intestine, kidney and blood vessels. Our results indicate that the contribution of PCSK9 on CAD risk is partly mediated by levels of subspecies of cholesterol and LDL particles, and partly mediated by independent pathways. Interestingly, the role of PCSK9 reducing LDL cholesterol was found to act via endosome/lysosome LDLR degradation [78]. However, it was observed that PCSK9 inhibitors reduced LDL cholesterol in more than 60% of initial levels on top of statin therapy, while the reduction in cardiovascular risk was about 15%, much less than expected [79]. This has been partially explained by potential less benefit in statin-treated patients, and by the lack of anti-inflammatory effect of PCSK9 inhibitors. In addition, an increasing number of studies suggested that PCSK9 also influences the haemostatic system, by altering platelet function and the coagulation cascade [80], and inflammatory, apoptotic and immune pathways [81]. Our results add evidence on this line where the relation between PCSK9 and CAD risk appears to transcend lipid levels.

In the case of the 1p13 region, TWAS results identified both CELSR2 and PSRC1 as genes whose expression would modify lipid levels and CAD risk. Genetic studies early identified this locus as involved in lipid metabolism and CAD risk [82, 83]. Interestingly, PSRC1 genetic variants showed significant variation among different populations, being both protective in East Asian populations but risk factors in European populations [84]. Beyond lipid metabolism, the role of PSRC1 gene in cell growth and formation and progression of atherosclerotic plaque is well established [67]. Also, integrative functional analyses identified that PSRC1 variants may regulate gene expression of PSRC1, CELSR2, SORT1 genes in the liver, which alter LDL cholesterol levels and CAD risk, but also circulating granulin [85]. In particular, that study observed how the eQTL used for PSRC1 in our study was inversely associated with protein levels of PLA2G12B, C1QTNF1, CA10 and GRN, which, in turn, can affect CAD risk. These potential mechanisms of action would explain our observations on the opposite effects of lipid mediated and non-mediated effects of PSRC1 gene expression on CAD risk, as well the variation among different populations or type of patients, being of risk when altering levels cholesterol particles but protective when acting on independent pathways. Further functional analyses are required to determine the multifaceted contributions of PSRC1 gene expression on CAD risk.

An interesting observation of this study is that gene expression associated with cardiovascular risk phenotypes often showed a stronger association in tissues with a high percentage of immune and mesenchymal cells. This is in accordance with atherosclerosis as a disease associated with immune-related pathways, where macrophages, monocytes and T cells play a central role in the inflammatory response, plaque development and thrombosis [86–88]. In addition, there is a growing evidence on the role of mesenchymal cells, beyond tissue structure, on epithelial cell differentiation or intestinal immune cell education [89], reinforcing the increasing evidences for the gut-cardiac paradigm in cardiovascular health [90]. Increasing evidence of immune cells action is reported. A higher cancer risk has been associated with a high immune cell tissue content [91]. Here, the gene expression profiles acting in immune cells and tissues with high content of immune and mesenchymal cells provided by this study will contribute to understanding the molecular processes driving cardiovascular risk. Gene expression of these genes can be an attractive therapeutic target to modulate atheroma progression, regression or stabilization.

Another interesting finding of this study is the enrichment of the fine-mapped genes in genes associated with nutritional and metabolic diseases. Accordingly, triglyceride-rich lipoproteins particles, LDL particles, tricarboxylic acids and branched-chain amino acids have been associated with obesity, beta-cell function, insulin resistance and incident type 2 diabetes (T2D) [92–95]. In this study, a considerably proportion of fine-mapped genes were associated with carboxylic acids, and none of these were associated with cardiovascular outcomes. Therefore, the identified carboxylic acid genes could be partially driving this enrichment.

Regarding limitations and future analyses, it should be noted that these results are restricted to European populations and cannot be generalized to other continental populations. For instance, non-synonymous SNPs in APOB gene identified in this study differ from SNPs identified in a Japanese study. LD patterns of European samples or co-regulating genes can originate false positives, confounded results, and complicated the interpretation of the results. Thus, transcription instrumental-based results from different ancestry populations are needed to tease out causal genes. In addition, our results are based on the prediction of gene expression values in a TWAS framework. The predictive performance of gene expression prediction models is limited and depends on the sample sizes of the eQTL datasets used for training the statistical models. Therefore, future trials and functional analyses are needed to definitely validate the causative role of gene expression of the identified genes in the context of the proposed key cell types and tissues for metabolite levels and cardiovascular risk. Finally, since plasma was not always obtained in fasted conditions, genome-wide analyses were adjusted correcting by chylomicrons as a closely related measure of the post-prandial phase. However, this correction could obscure the genetic variants involved in chylomicrons biology [96–98].

Conclusions

We provided evidence for 53 candidate genes whose expression is associated with metabolite levels. Out of these 53 genes, the gene expression of 6 was robustly associated with cardiovascular risk phenotypes showing stronger association in immune cells and tissues with high immune component, including the novel gene regulatory target HCG27. In addition, we estimate the extent to which gene expression affect cardiovascular risk mediated and non-mediated by metabolite levels. These results suggest novel potential gene-drug targets to reduce the risk for cardiovascular outcomes by modifying atherogenic blood metabolite levels.

Authors’ contributions

All authors contributed to feedback of the manuscript. Conception and design: RdeC, VM, MO-S. Generation and data collection: IG-F, RC-T, FM-N, VD-O. Statistical and bioinformatic analysis: IG-F, RC-T, XF, BC, NB. Writing, review and/or revision of the manuscript: RC-T, IG-F, FM-N, VD-O, XF, BC, NB, AC, SI-G, MO-S, RdeC, VM. Administrative, technical or material support: AC, SIG. Study supervision: RdeC, VM. All authors read and approved the final manuscript.

Funding

This work was partially supported by grants; GCAT was funded by Acción de Dinamización del ISCIII-MINECO and the Ministry of Health of the Generalitat of Catalunya (ADE 10/00026); BC was supported by National Grant PI18/01512 and XF by VEIS project (001-P-001647) (co-funded by European Regional Development Fund (ERDF), “A way to build Europe”). IG-F, XF, BC, NB, AC and SI-G are part of the Agency for Management of University and Research Grants (AGAUR) of the Catalan Government grant SGR 01537. RCT was supported by the Horizon 2020 Framework Programme of the European Union under the Marie Sklodowska-Curie grant agreement No 796216, by the Instituto de Salud Carlos III through the Miguel Servet Program CP21/00058, and is part of the AGAUR SGR 01366. VM was supported by the Spanish Association Against Cancer (AECC) Scientic Foundation grant GCTRA18022MORE. Consortium for Biomedical Research in Epidemiology and Public Health (CIBERESP), action Genrisk. MOS received a post-doctoral fellowship from the Spanish Association Against Cancer Scientific Foundation (AECC; POSTD037OBÓN). MOS, FMN and VM are part of the group 55 of CIBERESP and of the AGAUR SGR 723.

Data availability

The GCAT genetic data generated and employed in this study have been deposited in the European Genome-phenome Archive (EGA) under study accession EGAS00001003018 (https://ega-archive.org/studies/EGAS00001003018) [99]. Access to the GWAS data is free but available upon request through the GCAT Data Access website (http://www.gcatbiobank.org/investigadors/gcat-data-access_en/) [100]. Accession URL for GWAS summary statistics for metabolites blood levels are Kettunen et al. (https://www.ebi.ac.uk/gwas/publications/27005778), Gallois et al. (http://statgen.pasteur.fr/Download.html), Shin et al. (https://metabolomips.org/gwas/index.php?task=download), Rhee et al. (https://www.cell.com/cell-metabolism/fulltext/S1550-4131(13)00257-X), Long et al. (https://twinsuk.ac.uk/resources-for-researchers/access-our-data/) and Draisma et al. (https://doi.org/10.34894/JFWWS4). Detailed data can be found in Additional file 2: Table S1. Accession URL for GWAS summary statistics for cardiovascular outcomes are Nikpay et al. ( http://www.cardiogramplusc4d.org/data-downloads/), Van der Harst et al. (https://data.mendeley.com/datasets/2zdd), Jiang et al. (https://yanglab.westlake.edu.cn/resources/u) and FinnGen (https://finngen.gitbook.io/documentation/da). Detailed data can be found in Additional file 2: Table S4. Computer code is available at the GCAT project GitHub repository (https://github.com/gcatbiobank/metabolomics_gwas) [101].

Declarations

Ethics approval and consent to participate

All research was carried out in accordance with relevant national and European guidelines and regulations. The GCAT study was carried out using anonymized data provided by the Catalan Agency for Quality and Health Assessment, within the framework of the PADRIS Program. All participants provided informed consent to participate in the study and ethical approval was obtained by the Ethical Committee at the Hospital Universitari Germans Trias i Pujol (CEI no. PI-13–020). The research conformed to the principles of the Helsinki Declaration. Additionally, detailed information regarding consents, questionnaire contents and available data can be found at www.genomesforlife.com.

Competing interests

The authors declare no competing interests.