Human genetic studies have identified numerous proteins involved in coronary artery disease (CAD), type 2 diabetes (T2D) and other cardiometabolic diseases through a combination of genome-wide association studies (GWAS), fine-mapping, colocalization and Mendelian randomization by overlaying information at strong cardiometabolic disease loci9,10,11,12. However, cardiometabolic diseases are polygenic in architecture since they depend on many thousands of variants across the genome, nearly all exerting small lifelong effects13,14,15,16,17. These variants are spread across many different pathways and likely exert their effects through multiple levels of regulation, including gene expression, proteins and their interactions, cell morphology and higher-order physiological processes18,19,20. PGS aggregate these small effects into a single number for each individual that captures a fraction of their disease susceptibility. The use of PGS for risk stratification has shown potential clinical utility for disease prevention21, yet the specific molecular consequences that precede disease risk for these polygenic effects are unknown. For example, proteins that are pathway-level hubs through which polygenic effects converge could be particularly promising targets for pharmaceutical intervention22,23.

In this study, we demonstrated how PGS can be used to identify proteins with causal roles in disease aetiology. The INTERVAL cohort consists of approximately 50,000 adult blood donors in England24,25, of which 3,087 participants have linked electronic hospital records, imputed genome-wide genotypes and quantitative levels of 3,438 plasma proteins26 (Supplementary Data 1 and 2). A schematic of the study is given in Extended Data Fig. 1. The characteristics of the participants are given in Extended Data Fig. 2; participants with a history of any cardiometabolic disease were excluded (Supplementary Table 1), reducing the potential for reverse causality in downstream analysis.

To quantify each participant’s relative polygenic risk of atrial fibrillation (AF), CAD, chronic kidney disease (CKD), ischaemic stroke (IS) and T2D, we applied externally derived genome-wide PGS consisting of 1.8–3.2 million variants. Using PGS, we identified 49 proteins whose levels differed with respect to polygenic risk at a false discovery rate (FDR) of 5% (Fig. 1a,b, Extended Data Figs. 3 and 4 and Supplementary Tables 2 and 3): 31 proteins for the T2D PGS; 11 proteins for the CAD PGS; 8 proteins for the CKD PGS; and 1 protein for the IS PGS. PGS–protein associations included proteins previously associated with cardiometabolic disease, such as cystatin-C (CST3) and beta2-macroglobulin (B2M), which are biomarkers for CKD27, and fructose-1,6-bisphosphatase 1 (FBP1), which plays a key role in glucose regulation and is a target of T2D drugs28. Associated proteins belonged to multiple non-overlapping pathways (Supplementary Information) and many are relatively understudied in the context of their respective diseases (Extended Data Fig. 5) thereby warranting future study.

Fig. 1: Proteins associated with polygenic risk for cardiometabolic disease.
figure 1

a, Quantile–quantile plots of two-sided P values from linear regression testing associations between PGS and protein levels in n = 3,087 INTERVAL participants across all 3,438 tested proteins. Each plot compares the distribution of observed two-sided P values (y axes) to the distribution of expected two-sided P values under the null hypothesis for 3,438 tests (x axes) on a −log10 scale. Associations were fitted using linear regression adjusting for age, sex, ten genotype principal components, sample measurement batch and time between blood draw and sample processing. Full summary statistics including exact P values are provided in Supplementary Data 3a. The top five proteins by P value are labelled. b, Heatmaps showing the 49 proteins whose levels were significantly associated with at least one PGS after Benjamini–Hochberg FDR multiple-testing correction (FDR < 0.05) of the two-sided P values (statistical tests are as described in a). Each heatmap cell shows the s.d. change in protein levels per s.d. increase in PGS. Point estimates for the 49 FDR-significant proteins are detailed in Extended Data Fig. 3. Details about each protein are provided in Extended Data Fig. 4. c, Barplots showing the proportion of the genome (%) required to explain each PGS–protein association in n = 3,087 INTERVAL participants (polygenicity). Proteins are ordered from left to right by strength of PGS–protein association. Highlighted in red are PGS–protein associations that were explained by singular variants regulating protein levels, pQTLs, rather than polygenic. Percentages are detailed in Extended Data Fig. 3.

PGS–protein associations were robust to technical, physiological and environmental confounding. We observed directional consistency and strong correlation of effect sizes when utilizing an orthogonal proteomics technology in independent samples (Extended Data Fig. 6a–c and Supplementary Information). Protein levels and PGS–protein associations were also temporally stable over 2 years of follow-up (Extended Data Fig. 6c,d and Supplementary Information). PGS–protein associations were also robust to circadian and seasonal effects, inclusion of participants with any prevalent cardiometabolic disease and body mass index (BMI), with the exception of six T2D PGS–protein associations that were partially mediated by BMI (Extended Data Fig. 6f,g).

Most PGS–protein associations were not explained by protein quantitative trait loci (pQTLs) (Supplementary Table 4); instead, they were highly polygenic (Fig. 1c). Each protein required a median 12% of the genome to explain its association with a PGS. Only 4 associations could be explained by pQTLs and contributing loci were spread across the genome for the remaining 46 associations (Extended Data Fig. 7).

Three possible scenarios could explain a PGS–protein association29: (1) the protein plays a causal role in disease; (2) protein levels are changing in response to disease processes but are not themselves causal (reverse causality); and (3) protein levels are correlated with some other causal factor (confounding) (Fig. 2a). Utilizing a median of 7.7 years of follow-up in nationwide electronic hospital records, we examined whether levels of PGS-associated proteins were associated with risk of onset of the respective cardiometabolic disease, then performed mediation analysis30,31,32 to identify the proteins that mediate the PGS–disease associations and thereby play causal roles in disease pathogenesis.

Fig. 2: PGS-associated proteins influence 7.7-year risk of CAD and T2D.
figure 2

a, Possible models of causality for PGS–protein–disease associations. C, causal disease factor upstream of the protein that induces a correlation between protein levels and disease. b, Association of PGS-associated proteins with 7.7-year risk of hospitalization with CAD or T2D in n = 3,087 INTERVAL participants in Cox proportional hazards models adjusting for age, sex, sample measurement batch and time between blood draw and sample processing. The data shown correspond to the HR for CAD or T2D conferred per s.d. increase in protein levels (points) and its 95% CI (vertical bars). P < 0.0012 indicates that the association was significant after Bonferroni correction for the 42 tests. c, Comparison of associations between protein levels and the CAD PGS or T2D PGS from Fig. 1b (x axes) to HRs for protein levels for 7.7-year risk of hospitalization with CAD or T2D from Fig. 2b (y axes). Beta estimates (points; x axes) correspond to s.d. change in protein levels per s.d. increase in CAD PGS or T2D PGS in the linear regression described in Fig. 1b. HRs (y axes) are as described in b. Linear regression and Cox proportional hazards models were fitted in the same n = 3,087 samples. The point shape and colour correspond to the P value in b. d, Percentage of PGS–disease association mediated by each protein in causal mediation analysis in n = 3,087 INTERVAL participants adjusting for age, sex, ten genotype principal components, sample measurement batch and time between blood draws. The data shown are the percentage of association between the PGS and hospitalization with CAD or T2D after 7.7 years of follow-up in Cox proportional hazard models mediated by each respective protein (points) and the 95% CI of this percentage (vertical bars). Proteins are ordered from left to right by their HR in Fig. 1b and coloured red where protein levels increased with PGS or blue where protein levels decreased with PGS in Fig. 1b. P < 0.0012 indicates that mediation was significant after Bonferroni correction for the 42 tests. Full summary statistics including exact two-sided P values for bd are detailed in Extended Data Fig. 3.

During follow-up of the participants with PGS and plasma proteomics, there were 27 incident T2D events and 15 incident CAD events, enabling us to evaluate the CAD and T2D PGS and their corresponding 42 associated proteins. Ten of 31 (32%) T2D PGS-associated proteins were significantly associated (P < 0.0012, Bonferroni correction for the 42 tested proteins) and a further 15 proteins were nominally significantly associated (P < 0.05) with increased risk of T2D (Fig. 2b and Extended Data Fig. 3). For the CAD PGS, no proteins were Bonferroni significant and 3 of 11 (27%) proteins were nominally significant. Notably, there was clear directional consistency between the effects of PGS on protein levels and hazard ratios (HRs) for protein levels on incident disease risk (Fig. 2c). Using mediation analysis, we found one protein, insulin-like growth factor-binding protein 2 (IGFBP2), that was a significant mediator (P < 0.0012) of polygenic T2D risk (Fig. 2d and Extended Data Fig. 3), indicating a causal role in disease pathogenesis. A further 1 and 14 proteins were nominally significant mediators of polygenic CAD and T2D risk, respectively. Protein–disease associations and causal mediation for IGFBP1 and IGFBP2 on polygenic T2D risk were replicated in independent samples where four PGS-associated proteins (IGFBP2, IGFBP1, progranulin (GRN) and tissue inhibitor of metalloproteinases 4 (TIMP4)) were measured with orthogonal proteomics technology (Supplementary Information and Supplementary Table 5).

Since polygenic disease risk is estimated from population-level data, it is unlikely that any single protein explains polygenic risk. In this study, we found that IGFBP2 explained 13.4% of the association between T2D PGS and incident T2D (Extended Data Fig. 3). Across all nominally significant mediators, proteins explained a median of 6.6% of PGS–disease associations (Extended Data Fig. 3), with the 1 CAD PGS mediator (apolipoprotein E (APOE)) explaining 5.4% of CAD polygenic risk–incident CAD association and the 15 T2D PGS mediators explaining 27% of the T2D polygenic risk–incident T2D association.

A complementary approach for causal inference, Mendelian randomization33, also supported causal effects on T2D for one protein, sex hormone-binding globulin (SHBG) (Extended Data Fig. 9 and Supplementary Tables 6 and 7), which is consistent with our mediation analysis and previous Mendelian randomization analysis of SHBG on T2D34. Notably, only 11 (22%) of the proteins associated with PGS could be tested with Mendelian randomization due to a lack of cis-pQTLs as genetic instruments, highlighting the complementarity of our PGS–protein association and mediation approach for identifying causal proteins.

Our findings are also consistent with a previous observational study of plasma proteins and T2D risk in the Age, Gene/Environment Susceptibility-Reykjavik Study (AGES-Reykjavik), a cohort of 5,438 older Icelanders with 654 prevalent T2D cases and 112 incident T2D cases in 2,940 participants with 5 years of follow-up and free of T2D at baseline35. Of the 31 proteins associated with the polygenic risk of T2D in our healthy, pre-symptomatic cohort, 23 were associated with prevalent T2D and 16 were associated with incident T2D in the AGES-Reykjavik Study (Extended Data Fig. 10a and Supplementary Table 8). Notably, HRs for incident T2D in our INTERVAL analyses were directionally consistent and of similar magnitude to the odds ratios for incident T2D in the AGES-Reykjavik Study (Extended Data Fig. 10b), with a significant overlap between the significant proteins from our causal mediation analysis in INTERVAL and those previously associated with incident T2D (Extended Data Fig. 10a).

Finally, we examined the druggability of proteins mediating polygenic disease risk using the druggable genome22. We found that 12 of the 16 proteins mediating the polygenic disease risk were also druggable targets (Table 1). Nine of these were targets of, or interacted with, 76 compounds in the DrugBank database36 (Supplementary Table 9). These results suggest therapeutic potential for these proteins as modulators of risk for T2D or CAD and indicate high priority targets for further investigation.

Table 1 Druggable proteins that were nominally significant mediators of polygenic risk

Polygenic disease scores are explicitly constructed to maximize risk prediction, typically without consideration of the underlying biology. However, PGS also hold additional promise for identifying molecular pathways in the development and progression of disease8,29. In this study, we identified 49 plasma proteins significantly associated with PGS for cardiometabolic disease in a healthy pre-disease cohort. Twenty-eight of these proteins were associated with increased risk of future disease and 16 were nominally significant mediators of T2D or CAD, including 12 druggable targets, suggesting that their modulation may potentially attenuate disease risk.

The vast majority of PGS–protein associations were highly polygenic, including for several well-known cardiometabolic disease proteins. This polygenicity was driven by aggregate modest polygenic effects on protein levels from across the genome, which were independent of cis- and trans-pQTLs and also present for proteins without pQTLs or for which sample sizes have not yet been sufficient for pQTL detection. This highlights the complementarity of PGS to established approaches that utilize information at a single locus, such as Mendelian randomization, colocalization and fine-mapping. However, it is important to recognize that mediation analysis provides weaker evidence of causality than these established single-locus approaches since it is more difficult to rule out confounding (either from measured or unmeasured factors)37, especially since PGS by design capture horizontal pleiotropy.

Our findings identify new potential targets of cardiometabolic disease that both are supported by human genetic evidence of causality and may be amenable to pharmacological manipulation. Our strongest results were for IGFBP2, a druggable target22, as a replicable mediator of polygenic risk and incident T2D. IGFBP2 is involved in the regulation of glucose uptake into adipocytes and is associated with increased insulin sensitivity and decreased adipogenesis38,39,40,41. Increased plasma IGFBP2 levels have been associated with lower T2D risk35,42,43; our findings are directionally consistent and indicate that this association is likely causal. Ten additional druggable proteins were found to be new mediators between polygenic risk and incident T2D.

Twelve new protein associations were found for CAD, CKD or T2D. Among these, the strongest evidence was for alcohol dehydrogenase 4 (ADH4), which is involved in a number of metabolic pathways44 and was found to be both a mediator of polygenic risk and incident T2D and a druggable target. Furthermore, several new associations concerned proteins with sparse literature on their function; for example, crystallin zeta like 1 (CRYZL1) was associated with polygenic risk and incident CAD; however, little is known about CRYZL1 beyond its gene identification45.

Proteomic data are becoming increasingly available in cohorts of large sample sizes, such as the planned proteomic profiling of UK Biobank participants. Proteomic platforms are also increasing their coverage of the human proteome. Therefore, we anticipate that our PGS mediation analysis approach will enable the identification of further causal proteins for cardiometabolic and other polygenic diseases in future studies.

Overall, this study demonstrates that PGS can be utilized to elucidate new disease biology with therapeutic potential and provides a useful study design for future studies into the molecular drivers of polygenic disease.



INTERVAL is a cohort of approximately 50,000 participants nested within a randomized trial studying the safety of varying the frequency of blood donation24,25. Participants were blood donors aged 18 years and older (median 44 years of age; 49% women) recruited between June 2012 and June 2014 from 25 centres across England. The collection of their blood samples for research purposes was done using standard protocols24: blood samples for research purposes were collected in 6-ml EDTA tubes using standard venepuncture protocols. The tubes were inverted three times and transferred at ambient temperature to the UK Biocentre for processing. Plasma was extracted into two 0.8-ml plasma aliquots by centrifugation and subsequently stored at −80 °C before use. Participants gave written informed consent and this study was approved by the National Research Ethics Service (no. 11/EE/0538).

Electronic health records were obtained for all INTERVAL participants from the January 2021 release of the National Health Service (NHS) Hospital Episode Statistics database ( for all events up to 8 February 2020, before the onset of the COVID-19 pandemic in England. The median and maximum follow-up times were 6.9 years and 7.7 years, respectively. The earliest available hospital record for any INTERVAL participant was from 25 March 1999, with a maximum retrospective follow-up of 13.6 years. These records came in the form of International Statistical Classification of Diseases and Related Health Problems, 10th revision (ICD-10) codes46 and were subsequently made available to analysts after summarization into 301 end points using the CALIBER rule-based phenotyping algorithms47 ( ICD-10 codes contributed to each event regardless of whether they coded for primary or non-primary diagnoses in the hospital records.

Genotyping, quality control and imputation of INTERVAL participants were performed as described previously48: participants were genotyped using the Affymetrix UK Biobank Axiom array in ten batches. Samples were removed if they had sex mismatch, had extreme heterozygosity, were of non-European ancestry or were duplicate samples. Related samples were removed by excluding one sample from each pair of close relatives (first- or second-degree; identity by descent \(\hat \pi > 0.187\)). Genotyped variants were removed if they were monomorphic, bi-allelic and had Hardy–Weinberg equilibrium P < 5 × 10−6 or call rate <99%. SHAPEIT3 was used to phase variants; imputation to the UK10K/1000 Genomes panel was performed using the Sanger Imputation Server (

Protein levels in INTERVAL were quantified using SOMAscan assays, processed and quality-controlled as described previously26: relative concentrations of 4,034 SOMAscan aptamers were measured in 3,562 INTERVAL participants in two batches by SomaLogic using v.3 of the SOMAscan platform. Aptamer concentrations (relative fluorescence units) were natural log-transformed and then adjusted within each batch for participant age, sex, the first three genetic principal components and time between blood draw and sample processing (<1 or >1 day); the residuals were then inverse rank normal-transformed. In this study, we further adjusted the normalized protein levels used in previous studies for batch number and filtered to 3,793 high-quality aptamers targeting 3,438 proteins after obtaining the latest information about aptamer sensitivity and specificity from SomaLogic. Aptamers were excluded if, in v.4 of the SOMAscan platform, they (1) targeted non-human proteins, (2) measured the fusion construct rather than the target protein or (3) measured a contaminant. A curated information sheet for all 4,034 aptamers is provided in Supplementary Data 1. The distributions of aptamer levels and associations with covariates before and after quality control are given in Supplementary Data 2.

In total, 3,087 INTERVAL participants, without prevalent cardiometabolic disease (see below) and with matched genotype, proteomic and electronic health record data available for the primary analyses, passed quality control.

Prevalent disease exclusion

The NHS Blood and Transplant blood donation eligibility criteria ( meant that there were built-in exclusions for the INTERVAL cohort for people with a history of major diseases, recent illness or infection. Specifically, for cardiometabolic diseases, the blood donation eligibility criteria excluded individuals who had been diagnosed with AF, had a history of any stroke or a history of major heart disease, including heart failure, coronary thrombosis, myocardial infarction, cardiomyopathy, ischaemic heart disease and arrhythmia, or surgery for non-congenital heart conditions. Use of aspirin or other blood thinners to control elevated blood pressure (hypertension) also made people ineligible to donate blood and participate in the INTERVAL cohort. Individuals with T2D were ineligible unless their T2D was well controlled by diet alone, did not require regular insulin treatment and the individual had not required insulin treatment for at least 4 weeks before attempting blood donation. Extended details on the blood donation criteria eligibility for specific diseases, medications and lifestyle factors can be found at

In addition to intrinsic exclusion due to the blood donation eligibility criteria, participants were excluded from the analyses if they had any events relating to cardiometabolic disease before baseline assessment. Among the 301 CALIBER end points, we classified 48 as cardiometabolic disease or having potential to introduce reverse causality by modifying risk for incident AF, CAD, CKD, IS or T2D (Supplementary Table 1). In total, 87 participants (2.7%) were excluded, predominantly due to prevalent hypertension (n = 57 events; 66% of excluded participants) and prevalent diabetes (n = 11 events; 13% of excluded participants), with all others accounting for less than 5% of excluded participants (Supplementary Table 1).


PGS were derived in a consistent manner by linkage disequilibrium (LD) thinning, at an r2 threshold of 0.9, the latest GWAS summary statistics for each respective disease (Supplementary Information). The GWAS summary statistics used to derive the AF, CKD and T2D PGS were those published by Nielsen et al.13 (GCST006414), Wuttke et al.14 (GCST008065) and Mahajan et al.15 (GCST007517), respectively. The PGS for CAD and IS used in this study were our previously published CAD49 and stroke50 meta-PGS. The CAD PGS was derived from the meta-analysis of three PGS for CAD, including a PGS derived as described above from the GWAS summary statistics published by Nikpay et al.51. The IS PGS was derived from the meta-analysis of PGS for IS and its risk factors, including a PGS derived as described above from the GWAS summary statistics for IS published by Malik et al.16. Each PGS comprised 1.75–3.23 million single-nucleotide polymorphisms (SNPs) genome-wide and is available to download through the Polygenic Score Catalog52 ( with accession nos PGS000727 (AF), PGS000018 (CAD), PGS000728 (CKD), PGS000039 (IS) and PGS000729 (T2D). All PGS were derived from the GWAS summary statistics including only individuals with European ancestry. See Supplementary Information and Extended Data Fig. 8 for details on PGS validation.

The levels of each PGS (sum of dosages × weights) were computed in INTERVAL from probabilistic dosage data using PLINK v.2 (ref. 53) after mapping PGS variants to those available in the INTERVAL genotype data (Supplementary Information). The levels of each PGS were adjusted for the first ten principal components of the imputed genotype data and standardized to have mean of 0 and s.d. of 1 before downstream statistical analyses.

PGS–protein associations

Each of the five PGS was tested for association with each of the 3,793 aptamers using linear regression (Fig. 1a,b and Extended Data Fig. 3). PGS and proteins were adjusted for covariates and normalized before model fitting (see above). Linear regression coefficients were averaged where multiple high-quality aptamers targeted the same protein (Supplementary Information). FDR correction was subsequently applied across the 3,438 P values (1 per protein) for each PGS separately. Details on aptamer specificity and sensitivity are given in Supplementary Table 2 for the 54 aptamers targeting the 49 PGS-associated proteins; aptamer-specific estimates of PGS on protein levels are detailed in Supplementary Table 3 for the 5 PGS-associated proteins targeted by more than one aptamer (WFIKKN2, GPD1, IGFBP1, IGFBP2 and SHBG).

Polygenicity of PGS–protein associations

To quantify the polygenicity of PGS–protein associations (Fig. 1c and Extended Data Fig. 7), we performed a multistep experiment to determine the proportion of the genome required to explain that association. First, we split the given PGS into separate scores for each of the 1,703 approximately independent LD blocks estimated in Europeans from the 1000 Genomes reference panel by Berisa and Pickrell54 ( Next, we tested each of these 1,703 scores for association with the given protein (Supplementary Data 3e). Then, we retested the PGS to protein association, progressively removing independent LD blocks, at each step removing the LD block whose score had the strongest association with the protein. From this, we quantified polygenicity (Fig. 1c) based on the LD blocks needed to be removed from the given PGS to attenuate the PGS–protein association (so that association P > 0.05; Supplementary Data 3f) as the sum of removed LD block sizes/sum of all LD block sizes (that is, the proportion of the genome removed). Extended Data Fig. 7 shows the independent LD blocks contributing to the polygenicity of each PGS–protein association.

Independent contributions of PGS and pQTLs to protein levels

Multivariable linear regression models were fitted for each protein on PGS levels and pQTL dosages to estimate their independent contributions to protein levels (Supplementary Table 4). The pQTLs used for each protein were (1) conditionally independent pQTLs mapped in INTERVAL and published by Sun et al.26, which included both cis-pQTLs (within 1 Mb of the encoding gene) and trans-pQTLs passing the trans significance threshold of P < 1.5 × 10−11; (2) trans-pQTLs with P < 1.5 × 10−11 (lead variant only) for proteins not published in Sun et al.26 (B2M, DUSP26 and FTMT); and (3) hierarchically significant55,56 cis-pQTLs (lead variant only) mapped in this study (Supplementary Data 4 and Supplementary Information) for proteins without cis-pQTLs passing the trans-pQTL significance threshold above (ACY1, ADIPOQ, APOE, CST3, GPD1, PTPRU, SHBG and UST).

Incident disease associations

PGS and protein levels were tested for association with incident disease using Cox proportional hazards models adjusting for age and sex (Fig. 2b and Extended Data Fig. 8) using the survival package (version 3.2-7) in R. The timescale used was time from baseline to first event of the relevant disease or to the latest available date in the hospital records (8 February 2020). PGS and proteins were adjusted for covariates and normalized before model fitting (see above). Cox model coefficients were averaged where multiple high-quality aptamers targeted the same protein (Supplementary Information).

Incident disease events for AF, CAD, CKD, IS and T2D were defined as the first hospital episode for the closest matching CALIBER phenotype47. Incident AF events were defined as any hospital episode with the ICD-10 code I48. Incident IS events were defined as any hospital episode with the ICD-10 code I63 or I69.3. Incident CAD events were defined as any hospital episode with ICD-10 code I21–I23, I24.1 or I25.2 (CALIBER end point myocardial infarction). The closest matching CALIBER phenotype for T2D was for diabetes more broadly, including ICD-10 codes for any hospital episode for T1D or T2D or complications thereof: E10–E14, G59.0, G63.2, H28.0, H36.0, M14.2, N08.3 or O24.0–O24.3. However, we note that individuals with T1D are not eligible to donate blood and adult onset of T1D is relatively rare compared to T2D57. The closest matching CALIBER phenotype for CKD was for end-stage renal disease more broadly, defined as any hospital episode with the ICD-10 codes N16.5, N18.5, T82.4, T86.1, Y60.2, Y61.2, Y84.1, Z49.1, Z49.2, Z94.0 and Z99.2.

Mediation analysis

Mediation analysis was used to identify causal proteins by identifying the PGS-associated proteins that partially mediate the association of PGS on disease (Fig. 2d). This approach uses the counterfactual framework to infer causal effects30,31,32 and can be adapted to this setting because the arrow of causality between PGS and any associated phenotype can only flow in one direction since the PGS is fixed at conception (that is, the underlying alleles in each person cannot be modified later in life by protein levels or the development of cardiometabolic disease). In this study, we used the natural effects model developed by Vansteelandt et al.58, which is available in the medflex R package (version 0.6-7 used in this study)59, to estimate natural indirect effects (effects of PGS on disease through protein levels) on the log odds scale by imputing unobserved counterfactuals. Standard errors were computed using the robust sandwich estimator60, from which 95% confidence intervals (CIs) and P values were calculated. The percentage of PGS–disease associations mediated by each protein and 95% CIs were subsequently computed as the natural indirect effect and its 95% CI was divided by the total effect estimated by each mediation test. Multiple mediation analysis61 was performed using the R package mma (version 10.3.2)62 to quantify the proportion of PGS–disease association mediated by the 15 causal T2D proteins.

Mendelian randomization

Two-sample Mendelian randomization33 was also performed as an orthogonal approach to identify proteins that may play a causal role in disease (Extended Data Fig. 9 and Supplementary Tables 6 and 7). PGS-associated proteins were tested provided they had three or more independent, as determined by LD (r2 < 0.1), hierarchically significant cis-pQTLs after mapping cis-pQTLs to the GWAS summary statistics (Supplementary Information) using five different Mendelian randomization methods63,64,65,66, each of which makes use of information across three or more instruments to estimate causal effects, with each method differentially robust to different sources of bias, to obtain a consensus (median) estimate of causal effects of protein levels on disease risk (Supplementary Information). Hierarchically significant cis-pQTLs and tagging variants (LD r2 > 0.1) were excluded where they encoded changes to protein structure67 (for example, missense mutations) and therefore potentially reflected differences in aptamer binding affinity rather than regulation of protein levels (Supplementary Information). Aptamers were also excluded if they had similar affinity for/comparable binding to multiple proteins or differential binding to specific isoforms (Supplementary Table 3 and Supplementary Information).

In total, 11 of the 49 PGS-associated proteins could be tested. GWAS summary statistics were obtained from Nelson et al.17 for CAD (GCST004787), Wuttke et al.14 for CKD (GCST008065), Malik et al.16 for IS (GCST006906) and Mahajan et al.15 for T2D (GCST007518). In all cases, we used the GWAS summary statistics for the samples of recent European ancestry. For T2D, we used the BMI-adjusted GWAS summary statistics to avoid false positive causal estimates arising where pQTLs influence T2D risk through BMI rather than through the tested protein (horizontal pleiotropy). We considered there to be a significant causal effect where P < 0.05 along with no significant evidence that causal effects were due to associations of the pQTLs with some other causal risk factor (horizontal pleiotropy; Egger intercept66 P > 0.05). Analysis was performed using the R package MendelianRandomization (version 0.5.0)68. Colocalization analysis69 was also performed where the cis-pQTL instruments had P < 1 × 106 in the respective GWAS (Supplementary Table 7 and Supplementary Information).

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.