Multi-cohort proteogenomic analyses reveal genetic effects across the proteome and diseasome Authors: Mine Koprulu, Karl Smith-Byrne, Brian Richard Ferolito, Erin Macdonald-Dunlop, Jian’an Luan, Åsa K. Hedman, Chibuzor Franklin Ogamba, Jurgis Kuliesius, Linda Repetto, Anna Ramisch, Fahim Abbasi, Johan Ärnlöv, Themistocles L. Assimes, BeLOVE Study Group, Hanna M. Björck, Sophia Björkander, Morten Böttcher, Adam Stuart Butterworth, Zhengming Chen, Kelly Cho, Robert Joseph Clarke, Simon Riddington Cox, Kamila Czene, John Danesh, George Dedoussis, Sölve Elmståhl, Niclas Eriksson, Per Eriksson, Tõnu Esko, Estonian Biobank Research Team, Aida Ferreiro-Iglesias, Paul William Franks, Jingyuan Fu, J. Michael Gaziano, Mohsen Ghanbari, Christian Gieger, Arthur Gilly, Harald Grallert, Marc James Gunter, Stefan Gustafsson Published: 2026 ABSTRACT SUMMARYUnderstanding the genetic regulation of circulating protein levels can provide new insights into disease mechanisms. Here, we present the largest proteogenomic study to date (n = 78,664 participants across 38 studies), identifying >24,000 protein quantitative trait loci (QTLs) associated with 1,116 proteins, acting near to (n = 5,040) or distant (n = 19,698) from the cognate gene. Using machine learning-guided effector gene assignment, we provide genetic evidence for pathways, cell types, and tissues that modulate circulating protein levels, highlighting N-linked glycosylation as an important regulatory pathway. We demonstrate that genetic instruments of protein production/function (“cis”) versus modulation (“trans”) reveal distinct phenotypic insights. We identify proteins as candidates for drug targets and engagement (e.g., plasma furin and cardiovascular diseases) by comparing cis-based genetic evidence with protein-disease associations. Systematic triangulation of trans-protein QTLs (pQTLs) with genetic and protein associations across many diseases highlights potential drug repurposing opportunities, e.g., tyrosine kinase 2 (TYK2) inhibitors for rheumatoid arthritis. Our multi-cohort meta-analyses generate proteogenomic insights into disease mechanisms and new treatment opportunities. INTRODUCTION Most disease-predisposing variation in the human genome resides in non-coding regions, limiting inference about causal genes and mechanisms.1,2 Systematic functional characterization of disease-associated variants can guide the identification of clinically relevant mechanisms,3–5 but model systems are difficult to scale or translate directly to human biology.1–3 High-throughput, broad-coverage proteomic technologies that simultaneously target thousands of proteins in human blood can help to link non-coding variants to disease-relevant mechanisms via proteins.6–24 Although such proteogenomic studies have highlighted novel clinically relevant mechanisms and potential drug targets or indications,6–24 important limitations remain. First, broad-coverage studies to date have almost exclusively relied on proximally acting cis-variants (i.e., cis-protein quantitative trait loci [cis-pQTLs]) to derive disease insights. However, non-coding variants are likely to map to regulatory regions directly affecting multiple genes encoded close by (as also demonstrated for gene expression QTLs25) or indirectly regulate proteins produced by genes that are elsewhere in the genome. Second, a better understanding of the polygenic architecture affecting diagnostic, predictive, or prognostic protein biomarkers can be important, as recently demonstrated for prostate-specific antigen and prostate cancer prediction.26 Finally, robust and generalizable identification of pQTLs requires replication across diverse populations, which until now has rarely been pursued for broad-coverage proteomics.18,23,24 Here, we present multi-cohort proteogenomic meta-analyses of up to 1,161 protein targets in up to 78,664 individuals across 38 international cohorts (see STAR Methods and Figure S1A) to address these limitations and provide a conceptual advance by integrating and improving the biological interpretability of trans-pQTL effects, i.e., pQTLs that are not nearby the protein-encoding gene for the protein they associate with. We identify pathways, such as N-linked glycosylation, tissues, and cell types that regulate the plasma proteome based on human genetic evidence and demonstrate divergence of findings from cis- versus trans-focused genetic analyses across diseases. Systematically comparing cis-based gene-to-disease versus measured protein-to-disease associations in up to 1.3 million participants, we demonstrate limited convergence, with only a few examples supported by evidence from both approaches. We finally demonstrate the value of the many newly identified pleiotropic (trans-)pQTLs by illustrating how they can inform disease mechanisms, including druggable opportunities. RESULTS Multi-cohort genome-proteome-wide pQTL discovery We performed a proteogenomic meta-analysis of 1,161 protein targets measured in blood across 38 cohorts, including up to 78,664 participants of European ancestry (Table S1A; for study design see Figure S1). We identified 14,690 regional sentinel variants (n = 1,009 cis-pQTLs, n = 13,681 trans-pQTLs) for 1,116 protein targets at a Bonferroni-corrected significance threshold. We observed at least one pQTL close to the cognate protein-coding gene (i.e., cis-pQTL) or elsewhere in the genome (i.e., trans-pQTL) for 87.1% and 94.1% of the tested protein targets (n = 1,161), respectively. Bayesian fine-mapping27 of the 14,690 identified genomic regions revealed 24,738 independent sets of credible variants associated with 1,116 protein targets (n = 5,040 cis-pQTLs and n = 19,698 trans-pQTLs; Figure 1). We observed a median of 4 credible sets (interquartile range [IQR]: 2–7; Table S2), including independent genetic variants within/nearby cognate protein-coding genes. We observed considerable correlation (r = 0.6) of effect sizes for the identified regional sentinel variants in cohorts of non-European ancestry (n = 3 cohorts), where overlapping information was available (3,709 pQTLs for 501 protein targets; Figure S2). For about a third of the protein targets with at least one cis-pQTL, we observed evidence that at least one of the cis genetic signals was also colocalized with expression QTLs of the protein-coding gene in one or more of 43 tissues, indicating these as potential origins for given blood proteins25 (see STAR Methods and Table S3). Functional annotation of identified pQTLs further replicated the previous observation of high rates of functional, e.g., missense, variants among cis-pQTLs (10.6%), although we observed similarly high rates of functional variants among trans-pQTLs (10.4%), including those associated with many protein targets. The latter may imply that broad effects on the plasma proteome require disturbance of causative gene products rather than subtle effects on gene regulation, as has been seen for most genome-wide association study (GWAS) traits. Most fine-mapped pQTLs were “high confidence” (82.3% of the 5,040 cis- and 83.3% of the 19,698 trans-fine-mapped pQTLs), defined as (1) genome-wide significant (p < 5 × 10−8); (2) directionally consistent in the overall meta-analysis between SCALLOP cohorts and UK Biobank (UKBB) for cis- and trans-pQTLs; and, additionally, with (3) no or minimal heterogeneity (phet > 1 × 10−4) for trans-pQTLs (Figure 1; Table S2). Of those, 278 cis- and 4,013 trans-pQTLs or their proxies (r2 > 0.1) have not previously been reported for the same protein target6–24 (Table S2). Investigating the impact of cohort-level characteristics from SCALLOP cohorts (mean cohort age, mean cohort BMI, composition of sex in the cohort, whether any cases were included in the cohort, fasting status, and blood-based sample type), we observed that the heterogeneity was mostly explained by small variations in very strong genetic effects (Tables S1 and S4). We observed large variation among protein targets when contrasting variance explained (mean = 8.4%, IQR = 4.9%–10.4%) by loci identified in this study and the estimated remaining polygenic background (mean = 10.2%, IQR = 6.6%–13.6%; Figure S3A). Extreme examples ranged from monogenic proteins such as Fc receptor-like protein 3 (FCRL3, variance explained = 45.3%; polygenic background = 3.8%), mainly explained by identified cis loci, to polygenic ones with contributions from cis and trans loci, as well as the polygenic background, such as vascular endothelial growth factor receptor 2 (VEGFR-2 encoded by KDR, variance explained = 21.6%; polygenic background = 16.4%). pQTLs close to the protein-coding gene explained, on average, (i.e., cis), more variation in plasma protein levels compared with the cumulative impact of variation elsewhere in the genome (i.e., trans) (Figure S3A; Table S5). However, extreme outliers included intercellular adhesion molecule 2 (ICAM2), for which cis-pQTLs explained only 0.3% of variance compared with 52.7% explained by trans-pQTLs or alpha-L-fucosidase 1 (FUCA1; 6.3% by cis-pQTLs versus 68.4% by trans-pQTLs). We further observed 261 protein targets for which the linear dependency between explained variance by pQTLs and polygenic heritability estimates did not persist, providing evidence that our study may have already saturated pQTL discovery for these proteins (Figure S3B). Understanding characteristics of protein targets under genetic regulation A higher number of pQTLs among protein targets showed a significant association with the presence of disulfide bonds or a transmembrane domain (Figure 2; Table S6), likely explained by measuring those in the most relevant compartment (i.e., the circulation). In contrast, a higher constraint of the protein-coding gene (i.e., probability of loss-of-function intolerance [pLI]) was inversely associated with the presence and number of cis-pQTLs (p = 4.8 × 10−4) as well as the number of trans-pQTLs (p = 9.9 × 10−13). Proteins with a higher number of trans-pQTLs were further characterized by those enriched for characteristics of secreted proteins,28 such as glycosylation and sulfation, but depleted for structural features of intracellular proteins, such as zinc-finger and DNA-binding domains (Figure 2; Table S6). We note that these findings persisted upon controlling for missingness rates for a given protein; that is, our findings are unlikely to be driven by higher precision for proteins reliably measured with the assay technology (Table S6). Effector genes at trans-pQTLs inform pathway and cell-type contributions Most identified genetic regulation of the human proteome does not occur within or nearby a protein’s cognate gene but rather elsewhere in the genome. Mapping effector genes in distal genomic regions to pQTLs that associate with plasma protein levels has been challenging but can benefit from prior biological knowledge, for example, in ligand-receptor pairs.20 We identified at least one candidate effector gene with at least medium confidence (candidate gene score > 1 out of 3) for more than half of the trans-pQTLs (n = 11,261) by incorporating prior biological knowledge in a machine learning framework (see STAR Methods and Table S7). This included 1,534 high-confidence assignments with a candidate gene score ≥2 (maximum: 2.704 out of 3). For two-thirds of the loci (n = 13,881), the distribution of candidate gene scores across genes indicated a single causal gene as the most likely. As a partial external validation, we observed 552 trans-pQTLs mapping to effector genes that encoded high-confidence protein-protein interaction partners from the STRING network, which was not part of our gene annotation pipeline. We next sought to identify the pathways, cell types, and tissues involved in the trans regulation of the assayed protein targets. We identified at least one significantly enriched pathway among the top-prioritized trans-effector genes for 431 protein targets (Table S8A). Those segregated roughly into two categories. For 101 protein targets, effector genes were enriched for at least one pathway that also involved the cis-protein, providing evidence that higher/lower circulating protein levels were, at least partly, a result of generally higher/lower pathway activity. For example, trans-effector genes for the low-density lipoprotein (LDL)-receptor were more than 85-fold enriched for members of cholesterol metabolism (false discovery rate [FDR] = 9.8 × 10−15). In contrast, for most assayed protein targets (n = 330), pathways enriched among trans-effector genes rather pointed to distal, systemic regulators of (circulating) protein levels. These included pathways related to post-translational modifications, such as “asparagine N-linked glycosylation” (n = 143 protein targets; Figure 3B), as well as pathways related to platelet and blood cell biology more broadly, such as platelet aggregation (n = 41) (Figure 3A). N-glycosylation is the most common post-translational modification of secreted proteins, determining correct folding and secretion, as well as affecting half-life and signaling properties of protein targets.29 We note that the most pleiotropic trans-effector genes mapped to elongation, branching, and capping of the glycans, modulating stability rather than earlier processes, such as folding in the endoplasmic reticulum. We identified 95 and 97 protein targets for which trans-effector genes were significantly (FDR < 5%) enriched for the specific/enhanced expression of the protein-coding gene in one of 12 tissues and 29 cell types, respectively (Figure 3C; Tables S8B and S8C). This included 44 protein-tissue and 76 protein-cell-type pairs that were not the primary site of protein production of the cis-protein, potentially demonstrating cross-organ communication in protein homeostasis. Most prominently, effector genes with enhanced expression in the liver, specifically hepatocytes, were linked to many circulating protein targets. Effector genes expressed in immune cell populations that modulate interleukins (ILs), e.g., genes with enhanced expression in natural killer (NK) cells, were also more than 13-fold enriched (FDR = 3.4 × 10−6) for trans-effector genes of the NK cell activator IL-15. Other examples included the brain-specific extracellular matrix protein tenascin R, for which trans-effector genes were more than 20-fold enriched (FDR = 2.8 × 10−8) for genes with enhanced expression in endothelial cells, or neurotrophin 3, a neuronal growth factor, with trans-effector genes being strongly enriched for enhanced expression in alveolar type 2 cells in the lung. We also observed examples with trans-effector genes being enriched for multiple, different cell types for the same protein target. For example, trans-effector genes for IL-27 with enhanced expression in hepatocytes and Kupffer cells (resident macrophages in the liver) were linked to glycosylation and recognition of pathogens, respectively. Molecular versus phenome-wide pleiotropy Almost half (43.4%; n = 4,547) of all independent pQTLs we identified have been reported to associate with at least one non-proteomic trait, e.g., disease susceptibility, in the GWAS catalog30 (see STAR Methods and Figures 4A–4C). This included a >4-fold enrichment (odds ratio [OR]: 4.11; p < 1.1 × 10−230) of trans- compared with cis-pQTLs, demonstrating the importance of understanding the polygenic architecture of plasma proteins (Figure 4; Table S9A). We subsequently characterized the pleiotropic genetic variants into three categories: (1) “molecular pleiotropy” (associated with >5 proteins and ≤5 non-proteomic phenotypes), (2) “phenotypic pleiotropy” (associated with >5 non-proteomic phenotypes and ≤5 proteins), and (3) “unspecific pleiotropy” (associated with >5 proteins and >5 non-proteomic phenotypes) (Figures 4B and 4C). More than half (332 out of 533) of the pQTLs that were pleiotropic at the protein level also showed phenotypic pleiotropy in the GWAS catalog. Associated effector genes were >2-fold enriched for enhanced expression in hepatocytes (OR = 2.82; FDR = 3.8 × 10−6), aligning with the liver’s role in whole-body homeostasis. This included 229 trans-pQTLs consistently enriched to likely act on protein complexes (OR = 1.83; p < 2.7 × 10−2), ligand-receptor pairs (OR = 2.73; p < 3.8 × 10−5), or pathway partners (OR = 7.66; p < 3.5 × 10−13) of associated protein targets compared with other modes of pleiotropy (Figure 4D), illustrating that part of the plasma proteome may be a surrogate of critical disease-predisposing mechanisms in tissues. For example, we identified rs10849448 as a trans-pQTL for eight protein targets, likely acting via lymphotoxin beta receptor (LTBR) (score 1.8), which encodes for LTBR, the receptor for lymphotoxin (LTA), which was among the associated protein targets. The same variant, or proxies thereof (r2 > 0.8), has been reported to associate with >20 traits in the GWAS catalog, including celiac disease, chronic obstructive pulmonary disease, primary biliary cirrhosis, and different measures of lymphocytes. Although LTBR is not expressed on lymphocytes, signaling via LTBR is essential for the development of tertiary lymphoid structures that are associated with chronic inflammatory diseases.31 Pharmacological blocking of LTBR signaling has further been shown to revert airway fibrosis in mouse models of smoking-induced chronic obstructive pulmonary disease.32 In general, genetic variants associated with multiple proteins might guide the interpretation of otherwise cryptic GWAS loci. We identified 285 pQTLs previously reported for non-proteomic traits in the GWAS catalog and among which associated proteins were significantly (FDR < 0.05) enriched for one or more pathways (Figure 4E; Table S9B). For example, we observed that two independent trans-pQTLs were 38-fold enriched (FDR < 5.5 × 10−4) for proteins associated with keratinization and had been previously linked to the risk of acne. Although one of the putative effector genes, ERRFI1, had been linked to skin morphogenesis,33 the role of the second one, SEMA4B, has so far been elusive,34 and we provide evidence for a potential function in skin morphogenesis or homeostasis. Contrasting proximal versus distal genetic regulation of plasma proteins reveals discordant phenotypic consequences Successfully integrating proximal, i.e., cis-based, genetic regulation of plasma protein levels with disease risk variants has a high biological prior that the protein is involved in disease onset. We found evidence for 300 such protein-disease pairs, combining our results with >700 diseases from the FinnGen project35 through causal inference methods for disease follow-up. However, only 73 examples were supported by both a dose-response relationship, via Mendelian randomization (MR), and statistical colocalization of genetic risk signals, emphasizing the need for complementary evidence for the genetic prioritization of candidate causal genes underlying diseases (Table S10A). For one-third of the cis-based examples (n = 115 out of 300), we identified enough protein-specific variants acting in trans to systematically test for independent support by MR. We observed supporting evidence for a third (n = 31) of cis-based protein-disease examples by demonstrating a directionally consistent association when using only trans-pQTLs for causal inference (phet > 0.01; Figure 5; Table S10B). This included proteins with known roles in blood, such as known pharmacological targets that act at the blood-tissue interface (e.g., proprotein convertase subtilisin/kexin type 9 [PCSK9] and coronary artery disease). However, we found no evidence for an enrichment of proteins specific to blood cells (e.g., lymphoid tissue: OR = 1.30; p = 0.73) or being actively secreted into blood (OR = 0.78; p = 0.56) among those with consistent effects in cis and trans. Proteins likely conveying risk within tissues included sulfotransferase family 2A member 1 (SULT2A1), linked to intrahepatic cholelithiasis of pregnancy (ICP) (Figure 5). Genetic variation near SULT2A1, which encodes a sulfotransferase with a role in the metabolism of endogenous compounds and drugs, has previously been shown to increase the risk for ICP,36 and gallstones in general,37 mainly attributed to increased hepatic expression and hence activity of the enzyme to produce supersaturated bile that promotes gallstone formation.16 We provide evidence that potential distal regulators of SULT2A1 abundance, e.g., via the trans-pQTL rs16919533, which we assigned to PANX1 with moderate confidence (gene score: 1.0), may also contribute to disease risk. However, we also observed significant differences (phet < 0.01) between results from well-powered trans-pQTL MRs not supporting (n = 41), or even opposing (n = 14), robust cis-pQTL evidence (Figure 5). For example, the well-documented effect of changes in sclerostin (SOST), a negative regulator of bone formation, on fracture risk observed with cis-pQTLs (betacis = 1.34, pcis < 5.6 × 10−19) was null (betatrans = 0.02, ptrans = 0.84) when considering 16 specific trans-pQTLs that cumulatively explained a higher amount of variation in plasma SOST levels (Figure 5). We observed a similar null effect, when further excluding bone mineral density-associated trans-pQTLs in an additional sensitivity analysis.38 In general, naive incorporation of trans-pQTLs into MR-based causal inference resulted in even larger disagreement with <20% of findings having concordant support based on cis-instruments, likely driven by pleiotropy not captured by statistical techniques (15 out of 96 for cis-/trans-pQTL-based MR and 6 out of 230 for trans-pQTL-based MR; Table S10C). The few potentially true positive findings missed by our biologically informed approach included plasma levels of von Willebrand factor and von Willebrand disease (cis/trans-pQTL MR: beta = −3.42, p < 2.9 × 10−16). Limited concordance between protein-phenotype associations identified using cis-focused genetic and observational studies Blood proteins with a causal role in disease development are attractive targets for drug development and as markers of target engagement. We therefore systematically triangulated prevalent and incident plasma protein biomarker studies in up to 52,164 UKBB participants (using logistic regression and survival analysis) with high-confidence genetic inference in a pan-biobank effort, including up to 1,296,701 participants (using cis-MR and colocalization) for 517 diseases (Table S11). We observed 193 protein-disease pairs with high-confidence genetic support, of which less than a quarter (n = 52) showed directionally concordant—and at least nominally statistically significant—support in biomarker studies (Figure 6A). A similar number of protein-disease examples (n = 59) showed statistically significant support but with opposing effect directions. The latter might be explained by compensatory mechanisms, but it exemplifies the need to contextualize purely genetically inferred “causality.” Conversely, among the 52,887 protein-disease associations passing statistical significance (p < 9.3 × 10−8) in the survival analysis, only 0.06% (n = 33) had directionally concordant, high-confidence support from genetic analysis, and another 36 had significant but directionally discordant evidence (Figure 6A; Table S11). The overlap between the two approaches did not improve, and even led to a loss in, convergent examples when accounting for potential confounders in survival analyses or considering prevalent cases. For 44 protein-disease pairs, we identified coherent evidence from genetic, prospective survival, and prevalent disease analyses, suggesting that those protein biomarkers may not only be relevant for the onset of the disease but also its persistence (Figure 6B). Among those, plasma furin levels stood out, being consistently associated with hypertension, myocardial infarction, and arterial fibrillation (Figure 6). The genetic locus encoding furin was first linked to blood pressure39 and later to cardiovascular diseases, with recent proteomic studies also confirming our results in another ancestry.40–42 Furin is an endopeptidase expressed in almost all tissues, with expression highest in the liver and with a broad substrate profile relevant for cardiovascular disease. Our results collectively point to the role of extracellular rather than intracellular furin. Almost all previously reported furin effects were described to occur in the trans-Golgi network, and extracellular furin has been mostly investigated in relation to the cleavage of pathogen products.43 The convergence of genetic and survival analysis (Figure 6B) may provide evidence for an extracellular role of furin in disease etiology, in contrast to the hypothesis that associations with the risk of hypertension and coronary artery disease are mediated via intracellular cleavage of brain natriuretic peptide (BNP)44 or modulation of the LDL-receptor and cholesterol metabolism.45 This hypothesis is supported by a lack of association between the lead cis-pQTL for furin (rs8027450-T) and circulating levels of both BNP (beta = 0.006; p = 0.2) and N-terminal prohormone of BNP (NT-proBNP, beta = 0.0058; p = 0.3). Pharmacological inhibition of furin in mouse models provided preliminary evidence for reduced atherosclerotic lesions and reduced vascular remodeling.46 trans-pQTL enrichment explains protein-disease signatures and can guide drug repurposing The low convergence of evidence for protein-disease links from observational associations versus genetic inference from cis-pQTLs might, in part, be explained by circulating proteins being merely surrogates for causal disease processes within tissues. Investigating the value of many newly identified trans-pQTLs, we identified that >90% (280 out of 307 diseases with ≥5 significantly associated proteins) of disease biomarker signatures were significantly enriched (FDR < 0.05) for proteins associated with one or more of 170 pleiotropic pQTLs (n = 139 trans-pQTL; Table S12). Triangulating the evidence from pQTLs with GWASs and biomarker analyses, we observed 58 examples where the protein targets associated with a given pleiotropic pQTL showed an enrichment for disease-associated plasma proteins, where the pQTL was also a GWAS locus for the same disease (Figure 6C). This supports the hypothesis that part of the circulating protein signature provides a readout of disease-predisposing processes within tissues.47 For example, we observed a >50-fold enrichment (OR = 56.3, FDR < 9.5 × 10−20) of proteins significantly associated with the onset of proteinuria among those proteins associated with a trans-pQTL, for which we prioritized SHROOM3 as a candidate causal gene (Figure 6C; Table S7). Knockdown of the mouse equivalent Shroom3 has been shown to induce proteinuria by affecting podocyte foot process effacement,48 impacting the cells forming part of the layer responsible for size-selective filtration of blood in the kidneys. Plasma protein signatures non-randomly linked to known disease variants may further guide drug repurposing and biomarker identification for patient selection. For example, a missense variant (rs34536443, p.P1104A) in tyrosine kinase 2 (TYK2) acted as a trans-pQTL for multiple proteins (bone marrow stromal antigen 2 [BST2], C-X-C motif chemokine 9 [CXCL9], CXCL10, CXCL11, interleukin-12 receptor subunit beta-1 [IL-12RB1], and programmed cell death protein 1 [PDCD1]) (Figure 7). Higher plasma levels of all six protein targets were associated with a higher disease risk for rheumatoid arthritis, hypothyroidism, and psoriasis. There was also strong evidence for a shared genetic signal across these proteins and disease risks (HyPrColoc posterior probability [PP] = 98.9%) at the TYK2 locus (Figures 7 and S4). Psoriasis is the approved indication for the TYK2 (encoded at TYK2) inhibitor deucravacitinib, which mitigates inflammatory burden in patients,49 supported by our genetic evidence (Figures 7 and S4). The associated circulating protein signature might therefore be indicative of early inflammatory dysregulation in patients at risk of these autoimmune diseases, and persistently elevated plasma levels among patients might help to identify those most likely to benefit from the treatment, as well as monitor treatment success, as demonstrated in a recent phase 2 trial of deucravacitinib.50 This example illustrates the potential value of trans-pQTLs to guide biomarker prioritization for diverse purposes, benefiting patients as well as informing drug repurposing opportunities. DISCUSSION Broad-coverage, high-throughput proteomic technologies applied to large-scale patient and population cohorts offer a comprehensive molecular view that can substantially deepen our understanding of human health and disease.6–24 Here, we present large-scale antibody-based proteogenomic meta-analyses consisting of 38 cohorts and up to 78,664 individuals to derive an expansive pQTL catalog of 24,738 fine-mapped pQTLs (including 5,040 in cis and 19,698 in trans). The multi-cohort design of our study provides consistent evidence across cohorts for >80% of the pQTLs identified in this study and also demonstrates that cohort and sample characteristics contribute only relatively minimally to observed heterogeneity of effects, in comparison to the overall strength of association, i.e., genetic effect size. We demonstrate the importance of distal (i.e., trans) genetic regulation for circulating protein targets by identifying pathways, tissues, and cell types with specific, but also broader, roles. By integrating effector gene assignments across protein targets, we were able to reconstruct entire pathways relevant to plasma protein abundances through a data-driven approach based on in-human evidence. N-linked glycosylation emerged as the most frequently enriched pathway, in line with its role in the secretory pathway, involving the correct folding, trafficking, and secretion of the protein into blood or other tissues.51–53 Enzymatic protein glycosylation is the most abundant posttranslational modification, and our observation of enriched members of N- but not O-linked (a hallmark of intracellular proteins) glycosylation reflects a general finding of our study: that transeffector genes are more likely to be directly linked to protein targets actively secreted into the blood. Many studies now screen for novel biomarkers or drug targets using publicly available data, either with individual-level access to resources such as UKBB or with summary statistics from large genetic consortia, the latter being pursued purely in silico with strong underlying assumptions. The discrepancies in the inference arising from both approaches may explain some of the limited visible success to date. Even the high-confidence examples identified here, such as furin’s role in diverse cardiovascular diseases, require follow-up before being evaluated as potential drug targets. We note that clinically useful protein biomarkers do not need to have genetic evidence to imply a causal role of the protein in disease development, but a clear and well-understood link to disease-predisposing mechanisms is essential for successful clinical application (e.g., neurofilament light chain used to diagnose neurodegenerative diseases). By integrating different lines of orthogonal evidence, we provide clear examples of trans-pQTLs guiding biomarker prioritization for specific diseases. As a proof of principle, our results highlight seven trans-pQTLs directly supporting the use of NT-proBNP as a biomarker for (future) heart failure, with all trans-variants being associated with both heart failure and plasma NT-proBNP levels. Similarly, we exemplified how trans-pQTLs, when integrated with prospective biomarker analysis and large-scale GWASs of different diseases, can point to protein biomarkers to support drug repurposing, such as TYK2 inhibitors for rheumatoid arthritis.54 Collectively, this provides clear conceptual advances beyond cis-focused causal inference techniques, which have predominantly been applied in proteo-genomic studies to date, highlighting how trans-pQTLs can inform protein biomarker detection for diseases and drug development as well as improve our understanding of identified susceptibility loci for these diseases. MR has become commonly used for genetically guided drug-target discovery. We observed discordant effects for well-curated trans-pQTL versus biologically plausible cis-pQTL disease associations. Although we acknowledge that limitations might remain for specific examples, our results did not reveal a systematic reason explaining the convergence or segregation of cis and trans effects on the diseases studied, and we propose a staged approach to guide the evaluation of trans-pQTLs in the context of MR studies. This includes biologically informed exclusion of instruments with evidence for pleiotropic effects on either the proteome or phenome, in addition to testing the alignment of cis- and trans-pQTL effect sizes and directions. Our study demonstrates the potential of larger multi-cohort studies in expanding our understanding of the genetic regulation of the circulating proteome and of the insights into disease mechanisms and therapeutic development that proteogenomics can deliver. Limitations of the study All proteomic technologies currently only provide a partial coverage of the proteins detectable in circulation, including their many isoforms and post-translational modifications. Similarly, we had incomplete coverage of the range of genomic variants across cohorts, as most cohorts had array-based imputed genomic data, limiting our ability to assess rare variants and utilize statistical approaches such as fine-mapping to their full extent. Our study was conducted using predominantly European samples due to measurement availability. We believe our study provides not only a valuable multi-cohort proteogenomic resource to the community but also a template for the biological and clinical insights that can be derived from such efforts, including future studies with even greater power, allele and protein coverage, and genomic diversity. RESOURCE AVAILABILITY Lead contact Further information and requests for resources should be directed to Claudia Langenberg (claudia.langenberg@qmul.ac.uk). Materials availability This study did not generate any new, unique reagents. Data and code availability Access to the UKBB genomic, proteomic, and phenotype data is open to all approved health researchers (http://www.ukbiobank.ac.uk/). This research has been conducted using the UKBB resource under application no. 44448. Access to the UKBB genomic, proteomic, and phenotype data is open to all approved health researchers (http://www.ukbiobank.ac.uk/). This research has been conducted using the UKBB resource under application no. 44448. Genome-wide summary statistics for all protein targets in this study will be available on https://omicscience.org/ upon publication. Genome-wide summary statistics for all protein targets in this study will be available on https://omicscience.org/ upon publication. All remaining data have been accessed via publicly available links provided in the key resources table. All remaining data have been accessed via publicly available links provided in the key resources table. Associated code and scripts for the analyses are available at https://github.com/comp-med/scallop-ukbb-ma. Associated code and scripts for the analyses are available at https://github.com/comp-med/scallop-ukbb-ma. STAR★METHODS EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS We performed a multi-cohort meta-analysis of genome-wide summary statistics of plasma levels of 1,194 protein targets from up to 37 cohorts (Table S1), referred to as SCALLOP meta-analyses. The participants of these cohorts were predominantly of European ancestry. Detailed information about the participant characteristics in each cohort, including mean age, mean BMI, percentage of females in the cohort and fasting status can be found on Table S1. All participants gave their informed written consent before entering the studies and each study was approved by their respective ethics committee. We additionally included proteogenomic analysis conducted using genomic and proteomic data from 48,017 participants of European ancestry in UK Biobank (UKBB). Further information about the UKBB cohort and proteomic measurements can be found elsewhere.20,55 METHOD DETAILS Proteomic measurements For the 37 cohorts included in this study, antibody-based proteomic measurements were generated through at least one of the 13 Target-96 panels offered by Olink measuring 92 protein targets (Cardiometabolic, Cardiovascular II, Cardiovascular III, Cell Regulation, Development, Immuno-oncology, Inflammation, Immune Response, Metabolism, Neurology, Neuro Exploratory, Organ Damage, Oncology II). Details regarding the assay have been described in detail elsewhere.61,62 Briefly, dual antibody based proteomic measurements are generated where each of the unique antibodies are labelled with complementary single stranded oligonucleotides, referred to as proximity extension assays (PEA).61 Hybridization occurs between the complementary oligonucleotides when both of the unique antibodies bind to their respective protein target and come into close proximity, which can subsequently be quantified through qPCR or next-generation sequencing (NGS) methods. Relative proteomic measurements are offered as normalized protein expression (NPX) units, provided on log2 scale. Each cohort has performed quality control on their proteomic measurements such as, but not limited to, removing samples that were extreme outliers using principal-component analysis from their entire proteomic profiles. Only blood-based (i.e. plasma or serum) proteomic measurements were included in this study. In UKBB, the proteomic measurements used in this study were generated through Olink Explore 1536 platform utilizing the same antibody-based technology but measuring a broader coverage of 1,463 protein targets. Details of proteomic measurements can be found elsewhere.20 Genome-wide meta-analyses of protein levels For 1,194 unique protein assays for each protein target from all 37 cohorts, we performed additional quality control measures and retained only biallelic variants with (i) a call rate above 95%, (ii) Hardy-Weinberg p-value above 1×10−5 , (iii) imputation INFO score above 0.8, (iv) standard error of SNP-effect less than 10, (v) minor allele count (MAC) above 3 and (vi) minor allele frequency above 0.1%. We filtered out any variant which did not have an existing rsID in the dbSNP database.63 Using all available summary statistics per protein target, we performed inverse-variance fixed-effects meta-analyses using METAL (v.2011–03-25),59 also referred to as SCALLOP meta-analyses. The sample sizes varied from 2,297 to 31,190 depending on protein target coverage across cohorts. In parallel, we have also conducted genome-wide association analyses for proteomic measurements of 1,463 protein targets in UK Biobank participants through REGENIE v.3.1.4.58 Finally, we performed an inverse-variance fixed-effects meta-analyses (with METAL59) for 1,161 protein targets (with overlapping UniProt ID across different antibody-based platforms offered by Olink) using summary level data from SCALLOP meta-analyses and UKB, reaching to a total sample size of up to 78,664 participants (n=17,602 – 78,864). Downstream analyses Following regional clumping for independent signal selection, we applied Bayesian fine-mapping to identify independent protein quantitative trait loci (pQTLs) using SuSie.27 We characterized variance explained by significant pQTLs and heritability by polygenic background (excluding any significant loci for a given protein) through LD-score regression implemented in LDSC.64 Leveraging the multi-cohort study design, we assessed the confidence for of each pQTL by their (i) statistical significance, (ii) directional concordance across cohorts and (iii) observed heterogeneity. We then tested the contribution of participant and cohort-level characteristics to the heterogeneity observed through a meta-regression model. We used zero-inflated Poisson regression models to test for associations between the presence and number of associated cis-/trans-pQTLs per protein and various protein characteristics. Using the output from machine learning models we trained for trans-pQTL effector gene assignment, we tested enrichment of the assigned effector genes among pathways, tissues, cell-types. Based on assessing the overlap of pQTLs or proxies (r2 >0.8) with phenome-wide associations from GWAS catalogue,30 we defined three categories of pleiotropy, (i) ‘molecular pleiotropy’ (associated with >5 proteins and ≤ 5 non-proteomic phenotypes), (ii) ‘phenotypic pleiotropy’ (associated with >5 non-proteomic phenotypes and ≤ 5 proteins), and (iii) ‘unspescific pleitropy’ (associated with >5 proteins and >5 non-proteomic phenotypes). For phenotypic follow-up, we first assessed the link between cis-pQTLs with 835 diseases from the FinnGen35 release 8 using fine-mapping augmented colocalization65 and two sample MR analyses.66 We further systematically tested for the relevance of transpQTLs by adopting the cis-MR workflow and testing for (i) cis and trans-pQTLs and (ii) trans-pQTLs only as genetic instruments for each protein target. To test for concordance between genetically inferred protein – disease relationships and plasma protein levels and disease onset and/or presence, we first run Cox-proportional hazard models and logistics regression models, respectively, adjusting for age, sex, and technical variables. We then integrated cis-based two-sample MR analyses from an internal pan-biobank project covering over one million individuals (MVP,67 the Pan-UK Biobank,68 and FinnGen35) and tested for concordance between the genetic and observational approaches using the relevant multiple-testing corrected significance thresholds. Finally, we tested whether the protein association profile of pQTLs that were reported for at least one non-proteomic trait in the GWAS Catalog were significantly enriched for any pathway and whether there was a non-random overlap between protein-biomarker signatures from prospective biomarker analyses and protein signatures associated with pQTLs using Fisher’s exact test. Detailed methods can be found on quantification and statistical analysis section. QUANTIFICATION AND STATISTICAL ANALYSIS Genome-wide meta-analyses of protein levels We did not observe evidence for genomic inflation for (i) SCALLOP only meta-analysis (mean=1.04, IQR=1.04–1.05) or UKBB proteogenomic analyses (mean=1.09, IQR=1.07–1.10). Using the summary statistics from SCALLOP meta-analyses and UKBB proteogenomic analysis, we performed inverse-variance fixed-effects meta-analyses using METAL (v.2011–03-25)59 for 1,161 protein targets, mapping to the same UniProt ID. We also did not see evidence for genomic inflation across the overall meta-analyses (mean=1.08, IQR=1.06–1.09). Genomic build GRCh37 was used throughout this study. Identification of regional sentinel variants Regional sentinel variant selection was performed by selecting the variant with the highest z-score within each 1 Mb window around each significant signal. The genomic regions were merged and considered as a single region if there were multiple significant variants within less than 500 kb from each other. Bonferroni corrected genome wide significance thresholds were used for each sub-study as follows, p<4.19×10−11 for SCALLOP meta-analyses (1,194 protein targets), p<3.42×10−11 for UKBB proteogenomic analyses (n=1,463 protein targets) and p<4.31×10−11 for the overall meta-analyses (n=1,161 overlapping protein targets). pQTLs were defined as cis-acting if they were located within 500 kb of a protein’s cognate gene and were defined as trans-acting otherwise. Where assays targeted more than one protein, cis-pQTLs were defined as those located within 500 kb of any potential cognate genes. Fine mapping We performed statistical fine-mapping on protein GWASs performed in the UKBB, as the largest single study, using the ‘sum of single effects’ model (SuSiE).27 Briefly, SuSiE implements variable selection under a Bayesian framework to identify credible sets of independent variants that likely contain the true underlying causal variant. We conducted these analyses using the R package susieR (v.0.12.16) with the default parameters and priors. We used a random subset of 20,000 unrelated participants of European ancestry from the UKBB as an LD reference. Additionally, we adopted a grid search approach that iterated the maximum number of credible sets from 2 to 10, selecting the number that ensured no variants within a credible set were in LD (r2> 0.1). Credible sets were taken forward where variants additionally met genome-wide significance. We have included regional summary statistics from the overall meta-analysis for any region where it was not possible to perform SuSie on UKBB summary statistics or regions on X-chromosome. We also clumped the independent credible sets into r2 groups across all protein targets based on linkage disequilibrium (r2>0.8). Concordance of the identified pQTLs Variants were characterized as ‘high confidence’ if the pQTL was (i) at least genome-wide significant (p<5×10−8), (ii) showed directional consistency in the overall meta-analyses between SCALLOP cohorts and UKBB for cis- and trans-pQTLs and (iii) additionally did not show substantial evidence of heterogeneity (phet >1×10−4) for trans-pQTLs. We tested for independent replication of identified lead variants from credible sets by investigating for the overlap for the pQTLs or any of their proxies (r2> 0.1) in previously published affinity-based proteogenomic studies, based on a common UniProt ID for the protein targets.6–24 We tested the correlation between the effect size for regional sentinel variants identified in this study with cohorts of predominantly European ancestry with effect size of the pQTL in cohorts of non-European ancestry, namely Immuno-oncology panel in China Kadoorie Biobank (CKB, n=816 participants of Chinese ancestry), Metabolism panel in Qatar Metabolomics study on Diabetes (QMDIAB, n=350 participants of Indian, Filipino or Arabic ancestry) and Inflammation, Cardiovascular III, Neurology and Oncology II panels in Shanghai Women and Men’s Health Study (SWMHS, n=548 participants of Chinese ancestry) (Figure S2). Variance explained by identified pQTLs We estimated the variance explained by the regional sentinel variants and heritability estimates by the polygenic background, excluding the regions harbouring any pQTLs18 for all 1,161 protein targets included in this study. In summary, we calculated variance explained for cis and trans loci by using the formula (2 × β2 × f × (1−f)), where f is the MAF. We calculated heritability estimates by the polygenic background, by excluding the regions harbouring a pQTL, through LD-score regression (LDSC).64 pQTLs and protein-related characteristics We tested the association of presence and also the number of high confidence cis- or trans- pQTLs with a wide range of protein-related characteristics, namely whether a protein (a) has sites for sulfation, glycosylation, phosphorylation, ubiquitination, s-nitrosylation, acetylation, palmitoylation, acetylglucosmination, glycosaminoglycan-chains, myristoylation, acylation and methylation, (b) has any disulfide bonds, DNA binding domains, alpha-helix domains, turn domains, transmembrane domains, zinc finger domains, coiled coil domains, beta-strand domains, or a protein’s (c) transcript count as obtained from UniProt.57 We also tested the association of the presence and also the number of high-confidence cis- or trans-pQTLs with the probability of loss of function intolerance (pLI), missense Z-score (i.e. the deviation of the observed from the expected number of missense variants), protein length and gene length as obtained from gnomAD v2.1.56 We used ‘pscl’ package in R v.4.2.2 which performs zero-inflated Poisson regression characteristics by including scaled sample sizes as a covariate. In summary, the package runs two set of tests for each characteristic which are (i) Poisson regression for the count data to test the association between the number of high-confidence cis or trans pQTLs and protein-related characteristics, (ii) logistic regression to test the association between the existence of high confidence cis or trans pQTLs and protein-related characteristics. The continuous traits were normalized to have a comparable effect size estimate with the binary ones. We applied a Bonferroni-corrected significance threshold of p<2×10−3 (corrected for the number of investigated characteristics, n=25). We also ran sensitivity analyses including missingness rate for proteins as a covariate in the model. Impact of cohort characteristics on heterogeneity To better understand cohort characteristics that might contribute to the heterogeneity, we performed meta-regression analyses in R v4.2.2 using ‘metafor’ package. The loci where we observed high levels of heterogeneity (heterogeneity p-value <1×10−4) across SCALLOP cohorts from the fixed-effects meta-analyses were taken forward and a meta-regression model was fit with between the effect size of these loci from each cohort and cohort characteristics and the overall genetic effect size of the loci (from the meta-analysis) as explanatory variables. The cohort characteristics that were included were mean cohort age, sex composition of the cohort (i.e. %females), mean BMI of the cohort, blood-based tissue type (EDTA plasma, citrate plasma, or serum), fasting status and whether any disease cases were included in the cohort (Table S1). Characterization of variant consequences We used Variant Effect Predictor (VEP, version 99) through Ensembl to characterize the consequence of the fine-mapped variants and their proxies.60 We used Graphql to query the Opentargets platform69 to obtain the closest gene for each unique fine-mapped pQTL variant. Colocalization with gene expression levels We systematically tested for a shared genetic signal between plasma abundance of a protein and gene expression levels (eQTL) of the protein coding gene in 49 tissues from the GTEx project (v8)25 assessed by posterior probability from statistical colocalization analysis using ‘coloc’ package in R 3.6.70 The build of our study was lifted over from b37 to b38 using LiftOver,71 to be compatible for the analyses. If any of the pQTLs did not have a corresponding genomic location in b38, they were excluded from the downstream colocalization analysis. GTEx25 variant-gene cis-eQTL associations from each tissue were downloaded from https://console.cloud.google.com/storage/browser/gtex-resources on January 2020. For the cis-variants, the region was defined as (±500 kb) around the protein-encoding gene of each protein-encoding gene with at least one cis-pQTL that were present in both datasets. For trans-pQTL - cis-eQTL colocalization analyses, the region was defined as the ±500 kb window around the trans-pQTL signal, regardless of the genes that reside in the region. The colocalization analysis was only performed if the region had at least suggestive evidence of association (p<10−6) in the defined region in any of the tissues and the lead variant was in high LD with the lead cis-pQTL in the region (r2 >0.6). A prior probability of a shared signal (p12) was defined as 1×10−5 and posterior probability above 80% was used to define a high likelihood of a shared genetic signal. Machine learning-based effector gene assignment We implemented a staged machine learning classifier to systematically assign candidate effector genes for all trans-pQTLs residing outside of the MHC region. We first collated comprehensive annotations for each genetic variant or proxies thereof (r2 >0.6), including 1) distance to the gene body in 1Mb window, and 2) putative functional consequences based on the variant effector prediction (VEP) tool. We further systematically collated for each gene within a 1Mb window: 1) evidence for colocalization between gene and protein expression levels based GTEx version 8,25 2) evidence for a rare gene burden association based on Dhindsa et al. (2023),22 3) evidence whether any trans gene encodes for receptor/ligand or protein complex of the cis-protein based on literature evidence using the OmnipathR package v3.10.1,72 and 4) whether any of the genes participate in the same biological pathway based on KEGG73 and REACTOME74 annotations. In the absence of generalisable gold-standard variant to gene assignments, we leveraged prior biological and genomic knowledge to derive three partly distinct sets of ‘putative true positive’ (PTP) sets: 1) Trans genes that encode ligand – receptor pairs of those with high-confidence evidence for forming a protein complex with the cis-protein (n=540 PTPs), 2) sentinel trans-pQTLs mapping to functional variants (n=1,747 PTPs), and 3) significant gene burden results for trans genes (n=1,049 PTPs). For each of the three PTP sets, we considered all other genes within a 1Mb window in the locus as negative examples. We pruned each set to contain at each locus only one cis-protein to avoid artificially good results due to pleiotropy. We then split each of the data sets 10 times in a 7:3 ratio to obtain training and test sets separating by genomic region. For each of the ten training sets, we trained a Random Forest classifier using repeated 3-fold cross-validation implementing subsampling to account for the unbalanced data sets. This was implemented using the R caret v6.0.94 package. We used the Kappa score to select the best performing forest within each training set. Eventually, we used each of the ten Random Forest classifiers from each PTP set to assign candidate scores for all putative effector genes across the entire set of trans-pQTLs. We thereby took the median score across all ten classifiers for a PTP set and finally summed the score across all three predictions. For each of the PTP sets, we omitted features used to define true positive sets. All three classifiers showed robust performance with median Kappa values of 0.54 – 0.57. The conception of our machine learning classifier was inspired by the ProGeM framework75 and with a rationale of integrating biologically relevant annotations such as ligand – receptor pairs or protein complexes. Pathway, tissue, and cell-type enrichment We obtained processed data for tissue and single cell expression from the Human Protein Atlas,76 and used the annotations ‘specific’ and ‘enhanced’ to define a common set of genes for each tissue/cell-type with enhanced expression. We used Fisher’s exact tests to test for a significant enrichment of assigned effector genes among enhanced gene expression sets. To avoid redundancies, we only took the effector gene with the highest score at each locus to test for enrichments. If required, we restricted the background of those tests to proteins covered on the Olink panels and computed the false discovery rate to account for multiple testing. We report only enriched, but not depleted tissues/cell-types, although those have been considered when performing multiple testing correction. For pathway enrichment, we used the R package gprofiler2 v0.2.2 restricting KEGG73 and REACTOME74 as data bases and FDR-correction. Phenotypic follow up We used two partially complementary approaches to systematically link cis-pQTLs to 835 diseases from the FinnGen release 8.35 Firstly, we used fine-mapped protein GWAS summary statistics to colocalize credible sets in ±500kB windows around protein encoding genes, as implemented in the R package coloc with a recent augmentation to relax the single variant assumption using the SuSiE method.65 Colocalization was only performed if there was at least a suggestive signal for a given region for the trait being tested. We used default priors apart from implying a more stringent prior to declare a shared genetic signal (p12 =5×10−6). To fine-map FinnGen outcomes, we used at most 5 credible sets and a random set of 30,000 UKBB participants to estimate an LD-backbone. We considered evidence of a shared signal between a protein and a FinnGen outcome, if at least one credible set was shared among both traits (PP H4 > 80%) provided the lead cis-pQTL was also in LD with the fine-mapped lead signal for the FinnGen trait. We used simple Wald ratios to assess effect directions, i.e., to understand whether an increase in protein levels is associated with an increase/decrease in disease risk. Secondly, we performed standard cis-based MR analysis,66 by aggregating Wald ratios across all independently selected, genome wide significant cis-pQTLs and any of the FinnGen outcomes using the inverse variance weighted method. We required overlap of at least two-thirds of identified cis-pQTLs to avoid artificial findings. Apart from correcting for multiple testing using Bonferroni correction, we further required consistent effect directions in sensitivity analysis, including weighted median, mode, and MR-Egger regression, as well as no strong evidence for heterogeneity (p>10−3) or pleiotropy (p>10−3). We further systematically tested for the relevance of trans-pQTLs by adapting the cis-MR workflow as follows. We considered all high-confidence trans-pQTLs associated with a protein target that were associated with less than five protein targets or phenotypes in the GWAS Catalog,30 i.e. specific trans-pQTLs. We first identified instruments strongly violating MR assumptions by performing leave-one-out analysis tracking Cochran’s Q statistic and omitted any instrument accounting for strong (median + 3 X SD) deviations compared to all other combinations of instruments - a procedure similar to recently proposed, but computationally more expensive, methods.77 We also tested combining cis- and trans-pQTLs. Similar to the cis-MR, we applied the Bonferroni correction to account for multiple testing and implemented several sensitivity analyses including performing MR-Egger to assess pleiotropy, assessing directionally concordance the across different MR methods and assessing the heterogeneity among the pQTLs. We further demonstrated the need to account for pleiotropy via biological insights by running MR analyses without pleiotropy filters for trans-pQTLs. Cis-based pan-biobank phenotypic follow up To harmonize phenotypes between MVP,67 the Pan-UK Biobank,68 and FinnGen35 (version 10), disease-based traits were mapped using codes provided by the biobanks. MVP used entirely phecodes (n=1,171), while the UKBB used phecodes (n=1,327) as well as ICD10 codes (n=915) to label their disease phenotypes. After restricting studies in the UKBB to those with European ancestry in their list of populations, all possible direct phecode to phecode (n=1,013) matches were made between the biobanks. UKBB traits with an ICD10 code were then mapped to phecodes using a conversion table (70), and if the derived phecode had not already been mapped to MVP in the previous step, then a match was made if possible (n = 10). For MVP phecodes that remained unmapped to UKBB, studies with case counts higher than 4,000 were reviewed and a decision was made on a case-by-case basis to map manually (n = 53) by a clinician before aligning with FinnGen. For all unmapped MVP phecodes with case counts lower than 4,000, MVP only data was used for analysis. To map MVP and UKBB phenotypes to FinnGen, we manually mapped all available R10 FinnGen phenotypes to those previously available from MVP or UKBB. All matches were double checked through clinical adjudication and comparison statistics using the number of cases per resource for each phenotype to identify significant outliers from this mapping strategy. Situations where a significant deviation from the overall relationship between MVP/UKBB/FinnGen case counts was observed were analyzed on a one-by-one case and a final harmonization decision was made. This led to FinnGen being mapped only with MVP (32 instances), only with UKBB (174 instances), and with both MVP and UKBB (501 instances). Pan-UK Biobank GWAS physical positions were converted from genome build GRCh37 to GRCh38 using LiftOver.71 Using METAL,59 we performed fixed effects inverse-variance weighted meta-analyses of MVP European results, UKBB European results, and FinnGen and obtained estimates of heterogeneity (option in METAL: ANALYZE HETEROGENEITY). If mapping was not possible between MVP, UKBB, or FinnGen, then only MVP results or select UKBB phenotypes were retained for further analyses. For very small p-values in MVP which were set to zero, we set to the lowest possible decimal place allowed in Python using sys.float_info.min() (2.2250738585072014e-308). To test for inflation of p-values following meta-analyses, we calculated the lambda values for all meta-analyses results. Any phenotype having a lambda value greater than 1.15 (n = 289) was rerun including the genomic control parameter in METAL. Phenotypes that were not meta-analyzed were also tested for inflated p-values. We found 81 phenotypes that needed correction, the vast majority being clinical biomarkers (75 studies). Due to highly inflated p-values, we removed eight height-based phenotypes from consideration. Two-sample Mendelian Randomization (MR) of each of the protein-coding genes were performed against all phenotypes using instruments from SCALLOP cis-pQTLs. In order to determine the correct ordering of alleles between the datasets we utilized the harmonise_data() function from the TwoSampleMR78 package in R. We used the Wald Ratio for instruments with one genetic variant and inverse variance weighted MR for instruments with multiple genetic variants. We additionally performed MR-Egger for proteins with three or more instruments to be used as a sensitivity analysis. We tested for heterogeneity across variant-level MR estimates, using the Cochrane Q method (mr_heterogeneity option in TwoSampleMR package) and the MR-Egger intercept. Comparison of observational and genetic analyses To test for concordance between genetically inferred protein – disease relationships and plasma protein levels and disease onset, we performed two analysis streams. We collated electronic health records, such as primary (45% of the population) and secondary care, death certificates, or cancer registry, but also self-reported information into 1,448 medical ontology terms called ‘phecodes’ (which we refer to as ‘diseases’ for simplicity) for all UKBB participants.79 We kept track of the first occurrence of each disease code in any of the resources and classified participants that had any code for a given disease prior to inclusion into UKBB or reported it at recruitment as ‘prevalent’, and any disease occurrence thereafter as ‘incident’. We then associated plasma protein levels with the risk of disease onset using Cox-proportional hazard models with age as the underlying timescale while adjusting for sex of the participant and technical variables, such as sample age and fasting time. We monitored the proportional hazard assumption using Schoenefeld residuals and applied different levels of multiple testing correction to declare significance (see below). For each disease outcome, we excluded participants with codes indicating prevalent diseases. We only tested diseases with at least 10 cases (n=765) among all 52,169 participants with overlapping proteomic and disease data. Secondly, we integrated the genetic analyses from pan-biobank genetic analyses as described above and restricted those protein–disease associations available in both pan-biobank genetic analyses and prospective analyses. For genetic discovery, we defined high confidence protein-disease associations as those passing correction for multiple testing across protein-disease pairs and with support from colocalization (PP4 > 0.8). We then sought support from prospective analyses for each high confidence protein-disease pair at p < 0.05. For prospective discovery, after correction for multiple testing, we investigated whether there was genetic support for each protein-disease pair at the maximum p-value where colocalization was performed (5×10−6) and PP4 > 0.8. Enrichment analysis anchored on pQTLs We performed two different sets of enrichment analysis utilizing pQTLs. Firstly, for pQTLs reported for at least one non-proteomic trait in the GWAS Catalog,30 we tested whether significantly associated protein targets were enriched for pathways using the same settings as described above using the Gprofiler software but setting the background for enrichment testing to the proteins captured in our study. Secondly, we tested for a non-random overlap between protein biomarker signatures from prospective biomarker analyses and protein signatures associated with pQTLs using Fisher’s exact test. We used the proteins included in all analyses as a background and proteins that associated with disease onset and the trans-pQTL as foreground. We tested only for enrichments if at least five proteins were associated with the trans-pQTL. We linked enrichment analysis to prospective biomarker analysis further by mapping GWAS Catalog entries to phecodes to assign convergence. Supplementary Material SUPPLEMENTAL INFORMATION Supplemental information can be found online at https://doi.org/10.1016/j.cell.2026.03.049.