Integrative analysis prioritizes proteins associated with renal cell carcinoma and its risk factors

Abstract

Background

Renal cell carcinoma (RCC), the predominant form of kidney cancer, is influenced by several risk factors, including obesity, hypertension, and smoking. The molecular mechanisms linking these risk factors to RCC, however, remain unclear.

Methods

We investigated plasma proteins as potential intermediates of the effects of risk factors on RCC using 2-step mendelian randomization. In step 1, we identified plasma proteins influenced by each risk factor, using genome-wide association study summary statistics from the UK Biobank Proteomics (n = 34 557) for plasma proteins and 19 risk factors encompassing anthropometric traits, blood pressure, smoking behavior, blood cell counts, and kidney function. In step 2, we evaluated the effects of these risk factor–associated plasma proteins on RCC, using the largest-to-date RCC genome-wide association study (29 020 cases). Proteins strongly associated both with risk factors and with RCC were further prioritized through convergent evidence from multiple downstream analyses.

Results

Among 2940 plasma proteins, 2339 were statistically significantly associated (P < 1.7×10^-05) with at least 1 of the 19 risk factors. Of these 2339 proteins, 33 showed a statistically significant effect on RCC (false discovery rate < 5%), with 28 mapping outside RCC genome-wide association study loci. Using multivariable mendelian randomized, we estimated the mediation effects of associated plasma proteins, finding that proteins such as APOL1 mediated 2.76% of white blood cells’ effect. Convergent evidence from multiple downstream analyses prioritized TYMP, UMOD, and USP28 as key protein intermediaries of the risk factor effects on RCC. TYMP and USP28, both inversely associated with RCC risk, showed immune-related and tumor-suppressive effects, while UMOD was positively associated, potentially linking kidney biomarkers to carcinogenesis.

Conclusions

Our analysis highlights plasma proteins as intermediates of the effect of distinct risk factors on RCC, thereby nominating molecular targets for further investigation.

Introduction

Kidney cancer poses a growing global health concern, with the majority of cases being renal cell carcinoma (RCC), which accounts for more than 90% of all kidney cancers. According to the European Renal Association, 434 419 new cases of kidney cancer were diagnosed in 2022 (and 155 702 related deaths),¹ with this number projected to rise to 475 400 diagnosed cases by 2030.2 With a growing population, an aging demographic, and projected increases in incidence, the rising burden underscores a pressing need to better understand RCC pathogenesis, identify early detection markers, and explore actionable therapeutic targets for this malignancy.

Epidemiological studies have established that obesity, smoking, and hypertension are associated with increased RCC risk, collectively accounting for approximately half of diagnosed cases.^3-7 In addition, several other suspected risk factors have been reported in some studies, such as alcohol consumption (associated with lower risk), blood cell counts, and markers of impaired kidney function and kidney disease, although these findings have not been consistently replicated.^8-11

The biological mechanisms that may underlie risk factor effects on RCC development and progression (assuming risk factor–RCC causality) remain poorly understood. With the recent advent of molecular studies, opportunities have emerged to integrate data on intermediate phenotypes, such as gene expression, protein levels, and metabolites,^12-15 to obtain a better etiological understanding of potential molecular intermediaries that lead to cascading risk factor effects on RCC. In particular, plasma proteins have emerged as valuable candidates in this context.^16,17 The plasma proteome encompasses thousands of proteins involved in diverse pathophysiological processes that could be targeted for drug development, including metabolism, vascular regulation, immune response, and cell signaling.^18,19 Several studies investigating blood proteins have successfully identified etiological and prognostic markers for various cancers—for instance, IGF-I in breast and prostate cancers and microseminoprotein-β in prostate cancer.^20-22

Few large studies linking risk factors with proteomic intermediaries have been conducted, however, due to the limited number of proteomic studies with detailed clinical and risk factor information available, necessitating an integrative, multistep analysis approach. Furthermore, due to the observational nature of epidemiological studies, it is difficult to pinpoint which factors directly influence risk of disease and which are merely correlated with the underlying causal factor. A potentially useful tool for addressing this question is mendelian randomization, which uses germline genetic variants as instrumental variables (or proxies) for exposures and subsequently tests their association with an outcome.^23-25 Compared with traditional observational epidemiological studies, mendelian randomization has the conceptual advantage of being less susceptible both to confounding and to reverse causation and can thus provide unbiased estimates of potential causal effects of exposures on outcomes.²⁴

Recent advancements in multiplex proteomics have enabled high-throughput profiling of thousands of circulating proteins in larger numbers of samples, along with the discovery of genetic variants that influence plasma protein levels on a proteome-wide scale.²⁶ These variants, referred to as protein quantitative trait loci, can be used as genetic instruments in mendelian randomization to estimate the effects of plasma protein levels on disease, including cancers. Thus, a proteogenetic mendelian randomization analysis can potentially prioritize critical proteomic targets associated with disease—in particular, RCC.^27-31

In this study, we used 2-step mendelian randomization to identify the role of plasma proteins and explain the potential biological networks through which RCC risk factors affect RCC development. In step 1, we estimated the effects of 19 established and suspected RCC-related risk factors on plasma proteins; in step 2, we estimated the effects of risk factor–associated plasma proteins on RCC. Taken together, this process can identify proteins through which the effects of risk factors form a cascade on RCC. Subsequently, we prioritized RCC-associated proteins using convergent evidence from multiple downstream analyses, arriving at 3 prioritized proteomic intermediaries. We further evaluated the relevance of each of these prioritized proteins in RCC etiology using in silico bioinformatic analysis of single-cell data, incidence rates, epigenomics, and transcription factor binding. Through this integrative proteogenetic approach, we aimed to highlight novel biomarkers for and generate mechanistic insights into how risk factors impact RCC etiology.

Methods

We employed a 2-step summary data–based mendelian randomization approach to identify proteomic intermediates of the effects of established and suspected risk factors for RCC. For mendelian randomization, we used the Genome‐wide Complex Trait Analysis–Generalized Summary Mendelian Randomization (GCTA‐GSMR),^32,33 which can use publicly available genome-wide association study (GWAS) summary statistics for exposures and the outcome of interest while accounting for mild linkage disequilibrium between genetic instruments.

This study consisted of 2 components (Figure 1): a 2-step mendelian randomization analysis followed by the prioritization of proteomic intermediates of using downstream in silico analyses. The steps were employed as follows:

Figure 1.

To identify proteomic intermediaries of the effects of risk factors on renal cell carcinoma (RCC), we employed 2-step mendelian randomization. In step 1, we assessed the impact of risk factors on 2940 plasma proteins. In step 2, we evaluated the effects of the proteins associated with (ie, affected by) risk factors on RCC. We subsequently prioritized proteins using additional analysis, including cis–mendelian randomization, co-localization, and differential expression analysis. Further, we performed multiple follow-up analyses of prioritized proteins to evaluate their relevance to RCC. Abbreviations: Cis-MR = Mendelian randomization based on genetic instruments in the cis-region; PheWAS = Phenome-wide association study; scRNA-seq = single-cell RNA-seq; TSMR = Two-step mendelian randomization.

1. Step 1: A 2-sample mendelian randomization to evaluate the causal effects of risk factors on the level of circulating plasma proteins

2. Step 2: A 2-sample mendelian randomization to assess the causal effects of risk factor–driven proteins on RCC

3. Step 3: Prioritizing proteomic intermediates based on in silico analysis, including cis–mendelian randomization, co-localization, and differential gene expression and conducting sensitivity analyses to assess the robustness of 2-step mendelian randomization analysis

Step 1: effect of risk factors on plasma proteins

In the first step, we used summary statistics–based 2-sample mendelian randomization to estimate the effects of established and suspected risk factors for RCC on circulating plasma protein levels. For this step, we selected 19 risk factors, including

• anthropometric measurements (body mass index [BMI], body fat percentage, whole-body fat mass, and waist to hip ratio);

• blood pressure metrics (diastolic blood pressure [BP], systolic BP, and pulse pressure);

• smoking-related factors (smoking initiation, smoking cessation, number of cigarettes per day, number of drinks per week, and age at start of smoking);

• blood cell counts (white blood cells [WBCs], red blood cells, hemoglobin, and hematocrit); and

• kidney-related biomarkers (estimated glomerular filtration rate [eGFR], cystatin-based eGFR, and serum urea nitrogen).

The details of mendelian randomization process are provided later in this section. Briefly, we identified genetic instruments for risk factors through linkage disequilibrium clumping, removed potentially pleiotropic instruments, and used the final list of genetic instruments to estimate the effect of the risk factor on a plasma protein using an inverse variance-weighted mendelian randomization estimator. We further account for mild linkage disequilibrium (r² < 0.1) between the variants beyond linkage disequilibrium clumping, estimated from an appropriate ancestry-matched reference panel. We used publicly available GWAS summary statistics for each of the 19 risk factors, with sample sizes exceeding 290 000 participants for each risk factor. The summary-level data for the outcome, which consisted of 2940 plasma proteins, were obtained from the discovery stage of the UK Biobank Pharma Proteomics Project (UKB-PPP), which was made up of 34 557 participants of European ancestry.¹⁶ Plasma proteins were declared to be associated with a risk factor for P < 1.7 × 10^‒5, reflecting a Bonferroni correction for 2940 proteins.

Step 2: effect of risk factor–associated plasma proteins on RCC

In the second step, we estimated the effects of risk factor–associated plasma proteins on RCC, again using summary statistics-based 2-sample mendelian randomization.^32,34 For this analysis, plasma proteins were considered the exposure, while RCC was treated as the outcome. For each risk factor–associated plasma protein, we identified the genetic instruments substantially and independently associated with levels of the plasma protein, and then assessed them for association with RCC. We used summary statistics from the largest-to-date GWAS on RCC, which consisted of 864 690 participants (29 020 cases) by Purdue et al.³⁵ Associations were considered statistically significant if the false discovery rate was below 0.05.

Step 3: prioritize proteomic intermediates

Proteins identified as being associated with at least 1 risk factor and RCC were subjected to further downstream analysis as follows (details are provided later in this section). First, we restricted mendelian randomization to the cis-region (±1 megabase pair from the transcription start site) of the protein. This substep can eliminate potential pleiotropic biases and provide a robust inference for the effect of the protein on RCC by relying solely on the cis-region, potentially reflecting direct regulatory associations between the genetic instruments and plasma proteins. Next, we estimated the posterior probability of the hypothesis that the plasma protein and RCC share the same causal variant in the cis-region of the protein, providing additional evidence of potential shared associations between plasma proteins and RCC. Then, we looked for differential expression in The Cancer Genome Atlas (TCGA) Kidney Renal Clear Cell Carcinoma Collection (KIRC) to evaluate whether the gene expression varied substantially between tumor and normal tissue. Collectively, these analyses reflect additional evidence in terms of the robustness of the mendelian randomization estimate, sharing potentially causal genetic variants and signatures in tumors compared with normal tissue. Proteins associated with risk factors and RCC were prioritized based on the convergence of findings from 3 of these analyses. Finally, we further investigated the relevance of the prioritized proteins in RCC using publicly available molecular and tissue profiling (see ”The 2-Step Mendelian Randomization Analysis”).

The 2-step mendelian randomization analysis

The UKB-PPP

Proteogenetic analysis in 2-step mendelian randomization steps 1 and 2 was based on GWAS summary statistics for 2940 plasma proteins (see “Data Availability” for web resources) profiled using the Olink platform in 34 557 individuals with genetic ancestry similar to Europeans.¹⁶

Variant exclusion criteria

Variants were harmonized across exposure, mediator, and outcome GWAS by aligning effect alleles and removing variants with missing alleles, missing effect estimates, duplicated variant identifiers, or incompatible alleles. Ambiguous palindromic variants were excluded when a strand could not be resolved using allele frequency. Variants with a minor allele frequency of 0.01 or less were excluded to ensure robustness. Variants were also removed if they showed a large allele frequency discrepancy between GWAS summary statistics and the linkage disequilibrium reference panel. A large allele frequency discrepancy was defined as an absolute effect allele frequency difference greater than 0.20.

Genetic instrument selection and harmonization

After variant quality control, genetic instruments were selected from GWAS summary statistics (for risk factor and plasma protein) using prespecified significance and linkage disequilibrium filtering thresholds. For risk factor–plasma protein analyses, variants associated with each risk factor were selected at P < 5 × 10⁻⁸ and clumped using r² < 0.05. For plasma protein to RCC analyses, protein instruments were selected from the UKB-PPP protein quantitative trait loci summary statistics. Cis–protein quantitative trait loci were defined as variants located within ±1  megabase pair of the transcription start site of the corresponding protein-coding gene, and instruments were selected using the same specifications as above.

Linkage disequilibrium reference panel

Throughout the analysis, to reduce bias due to population misspecification, we used 1000 genome-unrelated Europeans (n = 498) as a linkage disequilibrium reference panel.³⁶

Covariates in the source GWAS

Mendelian randomization analyses used summary statistics from previously published or publicly available GWASs. Covariates were therefore those included in the source GWAS. The UKB-PPP protein GWAS adjusted for demographics; genetic ancestry; and technical covariates, including age, sex, genetic principal components, sample processing factors, assay batch, and plate effects. The RCC GWAS adjusted for sex and ancestry-based principal components.

Mendelian randomization assumptions and sensitivity analyses

Mendelian randomization relies on 3 core assumptions: (1) Genetic instruments are associated with the exposure, (2) instruments are independent of confounders of the exposure-outcome relationship, and (3) instruments influence the outcome only through the exposure. In our 2-step mendelian randomization framework, these assumptions applied to the risk factor–plasma protein and plasma protein–RCC steps. At each of steps 1 and 2, we performed mendelian randomization using GCTA-GSMR (see “Data Availability”), which can efficiently handle summary-level data from GWASs and uses the inverse variance–weighted method to estimate the effect of an exposure on the outcome. Using GWAS summary statistics from both the exposure and the outcome, GCTA-GSMR identifies genetic instruments via clumping, removes potentially pleiotropic single-nucleotide variations (SNVs; formerly single-nucleotide polymorphisms) through heterogeneity in dependent instruments (see below) and uses the final list of genetic instruments to perform mendelian randomization. The GCTA-GSMR further accounts for mild linkage disequilibrium (r² < 0.05 in this analysis) between the variants beyond linkage disequilibrium clumping, estimated from an appropriate ancestry-matched reference panel (1000 genome-unrelated Europeans here).^32-34

Sensitivity analyses

We performed additional sensitivity analyses to evaluate the robustness of the GCTA-GSMR estimates, assess potential pleiotropy, and identify any outlying effects of individual instruments. To investigate the robustness, we compared the effect estimates $\left(b_{xy}\right)$ of 33 proteins for each of the 19 risk factors obtained using the GCTA-GSMR method to effect estimates obtained using maximum likelihood method, simple and weighted median methods, simple and weighted mode methods, inverse variance weighted, and the mendelian randomization–Egger method.³⁷

We employed the heterogeneity in dependent instruments outlier test to detect and eliminate SNVs that may influence RCC through confounding variables unrelated to the risk factors and plasma proteins. The heterogeneity in dependent instruments–outlier test identifies pleiotropic SNVs by detecting statistically significant deviations in the causal estimate from what is expected under a causal model. The SNVs identified as pleiotropic were subsequently removed from the GCTA-GSMR analysis.³³ The Cochran Q test was used as a confirmatory assessment of the heterogeneity in instrumental variables.

To further assess horizontal pleiotropy, we used the Mendelian Randomization Pleiotropy RESidual Sum and Outlier method.³⁸ This approach tests the statistical significance of the intercept against zero; a statistically significant deviation indicates the presence of horizontal pleiotropy. We applied a statistical threshold of P < 05 to identify instrumental variables that exhibit horizontal pleiotropy.

We conducted a leave-1-out analysis to evaluate the influence of individual SNVs on the overall causal effect estimate. This method involves systematically removing 1 SNV at a time and re-estimating the causal effect. By doing so, we ensured that our results were not disproportionately influenced by any single outlier SNV.

Through these sensitivity analyses, we aimed to verify the robustness of our findings and minimize the impact of pleiotropy and outliers on our causal inference.

Protein-protein interaction enrichment

To identify biological pathways particularly enriched in a list of plasma proteins that showed statistically significant associations with risk factors and RCC in GCTA-GSMR, we conducted a series of pathway enrichment and protein interaction analyses using gene set cancer analysis. The STRING platform (https://string-db.org/) provides network-based visualization of protein-protein interaction.³⁷ We used this platform to identify the relationships among unique proteins associated with RCC and discover hub proteins that are engaged in multiple interactions with others.

Differential gene expression

Differential expression analysis for the KIRC cancer type is based on normalized and batch corrected RSEM messenger RNA expression. A total of 606 samples were analyzed, comprising 534 tumor samples and 72 normal samples. The fold change was determined based on the 72 paired samples, defined as the ratio of mean expression in tumor cells to the mean log expression in normal cells. The P value was estimated through a linear regression model that adjusted for age at diagnosis, sex, and pathological stage and was corrected for multiple testing by false discovery rate (FDR).

Bayesian colocalization analysis

Bayesian colocalization is a statistical method that can assess whether the same variants are responsible for 2 GWAS associations or distinct causal variants are close to each other.⁴⁰ Using simply single-variant summary statistics to test for colocalization of GWAS signals, we performed a Bayesian test for the co-localization of 2 traits using the coloc R package (R Foundation for Statistical Computing) to estimate the posterior probability of the hypothesis of interest that exposure and outcome are driven by the same causal variant (posterior probability of hypothesis 4 [PP.H4]).^40-42 When PP.H4 surpasses 0.8, it suggests a high probability of common genetic causal factors between the exposure and the outcome, indicating the existence of horizontal pleiotropy.

Observational association between plasma proteins and RCC

Using multivariable Cox proportional hazards regression analysis, we tested whether higher or lower levels of baseline plasma protein levels were associated with the cumulative incidence of RCC, adjusting for the same covariates (age, sex, recruitment center, and Olink measurement batch). The time to event was calculated by subtracting the date of event registry from the date of enrollment (data field: 53), focusing on events that occurred within 8 years from enrollment. We excluded prevalent cancer cases and those whose date of event was not recorded.

Multivariable mendelian randomization

Multivariable mendelian randomization is an extension of mendelian randomization that allows for the causal effects of multiple exposures on an outcome to be estimated. In this study, we used multivariable mendelian randomization to obtain the relative change in estimates of risk factors in the presence of 1 or more intermediate plasma proteins. At first, we estimated the association between a risk factor and RCC by regressing $Z_{\mathit{RCC}}\mathrm{ }\left(\frac{{\beta }_{\mathit{RCC}}}{S{E}_{\mathit{RCC}}}\right)$ on $Z_{RF}$. Then, we estimated the same association—between the risk factor and RCC—in the presence of a plasma protein associated with the risk factor. Here, we regressed $Z_{\mathit{RCC}}$ on both $Z_{RF}$ and $Z_{PP}$. Finally, we calculated the relative change in β estimate of the $Z_{RF}$, attributable to the effect of an intermediate plasma protein. We conducted this analysis using summary-level data on RCC, risk factors, and plasma proteins.

deCODE

We analyzed publicly available proteogenetic summary statistics from deCODE (deCODE Genetics; see “Data Availability”) to investigate the replicability of the associations with prioritized proteins. The overall data contained GWAS summary statistics for 4907 plasma proteins measured using Somalogic on 35 559 Icelanders. Given the differences in platforms (Olink in UK Biobank) and underlying study populations, along with previous reports of modest cross-platform replicability,⁴³ we expected limited replicability of results.

In silico analysis

Single-cell RNA sequencing

We processed single-cell RNA sequencing data from 79 kidney samples (from 77 participants) profiled as a part of the Kidney Precision Medicine Project (see the web resources in “Data Availability”) made available in the Kidney Tissue Atlas Explorer, version 1.5, release. The Kidney Precision Medicine Project samples are derived from kidney biopsy/reference tissues and include multiple enrollment categories, including healthy reference, chronic kidney disease (CKD), acute kidney injury, and diabetes mellitus-resistor samples. In particular, we used the data from 40 healthy participants only. The Kidney Precision Medicine Project used droplet-based 10× chromium single-cell RNA sequencing, followed by quality control filtering, normalization, clustering, integration with the broader kidney atlas, and expert cell-type/state annotation. We queried the processed gene expression in healthy samples for TYMP, UMOD, and USP28 to evaluate their cell-type localization in healthy adult kidney tissues, aiming to establish a cellular context for the plasma proteins.

TCGA immune infiltration analysis

We used TCGA-KIRC bulk RNA sequencing data through the Gene Set Cancer Analysis platform to evaluate additional kidney tumor relevance of genes encoding prioritized plasma proteins. For TYMP, we assessed correlations between tumor gene expression and immune cell infiltration estimates, estimated through ImmuCellAI,⁴⁴ across 24 immune cell populations using the Gene Set Cancer Analysis immune infiltration module. We also queried TCGA–Kidney Renal Papillary Cell Carcinoma as a related RCC cohort for comparison of immune expression patterns. These analyses were used for biological contextualization and prioritization, independent of the mendelian randomization analyses (see the web resources in “Data Availability”).

Phenome-wide association (PheWAS)

We performed phenome-wide lookups for the top cis–protein quantitative trait locus variants of the prioritized proteins using the Open Targets Genetics database (see the web resources in “Data Availability”). For each prioritized protein, we queried the lead cis-acting variant and extracted reported associations across available GWAS traits and disease outcomes, including the reported trait, mapped disease or phenotype label, source study identifier, P value (where available), and effect estimate. These analyses were used to contextualize whether protein-associated genetic variants were linked to biologically relevant hematological, immune, metabolic, kidney function, hypertension, or kidney disease phenotypes. Phenome-wide associations were interpreted as supportive biological annotation only and were not used as causal evidence in the mendelian randomization models.

Results

Step 1: effect of risk factors on plasma proteins

We used summary statistics-based mendelian randomization to estimate the effects of 19 known and suspected risk factors on each of 2940 circulating plasma proteins using GWAS summary statistics primarily from participants of European genetic ancestry (n = 34 557) (see “Methods”). At a threshold of P < 1.7 × 10^‒5, reflecting the Bonferroni correction for 2940 plasma proteins for each risk factor (Figure 2), we found that on average, each risk factor was associated with approximately 547 plasma proteins, while 2339 plasma proteins were associated with at least 1 of the risk factors (Table 1), indicating a wide spectrum of potential biological functions the risk factors affected.

Figure 2.

Effect of 19 risk factors on plasma proteins in 2-step mendelian randomization step 1. Volcano plot illustrating the effect of 19 risk factors on 2940 plasma proteins across 5 different categories, as identified in step 1: (A) anthropometric, (B) blood cell types, (C) kidney markers, (D) smoking and alcohol intake, and (E) blood pressure. The horizontal axis shows the log odds ratio of the effect of the risk factor on the plasma proteins, while the vertical axis shows the ‒log₁₀(P value) for the same. For each category, the statistically significant associations at a P value threshold (dashed line) of $1.7\times {10}^{-5}$ are colored. Abbreviations: BFP = body fat percentage; BMI = body mass index; BUN = blood urea nitrogen; CPD = cigarettes per day; DBP = diastolic blood pressure; DPW = drinks per week; EGFR = estimated glomerular filtration rate; EGFR_CYS = estimated glomerular filtration rate based on cystatin C; HMGLBN = hemoglobin; HMTCRT = hematocrit; PP = pulse pressure; RBC = red blood cells; SBP = systolic blood pressure; SMKC = smoking cessation; SMKI = smoking initiation; WBC = white blood cell; WBFM = whole body fat mass; WHR = waist-to-hip ratio.

Table 1.

Number of plasma proteins associated with 19 different risk factors analyzed here, as identified in 2-step mendelian randomization step 1.

Risk factor	PMID/study	Sample size	Significant number of plasma proteins	Significant number of plasma proteins a , No.
Body mass index	UK Biobank + Genetic Investigation of Anthropomorphic Traits	683 365	1042	944
Body fat percentage	UK Biobank	354 628	1005	859
Waist to hip ratio	UK Biobank + Genetic Investigation of Anthropomorphic Traits	485 486	1258	1055
Whole-body fat mass	UK Biobank	354 244	968	862
Age at which the individual started smoking	GWAS & Sequencing Consortium of Alcohol and Nicotine use + UK Biobank	293 995	0	0
Cigarettes/d	GWAS & Sequencing Consortium of Alcohol and Nicotine use + UK Biobank	295 971	19	8
Drinks/wk	GWAS & Sequencing Consortium of Alcohol and Nicotine use + UK Biobank	621 414	55	27
Smoking cessation	GWAS & Sequencing Consortium of Alcohol and Nicotine use + UK Biobank	355 406	14	6
Smoking initiation	GWAS & Sequencing Consortium of Alcohol and Nicotine use + UK Biobank	743 083	207	139
Red blood cell	UK Biobank	407 995	571	461
White blood cell	UK Biobank	439 516	1432	1323
Hemoglobin	UK Biobank	407 994	666	537
Hematocrit	UK Biobank	407 995	604	476
Estimated glomerular filtration rate	CKDGen + UK Biobank	1 097 041	889	810
Cystatin-based eGFR	CKDGen + UK Biobank	460 228	689	626
Serum/Blood urea nitrogen	CKDGen + UK Biobank	771 044	433	362
Diastolic blood pressure	UK Biobank	998 049	158	102
Pulse pressure	UK Biobank	980 638	173	76
Systolic blood pressure	UK Biobank	980 954	221	145

Abbreviation: GWAS = genome-wide association study.

^aAn updated ɑ threshold of .05 / (19 × 2940) that additionally control for the 19 separate risk factor tests.

Across these risk factors, WBC count was associated with the highest number of plasma proteins (1432) (Figure 2, B), while age at which the patient started smoking had no associations. In particular, highly related risk factors had similar patterns of associations. For example, each anthropometric risk factor, all of which are associated with obesity, is associated with roughly one-third of the 2940 plasma proteins (Table 1). In contrast, smoking-related risk factors had the lowest number of associated plasma proteins (eg, 14 for smoking cessation, 19 for cigarettes per day) (Figure 2, D). We noted a high correlation between magnitude and direction among the effects of related risk factors on plasma proteins (Figure S1).

Step 2: effect of plasma proteins on RCC

Next, we estimated the effects of the 2339 risk factor–associated plasma proteins on RCC, again using summary-based 2-sample mendelian randomization. At a false discovery rate below 0.05, we found that 33 plasma proteins were associated with RCC (Figure 3, A; Table 2). Collectively with step 1, these associations can be interpreted as proteomic intermediaries through which the effects of risk factors have a cascading effect on RCC. The identified plasma proteins included several proteins previously reported to be associated with RCC. For example, CRELD1 (β_{CRELD1 ⇾ RCC} = .07; P < 3.7 × 10^‒12), associated with anthropometric risk factors and kidney functions, was reported to be associated with RCC in a recent proteome-wide association study⁴⁵ and maps to a GWAS-significant susceptibility locus.³⁵ Overall, among the 33 plasma proteins identified, 23 had plasma levels inversely associated with RCC (eg, EDA2R: β_{EDA2R ⇾ RCC} = ‒.37; P < 3.41 × 10^‒12), with positive RCC associations observed for the remaining 10 plasma proteins, (eg, TNFRSF6B: β_{TNFRSF6B ⇾ RCC} =.22; P < 2.05 × 10^‒18) (Table 2). Thirty proteins were associated with multiple risk factors across different categories (eg, ANGPTL7, CD1C, GZMA) (Figure 3, A), potentially indicating pleiotropic functions.

Figure 3.

Heat map (left) showing the direction and magnitude of associations between 19 risk factors and 33 plasma proteins. The color scale indicates effect direction and strength, with positive associations in red and negative associations in blue. A violin plot (right) illustrates the distribution of absolute relative changes (%) in risk factor effect sizes after adjusting for individual mediator plasma proteins, highlighting the extent to which these proteins may potentially mediate risk factor associations. Green violins represent plasma proteins statistically significantly associated with the risk factor, while the red violins represent unassociated plasma proteins. Statistical significance levels are denoted as follows: ***P < .001, **P < .01, *P < .05, and “ns” indicates nonsignificant associations (P ≥ .05). Abbreviations: AgeSMK = Age at which smoking started; BFP = body fat percentage; BMI = body mass index; BP = blood pressure; BUN = blood urea nitrogen; CPD = cigarettes per day; cys_eGFR = estimate glomerular filtration rate using cystatin C; DPW = drinks per week; eGFR = estimated glomerular filtration rate; HMGLBN = hemoglobin; HMTCRT = hematocrit; PP = pulse pressure; RBC = red blood cells; SBP = SBP; SMKC = smoking cessation; SMKI = smoking initiation; WBC = white blood cell; WBFM = whole body fat mass; WHR = waist-to-hip ratio.

Table 2.

Proteins associated with RCC and at least 1 risk factor, as identified after 2-step mendelian randomization step 2.^a

Plasma protein		Overall mendelian randomization		Cis–mendelian randomization		Co-localization	Differential gene expression (KIRC)		Nearest GWAS signal b
Name	Region	β	P	β	P	PP.H4	Fold change	P (false discovery rate)	rsID	P	Distance, kilobase pair
AAMDC	11q14.1	−.04	6.7×10-04	−.05	4.8×10-04	0.08	1.14	0.04	rs7128118	3.5×10-04	715.4
ANGPTL7	1p36.22	−.07	1.6×10-04	−.06	0.01	0.15	0.13	5.0×10-04	rs61775595	3.8×10-04	180.7
APOL1	22q12.3	−.05	7.1×10-04	−.04	0.21	9.2×10-03	6.16	1.0×10-04	rs56777294	5.2×10-04	317.3
CA4	17q23.1	−.07	2.3×10-04	−.06	0.03	0.03	0.29	1.0×10-13	rs199797429	1.7×10-04	777.0
CD1C	1q23.1	−.06	3.4×10-04	.01	0.87	0.03	0.85	0.47	rs200246741	5.4×10-05	915.6
CDA	1p36.12	.07	7.3×10-06	.09	2.5×10-06	0.99	1.41	0.23	rs2072671	3.9×10-06	1.22
CDH5	16q21	−.04	1.2×10-04	−.05	0.32	6.0×10-03	2.52	9.8×10-12	rs112569130	4.7×10-07	461.1
CLEC4C	12p13.31	−.09	1.6×10-05	NAN	NAN	4.3×10-03	1.36	0.45	rs143365590	4.3×10-05	953.4
CRELD1	3p25.3	.07	3.7×10-12	.07	7.3×10-11	3.0×10-03	0.86	2.0×10-03	rs7629500	1.6×10-13	205.4
CRH	8q13.1	−.05	4.3×10-04	−.03	0.31	0.15	0.13	2.0×10-04	rs144794396	3.1×10-04	184.5
CTSE	1q32.1	.11	1.0×10-05	NAN	NAN	1.4×10-05	7.12	0.04	rs12118203	1.2×10-08	927.1
EDA2R	Xq12	−.37	3.4×10-12	NAN	NAN	0.06	4.85	1.5×10-21	rs145790884	2.8×10-04	946.1
ENPP6	4q35.1	−.06	5.7×10-04	.01	0.83	0.01	0.05	8.1×10-09	rs12499037	2.9×10-04	188.8
ERBB4	2q34	−.08	1.0×10-04	−.07	0.03	0.02	0.05	7.5×10-24	rs140489488	4.4×10-05	305.8
GZMA	5q11.2	−.29	0.0E + 00	.07	0.54	0.04	8.90	1.8×10-09	rs113459169	3.3×10-04	947.2
IGFBP6	12q13.13	.15	3.2×10-04	.05	0.62	7.8×10-03	1.14	0.58	rs7962345	2.3×10-04	266.8
IL17C	16q24.2	−.18	6.4×10-05	−.17	0.02	0.02	0.90	0.69	rs4374157	3.1×10-05	109.4
KLK1	19q13.33	−.03	4.2×10-04	−.03	0.19	0.01	0.04	3.4×10-07	rs144114333	8.9×10-05	638.7
OBP2B	9q34.2	.05	3.7×10-05	.05	1.8×10-04	0.35	5.55	0.21	rs6479631	4.3×10-05	808.5
PDCD6	5p15.33	−.04	4.8×10-06	−.04	3.6×10-06	7.4×10-06	1.42	1.9×10-07	rs7734992	4.1×10-14	965.0
PILRA	7q22.1	.04	3.1×10-07	.04	1.5×10-07	0.01	3.40	9.4×10-16	rs554177256	7.6×10-04	826.3
PILRB	7q22.1	.03	3.4×10-05	.02	4.6×10-03	0.01	1.99	1.5×10-06	rs554177256	7.6×10-04	811.0
PXN	12q24.23	−.05	5.8×10-04	−.05	9.4×10-04	0.18	1.13	0.06	rs573529608	2.8×10-05	157.6
RBP2	3q23	−.16	3.2×10-04	.02	0.88	6.1×10-03	0.01	1.6×10-07	rs144633990	1.9×10-04	376.2
SCGN	6p22.2	−.15	1.8×10-04	−.07	0.51	0.01	25.29	1.7×10-07	rs200686463	2.0×10-04	808.8
SELE	1q24.2	.04	5.7×10-04	.02	0.87	4.7×10-03	0.26	1.0×10-04	rs79562616	3.2×10-04	711.7
TEK	9p21.2	−.04	2.6×10-04	−.03	0.06	6.9×10-03	0.42	1.3×10-19	rs13297316	4.2×10-06	787. 8
TFF3	21q22.3	−.07	2.0×10-04	.00	0.95	0.05	0.64	0.03	rs139980534	2.3×10-04	248.9
TNFRSF6B	20q13.33	.22	2.0×10-18	.22	9.9×10-14	6.6×10-06	10.27	1.4×10-08	rs200110285	2.2×10-11	587.5
TYMP	22q13.33	−.17	2.6×10-06	−.29	4.9×10-08	0.99	6.81	1.7×10-14	rs131813	5.4×10-09	30.12
UMOD	16p12.3	.05	1.4×10-06	.06	1.3×10-08	0.89	0.00	5.1×10-15	rs7203642	8.2×10-06	4.93
USP28	11q23.2	−.31	4.5×10-05	−.46	4.1×10-05	0.99	0.85	5.0×10-05	rs4288784	1.2×10-05	6.86
VASN	16p13.3	−.06	3.5×10-04	−.05	0.02	0.05	1.06	0.53	rs576212384	1.2×10-05	715.0

Abbreviations: GWAS = genome-wide association study; KIRC = Kidney Renal Clear Cell Carcinoma; NAN = Not available (due to lack of sufficient genetic instruments) ; PP.H4 = posterior probability of hypothesis 4; RCC = renal cell carcinoma.

^aResults from cis–mendelian randomization, co-localization, and differential gene expression of the associated plasma proteins for each. The prioritized proteins (see “Methods” and “Results”) are shaded and/or bold.

^bStrongest GWAS association within ±1 megabase pair of the protein’s transcription start site.

To identify whether the plasma proteins pertain to 1 or more histologic subtypes, we repeated the mendelian randomization analysis with plasma proteins as exposures for the 2 main histologic subtypes: clear cell RCC and papillary RCC. In a subtype-specific analysis, 19 of the 33 plasma proteins were nominally associated (P < .05) with clear cell RCC, and 8 were associated with papillary RCC, with 3 plasma proteins (EDA2R, UMOD, and VASN) exhibiting nominal associations with both (Table S1). To directly contrast the effects of plasma proteins on the subtypes, we further performed a case-only mendelian randomization analysis,¹¹ which provided us with a test for the heterogeneity of plasma protein effects on subtypes. We identified 8 proteins (Table S1) that have nominally significant heterogeneity of plasma protein effects on clear cell and papillary RCC. The 33 identified plasma proteins were further enriched for protein-protein interactions (P < 3.25 × 10^‒4), as annotated in the STRING database.³⁹ Four of the genes encoding these 33 proteins—ERBB4, KLK1, PDCD6, and USP28—are putative cancer driver genes,⁴⁶ of which ERBB4 is classified as a tier-1 cancer gene in the Catalogue of Somatic Mutations in Cancer database.⁴⁷

Notably, the majority of the identified plasma proteins do not map to substantial genome-wide RCC loci (28 of 33), offering novel mechanistic insights beyond GWAS due to the “risk factor–first” approach of the analysis. For instance, our analysis identified that APOL1 on 22q12.3, which does not map to any GWAS-significant region, is impacted by WBC count (β_{WBC ⇾ APOL1} = .08; P < 1.21 × 10^‒9) and inversely associated with RCC (β_{APOL1 ⇾ RCC} =.05; P < 7.1 × 10^‒4), indicating that APOL1 can cascade the effect of WBC count on RCC.

We further performed multivariable mendelian randomization for each of the 33 plasma proteins to evaluate their potential mediating roles. For each risk factor, we analyzed 2 models, including and excluding each of the 33 plasma proteins, and compared the changes in effect size estimates of each risk factor to quantify the potential mediating effects of the plasma protein. Seven risk factors (age at which the patient started smoking, number of cigarettes per day, diastolic BP, number of drinks per week, pulse pressure, smoking cessation, and smoking initiation) had fewer than 5 associated plasma proteins and were excluded from this analysis. Among the remaining 12 risk factors, 7 (body fat percentage, BMI, whole-body fat mass, systolic BP, hematocrit, red blood cell count, and WBC count) showed statistically significant (P < .05) mediation effects for corresponding associated plasma proteins compared with unassociated proteins (Figure 3, B). For instance, BMI shows an average absolute relative change of 8.82% in the presence of the associated plasma proteins vs only 1.75% with the unassociated proteins. In particular, we found that proteins such as PILRB and CDA alter the effect of BMI by 17.4% and 16.1%, respectively, indicating potentially strong mediating roles. Similarly, plasma proteins associated with WBC count exhibit an average effect of 4.6% compared with 1.5% for unassociated plasma proteins, highlighting the potential mediating role of these plasma proteins. In particular, we found that APOL1 mediates approximately 2.8% of the effect of WBC count on RCC.

Step 3: prioritize risk factor–associated plasma proteins that have an effect on RCC

For the 33 identified plasma proteins that have an effect on RCC, we conducted multiple downstream analyses to prioritize proteomic intermediates of risk factors, as mentioned above, including (1) cis–mendelian randomization to further eliminate potential pleiotropic effects, (2) co-localization to identify shared causal variants, and (3) differential gene expression in TCGA-KIRC. Using the above analyses, we prioritize 3 plasma proteins through convergent evidence across all these analyses (TYMP, UMOD, and USP28). Below, we describe in detail these 3 proteins and provide further context for their relevance in RCC.

TYMP

TYMP is a multifunctional enzyme involved in multiple processes intimately tied to tumor biology^48,49 and has been explored as a prognostic biomarker and therapeutic target in several cancers, including RCC.^48,50 In step 1, we identified that TYMP was associated with genetically predicted WBC count (β = .11; P < 1.5 × 10^‒13), confirmed across multiple methods (Table S2) with only nominal evidence of horizontal pleiotropy (P < 5.0 × 10^‒3), and heterogeneity (P < 3.4 × 10^‒3) but no statistically significant outliers. In step 2 (Table 2), we found TYMP to be inversely associated with RCC (β = ‒.17; P < 2.59 × 10^‒6). Taken together, this finding suggests that TYMP is affected by systemic inflammation and, in turn, can influence RCC risk. Using multivariable mendelian randomization, we estimated that TYMP can attenuate 3.03% of the effect of WBC count. The association was further supported by cis–mendelian randomization (β = ‒.29; P < 4.9 × 10^‒8), co-localization (PP.H4 = 0.99) (Figure 4, A), and statistically significant differential expression (false discovery rate = 1.69 × 10^‒14) of TYMP in TCGA-KIRC (Figure S2).

Figure 4.

TYMP as the proteomic mediator of white blood cell (WBC) count on renal cell carcinoma (RCC). (A) Regional association for the 22q13.33 region with RCC and TYMP. The variants are colored by the level of linkage disequilibrium with the strongest RCC-associated variant (rs131805) in the region. The lower panels show the H3 K27ac peaks in clear cell (CC2) and papillary (papRCC) subtypes in the region and binding sites for HIF1b and HIF1a in RCC-related cell lines (RCC4 and 786-O). (B) Replication in deCODE: associations between WBC count, TYMP, and RCC, as estimated from proteogenetic data in deCODE compared with that obtained from UK Biobank. (C) Cumulative incidence of RCC by TYMP levels, dichotomized at the median. (D) Single-cell expression of TYMP in a monocyte cluster from the Kidney Tissue Atlas (see the web resources in “Data availability”). Abbreviations: deCODE = Proteogenetic study in deCODE (See Methods); WBC = white blood cell counts.

Analysis in proteogenetic data reported by deCODE⁵¹ replicated the association between WBC count and TYMP (β = .08; P = 9.5 × 10^‒7), and the association between TYMP and RCC was also directionally consistent (β = ‒.03; P = .62), albeit not statistically significant (Figure 4, B). Quantile-based Cox analyses in UK Biobank echoed these findings at the phenotypic level: Higher baseline TYMP was associated with reduced cumulative RCC incidence (hazard ratio = 0.88; P = 9.1 × 10^‒03), consistent with step 2 (Figure 4, C).

TYMP resides in the 22q13.33 region, which has been identified through GWAS to be associated with RCC susceptibility.³⁵ Notably, the lead cis–protein quantitative trait loci (rs131805) increases TYMP and is protective for RCC, matching the mendelian randomization direction and supporting a model in which increased TYMP is antitumorigenic in RCC. This locus is also associated with hematological traits, immune function, and several blood metabolites (Table S3), indicative of its involvement in multiple distinct processes, some of which are potentially relevant to RCC.

TYMP expression in RCC tumors correlated with markers of immune cell infiltration, indicating its relevance in cascading the effects of WBC count on RCC. In bulk RNA sequencing data from TCGA-KIRC tumors, TYMP levels were strongly correlated with CD8⁺ T-cell abundance (r = 0.68; P = 1.7 × 10^‒86) and helper T type 1 cell signatures (r = 0.62; P = 1.3 × 10^‒66) (Table S4). Consistently, single-cell RNA sequencing of adult kidney tissue revealed that TYMP expression was predominantly localized to monocyte-lineage populations, identifying these immune cells as the primary source of TYMP in the kidney microenvironment and underscoring its WBC count–relevant role in kidney tissues (Figure 4, D; Table S5). Tissue profiling showed that the transcription start site of TYMP was proximal to RCC-relevant cis-regulatory elements, including H3K27ac-marked enhancers in RCC tumor samples and HIF transcription factor binding in RCC cell lines. Specifically, the region overlapped HIF1A binding peaks in RCC4 and HIF2A binding peaks in 786-O (Figure 4, A).

UMOD

UMOD is a glycoprotein highly overexpressed in kidneys and in human urine and has a role in preventing urinary tract infections.^52,53 Previous studies have found higher UMOD levels to be predictive of incidence of CKD and adverse kidney events, indicating its potential role in RCC.^54-56

In step 1, we identified UMOD as being negatively associated with anthropometric measurements, number of drinks per week, and smoking cessation and positively associated with eGFR and cystatin-based eGFR (Figure 3, A). Using multivariable mendelian randomization, we estimated that 2.96% to 16.98% of anthropometric trait effects, 13.45% of smoking effects, and 26.82% of drinks per week effects to be potentially mediated via UMOD. Sensitivity analyses showed that effect estimates were consistent, and no evidence of directional pleiotropy or clinically significant outliers were identified (Table S6). In step 2 (Table 2), we found UMOD to be associated with RCC (β = .05; P = 1.37 × 10^‒6) and prioritized through downstream analysis (Table 2), including cis–mendelian randomization (β = ‒.06; P = 1.3 × 10^‒8), co-localization (PP.H4 = 0.89) (Figure 5, A), and differential expression (Figure S2) in KIRC tumor samples (false discovery rate = 5.05 × 10^‒15). In particular, UMOD showed strong risk factor–plasma protein co-localization (Table S7) with adiposity-related traits, including whole-body fat mass (PP.H4 = 0.992) and body fat percentage (PP.H4 = 0.928). These traits also showed strong risk factor–RCC co-localization (Table S8) at the UMOD locus, with a PP.H4 of 0.991 for whole-body fat mass and a PP.H4 of 0.910 for body fat percentage, supporting a shared local genetic architecture linking adiposity-related traits, UMOD regulation, and RCC susceptibility. Notably, subtype-specific analysis and case-only mendelian randomization revealed UMOD had a statistically significantly stronger association with papillary RCC than clear cell RCC.

Figure 5.

UMOD as the proteomic mediator for RCC. (A) Regional association for the 16p12.3 region with RCC and UMOD. The variants are colored by the level of linkage disequilibrium with the strongest RCC-associated variant (rs12922822) in the region. The lower panels show the H3K27ac peaks in clear cell (CC2) and papillary (papRCC) subtypes in the region and binding sites for HIF2a in RCC4 cell lines. (B) Replication in deCODE: associations between UMOD, risk factors associated with UMOD, and RCC, as estimated from proteogenetic data in deCODE compared with that obtained in UK Biobank. (C) Single-cell expression of UMOD in the thick ascending limb (TAL) cell population from Kidney Tissue Atlas (see the web resources in “Data Availability”). Abbreviations: BMI = body mass index; deCODE = Proteogenetic study in deCODE (See Methods); eGFR = estimated glomerular filtration rate; RCC = renal cell carcinoma.

Replication in deCODE⁵¹ showed inconsistent results with risk factors, while the association between UMOD and RCC was only directionally consistent (β = .001; P = .89) (Figure 5, B). This finding could be due to the different proteomic profiling platforms used in UK Biobank (Olink) and deCODE (SomaLogic). Quantile-based Cox proportional hazards analysis suggested that higher UMOD levels may be associated with an increased cumulative incidence of RCC (hazard ratio = 1.03; P = .08), although the association was not statistically significant.

UMOD resides in the 16p12.3 locus, which has not yet been reported to be associated with RCC. Variants at this locus, however, are associated with multiple kidney biomarkers—diseases as well as blood cell counts (Table S9). In particular, the C allele of the top cis protein quantitative trait locus (rs12922822: C/T), which is associated with increased levels of UMOD and increased risk of RCC, is also associated with an increased risk of hypertension, CKD, kidney failure, and decreased calculus of kidney, highlighting its involvement in processes related to kidney diseases and associated risk factors.¹¹

In a 50-kilobase pair window in the cis region around the transcription start site of UMOD, we identified putative active enhancers marked by H3K27ac in multiple RCC subtypes, indicating its potential regulation in RCC tumors as well as binding of EPAS1, a member of the HIF family of transcription factors, using chromatin immunoprecipitation sequencing data in RCC cell lines (RCC4). Single-cell profiling in adult kidney identified that UMOD was expressed particularly in loop of Henle thick ascending limb epithelial cells as a marker gene (Figure 5, C; Table S10).

USP28

USP28 plays a critical role in protein homeostasis, with effects on cell proliferation, DNA repair, and tumor development.⁵⁷ In step 1, we found USP28 to be associated with 6 risk factors (BMI, whole-body fat mass, hematocrit, WBC count, eGFR, and cystatin-based eGFR), with consistent associations across multiple methods and no evidence of heterogeneity, directional pleiotropy, or influential outliers. In step 2, we identified increased levels of USP28 as being negatively associated with RCC risk (β = ‒.31; P = 4.5 × 10^‒5), and it was prioritized due to convergent evidence from cis–mendelian randomization (β = ‒.46; P = 4.1 × 10^‒5), co-localization (PP.H4 =0.99), and differential gene expression (false discovery rate = 4.97 × 10^‒5). Overall, we estimated that 2.26% to 4.62% of anthropometric trait effects, 1.94% to 2.36% of kidney biomarker effects, and 0.45% of drinks per week effects on RCC can be attributable to USP28 using multivariable mendelian randomization.

Analysis in deCODE replicated the association of USP28 with anthropometric and blood cell count risk factors but not for kidney markers; however, we did not identify an association with RCC (β = .03; P = .71). Multivariable Cox proportional hazards analysis in 31 091 individuals (83 RCC cases) did not show any difference in the incidence rate of RCC (hazard ratio = 0.99; P = .31) for different quantiles of USP28 levels.

The cis window (50 kilobase pair) around the transcription start site of USP28 contained putative active enhancers in multiple RCC subtypes as well as evidence of possible binding of multiple members of the HIF family of transcription factors, indicating potential hypoxia-induced regulation in the region. Single-cell profiling in adult kidney identified that USP28 was particularly expressed in several distinct cell populations, including natural killer cells, inner medullary collecting ducts, and intercalated cell type A and T cells, indicating a broad range of potential functions (Table S11).

Discussion

Genetic epidemiological studies of RCC have revealed a complex interplay between established and suspected risk factors, including obesity, BP, and smoking, and genetic factors and molecular intermediates. The precise biological pathways linking these risk factors to RCC development, however, remain unclear. In this study, we implemented a 2-step mendelian randomization framework to uncover plasma proteins that may link the effects of RCC risk factors with tumorigenesis. This approach has previously been used to investigate the effect of obesity on disease outcomes^58,59 but, to our knowledge, has not yet been employed to study a wide spectrum of cancer risk factors. By using a large-scale GWAS of risk factors and RCC as well as recently published results from the UKB-PPP, we identified 33 plasma proteins whose levels were genetically influenced by at least 1 of 19 RCC risk factors (spanning adiposity, BP, smoking, blood cell counts, and kidney function markers) and that in turn showed evidence of an effect on RCC. This proteogenomic analysis revealed diverse intermediaries linking risk factors for RCC, highlighting convergent effects of risk factors through a smaller number of downstream candidate targets. Of these, 3 (TYMP, UMOD, and USP28) were prioritized through consistent evidence from cis–mendelian randomization, co-localization, and differential gene expression analyses. Through further in silico analysis, we established the putative relevance of these proteins for RCC and their possible roles in producing the cascading effects of the corresponding risk factors. Our findings are novel in demonstrating, for the first time, a proteome-wide landscape that links RCC risk factors to their downstream proteomic targets, revealing novel insights for RCC etiology.

One of the prioritized proteins, TYMP, proximal to a previously identified RCC GWAS locus, is a proangiogenic enzyme involved in nucleotide metabolism and vascular remodeling⁴⁹ and has been proposed as a prognostic marker.⁶⁰ Previous mendelian randomization studies have found divergent evidence of TYMP for different cancers (eg, protective for lung adenocarcinoma and prostate cancer but risk increasing for ovarian cancer).^33,61,62 Here, our findings that TYMP is associated with both WBC count and RCC provide novel mechanistic insights into how systemic immune activity might shape kidney carcinogenesis. In silico analyses suggest that TYMP highly expressed normal kidney monocytes; likely supports tissue maintenance, local immune defense, and immune surveillance; and may induce an antitumor immune environment, consistent with its correlation with greater CD8⁺ T-cell and helper T type 1 cell infiltration in RCC tumors. Epigenomic data showed hypoxia-inducible factors (HIF1A and ARNT) binding near the TYMP locus in RCC cells, linking TYMP to the hypoxia pathway central in RCC. As with many immune factors, TYMP may have a context-dependent role: Although protective before tumor initiation, it may later aid tumor angiogenesis once cancer is established.⁴⁸

Our study further prioritizes UMOD, a kidney-derived glycoprotein and the most abundant protein in human urine.⁶³ Genetically higher UMOD levels have been causally linked to increased risk of CKD and hypertension, conditions that themselves predispose humans to RCC.⁵²,⁶⁴ Consistently, our 2-step mendelian randomization analysis indicated that several risk factors—notably, obesity-related traits and kidney function measures—influence UMOD, and that, in turn, higher UMOD levels causally increase RCC risk. Mechanistically, the protumor effect of UMOD could relate to its role in kidney physiology and cell signaling. We found that UMOD’s gene region in tumor tissue bears active chromatin marks and binding sites for EPAS1 (HIF-2α), a master transcription factor in RCC pathogenesis, hinting that hypoxic tumors may drive UMOD expression. Furthermore, UMOD has been known to interact with proteins such as APOL1 and TP53, suggesting that aberrant UMOD levels could influence tumor suppressor pathways or immune evasion. Indeed, UMOD’s top cis-acting genetic variant (rs12922822), which lowers UMOD, was associated not only with lower RCC risk in our study but also with reduced risk of hypertension and kidney failure in phenome-wide analyses.⁶⁵ This genetic convergence supports the hypothesis that high UMOD expression contributes to a spectrum of kidney-related factors that favor carcinogenesis.

In our study, higher genetically predicted USP28 levels were linked to a lower risk of RCC. We also observed that USP28 gene expression is markedly reduced in RCC tumors, consistent with a protective role. This finding is noteworthy because USP28 has been characterized as a context-dependent protein in cancer: It stabilizes oncogenic factors such as c-Myc and HIF-1α but also tumor suppressors (eg, p53) and maintains genomic stability.⁶⁶ In RCC, a tumor driven heavily by HIF signaling, loss of USP28 in tumors might remove constraints on tumor growth (for instance, by disabling p53 or other restraining pathways),⁶⁷ aligning with our finding that a higher USP28 level has a protective effect. Epigenomic analysis indicating HIF binding at the USP28 locus further suggests that RCC cells may actively downregulate USP28 to gain a growth advantage under hypoxia. These findings broaden the spectrum of RCC biology, suggesting new avenues for therapeutic intervention (eg, USP28 agonism) while also underscoring that such interventions must account for the proteins’ multifaceted roles in normal physiology.

Beyond the 3 prioritized proteins, our analysis uncovered numerous additional proteins that provide further insight into RCC’s molecular underpinnings and merit investigation. For example, TNFRSF6B, which showed 1 of the strongest associations in our analysis, is often upregulated in tumors and can modulate apoptosis and angiogenesis. Our analysis shows that TNFRSF6B is affected by anthropometric and kidney trait risk factors, with higher levels associated with higher RCC risk, as expected. We also identified CRELD1, which has been detected in previous GWASs and transcriptome-wide association studies of RCC and may affect cell adhesion or growth factor signaling.^35,45 Our results linking CRELD1 to increased RCC risk align with a possible protumor role, creating a cascade of the effects of anthropometric and kidney-related traits. Notably, 28 of the 33 identified proteins lie outside known RCC GWAS loci, underscoring how our proteome-first approach can uncover novel biology beyond large-scale genetic studies. For example, APOL1, which is affected by WBC count, is known for its role in kidney disease and innate immunity, and recent studies have shown that it functions as a tumor suppressor in clear cell RCC by impeding metastasis and epithelial-mesenchymal transition.^68,69 Consistent with that finding, we showed that higher APOL1 levels are associated with lower RCC risk. Members of the tumor necrosis factor (TNF) superfamily also featured prominently in our results. Interestingly, several proteins with roles in vascular and metabolic biology were protective intermediaries—for instance, carbonic anhydrase IV and paxillin were associated with risk-lowering effects. Although we did not prioritize these proteins for detailed follow-up, their identification is important because it suggests that RCC risk is influenced by a broad network of biological processes, from immune surveillance (APOL1) and inflammatory signaling (TNF family receptors) to cellular adhesion (CRELD1, paxillin) and metabolic homeostasis (carbonic anhydrase IV, among others), each representing a hypothesis for further investigation.

The downstream tumor expression analyses used TCGA-KIRC, which represents clear cell RCC. Therefore, these expression-based findings are most directly relevant to the clear cell RCC histological subtype and might not be generalizable to all RCC subtypes. To address subtype heterogeneity, we conducted additional mendelian randomization analyses for clear cell RCC and papillary RCC separately as well as a case-only mendelian randomization. Several proteins prioritized in the primary analysis showed evidence of association with clear cell RCC, and a smaller subset showed evidence for papillary RCC, with EDA2R, UMOD, and VASN associated with both subtypes. Of the prioritized proteins, UMOD had a much stronger effect on papillary RCC than on clear cell RCC. These results indicate that some plasma proteins may have shared effects across RCC subtypes, while others may be subtype specific. Larger subtype-specific GWASs, particularly for less prevalent subtypes, such as papillary and chromophobe, will be needed to more precisely define subtype-specific proteomic associations.

Despite the robustness of our analytical framework, several limitations warrant discussion. Our analyses were restricted to individuals of European ancestry due to limited available proteomic data for African genetic ancestry (34 557 individuals of European ancestry vs 931 individuals of African ancestry).⁷⁰ Furthermore, our reliance on data from biobanks may introduce healthy volunteer bias because individuals with severe or rapidly progressing RCC may be underrepresented. We also caution that for many of the investigated risk factors, a causal relationship with RCC has not yet been established. Although our findings generate candidate hypotheses, detailed functional investigations or targeted intervention studies will be necessary to determine whether modulating the levels of these identified plasma proteins can effectively alter RCC risk. It should also be noted that although we implicitly aimed to study the pathway risk factor→plasma protein→RCC, we cannot claim that the plasma proteins are actual “mediators” of risk factor effects because we did not perform a formal mediation analysis. Overlap of the genetic instruments of risk factors and plasma proteins could complicate the interpretation of the identified proteins as mediators of risk factor effects.⁷¹ Finally, although some mendelian randomization mediation analyses estimate the proportion of mediation, we chose not to emphasize these estimates given methodological challenges and potential biases inherent in cis–mendelian randomization mediation frameworks.

In summary, we present a comprehensive discussion of how a 2-step mendelian randomization approach can help elucidate the potential mechanisms underlying epidemiological risk factors to cancer through molecular intermediates. To our knowledge, this is the first study to integrate large-scale proteogenomic data to propose proteomic intermediaries of risk factor effects on RCC, opening novel avenues for biomarker and therapeutic target exploration. Our work not only advances the field of RCC research by highlighting new targets and pathways beyond GWASs but also provides a more systems-level understanding of RCC causation. We emphasize that our objective was to explore the proteomic cascades of risk factor effects on RCC to generate newer hypotheses that merit further laboratory investigation. Our results help identify specific molecular processes critical for RCC development and show that the effects of a broad and diverse set of risk factors converge on a smaller set of core-like proteins. The novelty of our approach lies in uniting epidemiological evidence with proteomics, and the insights gained here can be used in future studies aiming to translate genetic discoveries into tangible health benefits.

Author contributions

Ibrahim Hossain Sajal (Data curation, Formal analysis, Investigation, Methodology, Software, Writing—original draft, Writing—review & editing), Andrew J. Song (Data curation, Formal analysis, Writing—review & editing), Kevin M. Brown (Validation, Writing—review & editing), Mitchell J. Machiela (Data curation, Resources, Writing—review & editing), Peter Kraft (Methodology, Writing—review & editing), Stephen Chanock (Investigation, Resources, Writing—review & editing), Mark P. Purdue (Data curation, Methodology, Resources, Writing—review & editing), and Diptavo Dutta (Formal analysis, Investigation, Project administration, Supervision, Writing—original draft, Writing—review & editing)

Supplementary material

Supplementary material is available at JNCI Cancer Spectrum online.

Funding

The current analysis was supported by the Intramural Research Program of the National Cancer Institute, NIH, Department of Health and Human Services.

Conflicts of interest

None declared.

Data availability

We used publicly available GWAS summary statistics from the following sources: BMI GWAS from GIANT (https://portals.broadinstitute.org/collaboration/giant/); deCODE (https://www.deCODE.com/summarydata/); UKB-PPP (https://www.synapse.org/Synapse:syn51364943/wiki/622119); GWAS datasets of kidney function biomarkers (https://ckdgen.imbi.uni-freiburg.de/datasets/Stanzick_2021); Generalized Summary Mendelian Randomization (https://yanglab.westlake.edu.cn/software/gcta/). The full code and results from 2-step mendelian randomization will be made available upon publication at the following GitHub repository: https://github.com/diptavo/RCC_TSMR. Readers are directed to the following web resources: OpenTargets (https://platform.opentargets.org/); Gene Set Cancer Analysis (https://guolab.wchscu.cn/GSCA/#/); and Kidney Precision Medicine Project (https://atlas.kpmp.org/).