Robust cis-by-trans epistasis in the human plasma proteome highlights an ABO-centered interaction network

Abstract

Detecting and interpreting epistatic effects in the human genome is essential for a complete understanding of the genetic architecture of gene regulation and complex traits. Progress in this area, however, has been hindered by limited discovery power and an abundance of false positives. Here, leveraging large-scale plasma proteomic data from the UK Biobank Pharma Proteomics Project (UKB-PPP), we systematically map high-confidence cis-by-trans epistatic effects regulating protein expression. To reduce the multiple-testing burden, we focus on protein-level rather than variant-level interaction effects, and we employ a sandwich variance estimator to control false positives, as validated by extensive simulations. We applied the approach to identify cis-by-trans interactions across all proteins assayed in UKB-PPP and detected seven significant epistatic effects. Incorporating these interactions into protein-wide association studies is highly prone to false positives. We thus developed a principled framework but found that they offered little improvement in trait mapping power, likely because their effects on protein expression variance are outweighed by additive effects. Analysis of the biological and evolutionary context of our detected epistatic interactions pointed to ABO as a central regulator that appears to modulate mean protein expression for several trans-chromosomal targets through its effects on blood type, potentially driving long-term balancing selection on loci within its epistatic network. Together, our results clarify both the biological relevance and the practical limitations of epistasis-aware analysis in human proteomic regulation.

eTOC blurb

Li et al. identify robust cis-by-trans epistatic interactions in the human plasma proteome. They uncover an ABO-centered epistatic regulatory network where blood type modifies cis-genetic effects on proteins, revealing a buffering mechanism consistent with long-term balancing selection and highlighting how genetic background shapes protein regulation.

Graphical Abstract:

Introduction

Genetic epistasis—non-additive interactions between genetic variants—has long been recognized as a fundamental component of the genetic architecture of complex traits. In epistatic systems, the phenotypic effect of a genetic variant depends on the genetic background in which it occurs, violating the assumptions of purely additive models. The role of epistasis in non-human model organisms has been empirically and computationally studied and validated^1–3, and recent work has demonstrated that genome-wide interaction effects collectively explain higher trait variance than pure additive effects in human complex traits⁴. The genetic loci exhibiting epistasis have been reported to be associated with balancing selection, which highlights the impact of genetic epistasis on evolutionarily shaping human complex traits^5,6. Nonetheless, individual epistasis effects remain challenging to detect and interpret in human populations due to small marginal effects, the severe multiple-testing burden and pervasive false positives^7,8.

Molecular phenotypes, including gene expression and protein abundance, provide a powerful framework for studying epistasis because they lie closer to the direct action of genetic variation than organismal traits. Large-scale studies have mapped thousands of quantitative trait loci (QTLs) for molecular traits across tissues and cell types. While most studies focus on additive genetic effects, growing evidence highlights the role of epistasis in molecular regulation. For instance, multiple studies have reported epistatic effects on gene expression^9,10, i.e., the impact of genetic variants on gene expression can depend on other genetic variants or the expression levels of other genes. Such interactions often reflect biological mechanisms, such as transcription factor (TF) binding dynamics, where the regulatory impact of a variant varies with TF abundance^11,12. Previous studies have also reported a few epistatic effects on protein expression^13,14. These findings highlight that gene regulation is inherently non-additive and that epistatic effects can help identify biologically meaningful regulatory interactions.

The recent UKB-PPP study measured plasma protein levels in 54,219 participants and identified extensive cis- and trans- genetic associations with protein abundance¹³. Such an expansive dataset enables us to conduct a comprehensive search for genetic epistasis at the protein-expression level with substantially improved statistical power. Nonetheless, detecting epistatic effects in proteomic data poses substantial statistical challenges. Genome-wide variant–variant interaction scans are computationally prohibitive and suffer from extreme multiple-testing burdens. Moreover, interaction tests are particularly susceptible to false positives arising from linkage disequilibrium (LD), model misspecification, and heteroskedasticity^15,16, complicating the interpretation of reported signals.

In this study, instead of performing genome-wide pairwise SNP-by-SNP interaction scans, we adopt a robust, protein-based framework to systematically identify cis-by-trans epistatic effects regulating plasma protein expression. Restricting the analysis to cis-by-trans interactions offers two distinct advantages. First, this approach is biologically motivated, as it captures interactions between cis-regulatory variation and trans-acting factors such as transcriptional or post-transcriptional regulators, which are particularly plausible drivers of epistatic effects given their inherent ability to exert downstream influences across the genome. Second, this restriction substantially reduces the search space and multiple-testing burden. We applied the approach to identify cis-by-trans interaction effects across all proteins assayed in UKB-PPP and identified a central role for ABO (MIM: 110300), which participates in multiple inter-chromosomal interactions. We assessed the implications of epistatic effects for complex trait analysis by evaluating whether incorporating interaction effects into protein-wide association studies (PWAS) improves trait mapping. Finally, we further examined the evolutionary context of these interactions by testing whether genes involved in epistatic regulations are enriched for signatures of long-term balancing selection.

Material and methods

Predicting protein expression level based on cis-regulatory SNPs

The individual level phenotypes, genotype and protein expression data were accessed through the UK Biobank under the Application ID 89052. UKB-PPP includes 49,484 individuals of European ancestry. After excluding individuals with missing genotype or covariate data (age, sex, UK Biobank assessment center, or genotyping array), 49,024 individuals remained for analysis. We randomly selected 10,000 individuals to train the FUSION¹⁷ prediction model for each protein and used the remaining individuals for downstream analyses, thereby ensuring that protein prediction based on cis-genetic effects was performed in an independent dataset. Specifically, we used cis-SNPs (≤ 500 Kb from the transcription starting site, TSS) of the gene encoding the protein as the model input and obtained the SNP weights calculated from elastic net after adjusting for age, sex, UKB center, UKB genetic array and the first 20 genetic principal components. We then predicted the protein expression for the rest of the individuals based on the SNP weights. The cis-heritability (cis-$h^2$) of each protein was estimated based on the same cis-SNPs using the REML algorithm implemented in GCTA¹⁸.

Inflated interaction effects estimated through OLS at the SNP level

To show the inflation of interaction effects estimated through ordinary least squares (OLS) at the SNP level when there is no real interaction effect, we conducted a simulation study as follows. For simplicity, we simulated three SNPs ($SN{P}_1$, $SN{P}_2$, and $SN{P}_3$) with the same MAF = 0.2, and a protein (A) impacted by only $SN{P}_1$ for 1,000 samples:

$$ob{s}_A={\beta }_{1}SN{P}_1+\epsilon ,$$

where $ob{s}_A$ is the observed expression of protein A, ${\beta }_1$ is the effect size of $SN{P}_1$ and $\epsilon$ is a random error generated from a standard normal distribution. We generated $SN{P}_1$ and $SN{P}_2$ from a multivariate binomial distribution with 5 different correlations ($r=0,0.3,0.6,0.9,\text{or}1$). ${SNP}_3$ is generated from an independent binomial distribution. For each correlation scenario, we simulated 4 heritabilities of $y$ ($h^2=0.05,0.1,0.2,\text{or}0.5$), and ${\beta }_1$ is calculated accordingly. The interaction $\left({\beta }_I\right)$ between $SN{P}_2$ and $SN{P}_3$ on protein A is tested through OLS using the following model:

$$ob{s}_A={\beta }_{2}SN{P}_2+{\beta }_{3}SN{P}_3+{\beta }_{I}SN{P}_2\times SN{P}_3+e.$$

For each combination of SNP $r$ and phenotype $h^2$, we repeated the above simulation 1,000 times and calculated the genomic inflation factor $\left(\lambda \right)$ as the ratio of the median of the observed ${\chi }^2$ test statistics (based on the $P$ values) to the expected median of the ${\chi }^2$ distribution with 1 degree of freedom (0.456). We recalculated the $P$ value of each interaction effect using sandwich variance estimator (SVE)¹⁹, which reestimates the variance of the effect coefficients, to remove the inflation. We then conducted 20 such 1,000-time simulations to obtain the distribution of $\lambda$.

Inflated interaction effects estimated through OLS at the protein level

To further show the inflation of interaction effects estimated through OLS at the protein level when using only SNPs in LD with the true causal SNPs to predict the protein expression, we simulated a protein A with three causal SNPs:

$${prot}_A={\beta }_{1}SN{P}_1+{\beta }_{2}SN{P}_2+{\beta }_{3}SN{P}_3+\epsilon ,$$

where each SNP effect $\beta$ is generated from a standard normal distribution. We then simulated three SNPs ($SN{P}_4$, $SN{P}_5$, and $SN{P}_6$) in LD with the causal SNPs ($SN{P}_1$, $SN{P}_2$, and $SN{P}_3$), respectively. Each of the three pairs of SNPs are generated from a multivariate binomial distribution with 4 different correlations ($r=0,0.3,0.6,\text{or}0.9$), and MAF of all SNPs were set to be 0.2. The residual is generated from a normal distribution with zero mean and standard deviation based on the $h^2$ of protein A, which ranges among 0.05, 0.1, 0.2, and 0.5. We simulated 1,000 individuals for each combination of LD and $h^2$, and used 500 individuals to estimate the effect of the three SNPs in LD with the causal SNPs:

$${prot}_A={\beta }_{4}SN{P}_4+{\beta }_{5}SN{P}_5+{\beta }_{6}SN{P}_6+e.$$

We then predicted the protein expression of the remaining 500 individuals. We simulated a protein B with 3 causal SNPs in a similar fashion as protein A, except that we estimated the SNP effect and predicted the protein expression using the true causal SNPs, which are generated from independent binomial distribution with MAF = 0.2. The $h^2$ of protein B was set to be the same as protein A. Then the interaction $\left({\beta }_I\right)$ between the predicted expressions of the two proteins on protein A is tested through OLS using the following model:

$${\mathit{obs}}_A={\beta }_A{\mathit{pred}}_A+{\beta }_B{\mathit{pred}}_B+{\beta }_I{\mathit{pred}}_A\times {\mathit{pred}}_B+e.$$

For each combination of LD and $h^2$, we repeated the above simulation 100 times, and recalculated each interaction $P$ value using SVE. We then conducted 20 such 100-time simulations to obtain the distribution of $\lambda$.

To show the efficacy of SVE under more realistic settings, we performed additional simulations using the genotypes of 3,000 European individuals randomly selected from the UK Biobank. In each round of the simulation, we randomly selected a protein-coding gene, and then randomly selected 5 causal SNPs, within the 20 Kb of the TSS. The 5 tagged SNPs are selected such that each is correlated ($\left|r\right|>0.2$) with at least one of the causal SNPs. The overall LD between the tagged SNPs and causal SNPs is calculated as the average $r^2$ of the pairwise causal and tagged SNPs. The 3,000 individuals are divided into two equal sets for training and testing, respectively. The rest of the simulation procedures are similar as above.

Detection of epistasis at the protein level

To detect protein-based cis-by-trans epistatic effects, we tested interactions between the cis-predicted expression levels of protein A and protein B, where A and B are encoded by genes located on different chromosomes. For each protein A, we tested the interaction effect using the following model plus SVE correction:

$${\mathit{obs}}_A={\beta }_A{\mathit{pred}}_A+{\beta }_B{\mathit{pred}}_B+{\beta }_I{\mathit{pred}}_A\times {\mathit{pred}}_B+e,$$

where ${\text{obs}}_A$ is the observed expression of protein A adjusted for age, sex, UKB center, UKB genetic array and the first 20 genetic principal components, ${\mathit{pred}}_A$ is the cis-predicted expression of protein A, ${\mathit{pred}}_B$ is the cis-predicted expression of protein B, which is trans to protein A and on different chromosomes. ${\beta }_A$, ${\beta }_B$, and ${\beta }_I$ are the additive effects of ${\mathit{pred}}_A$ and ${\mathit{pred}}_B$, and the interaction effect, respectively. $e$ is the residual. Significance was assessed using a two-sided $t$-test. To account for multiple testing across all tested protein pairs, we calculated the $q$ value²⁰ to control the false discovery rate (FDR) at 5%. The epistasis detection was performed exclusively in the individuals not used in the FUSION model training.

Detection of epistasis at the SNP level

For each of the seven significant protein–protein epistatic pairs, we performed SNP-by-SNP interaction testing between all pairwise cis-regulatory SNPs of the target protein and cis-regulatory SNPs of the regulator protein. SNPs were those selected by FUSION within 500 kb of the TSS. We applied the following model plus SVE to control inflation:

$$ob{s}_A={\beta }_{sA}SN{P}_A+{\beta }_{sB}SN{P}_B+{\beta }_{sI}SN{P}_A\times SN{P}_B+e,$$

where ${SNP}_A$ and ${SNP}_B$ are the cis-regulatory SNPs of target and regulator proteins, and ${\beta }_{sA}$, ${\beta }_{sB}$, and ${\beta }_{sI}$ are the effects of $SN{P}_A$, $SN{P}_B$, and the interaction, respectively. We calculated the $q$ value to control the FDR at 5% for all tested SNP pairs.

Enrichment of epistatic in chromatin states

We annotated all cis-regulatory SNPs used in the SNP-level epistasis analysis with chromatin states of peripheral blood mononuclear cells from the Roadmap Epigenomics Consortium²¹ (https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/imputed12marks/jointModel/final/). We matched the effect size of each epistatic SNP and a non-epistatic SNP, i.e., the difference of the effect size is smaller than 10% to avoid the bias. Then the enrichment is calculated as the ratio between the proportion of epistatic SNPs and the proportion of non-epistatic SNPs in each chromatin state. To obtain the significance of the enrichment, we first generated 1,000 bootstrapped samples using the R package ‘boot’ (v.1.3.28.1)²², then used the following equation to calculate an empirical $P$ value:

$$P_{\mathit{empirical}}=2\times \mathrm{m}\mathrm{i}\mathrm{n}\left(P_{\mathit{1side}},1-P_{\mathit{1side}}\right),$$

where $P_{\mathit{1side}}$ is the proportion of the enrichment > 1 across 1,000 bootstrapped samples. When $P_{\mathit{1side}}=0$, we declare a $P<0.001$.

Simulations of PWAS incorporating epistasis

We performed simulations to evaluate whether including epistatic components in protein expression prediction could lead to spurious associations in PWAS. We simulated a target protein and a regulator protein for 20,000 individuals. We assumed that the regulator protein B was only impacted by three independent cis-SNPs drawn from a binomial distribution with equal MAF of 0.2:

$${prot}_B={\beta }_{4}SN{P}_4+{\beta }_{5}SN{P}_5+{\beta }_{6}SN{P}_6+\epsilon .$$

The effect sizes of each SNP were drawn from a standard normal distribution. The Gaussian noise was scaled such that the cis-$h^2$ is 0.2. We assumed that the target protein was impacted by a cis-regulatory component, a trans-regulatory component, and a cis-by-trans interaction effect:

$${prot}_A={cis}_A+{\beta }_{\mathit{trans}}{prot}_B+{\beta }_I{cis}_A{prot}_B+\epsilon .$$

The cis-regulatory component was simulated to be impacted by three independent cis-SNPs drawn from a binomial distribution with equal MAF of 0.2:

$${cis}_A={\beta }_{1}SN{P}_1+{\beta }_{2}SN{P}_2+{\beta }_{3}SN{P}_3.$$

The effect sizes of each SNP were drawn from a standard normal distribution. The effect size of the regulator protein B and the interaction were scaled such that the proportion of heritability explained by cis, trans, and interaction is 1:8:1. The cis to trans proportion reflects the average proportion of heritability explained by cis and trans effects in the plasma proteome²³. The Gaussian noise was scaled such that the overall $h^2$ is 0.2. We standardized both target protein A and regulator protein B to simulate a phenotype:

$$y={\beta }_A{\mathit{prot}}_A+{\beta }_B{\mathit{prot}}_B+\epsilon .$$

In the null case where the effect of target protein A on the phenotype is zero, we set ${\beta }_A$ equal to 0, ${\beta }_B$ equal to 5 and the variance of the independent Gaussian noise equal to 100, such that the proportion of phenotypic variance explained by the protein expression is 0.2.

Then the data were divided into a training set ($n=\mathrm{5,000}$), a testing set ($n=\mathrm{5,000}$), and an independent PWAS set ($n=\mathrm{10,000}$) for association testing. We first fitted separate linear models to predict target and regulator proteins from their respective cis-SNPs using the training set:

$${prod}_A={\beta }_{1}SN{P}_1+{\beta }_{2}SN{P}_2+{\beta }_{3}SN{P}_3+e,$$

$${prot}_B={\beta }_{4}SN{P}_4+{\beta }_{5}SN{P}_5+{\beta }_{6}SN{P}_6+e.$$

In the testing set, we obtained cis-predicted components of proteins A (${pred}_A$) and B (${pred}_B$) using the SNP effect size estimated from the training set, and then fitted the following models to obtain the effect of each genetic component on the target protein A:

$${prot}_A={\beta }_{cis}{pred}_A+{\beta }_{trans}{pred}_B+{\beta }_I{pred}_A\times {pred}_B+e.$$

Note that in this step, centering the cis-predicted components before fitting the model could minimize the PWAS inflation. In order to show this, we fitted the model using ${pred}_A$ and ${pred}_B$ with or without centering and performed the following PWAS.

In the PWAS set, we obtained the cis-predicted components of proteins A (${pred}_A$) and B (${pred}_B$) using the effect estimated from training set, and the predicted target protein A with or without the trans effect:

$${pred}_{Aepi}=\widehat{{\beta }_{cis}}{pred}_A+\widehat{{\beta }_{trans}}{pred}_B+\widehat{{\beta }_I}{pred}_A\times {pred}_B,$$

$${pred}_{Aepi}=\widehat{{\beta }_{cis}}{pred}_A+\widehat{{\beta }_I}{pred}_A\times {pred}_B.$$

Finally, we performed PWAS by regressing the phenotype $y$ on the predicted target protein value (${pred}_{Aepi}$) in the PWAS set and recorded the $P$ values. We repeated the above simulations 1,000 times to obtain the genomic inflation factor, $\lambda$. Then 100 such 1,000-time simulations were conducted to obtain the distribution of $\lambda$. We calculated the false positive rate as the proportion of simulations with $P<0.05$.

We also performed causal simulations where the effect of target protein on the phenotype is equal to 2.

PWAS across UK Biobank traits

We performed protein-wide association studies for 124 complex traits from the UK Biobank (Table S1) using both standard cis-only protein predictions and epistasis-adjusted predictions. For each protein–trait pair, we first removed individuals used in the protein prediction to avoid data leakage, and then fitted a linear (for continuous traits) or logistic regression (for binary traits) adjusted for age, sex, and the top 20 genetic principal components.

Prediction of protein expressions

To investigate whether the incorporation of epistatic effects could improve the accuracy of protein expression predictions, we compared the accuracy of protein levels predicted using cis-regulatory components alone with that of epistasis-adjusted predictions using both cis-regulatory and epistatic components. To avoid data leakage during epistatic effect estimation, we performed a 10-fold cross-validation using only the individuals not included in the FUSION model training. For each fold, the interaction model was trained on nine folds and applied to the held-out fold to generate out-of-fold predicted protein expression for the downstream PWAS. This process was repeated 10 times for each held-out validation set. Specifically, for prediction using cis-regulatory components alone, we estimated the additive effects of cis-regulatory components using the training set:

$${\mathit{obs}}_A={\beta }_A{pred}_A+{\beta }_B{pred}_B+e,$$

where ${\text{obs}}_A$ is the observed expression of target protein, ${pred}_A$ and ${pred}_B$ are FUSION predictions of regulator and target protein. We then use the estimated ${\beta }_A$ to predict protein expression in the validation set as $\widehat{{\beta }_A}{pred}_A$.

Similarly, for the epistasis-adjusted prediction, we estimated the additive and epistatic effects in the training set as follows:

$${\mathit{obs}}_A={\beta }_A{\mathit{pred}}_A+{\beta }_B{\mathit{pred}}_B+{\beta }_I{\mathit{pred}}_A\times {\mathit{pred}}_B+e.$$

We then predicted protein expression in the validation set as $\widehat{{\beta }_A}{pred}_A+{\widehat{\beta }}_I{pred}_A\times {pred}_B$. The prediction accuracy was evaluated as the squared Pearson correlation between the predictions and observations in the validation set. Note that we performed the cross-validation conditional on the trans-additive effects of the regulator protein to ensure a fair comparison.

Enrichment of ABO interacting genes in Tajima’s D scores

We obtained the Tajima’s D track from the UCSC Genome Browser (https://hgdownload.soe.ucsc.edu/goldenPath/hg18/database/), which reports the Tajima’s D measure of each 10-kb window²⁴. We computed Tajima’s D for each gene as the average of the 10-kb windows overlapping with the gene body based on hg18 coordinates.

To calculate the enrichment of ABO-interacting genes in high Tajima’s D (> 1.5), we compared the proportion of significant ABO-interacting genes with high Tajima’s D to the proportion among non-significant ABO-interacting genes. We used five different interaction $P$ value thresholds (0.1, 0.05, 0.01, 0.001, and 1.7 × 10⁻⁷), and the last threshold corresponds to FDR < 0.05 used to define the significant epistasis at the protein level in this study.

Results

Detecting cis-by-trans interaction effects in UKB-PPP

A genome-wide search for variant–variant interaction effects across all proteins entails a substantial multiple-testing burden, severely limiting statistical power. To address this challenge, we adopted two complementary strategies to reduce the search space while preserving biologically meaningful interaction signals (Figure 1A). First, for each protein, we restricted our analysis to interactions between cis-regulatory and trans-regulatory genetic components. This interaction structure can avoid spurious interactions arising from LD between the interacting variants. More importantly, it is biologically motivated: trans-acting influences such as transcription factors are known to exert genome-wide effects that can potentiate, silence, or modulate the activity of the regulatory elements impacted by cis-acting genetic variation (e.g., promoters or enhancers). Variants affecting these trans-regulatory mechanisms are thus natural candidates for the genetic background effects implicated in epistasis. Focusing on cis-by-trans interactions therefore targets a class of epistatic effects that are both interpretable and mechanistically plausible, while substantially reducing the number of tested interactions. Second, rather than testing interactions between individual variants, we implemented a protein-based interaction framework that aggregates the effects of multiple variants within cis-regulatory regions. This strategy parallels the rationale of transcriptome-wide association studies (TWAS)^17,25, in which gene-level genetic predictors are used to improve power and robustness. By operating at the protein level, we further reduce the multiple-testing burden and mitigate the impact of weak, noisy variant-level interaction effects.

Figure 1. Protein-based and inflation-controlled detection of cis-by-trans epistatic effects of protein expression levels.

(A) Schematic of our pipeline to identify cis-by-trans epistatic effects of protein expression.

(B) The genomic inflation factor ($\lambda$) of the interaction $P$ values with or without applying sandwich variance estimator (SVE) using simulated and UK Biobank (UKB) genotypes. The dots represent the average inflation across $4h^2$ groups × 20 replicates (inflations of each $h^2$ group are shown in Figure S3), and error bars represent the 95% confidence interval (Methods). Detailed summary statistics used to plot the figure are shown in Table S2.

(C) Seven significant pairs of epistatic effects. The arrow points from the trans regulator protein to (red) the target protein (blue).

Specifically, we first predicted the expression level of each protein assayed in UKB-PPP European individuals (N = 49,024)¹³ using FUSION¹⁷ based on cis-SNPs (SNPs within 500 Kb from the TSS of the gene encoding the protein). We conducted a training-testing split to predict protein expression (Methods). Among the 2,818 proteins we analyzed, 1,208 had a Pearson’s correlation between prediction and observation greater than 0.1 ($R^2>0.01$), and were retained for subsequent analyses. On average, cis-$h^2$ across those proteins is 0.13, and $R^2$ is 78% of the cis-$h^2$ (average $R^2/cis-h^2$ across all genes; Figure S1A), which suggests a decent predictive accuracy. As a secondary analysis, we compared the prediction performance of the 1,208 proteins between the training and testing sets (Figure S1B). As expected, the average R² was slightly higher in the training set than in the testing set (0.13 vs. 0.12; Wilcoxon P = 5.3×10⁻⁵), indicating that the prediction models generalize well to independent individuals with minimal overfitting. Next, for each protein A, we test the interaction between the cis-genetic component of A and a potential trans-regulatory effect on A—in this case, the cis-genetic component of another protein B located on a different chromosome (Figure 1A):

$${\mathit{obs}}_A={\beta }_A{\mathit{pred}}_A+{\beta }_B{\mathit{pred}}_B+{\beta }_I{\mathit{pred}}_A\times {\mathit{pred}}_B+e,$$

where ${\text{obs}}_A$ is the observed expression of protein A adjusted for age, sex, UKB center, UKB genetic array and the first 20 genetic principal components, pred_A is the cis-predicted expression of protein A, ${\mathit{pred}}_B$ is the cis-predicted expression of protein B, ${\beta }_A$, ${\beta }_B$, and ${\beta }_I$ are the additive effects of ${\mathit{pred}}_A$ and ${\mathit{pred}}_B$, and the interaction effect, respectively, and $e$ is the residual.

The detection of epistatic effects is highly susceptible to false positives, of which a major source is LD¹⁵, particularly when the SNPs included in the model are tagging variants rather than causal variants. Using simulations, we show that naive ordinary least squares (OLS) estimation of the interaction coefficient ${\beta }_I$ can be severely inflated under these conditions. In our simulations, a protein is causally regulated by a single SNP ($SN{P}_1$). We also simulated a tagging variant ($SN{P}_2$) in LD with $SN{P}_1$ at varying LD levels and an independent variant ($SN{P}_3$), with no true interaction effects among variants. When $SN{P}_1$ was excluded from the regression model, we observed substantial inflation of the estimated interaction effect between $SN{P}_2$ and $SN{P}_3$ under OLS, quantified by the genomic inflation factor $\lambda$, despite the absence of any true epistasis (Figure S2). We further demonstrate that this LD-induced inflation extends to protein-level interaction analyses based on predicted protein expression. Specifically, interactions between predicted expression levels of proteins A and B—constructed from cis-regulatory SNPs—exhibited inflated interaction effects on the observed expression of protein A when the causal cis-SNP was not explicitly modeled (Figure 1B; Table S2; Methods).

To control false positive inflation due to LD, we used the SVE method¹⁹, instead of the OLS estimator. The SVE method effectively controls the inflation by accounting for heteroskedasticity, which is a common feature of G×G and G×E interaction models¹⁶. Using simulations, we demonstrate that the inflation of OLS estimates becomes more pronounced as SNP effect sizes increase (increasing $h^2$). We also observed inflation of the OLS estimates across different LD ($r^2$ between 0.3 and 0.9) between the tagging and causal SNPs (Figure S2). SVE successfully minimizes the inflation in both SNP-by-SNP interactions (Figure S2) and predicted expression by expression interactions using both simulated genotypes and real genotypes from the UK Biobank (Figure 1B and Figure S3; Table S2; Methods).

Using all the pairwise protein-level interactions from the 1,208 proteins with good prediction accuracy, we identified seven significant pairs of epistatic effects using SVE after controlling the FDR at 5% across all tested pairs (Figure 1C; Table S3). All but one of these interactions fell within a single connected network, five of which involved ABO as the epistatic regulator (Figure 1C). In contrast, and consistent with the inflation observed for the standard OLS model in our simulations, OLS identified 2,248 significant interactions (Figure S4A and Figure S4B), accompanied by a clear test statistic inflation (genomic inflation λ = 1.1). Given the agreement between empirical and simulated inflation under OLS, the vast majority—if not all—of these additional signals are likely false positives. This pronounced discrepancy underscores the necessity of careful statistical modeling to avoid spurious detection of epistatic effects. We next sought to replicate these seven associations in UKB-PPP samples of African ancestry (n = 1,204; Table S3). The most significant interaction from the discovery analysis, between ABO and MBL2 (MIM: 154545), was replicated (P = 0.0031, replication P threshold = 0.05/7). The lack of replication for the remaining six effects is likely attributable to the considerably smaller sample size of African ancestry cohort, making the replication of the ABO-MBL2 interaction a notable finding. We examined the sign consistency of the interaction effects between the two populations (Figure S4C). Four of the seven interactions (MBL2–ABO, CD209 [MIM: 604672]–ABO, ICAM5 [MIM: 601852]–ABO, and MBL2–OBP2B [MIM: 604606]) showed consistent directions of effect. The remaining three interactions were far from nominal significance (P > 0.1) in the African ancestry dataset, making it difficult to meaningfully assess sign consistency because their effect estimates are highly uncertain.

To examine whether different definitions of trans regions have impacts on the result, we performed additional epistasis analysis to include intra-chromosomal cis-trans pairs that are at least 10 Mb apart. We found no additional significant signals (multiple-testing correction across all protein pairs), and all seven inter-chromosomal pairs remained significant (FDR < 0.05).

Epistatic effects at the SNP level

After identifying protein-by-protein epistasis, we sought to determine which SNPs are responsible for these epistatic effects and if these SNPs are enriched for regulatory functions, such as enhancers or promoters. We performed pairwise SNP-by-SNP interactions underlying the seven significant protein-protein epistatic pairs, and identified a total of 1,304 significant pairs of SNP-SNP interactions (FDR < 0.05; Table S4), where the effect of a target protein’s cis-SNP varied substantially depending on the dosage of the regulator cis-SNP, and vice versa.

Next, we examined the functional context of epistatic SNPs based on our observation that SNPs exhibiting significant epistasis had larger additive effects on protein abundance than other cis-regulatory SNPs without significant epistasis (Figure 2A). This pattern suggests that epistatic effects involving cis-variants with small additive effects may be underpowered in our study, and that detectable epistasis preferentially involves regulatory loci with substantial functional impact. To characterize the functional context of detected variants, we analyzed the chromatin states of the genomic loci in which they reside²⁶, looking for enrichment in annotations linked to regulatory functions. Specifically, we examined annotations in blood mononuclear cells²¹, performing enrichment analysis on the chromatin states of epistatic cis-variants against a background of additive cis-variants matched for marginal effect size. Under this controlled comparison, we observed a modest but significant depletion of epistatic SNPs in the “transcription” state (Figure 2B; enrichment = 0.18, bootstrapping P < 0.001), corresponding to ongoing transcription rather than to regulatory elements controlling transcription initiation. This indicates that cis-variants involved in cis-by-trans epistasis are less likely than additive cis-variants to act through variation in gene-body regions associated with transcriptional elongation, which is generally less sensitive to trans-regulatory variation than upstream regulatory steps²⁷. We also observed a 1.2-fold enrichment (Figure 2B; bootstrapping P = 0.04) of epistatic SNPs in Quiescent/low annotation, which correspond to regions lacking any chromatin marks in blood cells. This finding suggests that epistatic regulation in plasma proteins may involve regulatory mechanisms that are not captured by baseline chromatin activity in blood cells, potentially including cell-type–specific regulatory elements or co- or post-transcriptional processes. Contrary to our expectations, we did not observe significant enrichment or depletion in any regulatory functional elements, such as enhancers or promoters, relative to non-epistatic cis-regulatory SNPs (Figure 2B; Table S5). However, given the limited number of detected epistatic variants, this analysis may be underpowered to detect moderate enrichment in these categories. Importantly, the absence of enrichment does not necessarily imply that promoters or enhancers are uninvolved, but rather that it is difficult to distinguish epistatic cis-QTLs from additive cis-QTLs of similar effect size by canonical regulatory annotations alone.

Figure 2. SNP-by-SNP interactions on protein expression.

(A) The absolute effect size of cis-regulatory SNPs with significant SNP-level epistasis and those without. The $P$ value is based on a two-sided Wilcoxon test. In each box, the center line represents the median, box limits represent the upper and lower quartiles, whiskers represent 1.5 × interquartile range, and individual points are outliers.

(B) Functional enrichment of cis-regulatory SNPs with significant SNP-level epistasis compared to those without significant epistasis after matching the effect size (Methods). Chromatin states are annotated based on blood mononuclear cells from Roadmap²¹. The error bar represents a 95% confidence interval based on bootstrapping over 1,000 sampling iterations. * denotes bootstrapping P < 0.05 after Bonferroni correction, and *** denotes P < 0.001 (Methods). Detailed enrichment values are shown in Table S5.

Limited PWAS performance after incorporating epistatic effects

In a typical PWAS framework, protein levels are predicted with cis-SNPs only, and the predicted expression levels are used to test for associations with complex traits. In recent years, we noted emerging approaches in the field to incorporate SNPs beyond cis-SNPs^28–30. The detected cis-by-trans epistasis effects can also be incorporated in the PWAS framework. However, we suspect that naively extending the PWAS framework to include trans and/or epistatic effects introduces additional modeling challenges. If not carefully handled, these issues can lead to spurious associations. One major potential source of false positives is the inclusion of additive effects from trans regulators in expression prediction models. Because trans regulators can directly influence complex traits independently of the target protein, their inclusion can induce horizontal pleiotropy, leading to false positive protein-trait associations. We modeled this effect using null simulations in which the phenotype was affected only by the trans regulator but not by the target protein (Methods). In this setting, including trans-regulatory additive effects resulted in biased effects between the target protein and the phenotype (Figure 3A) and severe inflation of false-positive associations (Figure 3B). We performed additional simulations of different causal effects between the target protein and the phenotype, and consistently observed biased estimates (Figure 3A).

Figure 3. Epistasis-adjusted protein-wide association studies (PWAS).

(A) Adding trans-regulator additive effects to the PWAS model leads to false positives in 1,000 simulations. In the null simulation, the phenotype is affected only by the regulator protein (true effect of target protein = 0). In the causal simulation, the phenotype is affected by both the regulator protein and the target protein (true effect of target protein = 2). The association coefficient between the predicted target protein and phenotype is not biased when predicting the target without the trans-regulator additive effects in both simulations, but severely biased when including them.

(B) The genomic inflation factor of different null simulation settings (true effect = 0) with 100 replications. Predicting the target adding the trans effects (left bar) or excluding the trans effects but without centering leads to inflation of P values (middle, blue bar). Mean-centering the cis-predicted expression removes the inflation of PWAS association P values (right, orange bar).

(C) Comparison of P values between PWAS with epistatic components ($P_{\text{epi}}$) and PWAS without ($P_{\text{raw}}$) across 124 UK Biobank traits. Detailed P values of PWAS after excluding trans additive effects and mean-centering are shown in Table S6

(D) Prediction accuracy of target proteins using only cis-regulatory SNPs and using both cis-regulatory SNPs and interaction effects (Methods).

In each box of (A) and (B), the center line represents the median of simulations, box limits represent the upper and lower quartiles, whiskers represent 1.5 × interquartile range, and individual points are outliers.

A second source of inflation is the failure to mean-center predicted expression levels when testing for associations. In the same set of null simulations, we removed trans-regulators from the prediction model and predicted expression of the target protein with only the additive cis-component and interaction as follows:

$${pred}_{Aepi}={\beta }_A{pred}_A+{\beta }_I{pred}_A\times {pred}_B,$$

where ${pred}_A$ and ${pred}_B$ denote cis-predicted expression levels of proteins A and B, respectively. When ${pred}_A$ and ${pred}_B$ were not mean-centered, we observed unbiased estimates (Figure 3A), but systematic inflation of PWAS association P values (Figure 3B), leading to false positive discoveries (Figure S5A). Mean-centering ensures that the interaction predictor represents a pure non-additive component. Specifically, the uncentered interaction term can be written as ${pred}_A\times {pred}_B=\left({pred}_A-{\mu }_A\right)\left({pred}_B-{\mu }_B\right)+{\mu }_A\left({pred}_B-{\mu }_B\right)+{\mu }_B\left({pred}_A-{\mu }_A\right)+{\mu }_A{\mu }_B$, where ${\mu }_A$ and ${\mu }_B$ are the means of ${pred}_A$ and ${pred}_B$. Thus, the uncentered interaction term contains additive cis $\left({\mu }_B\left({pred}_A-{\mu }_A\right)\right)$ and trans $\left({\mu }_A\left({pred}_B-{\mu }_B\right)\right)$ components in addition to the interaction component $\left({pred}_A-{\mu }_A\right)\left({pred}_B-{\mu }_B\right)$. Because the interaction term is subsequently incorporated into the PWAS predictor, these additive components can induce inflation in downstream association tests by allowing additive trans genetic effects to be incorporated into the interaction-based predictor. Mean-centering removes these additive components and isolates the deviation-from-additivity component.

To address these issues, we implemented a conservative epistasis-aware PWAS framework that first predicts the expression of protein A without the trans regulatory component to avoid horizontal pleiotropy, and then mean-centers the cis-predicted expression to obtain the predicted expression of A (${pred}_{Aepi}$). Under this specification, PWAS effect size estimates in simulations were unbiased (Figure 3A), and P values were well calibrated, indicating effective control of false positives (Figure 3B; Figure S5A). We did not pursue more complex Mendelian randomization strategies to account for horizontal pleiotropy, as our focus was on interaction effects rather than trans-mediated causal inference.

We then applied this framework to real data, incorporating the seven epistatic effects we identified through our genome-wide search. We tested associations between epistasis-adjusted protein expression predictions and a phenome-wide collection of 124 complex traits from the UK Biobank (Table S1), including a wide range of phenotypes from those closely related to plasma proteins (blood cell, disease antigens in blood, and blood biochemistry) and those less related (anthropometric and mental health). Models that included additive effects from trans-regulators or omitted mean-centering both identified multiple novel significant associations between proteins and complex traits, compared to cis-only PWAS (Figure 3C). However, given that both approaches exhibited inflated test statistics in simulations, these additional associations are likely driven by model misspecification. In contrast, our epistasis-aware and well-calibrated model did not identify any novel protein–trait associations beyond those detected by standard PWAS (Figure 3C; Table S6).

This absence of additional discoveries is likely attributable to the limited contribution of detectable epistatic effects to overall protein expression variance. Although statistically significant, individual epistatic interaction effects are much smaller relative to additive genetic effects (Table S3), resulting in minimal improvements in prediction accuracy when epistasis was incorporated (Figure 3D). Consistent with this interpretation, we found in our simulations that the power gain obtained by including epistatic interactions increases with the proportion of heritability explained by epistasis (Figure S5B). Together, these results indicate that while epistatic interactions can exert measurable effects on protein regulation, their impact does not necessarily propagate to the examined complex traits at a detectable level. Naively incorporating such interactions into a PWAS framework may appear to improve discovery power, but this apparent improvement will consist mostly or entirely of false positives except in cases where epistasis accounts for a substantial fraction of heritability. However, the use of a properly controlled PWAS model that excludes additive trans-regulator effects and mean-centers cis-predicted expression eliminates these spurious associations.

ABO-centered epistatic regulation of protein expression

Six of our seven significant interactions appear to be linked within a discrete network, and five involve ABO as the trans-acting influence, nominating it as a master regulator that modifies the cis-effects of multiple proteins. ABO is a glycosyltransferase enzyme whose canonical function is creating the red blood cell antigens that form the basis of the ABO blood system. However, the epistatic network centered on this protein suggests that it might also serve a broad regulatory role within the plasma proteome.

Alleles of ABO can be categorized as A, B, or O, and an individual’s specific combination of alleles determines their ABO blood type. We observed that the number of O alleles carried by individuals is negatively associated with ABO expression levels in UKB-PPP (Figure 4A). Stratifying individuals by O blood type, we further found that the cis-effects on expression of the target proteins (CD209, FCGR2B, ICAM5, MBL2, and NPTX1) differ significantly across blood-type backgrounds (Figure 4B; Table S7). This observation suggests that ABO genotype modulates the magnitude of cis-regulatory effects of target proteins rather than acting through purely additive trans effects.

Figure 4. ABO blood type changes ABO protein expression level and cis-effects of multiple target proteins.

(A) ABO expression levels by blood type haplotypes across 32,323 individuals. Predicted expression level is based on FUSION predictions. In each box, the center line represents the median, box limits represent the upper and lower quartiles, whiskers represent 1.5 × interquartile range, and individual points are outliers.

(B) Five proteins with significant epistatic effects with ABO exhibit different cis-effects between O and non-O blood types. The cis-regulatory effect is the additive effect of the predicted protein expression on the observed expression (estimated ${\beta }_A$ in Figure 1A). The points show the estimation of cis-regulatory effects, and error bars represent the standard error. Detailed cis-effects are shown in Table S7.

(C) The variance of MBL2 expression differs across ABO allele dosages.

(D) Proteins having smaller interaction P values with ABO are more enriched in genes with high Tajima’s D (Methods). P < 1.7 × 10⁻⁷ (FDR < 0.05) is the P value threshold to identify significant epistasis in this study.

The error bars in (C) and (D) show 95% confidence intervals determined by bootstrapping over 1,000 sampling iterations, and points show mean values of bootstrapped samples.

The effects of blood type can also be observed at the level of SNP-by-SNP interactions. For example, the SNP rs73008220 (cis to CD209) exhibited modified cis-effects when interacting with the ABO SNP rs8176719 (interaction P = 3.1 × 10⁻¹⁶; Figure S6). The latter is one of the primary loci that determines blood type—when both alleles have no insertion, the blood type is O, while a C insertion creates a non-O blood type³¹. In individuals with an SNP dosage of 0 for rs8176719 (i.e., blood type O), the cis-effect strength of rs73008220 is −0.01 (P = 8 × 10⁻⁴); in contrast, when the SNP dosage is 1 or 2 (non-O blood type), the cis-effects are much stronger at −0.06 (P = 5.7 × 10⁻²⁹) and −0.08 (P = 3.4 × 10⁻¹⁷), respectively.

In addition to blood type–dependent differences in the strength of cis-effects on mean expression, we observed significant differences in expression variance across ABO genotypes for proteins in epistasis with ABO. For example, the aforementioned ABO SNP rs8176719 stratified expression variance of MBL2, a key component of the lectin complement pathway that is involved in pathogen recognition and innate immune activation. Specifically, the expression variance of MBL2 was markedly reduced in individuals with an rs8176719 genotype associated with blood type O (SNP dosage = 0) compared with others (SNP dosage = 1 or 2) (Figure 4C). Together with our finding that the O blood type attenuates cis-effects on MBL2, this constraint on protein expression variance indicates that the O blood type defines a regulatory context in which MBL2 expression levels are buffered and robust to cis-genetic variants. In contrast, non-O blood types permit larger cis-effects and greater expression variability, corresponding to a more responsive but less constrained regulatory regime. We observed a similar pattern for the other four proteins that displayed significant interaction with ABO (Figure S7). These results imply a trade-off between robustness and responsiveness, suggesting that both modes of cis-regulation—and thus both O and non-O genetic backgrounds at the ABO locus—may be advantageous under different environmental contexts.

Intriguingly, this pattern is compatible with models of balancing selection, of which the ABO locus is a canonical example³², with multiple alleles maintained at intermediate frequencies across diverse human populations^32–35. We therefore asked whether these epistatic buffering effects are accompanied by signatures of long-term balancing selection. We determined that the ABO locus shows elevated Tajima’s D (Tajima’s D = 1.32 based on the entire gene body), consistent with balancing selection at this locus. We then examined whether genes affected by ABO’s epistatic buffering effects on protein expression were similarly enriched for high Tajima’s D values. After stratifying all proteins into potentially ABO-interacting and non-interacting using different interaction P value cutoffs, we used the non-interacting proteins as the background to estimate the enrichment of interacting proteins for high Tajima’s D scores (> 1.5). We found that ABO interacting proteins were increasingly enriched for high Tajima’s D scores with an increasingly stringent P value cutoff (Figure 4D). Overall, we observed a 1.9-fold enrichment of the five significant ABO interacting proteins for elevated Tajima’s D scores (bootstrapping P < 0.001, Figure 4D), consistent with long-term balancing selection^32,36. Together, these results indicate that ABO participates in epistatic regulation that modulates both the mean and variance of protein expression of its interacting partners, potentially giving rise to a two-locus balancing selection framework.

The precise molecular mechanisms underlying these interactions remain unknown, and elucidating them will require targeted functional studies that are beyond the scope of the present work. However, we note that several ABO interaction partners—including MBL2 (Tajima’s D = 1.88), CD209 (Tajima’s D = 2.27), and FCGR2B (MIM: 604590; Tajima’s D = 1.74)—are immune-related genes encoding proteins involved in pathogen recognition or immune regulation^37–39. These processes are sensitive to glycan-mediated interactions. As ABO is a major determinant of cell-surface glycosylation, epistatic interactions between ABO and these immune genes provide a plausible biological framework through which glycosylation variation may modulate immune function in a context-dependent manner.

Discussion

Our study provides a systematic investigation of epistasis in the human plasma proteome and its role in protein regulation and trait association. By narrowing the search to cis-by-trans interactions and developing a robust analytical framework that accounts for inflation in epistasis detection, we identified seven significant protein-protein epistatic pairs with 1,304 SNP-SNP interactions, uncovering non-additive genetic effects in the plasma proteome.

We did not find significant enrichment of our interaction pairs in protein-protein interactions (PPI; data not shown), which was suggested by a recent publication⁴⁰. This is likely due to the small number of interaction pairs identified, which may reflect a lack of statistical power.

We demonstrated that incorporating trans and interaction effects into PWAS without accounting for certain key sources of error results in numerous false positives. For example, including trans-additive-effects in predictions of protein expression levels violates the horizontal pleiotropy assumption of causal inference and leads to biased effect estimates. The inflation is not only limited to epistasis-adjusted PWAS, but also applies to additive-effects-only PWAS. Several researchers have explored adding trans-effects to their prediction models in recent PWAS studies^28–30. Without properly adjusting for horizontal pleiotropy, this type of inference can be prone to false positive association between proteins and complex traits.

When using a principled framework that corrects for these sources of error, we found that incorporating our observed epistatic effects did not improve PWAS performance across the 124 complex traits examined. Several factors may explain this result. First, although statistically significant, the detected epistatic effects provided only minimal improvements in protein expression prediction accuracy, which likely limits their impact on downstream PWAS discovery. While recent findings suggest that genome-wide interaction effects can collectively explain more heritability than additive effects⁴, they are typically weak and widely distributed across the genome. Thus any individual interaction explains only a limited proportion of variance. Second, the set of complex traits from the UK Biobank may not include the specific phenotypes most strongly influenced by the detected epistatic effects. Although many of the analyzed traits are immune-related and therefore biologically relevant to the interacting proteins identified here, it remains possible that the true downstream traits were not captured. Third, the lack of PWAS improvement may reflect a more general attenuation of epistatic effects at the complex-trait level. Even when epistatic interactions exert measurable effects on protein regulation, their influence may be buffered through downstream biological networks, resulting in substantially weaker effects on organismal phenotypes. This interpretation is consistent with prior observations that machine-learning models designed to capture non-additive genetic effects, such as neural networks⁴¹, often yield only minimal performance gains over additive polygenic risk score models.

Through our investigation of the specific proteomic interactions detected by our model, we demonstrated that genetic background can modify cis-regulatory effects and modulate protein expression levels. In interactions involving ABO and many immunorelevant proteins, modulation of protein expression levels is consistent with the previously reported evolutionary context of balancing selection. In the example of ABO and MBL2 interaction, while high MBL2 activity can enhance pathogen recognition and clearance, excessive or dysregulated activation of the lectin complement pathway can increase the risk of inflammatory damage and autoimmune pathology³⁶. By stabilizing MBL2 expression and dampening regulatory perturbations, the O allele may reduce the risk of harmful overactivation of innate immune responses. Conversely, non-O alleles, which allow greater variability and stronger regulatory effects, may facilitate more dynamic immune responsiveness in environments where pathogen pressure favors stronger or more flexible lectin pathway activity. The epistatic buffering effects observed here provide a regulatory mechanism that is compatible with long-standing balancing selection, where no single ABO allele confers uniformly optimal fitness across all regulatory or environmental contexts, thereby reducing directional selection and favoring the long-term maintenance of multiple alleles. This is potentially driven by host-pathogen interactions, immune recognitions, or pleiotropic effects across multiple physiological traits.

Our study has several limitations. First, we focused only on cis-by-trans interactions to reduce multiple testing and thereby improve detection power. This choice is also biology-motivated, as the interaction of trans-regulatory molecular factors with cis-regulatory genomic elements represents a plausible mode of action for epistatic effects. This restriction might be unnecessary with access to datasets with much larger sample sizes or more powerful methods, in which case more exhaustive mapping of epistasis effects would be possible. Second, we only detected seven significant interaction pairs. However, comparable studies of epistasis effects suggest that this is not an abnormally small number of signals. Recent work using deep learning models identified only 15 dominance and epistasis effects in the UKB-PPP (and also identified an interaction signal with ABO)¹⁴. Another study identified only 16 marginally significant epistasis signals across 53 complex traits in the UK Biobank⁴. Third, UKB-PPP includes fewer than 3,000 proteins, a small subset of the entire proteomic profile. Therefore, the interaction effects identified in this study do not reflect the full range of whole-genome epistasis. Additionally, the UKB-PPP composition—primarily consisting of individuals with European ancestry—does not represent the full genetic diversity of the human population. Future studies investigating proteomics data with larger sample size, more comprehensive protein profiles, and more diverse genetic background would be expected to detect additional cis-by-trans interactions. Fourth, the plasma proteome includes proteins from multiple tissue sources, such as blood and liver, thus the epistasis in plasma does not reflect tissue specificity. These limitations are likely to be overcome with more comprehensive proteomic datasets across different tissues and cell types in the future. Fifth, our cis-focused approach to the prediction of protein expression overlooks trans-regulatory effects, which play an important role in protein regulation²³. The incorporation of additive trans-effects into protein prediction requires a more robust statistical framework that is beyond the scope of this work.

In conclusion, we show here that mapping cis-by-trans interaction effects in proteomic data represents a promising framework for discovery and functional interpretation of genetic epistasis. We identify and account for previously unrecognized sources of statistical inflation in the detection of epistatic effects and their application to trait association studies. Finally, we demonstrate that our inflation-corrected approach preferentially detects biologically plausible regulatory interactions, including a suite of trans-effects implicating the ABO locus as master regulator within the immunological proteome.

Data and code availability

Codes used to conduct the analyses in this study are available on GitHub (https://github.com/liangend/proteome_epistasis.git).