Separating Real Post-Transcriptional Regulation from Assay Artifacts in Protein Genetics

When a DNA variant changes how much of a protein circulates in blood, geneticists call it a protein quantitative trait locus, or pQTL. When the same variant fails to change the gene's messenger RNA (mRNA), the usual reading is that regulation happens after transcription: in translation, protein folding, or degradation. That reading underpins a growing body of drug-target work. The liver study of He et al. found that only about half of pQTLs were also expression QTLs (variants that change mRNA levels), and concluded that roughly half of genetic control acts post-transcriptionally.

A second explanation fits the same data. Most large protein studies measure abundance with affinity reagents, meaning antibodies or aptamers that bind the protein. A coding variant that changes an amino acid can change how well the reagent binds without changing how much protein exists. Raffield et al. showed that one-third of pQTL signals in a multi-ethnic cohort are driven by such protein-altering variants. When the two leading platforms disagree on the direction of a genetic effect, a protein-altering variant is involved 84% of the time.

We propose a statistical fine-mapping model that resolves this ambiguity variant by variant. Fine-mapping means assigning each variant a probability of being causal while accounting for linkage disequilibrium, the tendency of nearby variants to be inherited together. Our model jointly analyzes summary statistics from mRNA, from two affinity platforms and, where available, from mass spectrometry, which reads peptides directly and does not depend on binding. Each causal signal it finds carries a posterior label: "transcriptional" (all layers move together), "post-transcriptional" (all protein measurements agree, mRNA does not move), or "measurement artifact" (one platform moves alone). The point is to give a calibrated estimate of how much reported transcript-protein discordance is biology and how much is assay behavior. Low-abundance proteins are the regime where this matters most, because there weak signals, missing values and binding artifacts are hardest to tell apart.

Core Claim

Treating "which measurement layers does this causal variant affect?" as a latent variable inside a sum-of-single-effects fine-mapping model separates two things that current pipelines merge. The first is a coding variant that truly destabilizes a protein, which should shift every protein assay, including mass spectrometry, in the same direction. The second is a coding variant that only disrupts a reagent's binding site, which should shift one platform alone. The model can also tell either case apart from a nearby non-coding regulatory variant in the same linkage block.

The fix used today is to regress out protein-altering allele count, as in Raffield et al.. That fix cannot make these distinctions. It removes true destabilizing effects along with artifacts, and it implicitly assumes the coding variant is the only causal variant at the locus.

Relation to Prior Work

Two lines of work motivate the proposal by exposing the gap. He et al. measured liver proteins with mass spectrometry in 287 samples and quantified the mRNA-protein disconnect. However, they used single-variant regression with no credible sets, and they covered only 1,508 mostly abundant proteins. Raffield et al. compared SomaScan 7k (an aptamer platform) with Olink Explore 3072 (an antibody-pair platform) across 2,157 shared proteins in the Multi-Ethnic Study of Atherosclerosis (MESA). They found a median inter-platform correlation of only 0.30 and traced the discordance to protein-altering variants. They did not fine-map, did not use mass spectrometry, and did not look at mRNA. Our model is designed to answer the question both papers leave open: of the observed discordance, how much is post-transcriptional and how much is epitope artifact?

A second group of papers applies fine-mapping to proteins but ignores platform behavior. Pozarickij et al. ran SuSiE on Olink data from 3,974 Chinese adults and found that 36% of credible sets differed from European ones. Li et al. fine-mapped six cancer GWAS and colocalized them with Olink and SomaScan pQTLs. Neither study corrected for epitope effects. This is a concrete risk: some ancestry-specific credible sets may reflect ancestry-specific allele frequencies of protein-altering variants. Raffield et al. reported 80 such proteins with ancestry-dependent measurement agreement. The BLISS imputation models of Wu et al. pool SomaScan and Olink and report that the two platforms capture complementary associations. We would test whether part of that complementarity is artifact.

The method itself builds on the SuSiE family. SuSiE-RSS fits a sum of single-effect regressions to z-scores and a reference correlation matrix. SuSiE 2.0 adds a modular implementation, the adaptive-shrinkage variant SuSiE-ash, and the simxQTL simulator. Funmap and SparsePro show that annotation priors help fine-mapping. Funmap also warns that fixed-effect annotation priors inflate false discoveries. We take that warning as a constraint, so our coding-variant prior is a single random-effect weight rather than a high-dimensional annotation set.

The mass spectrometry arbiter relies on library-free deconvolution of data-independent acquisition spectra. In this acquisition mode the instrument fragments all peptides in a mass window at once, and software has to untangle the mixed signal afterwards. Dear-DIA and DIAmeter both recover low-intensity peptides this way. The single-cell translation model scChord addresses discordance at the cell level with no genetics. We use it only as a reminder that dropout zeros mimic biological discordance, a confound our likelihood must handle through per-layer standard errors.

Research Program

Let $z_k$ be the vector of association z-scores at a locus for measurement layer $k$. The layers are mRNA ($m$), SomaScan ($s$), Olink ($o$) and, optionally, mass spectrometry ($ms$). Let $R$ be the linkage correlation matrix. Each of $L$ single effects $l$ selects one causal variant through an indicator vector $\gamma_l$ and one configuration $c_l$ from a small fixed set. The set contains transcriptional, post-transcriptional, SomaScan-only artifact, Olink-only artifact, and mass-spectrometry-only artifact. Each configuration fixes a binary layer mask $w_{c}$. The effect of single effect $l$ on layer $k$ is

$$ b_{lk} = \gamma_l \, w_{c_l k} \, \beta_{lk}, \qquad z_k \approx \sum_{l=1}^{L} R\, b_{lk} / s_k, $$

where $\beta_{lk}$ is a layer-specific effect size, $s_k$ is that layer's standard error vector, and the approximation is the SuSiE-RSS summary-data likelihood applied per layer.

The effect sizes are not left fully free. In concordant configurations, the $\beta_{lk}$ across layers share a correlated normal prior, so the layers are expected to move together. The prior probability of the artifact configurations is raised for variants annotated as protein-altering, through one learned random-effect weight. Because the standard errors $s_k$ enter the likelihood, a weak protein signal for a low-abundance protein is treated as uninformative rather than as evidence that the protein does not move. This is what keeps measurement noise from being counted as discordance.

Fitting extends iterative Bayesian stepwise selection: each update sums over variants and configurations jointly, which costs about five times plain SuSiE per locus. We would implement it on the SuSiE 2.0 modular code.

The outputs per locus are credible sets, each annotated with configuration probabilities. Aggregating these gives a protein-wide estimate of the fraction of discordant signals attributable to artifact, stratified by protein abundance. Layers estimated in different cohorts need linkage matrices of matching ancestry. Linkage mismatch is a known failure mode of summary-data fine-mapping (SuSiE-RSS), so we run the standard allele-flip and consistency diagnostics before fitting.

Baselines

The naive reference point is what He et al. effectively did: call a pQTL "post-transcriptional" whenever its lead variant is not a significant expression QTL. Running this on the same summary statistics shows how much the headline "half of pQTLs are not eQTLs" changes once power and artifacts are modeled. We would compute it with the same significance thresholds and the same loci as our model.

The strongest existing correction is PAV adjustment from Raffield et al.. They regressed protein measures on protein-altering allele count, which raised mean inter-platform correlation across 15 discordant proteins from 0.38 to 0.51, and raised suPAR correlation in African-ancestry participants from 0.24 to 0.59. We would apply that adjustment to MESA, then run per-layer SuSiE-RSS followed by pairwise colocalization, the pipeline of Pozarickij et al. and Li et al.. This pipeline uses the same summary statistics, the same linkage matrices, the same maximum number of effects $L$, and the same prior grid as our model, so the comparison isolates modeling rather than data.

No reported result was found on Lacuna for any method's accuracy at classifying pQTL mechanism. For general fine-mapping, the reference figures are AUPRC 0.4042 for SuSiE and 0.4636 for SparsePro in UK Biobank chromosome 22 simulations (SparsePro). That is a different task, so those numbers set expectations for causal-variant recall, not for mechanism labels.

We count the proposal as beating PAV adjustment if it meets three conditions. Its inter-platform correlation after removing artifact-labeled effects must match or exceed 0.51 on the same 15 proteins. It must outperform PAV adjustment on mechanism AUPRC in simulation. And its labels must agree with mass spectrometry more often on held-out proteins.

Two ablations isolate the key ingredient. The first removes the artifact configurations but keeps the joint multi-layer prior, which tests whether gains come from layer typing rather than from pooling data. The second keeps the configurations but drops the coding-variant prior weight, which tests whether the model learns artifacts from the data rather than from the annotation.

Why not just…?

Why not just regress out protein-altering allele count and fine-map each platform separately? This is the real threat, and for loci with a single coding causal variant it may well work as well as our model. It fails in two specific cases. If a missense variant genuinely destabilizes the protein, adjustment erases a true abundance effect, even though that effect should appear concordantly on both platforms and in mass spectrometry. If the coding variant is in linkage with a separate non-coding regulator, adjusting for allele count partially absorbs the regulator's signal. We will measure how often these cases occur in simulations built with simxQTL from SuSiE 2.0, and empirically in MESA. If they account for a negligible share of loci, we will report that PAV adjustment suffices and narrow the claim to the loci where it fails.

Why not just run SuSiE per layer and colocalize? Pairwise colocalization asks whether two traits share a causal variant. It does not ask which layer combination a variant affects. With three or four layers it multiplies tests and handles low-power layers poorly, because a weak low-abundance protein signal simply fails to colocalize. The first ablation tests this directly: if per-layer fitting plus colocalization matches our model's mechanism accuracy in simulation, the joint likelihood is not earning its cost.

Why not just use mass spectrometry and ignore affinity platforms? Mass spectrometry avoids binding artifacts, but even with deconvolution it misses many low-abundance proteins (He et al. quantified only 1,508). Affinity platforms reach thousands of proteins at far lower concentrations. Mass spectrometry is therefore most valuable as a partial arbiter that calibrates labels on the subset it covers, and that is how we use it.

Proposed Experiments

The first experiment is simulation. Using simxQTL on UK Biobank-like linkage structure, we plant four kinds of causal variant: transcriptional, post-transcriptional, epitope-artifact, and true-destabilizing coding. Some of these sit in linkage with each other, and we vary the per-layer noise to mimic low-abundance proteins. Metrics are causal-variant AUPRC, credible-set coverage, and mechanism-label AUPRC and calibration, compared against both baselines and both ablations.

The second experiment is MESA, which has SomaScan 7k, Olink 3072 and whole-genome sequencing across four ancestries. Here we test whether removing artifact-labeled effects restores platform concordance at least as well as the reported 0.38 to 0.51 gain. We also test whether artifact labels concentrate in the 80 proteins whose platform agreement varies by ancestry, which is a prediction the model makes without being given those proteins.

The decisive check uses proteins with binding-independent measurement: the He liver mass spectrometry cohort, re-fit with SuSiE-RSS if summary statistics are obtainable, and plasma mass spectrometry processed with DIAmeter or Dear-DIA. On these proteins we focus on loci where PAV adjustment and our model disagree. If mass spectrometry tracks our labels at those loci, the model is identifying something adjustment misses; if it tracks adjustment, our extra structure is unjustified.

Finally, we re-estimate the post-transcriptional fraction of pQTLs by abundance tertile. The prediction is that the naive "about 50%" shrinks once artifacts are removed and power is modeled, with the largest shrinkage in the lowest-abundance tertile. A null result there would itself be informative: it would mean the discordance reported by He et al. is mostly real biology.