Silent Variants in Hidden Cells: Methylation-Gated Splicing QTLs from Bulk Tissue

Many disease-linked DNA changes are "synonymous." They swap one letter of a gene without changing the protein it encodes, so standard pipelines discard them. Yet synonymous variants are about as often disease-associated as protein-changing ones, and many act by changing how the gene's RNA is cut and pasted together (splicing). This proposal argues that a large share of their heritable effect is invisible for a specific, testable reason. The effect occurs only in the minority cell types where the gene is switched on, typically infiltrating immune cells. In the majority cell types the gene is chemically silenced by DNA methylation, a tag on the DNA that shuts a gene off without changing its sequence. Bulk tissue RNA sequencing averages over all these cells, so a strong effect confined to 10% of cells that vary in abundance across people looks like weak noise.

We propose to model this explicitly. For each gene, a reference methylation atlas tells us in which cell types the gene can be active at all. Computational deconvolution, which estimates each sample's cell-type proportions from its bulk RNA profile, tells us how many of those cells each sample contains. Sequence models of splicing tell us which synonymous variants are likely to act at all. Combining these three pieces into one genotype-by-cell-fraction splicing model should recover immune-cell splicing effects of synonymous variants from large existing bulk cohorts. Aggregating those effects into a gene-level score should then explain trait heritability that expression-based methods miss. The idea has two parts: the variant's effect is confined to cells that make the transcript, and the methylation atlas says which cells those are, so the model no longer has to search blindly across every cell type.

Core Claim

A synonymous splice-altering variant's measured effect on the fraction of transcripts that include a given exon in bulk tissue should scale with the fraction of the sample made up of cells in which the gene is unsilenced, weighted by how much RNA each cell type contributes. If that holds, a gated model finds more synonymous splicing quantitative trait loci (sQTLs: DNA variants whose genotype predicts a splicing ratio) that replicate in purified immune cells than two alternatives: bulk mapping, and ungated cell-type-interaction mapping. The resulting immune-cell splicing scores should also add trait associations beyond expression-based transcriptome-wide association studies (TWAS). The methylation gate is the new ingredient. Deconvolution-based QTL mapping exists for expression, and sequence models of synonymous splicing exist for bulk tissues, but no work in the file joins them, or uses epigenetic silencing to constrain where a genetic effect can occur.

Relation to Prior Work

The closest precedent is Single-cell deconvolution of 3,000 post-mortem brain samples. It deconvolved 3,387 bulk brain samples with SPLITR, a method that corrects for phenotype-dependent marker shifts, nuclear-versus-whole-cell platform differences, and the roughly fourfold higher mRNA output of neurons. Its Bayesian model then found 4,757 cell-type-specific eQTLs, including 1,182 eGenes (genes with a detected eQTL) that were invisible in bulk analysis. This work motivates ours and supplies part of the method: we reuse its insight that per-cell RNA yield must be modeled. It differs from our proposal in three ways. It models only total gene expression, it asks nothing about synonymous or splicing variants, and it allows every cell type to carry an effect for every gene.

MTWAS partitions genetic effects on expression into shared and tissue-specific parts. It reports an average 47.4% improvement in prediction R² over PrediXcan across 47 GTEx tissues and was also evaluated on DICE sorted immune cells and OneK1K single-cell data. It needs pre-sorted or single-cell input and models expression only. S-PrediXcan predicts expression from nearby DNA variants and tests it against GWAS summary statistics, reporting r² > 0.99 concordance with individual-level analysis. It is the standard expression TWAS we must beat on heritability, and it is blind to splicing by construction.

On the variant side, SpliceTransformer predicts tissue-specific splicing from sequence. It reports about 85% top-k accuracy and 91% PR-AUC across 15 GTEx tissues, and finds that splicing alterations account for roughly 60% of intronic and synonymous ClinVar pathogenic mutations. We use it as a prior over which synonymous variants can affect splicing. Its tissue resolution is bulk, so it cannot say that an effect lives in macrophages rather than in parenchymal cells, the main tissue-specific cells of an organ. Chen et al. supplies the motivation: synonymous and protein-changing SNPs show similar disease-association rates (1.26% vs 1.46% in HapMap 3) and median odds ratios (1.70 vs 1.72). The same paper also raises the main threat, that synonymous hits may simply tag nearby causal variants they are inherited with (linkage disequilibrium, LD); we address this in the design. cdsFM shows that synonymous changes can also act through translation. Those effects leave no trace in RNA splicing, so we treat them as out of scope and use cdsFM scores only to flag variants whose effect we would not expect to see.

The silencing literature shapes the gate. Rao et al. found that regulatory switches for CAV1 sit in CpG island shores rather than the islands themselves (r < −0.7 between shore methylation and mRNA). CpG islands are dense clusters of CG letter pairs in gene promoters, and shores are the lower-density regions within about 2 kb of them. The gate therefore uses promoter and shore methylation, not island methylation alone. Promoter hypermethylation acting as a functional knockout, as with cyclin D2, justifies treating heavy methylation as near-zero activity. For the reference itself we rely on methylation deconvolution resources. cfDecon works with an atlas-level reference of 35 cell types and reports over 33% improvement in concordance correlation coefficient (CCC) over CelFEER. Long-read enhancer-promoter co-methylation shows that joint enhancer-promoter methylation separates cell types better than promoter methylation alone (0.69% vs 0.86% mixture error). We use these only as sources of cell-type methylation profiles, not as deconvolution engines.

Research Program

The model for a splice event $e$ (an exon inclusion or junction-usage ratio) in sample $i$ with genotype $G_{ij}$ at variant $j$ is

$$\mathrm{PSI}_{ie} = \sum_{c} \frac{w_{ic}\, r_c\, a_{cg}}{\sum_{c'} w_{ic'}\, r_{c'}\, a_{c'g}} \left(\alpha_{ce} + \beta_{cej}\, G_{ij}\right) + \gamma^\top z_i + \varepsilon_{ie}.$$

Here $\mathrm{PSI}_{ie}$ is the measured "percent spliced in," the fraction of the gene's transcripts in sample $i$ that include exon $e$. The weight $w_{ic}$ is the deconvolved proportion of cell type $c$. The factor $r_c$ is its relative mRNA yield per cell, which SPLITR-style correction estimates. The gate $a_{cg} \in [0,1]$ is the activity of gene $g$ in cell type $c$, set from reference promoter and shore methylation. The parameters $\alpha_{ce}$ and $\beta_{cej}$ are the cell-type baseline splicing and genotype effect, and $z_i$ holds covariates. The weighting follows from the fact that bulk splicing ratios mix transcripts, not cells: each cell type contributes in proportion to how many copies of the transcript it makes. The gate prunes the $\beta_{cej}$ that silencing forbids and redistributes the mixture weights, which is where we expect the power gain.

A hard gate would be brittle, because a methylation atlas can be wrong for a given tissue or disease state. We therefore place $a_{cg}$ under a prior centered on the atlas value and let the data move it, and we report how often the posterior overrides the atlas. SpliceTransformer scores enter as prior inclusion probabilities on $\beta_{cej}$ for synonymous and near-exon variants. Fitting uses sparse Bayesian regression with the SPLITR proportions held fixed. We will also propagate deconvolution uncertainty by refitting over bootstrap proportion estimates.

The discovery cohorts are GTEx v8 tissues with meaningful immune content and full-length RNA sequencing, whole blood, spleen, and lung, where splice junctions are observable. Deconvolution uses an immune-aware reference. To guard against LD confounding, every synonymous sQTL is conditioned on the lead expression QTL and on all missense variants in LD above a set threshold. We keep only credible sets whose top variant is synonymous.

For heritability, the fitted immune-cell $\beta$ values define per-gene, per-cell-type splicing predictors. These are tested against immune-related GWAS summary statistics in the S-PrediXcan style. We then ask how much trait variance they explain conditional on expression predictors. We measure total variance explained with SumVg and use UK Biobank phenome-wide heritability estimates as the ceiling defining the "missing" portion.

Baselines

The naive reference point is bulk sQTL mapping on the same GTEx samples, with the same covariates and variant set but no cell-type terms. Any gated gain must be measured against the number of synonymous sQTLs it finds that replicate in sorted cells.

The strongest existing method for the discovery step is a deQTL-style cell-type interaction model from Park et al., transplanted from expression to splicing ratios. It uses the same SPLITR proportions and RNA-yield correction, but every cell type is free to carry an effect. No reported result exists for this method on splicing: Park et al. report 4,757 cell-type eQTLs on brain expression, and the file contains no reported result for cell-type sQTL recovery from deconvolved bulk splicing. We will therefore run it ourselves under the same sparse prior, the same hyperparameter search budget, and the same false discovery control. Beating it means a higher count of DICE-replicating synonymous sQTLs at matched false discovery rate, together with a higher replication rate. More hits at a lower replication rate would not count.

The ablation that isolates the key ingredient is the full model with the gate set to one everywhere, which differs from the deQTL-style baseline only in the SpliceTransformer prior. A second ablation keeps the gate but drops the prior. Together they separate the contributions of silencing and sequence. A third control replaces the atlas gate with a cell-type-shuffled gate. If it performs as well as the true gate, the gain comes from sparsity, not biology.

For heritability, the strongest competitors are S-PrediXcan and MTWAS expression predictors trained on the same GTEx tissues. MTWAS's reported average 47.4% R² improvement over PrediXcan is the published bar for expression prediction, though our target is conditional trait association, not expression R². Beating them means significant additional gene-trait associations and variance explained after conditioning on their predictors. We will also run a bulk splicing TWAS without deconvolution, so that the cell-type step is tested separately from the move from expression to splicing.

Why not just…?

Why not just map sQTLs directly in DICE or OneK1K sorted and single-cell data? For purified cells this is the cleanest design, and it is a real threat. We use DICE as the ground truth for exactly that reason. The case for deconvolution is scale, since bulk tissue cohorts can reach thousands of samples, and tissue context: tissue-infiltrating macrophages are not blood monocytes. OneK1K is described in the file only as single-cell PBMC data. If, as is typical for cohorts of that size, it measures only transcript ends, most splice junctions will be unobserved; we will check its junction coverage before relying on it. We settle the question empirically. If direct DICE mapping finds as many synonymous immune sQTLs as gated deconvolution of larger bulk cohorts, the discovery claim fails, and we would reduce the project to the heritability analysis using DICE-trained predictors.

Why not just test whether SpliceTransformer-flagged synonymous variants are enriched for heritability, with no cell types or methylation? Annotation-level heritability partitioning is cheap and could show that synonymous splice variants carry excess heritability. It cannot say which gene or cell type is responsible, and it cannot distinguish effects masked by composition from effects that are simply small. We will run it as a reference. If its aggregate enrichment already accounts for all the conditional variance our cell-type predictors add, the cell-resolved model is not needed for the heritability claim, and we will say so.

Why not just add measured bulk methylation as a covariate or interaction term in bulk QTL mapping? Bulk methylation has the same composition problem as bulk RNA, as the file's own gap analysis notes. A drop in promoter methylation can mean either immune influx or demethylation in the tissue's own cells, so a genotype-by-bulk-methylation term mixes the two. The cell-type atlas gate avoids this by asking which cells can express the gene, not how methylated the bulk sample is. Where a GTEx-matched cohort has paired methylation, we will fit the bulk-interaction model alongside ours and compare DICE replication directly.

Why not accept that synonymous hits are LD tags of missense or promoter variants? Chen et al. acknowledge exactly this confounding. Our conditioning and credible-set filter handle the statistics, and the functional test is independent. A tagging variant has no reason to show its effect only in the atlas-permitted cell types or to agree in direction with SpliceTransformer's prediction; a causal splice variant should do both.

Proposed Experiments

The first experiment is a controlled simulation. We build pseudo-bulk mixtures from DICE sorted-cell profiles with known genotypes, vary immune fraction and gate accuracy, and measure sQTL recovery for gated, ungated, and bulk models. This establishes when the gate helps and how badly a wrong atlas hurts.

The second is discovery in GTEx whole blood, spleen, and lung, with replication in DICE sorted cell types. The primary metric is the number of synonymous sQTLs replicating in the correct cell type at 5% false discovery rate. Secondary metrics are replication rate and agreement in effect direction.

The third is a mechanistic check: among replicated hits, the direction of the predicted splicing change should agree with SpliceTransformer more often than for unreplicated hits. Variants cdsFM flags as translation-acting should not replicate as sQTLs.

The fourth is the heritability test on immune-related UK Biobank traits. It measures variance explained by immune-cell splicing predictors conditional on S-PrediXcan and MTWAS expression predictors, as a fraction of the gap between SNP heritability and what expression TWAS explains.

The decisive outcome requires two results. The gate must beat both its shuffled control and the gate-free ablation on DICE replication, and the conditional heritability gain must survive the annotation-only baseline. If the shuffled gate performs as well, the silencing hypothesis is wrong. If DICE-direct mapping performs as well, deconvolution is unnecessary.