Studies that ask how an exposure reaches a disease increasingly route the question through DNA methylation, the chemical tags on CpG sites (cytosine–guanine pairs) that switch genes up or down. The tool is mediation analysis: split the exposure's total effect on the outcome into a part that passes through a measured intermediate, here methylation at some site, and a direct remainder. Methylation platforms have grown denser over time, from the 450K array to the EPIC array to sequencing. The working assumption is that more measured sites can only help. We argue that for joint, high-dimensional mediation this assumption is wrong in a specific, predictable way, and that wireless engineers already know the shape of the failure.
In a dense wireless network, adding transmitters first helps because each receiver ends up closer to some transmitter. Past an optimal density it hurts, because every new node also interferes with its neighbors. That regime is called interference-limited scaling. Methylation behaves the same way. A denser array places a probe closer to the unmeasured site that actually carries the effect, so the signal captured goes up. But neighboring sites are strongly correlated, so a joint model must divide one biological effect among many near-copies, and the uncertainty about each one grows. Our core claim is that power to detect a mediating locus rises and then falls as probe density increases. We further claim the peak sits at a density set by how far correlation between neighboring sites extends along the chromosome. The practical consequence is that mediation should be tested at a data-chosen resolution: groups of correlated sites, not single CpGs. Because the resolution rule needs only within-window correlation summaries, it can run across cohorts that cannot pool individual-level data.
Core Claim
Suppose a true mediating site has indirect effect $\theta$, and methylation correlation between two sites decays with their genomic distance $d$ as $\rho(d)$, with a characteristic correlation length $\ell$. At probe density $\lambda$ (probes per kilobase), the nearest probe lies about $1/(2\lambda)$ away. A joint model also includes roughly $2\lambda\ell$ correlated neighbors. The test statistic for the best proxy then behaves like
$$z(\lambda) \approx \sqrt{n}\,\frac{\theta\,\rho\!\left(1/(2\lambda)\right)}{\sigma\sqrt{\mathrm{VIF}(\lambda)}},$$
where $n$ is sample size and $\sigma$ is residual noise. $\mathrm{VIF}(\lambda)$ is the variance inflation factor: how much collinearity with neighbors in the joint fit inflates the estimate's variance. The numerator is the "proximity gain" and saturates once probes sit inside the correlation length. The denominator is the "interference" and keeps growing with $\lambda$. On top of this, the multiple-testing threshold rises slowly with the number of tests. The prediction is an interior optimum in $\lambda$ for joint estimators such as de-biased Lasso, located near $\lambda \sim 1/\ell$. Marginal one-site-at-a-time tests should show no such peak, because they pay no collinearity penalty. In exchange, they cannot say which of the correlated sites carries the effect.
The analogy is a quantitative scaling hypothesis we will test, not a decorative metaphor. Percolation and Connectivity on the Signal to Interference Ratio Graph proves that connectivity in an interference-limited network exists only within a bounded window of node densities. It also shows that giving nodes a modest number of separate channels, $O(\log n)$, restores connectivity. The analogue of separate channels here is to merge correlated neighbors into one tested unit. That paper motivates the shape of our prediction. We borrow none of its proofs, since methylation correlation is not a sum of power-law interference terms.
Relation to Prior Work
The closest single-site-to-genome-wide method is HIMA2. It works in three steps. First it screens CpGs by the product of the exposure→methylation coefficient and the methylation→outcome coefficient. Then it fits the survivors jointly with de-biased Lasso, a penalized regression corrected so that its p-values are valid. Finally it controls the false discovery rate with a mixture model built for the composite null "either path is zero." Applied to the CARDIA cohort (smoking → blood methylation → lung function), it found two CpG mediators. HIMA2 is the method our hypothesis predicts will degrade under densification, because its joint step is exactly where collinearity bites. We reuse its screening and testing machinery and change the unit being tested. HIMA2's simulations do vary mediator correlation, but they do not vary probe density over a fixed underlying biology. That density variation is the experiment our hypothesis requires.
Region-level association already exists. The ChAMP pipeline wraps DMRcate and Bumphunter to call differentially methylated regions and blocks. It also provides normalization, including BMIQ, and cell-type deconvolution. These regions are defined by association with a single variable, not by the two-path structure of mediation, and they come with no mediation-aware inference. They supply a baseline and preprocessing, not our method.
The distributed setting comes from Distributed mediation analysis with communication efficiency. For a single mediator, it shows that sites should send intermediate linear t-statistics rather than local p-values. Averaging local p-values drives power toward zero as the number of sites grows. Sending t-statistics matches pooled power in simulations up to 32 machines, provided the number of sites $K$ is $o(\sqrt{N})$, where $N$ is total sample size. That paper constrains our design: anything we aggregate across cohorts must be a linear summary. It also exposes a gap. Its two-scalars-per-site scheme cannot carry the correlation information a joint multi-site model needs.
We should be honest about where the coordination cost actually lies. Sending two numbers for each of 850,000 CpGs is a few megabytes, so bandwidth is not the bottleneck. The real costs are two. One is the number of synchronization rounds. The other is how much of each cohort's correlation matrix must be disclosed. Region-level testing shrinks the second cost directly.
Mendelson et al. used Mendelian randomization, which treats genetic variants as natural experiments, across Framingham, Lothian, ARIC, GOLDN and PIVUS. They found 83 BMI-associated CpGs and an upstream causal role for SREBF1 methylation (cg11024682). Das et al. tie CPT1A methylation to metabolic syndrome in GOLDN. These loci give us positive controls with independent causal support. Low reliability of DNA methylation across Illumina Infinium platforms in cord blood warns that 450K and EPIC measurements of the same probe disagree. That finding constrains us to vary density within a platform rather than across platforms.
Research Program
The first component is theory and calibration. We estimate the correlation decay $\rho(d)$ and the length $\ell$ empirically from 450K data, separately in CpG islands, shores, and open sea, since Rao et al. show regulation differs sharply between islands and shores. Plugging these into the scaling relation gives a predicted optimal density for each genomic context. That prediction is written down before any power curves are computed.
The second component is the method. It is a resolution-adaptive mediation procedure. Each genomic window is cut into units by agglomerating adjacent CpGs whose pairwise correlation exceeds a threshold, chosen to minimize the predicted variance inflation. Each unit is summarized by its first principal component, an exposure-agnostic summary that avoids double-dipping. HIMA2's screening, de-biased Lasso and composite-null testing then run on units instead of CpGs. A unit's indirect effect is reported along with a within-unit attribution profile. The profile is descriptive rather than inferential, so we do not overclaim single-site resolution.
The third component is distributed execution. The unit boundaries depend only on local correlations within windows. These can be computed per cohort and agreed in one round, by averaging window correlation matrices across cohorts weighted by sample size. The joint step then needs each cohort's cross-product matrix of the selected units, exposure and covariates. For linear outcomes, these sum exactly to the pooled quantities. That makes the procedure two rounds and exact, with no dependence on $K$. For binary outcomes, where no finite summary is exact, we use the same intermediate-linear-statistic principle as the distributed mediation paper and measure the gap to pooled analysis. Cell-type deconvolution and batch correction are done locally in each cohort using ChAMP.
Baselines
The naive reference point is a standard per-CpG association scan. Each CpG is regressed on exposure and on outcome separately, and a site is called a mediator if both paths pass Bonferroni correction (the joint-significance rule). It requires no joint model and nothing beyond per-site scalars. It is run on identical preprocessed data. Our method must beat it on localization error and indirect-effect accuracy without losing more than a small, pre-stated amount of power.
The strongest existing method is HIMA2 at full native density. Its own paper reports lower error than HIMA, better power with false discovery rate held at the nominal 0.05, and two significant CpGs in CARDIA (cg26331243 near CCDC33, cg19862839 near TBX4). The file contains no numeric power value for HIMA2 on a named benchmark, so our comparison number is the CARDIA discovery count and the FDR calibration. HIMA2 gets the same screening size, the same tuning grid for the Lasso penalty, and the same compute. Beating it means one of two outcomes at matched FDR. Either we find strictly higher power in semi-synthetic data at native density, or we recover the CARDIA loci plus additional units that replicate in a held-out split.
The association-defined region baseline calls regions with DMRcate through ChAMP on the exposure, then runs HIMA2 on region summaries. It is the most direct competitor to our resolution rule. It differs only in how units are formed. Beating it means better FDR calibration, because its regions are chosen using the exposure, and equal or better power.
For the distributed comparison, the reference is the per-CpG distributed joint-significance test from the communication-efficiency paper, sending two t-statistics per CpG per cohort. That paper reports it matches pooled power up to $K=32$. We run it on the same cohort split.
The key ablation sweeps a fixed unit size, from single CpGs to blocks of increasing width, against the adaptive rule. This isolates whether data-chosen resolution matters or any coarsening suffices.
Why not just…?
Why not just keep using HIMA2 on every CpG, since its de-biased Lasso already fits mediators jointly? That is precisely the null hypothesis of this proposal. If the power-versus-density curves from probe thinning come out monotone for HIMA2, the interference story is wrong, and the right advice is "measure as densely as possible and use HIMA2." The thinning experiment below settles it directly. The variance-inflation argument gives a concrete reason to expect otherwise: de-biased Lasso variance grows with collinearity among included mediators.
Why not just call differentially methylated regions with DMRcate or Bumphunter and mediate on those? This is a real threat. It may capture most of the gain. Our two reasons to expect a difference are both testable. Regions chosen from exposure association reuse the data that later tests the exposure path, which risks inflated false discoveries. Association regions also have no reason to match the correlation length that controls variance inflation. If the DMR baseline matches our power and calibration in semi-synthetic data, we will report that the contribution is the diagnosis of the density effect, not a new unit-formation rule.
Why not just use marginal per-CpG tests, which pay no collinearity penalty and distribute trivially? Marginal tests should indeed show no power peak. The communication-efficiency paper shows they distribute with full pooled power. Their weakness is attribution. Every correlated neighbor of a true mediator also tests positive, so their localization and indirect-effect accuracy should degrade with density even as their power holds. If marginal tests match our method on localization in the semi-synthetic benchmark, the joint step is unnecessary and we will say so.
Why not just use Mendelian randomization at candidate loci? Mendelian randomization answers whether a methylation site is causally upstream of a trait, given a valid genetic instrument. It does not decompose an environmental exposure such as smoking into direct and methylation-mediated parts across the genome. Many exposures also lack instruments. We therefore use Mendelson et al.'s SREBF1 result as external validation, not as a competitor.
Proposed Experiments
The decisive experiment is probe thinning on real data. On the CARDIA 450K matrix used by HIMA2, we retain random fractions of probes from 10% to 100%. This keeps the biology and platform fixed while varying only density. Semi-synthetic mediators are planted by masking a true causal CpG and generating exposure and outcome from it with known effect sizes. That way the "true site" is often unmeasured, as it is in reality. The metrics are unit-level power at FDR 0.05, realized FDR, localization error in kilobases to the masked site, and mean squared error of the indirect effect.
The hypothesis predicts two things. HIMA2's power should peak in the interior of the density range, with the peak located near the pre-registered $1/\ell$ for each genomic context. Our adaptive method should be flat or rising across the range. The falsification condition is a monotone HIMA2 curve, or a peak location unrelated to $\ell$. Thinning cannot reach densities above 450K. For that side of the curve, we generate synthetic denser probes by interpolating from the estimated correlation model, and we label those results as model-based.
The real-data check reruns CARDIA (smoking, Year-15 methylation, Year-20 lung function) with all methods. The question is whether units containing CCDC33 and TBX4 are recovered and whether additional units replicate in a random half-split. A second analysis treats Framingham, ARIC and GOLDN as separate silos for the BMI→methylation→lipid or metabolic outcome question. Mediating units at SREBF1, CPT1A and ABCG1 serve as positive controls with independent support from Mendelson et al. and Das et al.
The distributed check compares the two-round procedure to a pooled analysis on the same cohorts. For linear outcomes, the agreement should be exact to numerical precision. For a binary obesity outcome, we report the power gap to pooled analysis and to per-CpG distributed joint significance. We also split the cohorts into progressively more, smaller pseudo-sites, past $K = \sqrt{N}/2$, to test whether region-level cross-products avoid the degradation the single-mediator method warns about.
The proposal succeeds if HIMA2 shows the predicted interior peak and the adaptive method removes it at matched FDR. If the curves are monotone, the outcome is a clean negative result with a direct recommendation to profile densely.
References
- Percolation and connectivity on the signal to interference ratio graph
- HIMA2: high-dimensional mediation analysis and its application in epigenome-wide DNA methylation data
- ChAMP: updated methylation analysis pipeline for Illumina BeadChips
- A beta-mixture quantile normalization method for correcting probe design bias in Illumina Infinium 450 k DNA methylation data
- Distributed mediation analysis with communication efficiency
- Association of Body Mass Index with DNA Methylation and Gene Expression in Blood Cells and Relations to Cardiometabolic Disease: A Mendelian Randomization Approach
- Association of DNA Methylation at CPT1A Locus with Metabolic Syndrome in the Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) Study
- Low reliability of DNA methylation across Illumina Infinium platforms in cord blood: implications for replication studies and meta-analyses of prenatal exposures
- CpG island shore methylation regulates caveolin-1 expression in breast cancer