Stress experiments on sheep, crops, or patient tissue usually measure a few dozen animals or samples. Researchers then build a gene co-expression network, which groups genes whose expression levels rise and fall together across samples, and nominate "hub" genes as stress regulators or biomarkers. With twenty samples, the correlation between any two genes is mostly noise. The resulting modules and hubs often fail to replicate. A Parkinson's signature built this way scored a perfect discovery AUC of 1.000 on 25 samples and fell to 0.900 on an independent cohort (Identification of hub genes of Parkinson's disease through bioinformatics analysis).
Many such studies also measure DNA methylation, a chemical mark on promoter DNA that can switch a gene down. Today they use it only as a separate list of hits. This note proposes using methylation as a second, independent noisy measurement of the same regulatory state, but only for the minority of genes whose expression is known to be controlled by their own promoter or flanking "shore" methylation. A large public reference cohort tells us which genes those are. The small cohort's own methylation then sharpens those genes' expression estimates before any network is built. The claim is narrow and testable. For epigenetically gated genes, this should make modules and hubs from tiny cohorts agree better with what a large cohort would have found, and it should do so beyond what generic statistical shrinkage achieves.
Core Claim
Weighted gene co-expression network analysis (WGCNA) turns gene-gene correlations into a continuous network by raising them to a power β. It groups genes by shared neighbours, using a quantity called the topological overlap matrix, and summarizes each group by its first principal component, the "module eigengene" (Identifying Functional Gene Modules via Co-Expression Networks). Every step inherits the sampling noise of the correlation matrix. That noise is severe below about 20 samples.
Our premise comes from two findings that seem to conflict. Genome-wide, methylation predicts expression poorly. LASSO models reach $R^2 > 0.1$ for only 10.29% of genes in TwinsUK adipose (Predicting gene expression using DNA methylation in three human populations). Yet for specific loci, methylation acts close to a switch. CAV1 shore methylation correlates with expression at $r < -0.7$ (CpG island shore methylation regulates caveolin-1 expression in breast cancer), and CCND2 hypermethylation mimics deletion (Absence of cyclin D2 expression is associated with promoter hypermethylation in gastric cancer). Together these imply a small, identifiable set of epigenetically gated genes. For those genes, methylation carries real information about expression. For everyone else it adds noise.
Our hypothesis is that the identity of the gated set is a fairly stable property of a gene within a tissue. The correlation structure among genes changes with condition, which is the point of stress studies. So we transfer only per-gene coupling strengths from a large reference cohort, never the reference network itself.
Concretely, a LASSO model is trained in the reference cohort to predict each gene's expression from CpGs in its promoter and shores, following the setup of the three-populations study. Its cross-validated $R^2$ becomes a reliability weight $w_i \in [0,1]$. In the small target cohort, we apply that model to the target's own methylation to get a methylation-implied expression score $\hat m_i$. We then form a combined profile
$$\tilde x_i = \sqrt{1 - w_i}\, z(x_i) + \sqrt{w_i}\, z(\hat m_i),$$
where $z(\cdot)$ standardizes across target samples and $x_i$ is measured expression. This is an inverse-noise-style average of two measurements. When $w_i = 0$ it reduces exactly to ordinary expression. Standard WGCNA then runs on $\tilde x$ unchanged, so every downstream tool still applies.
Relation to Prior Work
Several papers motivate the gap. The glioblastoma eQCM study found a survival-linked module enriched for histone acetylation ($p = 0.0016$) but built its network from expression alone (Predicting glioblastoma prognosis networks using weighted gene co-expression network analysis on TCGA data). We put the epigenetic measurement into the network input itself rather than reading it off enrichment afterwards.
The sheep tail-docking study measured DNA methylation, m6A, and RNA in the same animals. It nominated CITED4 and ZNF644 from 143 jointly altered genes but built no network and had no replication cohort (Concurrent Alterations in DNA Methylation and RNA m6A Methylation During Epigenetic and Transcriptomic Reprogramming Induced by Tail Docking Stress in Fat-Tailed Sheep). It is our stress application, not our validation.
The closest method is the pathway-extending approach of A network-based pathway-extending approach using DNA methylation and gene expression data to identify altered pathways. It weights edges of a fixed protein-interaction scaffold using sparse canonical correlation between methylation and expression. We infer the network fresh from the cohort, gate methylation's influence gene by gene using externally estimated reliability, and judge success by replication rather than by the number of pathways detected.
Two lines of small-sample robustness work supply the comparison methods. MVTLASSO estimates a shared sparse network across noisy expression datasets with heavy-tailed models (Robust Multi-view Co-expression Network Inference). DGBCox selects outcome-invariant genes across 13 breast cancer cohorts (Stable breast cancer prognosis). Both use expression only. Our method is complementary: it adds a second view of the same genes rather than more samples of one view.
Research Program
The work has three parts.
The first builds coupling maps. Using TwinsUK adipose (n = 856), inner-city children's PBMC (n = 194), and HapMap LCL (n = 280) from the three-populations study, plus TCGA tumour types with matched 450k methylation and RNA-seq, we estimate $w_i$ per gene per tissue. We then measure how well $w_i$ transfers across tissues and across tumour versus normal. This is the load-bearing assumption. If the gated set is not stable across conditions, the method can only be used within a matched tissue and condition, and we will report it that way. Cell-type composition differs between samples, and it drives methylation at many CpGs at once. That shared signal could create artificial modules among high-$w$ genes, so all methylation is first adjusted for estimated cell composition and batch. Batch correction uses pyComBat from InMoose.
The second runs subsampling experiments with ground truth. Inside each large cohort, a WGCNA built on all samples serves as the reference answer. We repeatedly draw disjoint subsamples of n = 10, 15, 20, 30, and 50. The reference LASSO models are refit each time excluding the subsample, so no information leaks. We compare the networks recovered by each method against the full-cohort network.
The third is the stress application. The method runs on the sheep tail-docking data with a sheep coupling map estimated from control animals. The pre-registered prediction is whether CITED4, ZNF644, and DGAT1 fall in methylation-anchored modules, which can be checked against any future sheep cohort. Plant stress datasets in the file, such as the rapeseed nutrient-stress RNA-seq, lack methylation, so they are out of scope.
Baselines
The naive reference point is standard WGCNA on expression alone, with β chosen by the usual scale-free fit, run on the identical subsamples. It is the $w_i = 0$ special case of our method, so any gain is attributable to the methylation term.
The strongest generic competitor is WGCNA on a shrinkage-regularized correlation matrix. This pulls noisy correlations toward a structured target, which is the textbook fix for few samples. Its shrinkage intensity will be tuned with the same inner cross-validation budget we give our single implicit choice, the reference model's LASSO penalty.
The strongest published competitor for small-sample network recovery is MVTLASSO (Robust Multi-view Co-expression Network Inference). We will run it with expression and methylation-implied scores as two views, giving its sparsity penalty and dimension the same tuning grid size. Its paper reports more true edges than Graphical Lasso on B. subtilis and S. aureus benchmarks, but there is no reported number on Lacuna for its exact TPR/FPR. We therefore establish its performance ourselves on our subsamples.
A transfer baseline imports the reference cohort's network as a prior, mixing reference and target correlations with a tuned weight. This tests whether we are merely smuggling in the large cohort's answer.
Two ablations isolate the key ingredient. In the first, gene-to-methylation assignments are permuted so each gene receives another gene's $\hat m$; this keeps the averaging but destroys the biology. In the second, a constant weight $w$ is used for every gene in place of the per-gene $w_i$.
Beating the baselines means two things. Our method must give higher agreement with full-cohort modules among high-$w$ genes than both shrinkage-WGCNA and MVTLASSO at n ≤ 20, with paired bootstrap intervals excluding zero. It must also show no loss on low-$w$ genes.
For the downstream biomarker test, the reported floor on the TCGA Meta-Dataset few-shot benchmark (50 training samples) is 71.19% mean accuracy for an MLP and 68.78% for logistic regression, against 66.18% for a majority classifier (The TCGA Meta-Dataset Clinical Benchmark). These figures cover all 174 tasks with expression only. We will rerun both on the subset of tasks with matched methylation, so that the comparison uses the same samples.
Why not just…?
Why not just run expression WGCNA and correlate module eigengenes with methylation afterwards? That tells you which modules associate with methylation. It cannot fix the module boundaries themselves, which are already wrong at n = 15. Our subsampling experiment measures this directly: post hoc annotation leaves module recovery identical to the naive baseline by construction.
Why not just shrink the correlation matrix? This is the real threat. Shrinkage reduces variance for every gene without any second assay, and it may capture most of the gain. The decisive comparison is the dose-response pattern. Generic shrinkage should help high-$w$ and low-$w$ genes roughly equally. True methylation information should help only high-$w$ genes. The permuted-assignment ablation should collapse to shrinkage-level performance. If shrinkage matches us on high-$w$ genes, we will conclude that methylation adds nothing beyond regularization and say so.
Why not just borrow the large cohort's network? Under a matched condition it may well win, and we expect it to. Stress studies, however, exist to detect rewiring. We will test this by deriving the reference from normal tissue and the target from tumour in TCGA, and from controls versus cases in PBMC. In those settings a borrowed network should drag the target toward the reference condition, while per-gene coupling weights should not. If the borrowed prior also wins under condition shift, our premise about stability is wrong.
Why not just predict the phenotype with LASSO on all expression and methylation features? Pooled classifiers on raw features can collapse out of distribution, as when 5-FU accuracy fell from 84.1% to 52.9% on independent STAD data (Leveraging TCGA gene expression data to build predictive models for cancer drug response). We include penalized logistic regression on concatenated features as a biomarker baseline. Network modules are only worth the trouble if their eigengene classifiers transfer better across cohorts.
Proposed Experiments
The main experiment uses TwinsUK, PBMC, LCL, and matched TCGA cohorts. We report several metrics as a function of n: adjusted Rand agreement of subsample modules with full-cohort modules, overlap of the top-k hubs by within-module connectivity, and the fraction of full-cohort eigengene-trait associations replicated in held-out samples. Traits are BMI in TwinsUK, asthma status in PBMC, and tumour stage in TCGA. Each metric is stratified by quintile of $w_i$.
There is no reported result on Lacuna for module stability under subsampling in any of these cohorts, so baseline curves are established here.
Cross-tissue transfer checks the stability assumption: coupling maps estimated in adipose are applied to PBMC and LCL targets, and vice versa.
The biomarker check trains eigengene logistic classifiers on 50 TCGA samples per task and tests on the held-out 100. We compare against rerun logistic regression and MLP baselines on the same methylation-matched subset, and also test on a different TCGA tumour type to probe distribution shift.
Three decisive outcomes would settle the claim. First, gains must concentrate in the top $w$ quintiles and vanish in the bottom ones. Second, the permuted-assignment ablation must lose those gains. Third, our method must beat the borrowed-network prior under condition shift. If all three hold, methylation is acting as a real second witness for gated genes. If the first two hold but not the third, the method is still useful within matched conditions. If the first fails, the idea is wrong and generic shrinkage is the right recommendation for small stress cohorts.
References
- Identification of hub genes of Parkinson's disease through bioinformatics analysis
- Session Introduction
- Predicting gene expression using DNA methylation in three human populations
- CpG island shore methylation regulates caveolin-1 expression in breast cancer
- Absence of cyclin D2 expression is associated with promoter hypermethylation in gastric cancer
- Predicting glioblastoma prognosis networks using weighted gene co-expression network analysis on TCGA data
- Concurrent Alterations in DNA Methylation and RNA m6A Methylation During Epigenetic and Transcriptomic Reprogramming Induced by Tail Docking Stress in Fat-Tailed Sheep
- A network-based pathway-extending approach using DNA methylation and gene expression data to identify altered pathways
- Robust Multi-view Co-expression Network Inference
- Stable breast cancer prognosis
- The TCGA Meta-Dataset Clinical Benchmark
- Leveraging TCGA gene expression data to build predictive models for cancer drug response