Gene co-expression networks are built on a simple bet: if two genes rise and fall together across many tissue samples, they probably belong to the same biological program. The dominant tool for this, weighted gene co-expression network analysis (WGCNA), turns pairwise correlations into a weighted graph, finds tightly connected groups of genes called modules, and links each module to clinical traits. The bet fails when genes co-vary for reasons unrelated to shared regulation. Examples include tumor samples differing in how many immune or stromal cells they contain, differing proliferation rates, or batches run on different days. These confounders produce strong, real correlations that survive every filter WGCNA applies. Particle physicists face the same structure when they hunt rare decays under backgrounds that mimic the signal. One of their most useful findings is that a neural network's advantage over hand-built selection rules often comes from learning the shape of the background, not a sharper picture of the signal. Learning to identify electrons showed this directly. The authors then distilled the network's advantage back into two interpretable observables.
We propose to use that two-step strategy on co-expression networks. First, train a neural classifier that looks at the full joint pattern of two genes across samples and sample covariates, not just their correlation coefficient, and asks whether the pair is co-regulated or merely co-confounded. We use it first as a measuring instrument: how much edge-level information does WGCNA's correlation-based adjacency throw away? Second, search a library of interpretable pair statistics for the few that reproduce the network's ranking of gene pairs. The deliverable is either a small, transparent correction to WGCNA that biologists can adopt without a neural network, or a clean measurement that no such gap exists beyond what covariate regression already removes.
Core Claim
The claim is that confounder-driven co-expression has a recognizable signature in the sample-level scatter of a gene pair. Such correlations should be concentrated in samples with extreme composition, track purity or library-size covariates, and be carried by a few high-leverage samples. A permutation-invariant network over samples can learn this signature from semi-synthetic data with known ground truth. The resulting gap over standard WGCNA, measured as edge-level and module-level recovery, can be closed by a handful of closed-form statistics.
The physics precedent supplies both the method and the expected shape of the result. In Learning to identify electrons, a convolutional network on calorimeter images reached an area under the ROC curve (AUC) of 0.972 versus 0.945 for seven physicist-designed features. Adding two background-sensitive observables found by decision-ordering search brought the features to 0.971. We expect co-expression to be analogous: existing correlation measures describe genuine co-regulation adequately, and what they miss is structure in the background.
How This Differs From Existing Work
According to the compiled file, no published work applies neural background suppression inside WGCNA. The closest method is Robust Multi-view Co-expression Network Inference. It models multi-study expression as a linear sum of a shared signal and study-specific noise, uses heavy-tailed distributions to resist outliers, and estimates a sparse network by expectation-maximization with graphical lasso. That method assumes linear noise and needs several independent studies. Its authors also report that the iterative fitting is expensive and makes hyperparameter selection hard. We differ in three ways. We work inside the standard WGCNA pipeline rather than replacing it. We model confounding as a learned, possibly non-linear pattern in each pair's sample-level data. And we score pairs with an amortized forward pass, meaning the network is trained once and then applied to new pairs cheaply, instead of solving an optimization problem for each dataset.
Applied WGCNA studies motivate the work because they show where the problem bites. Advancing Breast Cancer Biomarkers and the cutaneous squamous cell carcinoma study Integrative transcriptomic and single-cell analyses identify KLRD1 both run off-the-shelf WGCNA on bulk tumor tissue and correlate module summaries with traits. In neither study can one tell whether a module reflects regulation or tissue composition. The breast cancer hub-gene study CENPL, ISG20L2, LSM4, MRPL3 handles cross-study noise with rank aggregation, which discards the continuous expression magnitudes that WGCNA was designed to keep.
Two physics results constrain the design. The EXO-200 analysis in Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset shows the payoff of a learned discriminator: signal efficiency rose from 80.8% to 96.4±3.0% and sensitivity improved by about 25%. It also shows the risk: mismatch between simulated and real background limits trust in the result. Our training labels come from injected signals, so the same mismatch threatens us, and the evaluation plan is built around it.
Research Program
Semi-synthetic training data. Physicists train on Monte Carlo events with known labels, and we need an equivalent. We take real expression matrices from TCGA-BRCA (1,118 tumors) and plant synthetic modules into them. A planted module is a set of genes given a shared latent factor of controlled strength, added on top of each gene's real expression. Planting keeps every real confounder intact, including purity variation, immune infiltration, proliferation, and platform effects. The planted gene pairs are the signal. Background examples are gene pairs that are strongly correlated in the real data but whose correlation is mostly explained by measured covariates. These covariates include tumor purity estimates, immune and stromal scores, a proliferation score, library size, and batch. As a sharper background label, we also inject confounder-driven pairs: two genes both given a dependence on a composition covariate, with no shared factor. This ensures the network sees background that matches the signal on correlation strength. That property is exactly what Athar et al. identified as the failure mode of one-dimensional cuts in rare-decay searches.
The injection families vary in several ways: module size, factor strength, whether the factor is linear or saturating, and whether the planted program is restricted to a subset of samples. One family is held out entirely at test time, so we can detect a network that has learned injection artifacts rather than the background.
The pair discriminator. Each gene pair is represented as a set of sample tokens. Each token contains the two genes' standardized expression in one sample plus that sample's covariates. A permutation-invariant set encoder, which treats samples as an unordered collection the way a DeepSets model or a transformer without position encodings does, outputs a score $p_{ij}$ that the pair is co-regulated rather than co-confounded. True edges are rare among all pairs, roughly the extreme imbalance physics searches face. We therefore train on balanced batches and recalibrate to the true prior afterwards. To keep compute tractable, we restrict to the 5,000 most variable genes, about 12.5 million pairs, each scored in one forward pass.
Plugging into WGCNA. We change only the similarity step. Standard WGCNA uses $s_{ij} = |\mathrm{cor}(x_i, x_j)|$ and the adjacency $A_{ij} = s_{ij}^{\beta}$, where $\beta$ is the soft-thresholding power chosen so that the network's connectivity follows a power law. We instead use
$$A_{ij} = \left(s_{ij}\, p_{ij}\right)^{\beta},$$
so that correlation strength is kept but discounted by the probability that the correlation is background. The topological overlap matrix (a measure of how many neighbors two genes share), the hierarchical clustering, the dynamic tree cut, and the module eigengenes (each module's first principal component) all stay unchanged. Downstream biology therefore remains interpretable in WGCNA's own terms.
Distillation. Following the decision-ordering approach of Collado et al., we build a library of interpretable pair statistics. It includes Spearman and biweight midcorrelation, partial correlation given each covariate, correlation within purity strata compared with the pooled correlation, leave-top-samples-out correlation (which measures leverage), bootstrap stability of the correlation, and differences in variance across strata. We then greedily add the statistic that most increases agreement between a simple combined score and the network's ranking of pairs. The question is whether two or three such statistics recover most of the network's gain. If they do, the practical output is a closed-form "background-aware" adjacency that runs within WGCNA.
Baselines
The naive reference point is standard unsigned WGCNA on the same 5,000 genes. It uses Pearson correlation, $\beta$ chosen by the scale-free criterion, and default tree-cut settings, exactly as the applied studies above use it. Every other method must beat this to justify any added complexity.
The strongest simple competitor is covariate-adjusted WGCNA. It regresses the measured composition covariates and the top $k$ expression principal components out of each gene, then applies biweight midcorrelation and standard WGCNA. This is what a careful bioinformatician would do already. It directly targets the confounders we care about and needs no training. We tune $k$ and the covariate set on the same validation injections and with the same number of configurations we give the network's hyperparameter search. It is evaluated on the identical held-out injections.
The strongest existing published method is MVTLASSO from Robust Multi-view Co-expression Network Inference. We run it with TCGA-BRCA and METABRIC as two views on shared genes, using the authors' model with its latent dimension tuned under the same budget. If cost prevents it from running on 5,000 genes, we compare on a 1,000-gene subset where all methods run. The file reports only that MVTLASSO achieves higher true positive rate than graphical lasso and single-view variants on synthetic and Bacillus subtilis data. There is no numeric value: no reported result found on Lacuna for MVTLASSO on planted-module recovery or on human tumor data. We report its numbers fresh under our protocol.
Three ablations isolate the key ingredient, which is the network's access to sample-level structure and covariates. The first gives the same encoder only the scalar correlation; it should reproduce the naive baseline. The second gives it the sample scatter without covariate tokens, which tests whether the background is visible from the expression pattern alone. The third trains the full network with covariates permuted across samples. Beating a baseline means a gain in held-out-family module recovery, measured by adjusted Rand index between recovered and planted modules, that exceeds the seed-to-seed spread across five injection replicates. It also means a lower rate of composition-explained modules on real data, with no loss of module preservation in METABRIC.
Why not just…?
Why not just regress out principal components and composition covariates before WGCNA? This is the most serious threat, and the proposal is structured so that it can win. Its known weakness is that large genuine programs also load on the top principal components, so removing them can delete real modules while leaving non-linear confounding untouched. The planted-module experiments measure both effects directly. If covariate-adjusted WGCNA matches the full network on held-out injection families, the honest conclusion is that the background gap is linear and already closable. The distillation step should then select partial correlation as its first statistic, which confirms that conclusion. We would publish that as the result rather than argue around it.
Why not just use MVTLASSO or graphical lasso? These methods are principled, but they assume signal and noise add linearly, and MVTLASSO needs several independent studies. Its own authors flag the iterative fitting as a computational bottleneck that complicates tuning (Robust Multi-view Co-expression Network Inference). The saturating-factor and subset-restricted injection families test whether linearity actually costs accuracy. The single-cohort setting, where MVTLASSO cannot be applied, is the most common situation in practice.
Why not just use a robust correlation such as Spearman or biweight midcorrelation? These measures reduce the influence of outlier samples. However, a purity-driven correlation is spread across all samples, so it is not an outlier effect and robust correlation does not remove it. Both measures are included as baseline options and as entries in the distillation library, so if robustness alone is enough, the experiments will show it.
Why not just train a denoising autoencoder on the expression matrix? An autoencoder learns to reconstruct the dominant variance in the data, and in bulk tumor data that variance is largely composition. An unsupervised denoiser therefore tends to keep the background and discard weak programs. The physics lesson is that suppression needs a target that defines what counts as background. Our injected labels supply that target, and an autoencoder has no access to it.
Decisive Checks
The first check is on held-out injection families. The full network must beat covariate-adjusted WGCNA on module recovery. Its advantage must also disappear in the covariate-permuted ablation, which shows the gain comes from covariate structure and not from injection artifacts.
The second check concerns transfer to real data. Modules found in TCGA-BRCA should preserve in METABRIC, which uses a different platform, at least as well as baseline modules. In addition, the eigengenes of the discounted modules should correlate less with purity and immune scores. A secondary reference point is prognostic value: the WGCNA-plus-LASSO two-gene signature in Advancing Breast Cancer Biomarkers reports TCGA 5-year survival AUC of 0.624. We will compare it with signatures built from discounted modules by the same LASSO-Cox procedure. We treat this as descriptive, not as a pass/fail criterion, because survival prediction is not the goal.
The third check is a re-analysis. On GSE42677, we re-run the cutaneous carcinoma analysis and ask whether the MEblue module, reported at r = −0.76 with activated CD8⁺ T-cell infiltration in Integrative transcriptomic and single-cell analyses identify KLRD1, survives background discounting. Survival would suggest a regulatory exhaustion program. Dissolution would suggest that the module mostly tracks how many immune cells the sample contains. That case shows concretely what the method is for.
The last check is the distillation gap. We report the network's edge AUC against the best two- or three-statistic closed form, with the distance between them as the measured background gap, analogous to the 0.945, 0.972, and 0.971 sequence in the electron study.
References
- Learning to identify electrons
- Robust Multi-view Co-expression Network Inference
- Advancing Breast Cancer Biomarkers: A Centromere-Related Gene Signature Integrated with Single-Cell Analysis for Prognostic Prediction
- Integrative transcriptomic and single-cell analyses identify KLRD1 enrichment in exhausted CD8⁺ T cells in cutaneous squamous cell carcinoma
- CENPL, ISG20L2, LSM4, MRPL3 are four novel hub genes and may serve as diagnostic and prognostic markers in breast cancer
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- Search for radiative decays of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>Υ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>1</mml:mn><mml:mi mathvariant="normal">S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>into<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>η</mml:mi></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>η</mml:mi><mml:mo>′</mml:mo></mml:math>