Certifying Network Rewiring in DIA Proteomics Against Acquisition Artifacts

Biologists often ask whether a disease or treatment rewires protein networks: whether proteins that rise and fall together in healthy tissue stop doing so in disease, or start coordinating with new partners. The standard tool for finding such coordinated groups is weighted gene co-expression network analysis (WGCNA). It builds a network from correlations, groups proteins into "modules" and summarizes each module by one score. Data-independent acquisition mass spectrometry (DIA-MS) now measures thousands of proteins per sample. It does this by fragmenting every molecule inside fixed mass windows at once, which produces mixed spectra that software must later pull apart. Our concern is that a DIA protein correlation is partly a fact about the instrument and the software, not only about biology. Two proteins whose peptides land in the same mass window at the same moment share interference. When an abundant protein changes between conditions, that shared interference changes too, and a module can appear to rewire when nothing biological happened.

We propose to treat network rewiring the way power-grid engineers treat an unknown, changing grid topology. Instead of committing to one estimated network, the analysis keeps the whole set of networks consistent with the data across plausible processing choices and resamples. It also models a known physical cause of spurious connections, here co-elution in the same isolation window. A rewiring claim is reported only if it survives the entire set. The decisive test uses an existing DIA benchmark with built-in ground truth. In that benchmark, the human background should show no rewiring between groups at all, so any rewiring the analysis reports is false.

Core Claim

We claim two things. First, a measurable fraction of the module rewiring that standard per-group WGCNA reports in DIA data comes from acquisition and processing rather than biology. That spurious rewiring concentrates on protein pairs that share isolation windows and elution times with peptides whose abundance differs between conditions. Second, a set-valued rewiring test can remove most of it while keeping real rewiring detectable at the small cohort sizes typical of clinical DIA studies. The test requires a module change to hold across a family of DIA workflows and bootstrap resamples, after adjusting for an acquisition-proximity covariate.

Where This Sits

The DIA–WGCNA pairing is established and static. The study Brain proteomic analysis implicates actin filament processes and injury response in resilience to Alzheimer's disease ran one WGCNA network over 7,115 DIA-quantified proteins and found module M5, 181 actin and injury-response proteins raised in dementia but not in resilient brains. Temporal proteomic profiling of postnatal human cortical development correlated fixed module scores with age across 911 label-free proteins. Data-independent acquisition proteomics reveals circulating biomarkers of coronary chronic total occlusion in humans profiled plasma before and after revascularization but used only differential expression and STRING interaction maps. These papers motivate the question. None tests whether module structure changes, and none asks whether such a change could be an artifact.

Two papers model changing or multi-condition networks. Inferring Time-Varying Network Topologies from Gene Expression Data segments microarray time series into regimes and fits linear state-space networks inside each regime. That shows regime-specific edges exist, but the pipeline does not scale to proteome-wide data and it assumes measurements are clean. Robust Multi-view Co-expression Network Inference fits one shared sparse network across heavy-tailed, batch-affected studies and treats study-specific structure as noise. We take the opposite stance on structured differences: they are the object of interest, but they must be separated from technical structure that is itself condition-dependent.

The constraint on the whole proposal comes from Benchmarking of analysis strategies for data-independent acquisition proteomics using a large-scale dataset comprising inter-patient heterogeneity. That study showed that the choice among 1,428 workflows changes differential-abundance calls substantially. It combined software (DIA-NN, Skyline, OpenSwath, Spectronaut), library strategy, normalization and statistical test. If single-protein calls are this fragile, pairwise correlations, which are second-order statistics, are very likely more fragile.

The method comes from dynamic topology adaptation in power systems. Online Learning for Robust Voltage Control Under Uncertain Grid Topology keeps a "consistent set" of every grid topology compatible with the measurements and acts safely with respect to the whole set, rather than trusting one estimate. Online Energy Price Matrix Factorization for Power Grid Topology Tracking explains observed signals as a known structural operator times a sparse matrix of change events plus noise. We borrow these two ideas and not the algorithms. Proteins obey no circuit laws, and no one can actuate a proteome, so the finite-mistake guarantee and the ADMM factorization do not transfer. What transfers is the stance: report only conclusions that hold across the set of plausible networks, and model structured nuisance explicitly instead of hoping it averages out.

Research Program

The work has three pieces. The first is an acquisition-proximity score for every protein pair. From the run's isolation-window scheme and the retention times of each protein's quantifying peptides, we compute the fraction of fragment signal from protein $i$ that co-elutes in the same window as signal from protein $j$ or from a shared third interferer. We then weight that fraction by how much the interferers' abundance differs between conditions. The output is a pair-level covariate $\pi_{ij}$. It is cheap to compute from standard DIA software output.

The second piece is the set-valued network. For each condition $c$ we build WGCNA adjacency and topological-overlap networks under a family $W$ of workflows drawn from the benchmark's design space, and under $B$ bootstrap resamples of samples. Before soft-thresholding, each correlation is adjusted by regressing out its dependence on $\pi_{ij}$. Modules are defined once, on a pooled reference network, so that module identity does not change across the set. For module $m$ we compute a standard preservation statistic $Z_c(m; w, b)$ for each workflow $w$ and resample $b$, and define a certified rewiring score

$$R(m) = \min_{w \in W} \, q_{0.05,\,b}\big[\, Z_{\text{ref}}(m; w, b) - Z_{c}(m; w, b) \,\big],$$

Here $q_{0.05,b}$ is the 5th percentile over bootstrap resamples. A module counts as rewired only if $R(m)$ exceeds a threshold calibrated on null data. The minimum over workflows is the "consistent set" principle: one plausible workflow that shows no rewiring vetoes the claim. The cost is power, and measuring that cost honestly is part of the evaluation.

The third piece is a planted-rewiring generator for testing power. We take real DIA data and, within one group only, perturb the correlations of a chosen protein set along a low-rank direction. The perturbation is injected at the precursor level, before quantification, so it passes through the same software. This gives known positives at controlled effect sizes and group sizes.

Baselines

The naive reference point is what the field does now. We run WGCNA separately per condition on one default workflow and call a module rewired when its preservation $Z_{\text{summary}}$ drops below the conventional threshold. It is run on the same samples and the same pooled module definitions as our method, with $\beta$ chosen by the standard scale-free fit. It sets the false-rewiring rate we expect to beat.

The strongest existing competitor is per-module differential correlation with label-permutation testing on the best single workflow. This is the most rigorous form of the per-group approach. It controls sampling noise properly, and a reviewer would consider it sufficient. We give it the same samples and the same bootstrap budget, used as permutations. It is also allowed to pick its workflow by oracle, as whichever workflow the benchmark finds most accurate for differential abundance. On the spike-in benchmark there is no reported result found on Lacuna for co-expression rewiring by any method. The benchmark paper reports workflow counts (1,428) and 2,100 bootstrap datasets with group sizes from 3 to 23, but no F1 or AUC values. So we will produce this baseline's numbers ourselves.

We also include the shared-network model from Robust Multi-view Co-expression Network Inference, with conditions treated as views, as a robust-estimation competitor. It runs on a protein subset small enough for its EM procedure, and differences in the view-specific component are read as rewiring. The paper reports only that ROC curves improve as the number of views grows from 2 to 10, so there is no reported number to quote.

Two ablations isolate the ingredients: the workflow set without the proximity adjustment, and the proximity adjustment on a single workflow. Beating the strongest baseline means two things. On the null benchmark, our method must have a lower false-rewiring rate at matched nominal level. On the planted benchmark, at group sizes of 10 to 23, it must recover planted rewiring at no worse than modestly reduced power, which we define in advance as at least 80% of the baseline's true-positive rate.

Why not just…?

Why not just run WGCNA per group and compare module preservation? This is the naive baseline, and in large, clean cohorts it may be adequate. Our hypothesis is that it fails specifically because interference is condition-linked. When an abundant interferer shifts between groups, preservation genuinely drops in the measured data, and no per-group method can tell that drop apart from biology. The spike-in null shows directly whether this happens. If false rewiring under the naive method is already near nominal, the premise is wrong and we will report that.

Why not just use differential correlation with permutation tests? Permutation protects against chance. It does not protect against a systematic artifact that tracks the condition labels, because shuffling labels destroys the artifact in the null distribution, and the real-label artifact then looks significant. The ablation "proximity adjustment, single workflow" tests whether a one-workflow permutation method plus our covariate is enough. If it is, the workflow set is unnecessary and the proposal collapses to the covariate. That would be a smaller but still useful result.

Why not just pick the best DIA workflow? This is the most serious threat. The benchmark paper shows workflows differ, so one may simply be clean enough. We will compute the false-rewiring rate for every workflow in $W$ separately. If some workflow matches the certified score's false-rewiring rate with more power, we recommend that workflow and drop the minimum-over-workflows rule. Our method only wins if no single workflow is uniformly clean, so that the identity of the best workflow depends on the module.

Why not just fit a time-varying graphical model? The regime-based approach in Inferring Time-Varying Network Topologies from Gene Expression Data and shared-network models assume measurement noise is unstructured. Structured interference would be absorbed as real edges. These models also do not scale to thousands of proteins. We include the multi-view model on a subset to measure the first failure directly, not merely to assert it.

Decisive Experiments

The central experiment uses the inter-patient heterogeneity benchmark from Benchmarking of analysis strategies for data-independent acquisition proteomics using a large-scale dataset comprising inter-patient heterogeneity. It contains formalin-fixed lymph node tissue from 92 individuals, with E. coli digest spiked in at human-only, 1:25, 1:12 and 1:6. Human background proteins have no biological reason to rewire across spike levels. Every human-module rewiring call between spike groups is therefore a false positive, and E. coli peptides provide a graded, known source of new interference. The first prediction is that false rewiring under the naive and permutation baselines rises with spike ratio. The second is that it concentrates in modules with high acquisition proximity to E. coli peptides. The third is that the certified score holds false rewiring near nominal across bootstrap group sizes from 3 to 23. If false rewiring does not track proximity, the mechanism is wrong. The planted-rewiring generator applied to the same data then measures power against effect size and group size.

Two applications follow, and only methods that pass the benchmark are run on them. The paired coronary occlusion plasma cohort (ProteomeXchange PXD034267, 10 patients before and after revascularization plus 10 controls) is an extreme small-sample case dominated by abundant plasma proteins. We will ask whether any module rewiring around the seven reported candidate proteins survives certification, and we expect most of it may not. If the Alzheimer's resilience data are accessible, we will test whether the M5 module's elevation in dementia comes with certified internal rewiring or is a shift in level only. That distinction the original static analysis could not make. The label-free developmental cortex cohort serves as a non-DIA contrast. Where there are no isolation windows, the proximity adjustment should change little, which gives a check that the covariate is specific to DIA.