Calibrating Co-Expression Tipping-Point Signals on Power Grids

Biologists who want to know when a tissue is about to tip into disease (a stable artery wall rupturing into infarction, a tumour starting to metastasize) usually look for a warning sign in gene expression. Near the tipping point, a small group of genes starts to fluctuate more and to move together. The standard way to score this is an ad hoc ratio. It can flag the tipping point after the fact, but it has no unit, no threshold, and no way to say how far away collapse is. Power engineers face the same question about voltage collapse, and they answer it with margins measured in units of stress: how much more load the grid can take before its equilibrium disappears. This note proposes to give co-expression analysis a margin of that kind. The margin is read from how gene fluctuations scale as the system is pushed toward the tipping point, and it is extrapolated forward to predict where collapse will occur.

The main obstacle is not the mathematics. Biology has no ground truth: nobody knows the true distance to a disease bifurcation in a patient cohort, so no tipping-point index has ever been properly calibrated. Power grids do have ground truth. We therefore propose to build a calibration benchmark from standard grid test cases, where the exact collapse point is computable, and to validate the margin there before applying it to staged transcriptomes. We treat stability margin quantification as the source of both the method and the ground truth, and weighted gene co-expression network analysis as the application setting.

Core Claim

Consider a system of coupled variables (bus voltages, gene expression levels) jittered by small random noise around a stable equilibrium. Its fluctuations are governed by the Jacobian $J$, the matrix of how each variable's rate of change responds to every other variable. In the small-noise regime, the stationary covariance $C$ of the fluctuations satisfies the Lyapunov equation

$$J C + C J^{\top} + D = 0,$$

where $D$ is the noise covariance. As a control parameter $\mu$ (load level, disease stress) approaches a saddle-node bifurcation at $\mu_c$, the eigenvalue of $J$ closest to zero shrinks like $\sqrt{\mu_c-\mu}$; this follows from the normal form of the saddle-node. The leading eigenvalue of $C$ grows like the inverse of that eigenvalue, so

$$m(\mu) \equiv \left(\frac{\sigma^2}{\lambda_{\max}(C(\mu))}\right)^{2} \;\approx\; a\,(\mu_c-\mu),$$

where $\sigma^2$ is a noise-scale normalizer estimated from genes (or buses) outside the critical group, and $a$ is a system-dependent constant. The quantity $m$ is our proposed margin. It is dimensionless and falls roughly linearly to zero at collapse. Given measurements at a few early stages, fitting the line and extrapolating to $m=0$ predicts $\mu_c$. This mirrors how the power reserve coefficient $K_p$ measures distance to the nose of the voltage-versus-load curve in Static Voltage Stability Analysis of Power System Using the Combination of P-V Curve and the Modal Analysis. That paper shows voltage magnitudes stay deceptively flat because local controllers hold them up, while load-based margins reveal the true reserve.

The claim is that this linearized, noise-normalized, extrapolated margin predicts the transition stage from pre-transition data, and does so earlier and more accurately than peak-detection indices. We further claim that the leading eigenvector of $C$ identifies the weak nodes that the true Jacobian's participation factors identify. Both claims are testable exactly on grids and approximately on biological time courses.

Relation to Prior Work

The closest work is Identification of Key Genes as Early Warning Signals of Acute Myocardial Infarction Based on Weighted Gene Correlation Network Analysis and Dynamic Network Biomarker Algorithm. It runs weighted gene co-expression network analysis (WGCNA), a method that clusters genes into co-expressed modules. Separately, it runs the dynamic network biomarker (DNB) index, which is within-group variance times within-group correlation divided by correlation to outside genes, on 24 patients across three stages. The DNB index peaks at the intermediate-lesion stage, and the authors intersect the resulting genes with WGCNA modules. Our proposal differs in three ways. First, WGCNA there is a static pooled network, whereas we compute the spectral margin per stage inside each module. Second, the DNB index can only be read as a peak, whereas our margin has a predicted zero crossing. Third, no index in that line has been checked against a known bifurcation. The broader DNB literature motivates our view of disease onset as a bifurcation: hepatocellular carcinoma metastasis, landscape DNB in atherosclerosis, skin photodamage, MCF-7 differentiation, and Arabidopsis floral transition. All of these studies score tipping points by peaks of heuristic statistics.

On the grid side, three papers supply ground truth and constraints rather than the method. Voltage collapse in complex power grids derives a closed-form margin from an inverse network stiffness matrix and reports that its $\Delta=1$ condition predicted collapse within 2% for systems with low power factors. We use it and continuation power flow as the oracle distance. We deliberately do not transplant its stiffness matrix onto the topological overlap matrix, because a co-expression graph has no analogue of the reactive load vector that the formula needs. Hierarchical and Distributed Monitoring of Voltage Stability in Distribution Networks shows that a diagonal Jacobian approximation works on tree-shaped feeders, with error rising to 2–7% near collapse. It constrains us: dense co-expression modules will not admit such local approximations, so we work with the full module covariance. PMU-Based Distributed Non-Iterative Algorithm for Real-Time Voltage Stability Monitoring supplies bounded index conventions and meshed test systems. All three grid papers assume the Jacobian is known. We assume only that fluctuations are observed, which is the biological situation.

Research Program

The first component is the grid calibration benchmark. On MATPOWER systems (IEEE 30-, 300-bus, and the 2,383-bus Polish grid, as used in the papers above) and on the IEEE 123-bus feeder, we step the load level toward collapse. At each level we inject small stochastic load fluctuations, linearize the dynamics, and draw "samples" of bus voltages. We record the exact distance to collapse and the Jacobian's participation factors. We then run the unmodified WGCNA pipeline on each loading level: soft-thresholded correlation, topological overlap, modules, and module eigengene. For each module we compute the margin $m$.

We add two biological realism knobs. One sets the sample count to 8–30 per stage, the regime of the AMI cohort. The other adds sample-to-sample parameter heterogeneity, because cross-sectional patients are different systems, not repeated draws from one system. We also switch voltage-regulated buses on and off. These buses play the role of homeostatic control that masks stress. They let us test whether masking suppresses variance-based warnings, which the modal-analysis paper shows it does for voltage magnitudes.

The second component is estimation. With few samples the top eigenvalue of a sample covariance is biased upward. We estimate $\lambda_{\max}$ within modules with shrinkage and subtract the value expected from pure noise at the same size, estimated by permuting samples within stage. We also fit the margin line with bootstrap intervals so that every predicted $\mu_c$ carries uncertainty.

The third component is transfer to biology. We apply the calibrated procedure to staged transcriptomic datasets in which prior DNB studies located a transition, and ask whether extrapolation from stages before the reported transition predicts it. Candidate cohorts are the AMI cohort, the Arabidopsis floral time course, the MCF-7 differentiation series, and the photodamage series. We will also rank genes by their loading on the leading eigenvector. The goal is to separate genes that drive the critical direction from hubs that are merely central, which the standard module-membership ranking conflates; see RNF181 and breast-cancer hub genes.

Baselines

The naive reference is mean variance of module genes per stage, and on grids also the mean voltage magnitude. Voltage magnitude is known to be deceptively flat under control, per the P-V/modal analysis paper. Beating either one means a smaller error in predicted collapse location at equal numbers of pre-transition stages.

The strongest existing method is the DNB index as defined in the AMI paper, computed on the same modules, samples, and stages. We will give it the most favourable reading. It "predicts" the transition at the stage after its largest increase, and we will also try a tuned threshold on its growth rate, fit on the same training trajectories our margin uses. Neither the file nor Lacuna gives a quantitative lead-time or localization-error result for DNB: no reported result found on Lacuna for DNB tipping-point localization error. We will therefore establish this baseline ourselves on the grid benchmark, where truth is known. Beating it means lower absolute error in predicted $\mu_c$ on grids when only stages at least 30% of the load range before collapse are used. It also means correct stage prediction in more biological series than DNB when DNB is denied the transition stage itself.

The key ablation is the raw leading eigenvalue of the module correlation matrix, which is essentially WGCNA's eigengene variance explained. We run it with the same shrinkage and the same linear extrapolation, but without the noise normalization and without the square-root linearization. This isolates whether the physics-derived transform matters. A second ablation replaces the full module covariance with its diagonal, the analogue of the tree-feeder approximation, to test whether dense coupling matters. For eigenvector-based gene ranking, the baseline is intramodular connectivity and module membership. Success means higher rank correlation with the true Jacobian participation factors on grids.

A supervised baseline regresses stage-to-transition on expression, trained leave-one-trajectory-out with the same tuning budget. For context on how weak such models tend to be with around 50 training samples, The TCGA Meta-Dataset Clinical Benchmark reports 68.78% average accuracy for logistic regression versus 66.18% for a majority classifier. That is a different task, so it is context only, not a target.

Why not just…?

Why not just use the DNB index? It is the incumbent, and in data sets with many stages its peak may appear just as early as our zero-crossing prediction. If so, the extrapolation adds nothing but a confidence interval. The grid benchmark settles this directly: it measures both methods' localization error as a function of how much pre-transition data each sees. We will report honestly if DNB peak detection matches the margin.

Why not just track the module eigengene's variance explained over stages? This is the most serious threat, because our margin is a transform of essentially the same spectral quantity. If the raw eigenvalue with a linear trend predicts collapse as well as $m$ does, the contribution shrinks to the calibration benchmark and the masking and heterogeneity findings. That would still be useful, but the claim would be narrower. The saddle-node scaling argument predicts that raw eigenvalue trends curve sharply only near collapse and so extrapolate poorly from early stages. The ablation tests exactly this.

Why not just port the stiffness-matrix margin $\Delta$ onto the topological overlap matrix? It looks like the obvious bridge, but the formula needs a load vector and a physical stiffness. A co-expression graph supplies neither, so the resulting number would have no reason to hit 1 at a transition. We include it as a naive transplant baseline on the grid benchmark to show this rather than assert it.

Why not just train a classifier on expression? A classifier needs labelled transition-stage samples, which are exactly what is missing before a transition happens. The leave-one-trajectory-out regression baseline measures how far such models get.

Decisive Checks

The first check is on grids. Using only loading levels far from collapse, the extrapolated margin must predict the collapse load within a few percent in the regime where Voltage collapse in complex power grids itself achieves 2%, and it must beat both DNB and the raw-eigenvalue ablation. If it fails even with abundant samples, the approach is dead.

The second check is robustness. As samples drop to 8 per stage and patient heterogeneity rises, we find the point where the margin's advantage vanishes. We also measure how much voltage-regulated buses delay every variance-based signal. This yields a concrete sample-size requirement for biological studies.

The third check is on biology. In series with at least four pre-transition stages, extrapolation must predict the transition stage that prior DNB studies located, without seeing that stage. The leading-eigenvector genes should also overlap reported DNB genes more than top-connectivity hubs do.

A negative result on the grid, or a margin that matches its raw-eigenvalue ablation, would be reported as the answer.