Finding "Dark Channels": Detecting Drivers That Rewire Without Changing Value in Multivariate Time Series

In gene-expression biology, some of the most important regulators are "dark genes": their expression level stays flat across conditions, but whom they interact with changes sharply. Standard differential-expression tests miss them by construction. Network methods catch them only in static snapshots, because they compare one correlation graph per stage. Multivariate time-series anomaly detectors have the same blind spot, from the other side. They flag a sensor or variable when its values become hard to reconstruct or forecast. A channel whose mean and variance stay normal while it silently changes which other channels it drives therefore looks healthy, even when it is the root of the fault.

We propose to measure, for every channel and every time window, how much that channel's past helps forecast the other channels' future, and the reverse. We then flag channels whose influence pattern shifts while their own value distribution does not. The measurement uses a forecaster pretrained to infer cross-channel couplings from its context window. That lets the influence map be recomputed cheaply in each window as the couplings change, without labels, without a protein-interaction database, and with one masked forward pass per channel rather than one per channel pair. We test the approach on simulated gene networks with planted bifurcations and known drivers, on industrial sensor streams with injected "dark" faults that preserve each channel's value distribution, and on a human influenza time course where a static tipping-point detector has a published early-warning horizon.

Core Claim

The claim is that a time-lagged, directed, per-window influence map localizes value-invariant drivers of a regime change better than three simpler alternatives: reconstruction or forecast error, sliding-window correlation graphs, and windowed linear Granger tests. This advantage should hold specifically when coupling is lagged, nonlinear, or confounded by a common driver, and it should come at a cost that grows linearly in the number of channels. We also expect the claim to fail in one identifiable regime, instantaneous linear coupling with long windows. There, rolling correlation should do as well, and we will report that.

For a pretrained forecaster $f$, a window $w$, and channels $i, j$, define the influence of $j$ on $i$ as the increase in forecast loss on channel $i$ when channel $j$'s history is replaced by an uninformative mask:

$$I_{j\to i}(w) = \mathcal{L}_i\big(f(X_w^{\setminus j})\big) - \mathcal{L}_i\big(f(X_w)\big).$$

Here $X_w$ is the multichannel context in window $w$, $X_w^{\setminus j}$ is the same context with $j$ masked, and $\mathcal{L}_i$ is the forecast loss on channel $i$ over the next horizon. One masked pass yields a whole column $I_{j\to\cdot}$, so the full map costs $C$ passes for $C$ channels.

The dark-channel score for channel $i$ in window $w$ is the standardized deviation of its combined incoming and outgoing influence vector from a reference period, with deviation measured in units of its reference-period variability. We report this score alongside a marginal-shift statistic for the same channel. A channel is "dark" when the first is high and the second is not.

Where This Sits Relative to Prior Work

The biological literature supplies the phenomenon and the evaluation targets. Cell-specific network constructed by single-cell RNA sequencing data builds a dependency network per cell and showed that genes with static expression can shift degree across states. Potential network markers and signaling pathways for B cells of COVID-19 based on single-cell condition-specific networks found such rewired but non-differentially-expressed genes in infected B cells. scGIR weights these graphs by expression and ranks genes with PageRank. All three motivate the proposal. All three are also undirected and have no notion of time lag.

The tipping-point line is the closest in spirit. Landscape dynamic network biomarkers score single samples by rising within-module correlation and falling outside correlation. Single-sample Jensen-Shannon divergence scores divergence from a reference. DNFE adds edge direction through an empirical index on co-expression. CCNE infers "causal" edges by continuity over a protein-interaction scaffold along pseudotime. These methods constrain our design in two ways. First, their success at flagging pre-transition states means our influence map must at least match their early-warning horizons. Second, their reliance on instantaneous correlation or a fixed scaffold is exactly what we drop. We replace it with directions that come from time order ($X_{t-k}\to Y_t$) and are learned from the data itself.

The time-series literature supplies the machinery and the counterexamples. CCM-TAD contributes a causal mixing layer that prevents a timestep from seeing its own future, and we reuse this masking in our forecaster. Its channel clusters, however, are computed offline, so a channel that migrates between clusters is invisible to it. MTAD-GAT learns a feature-level attention graph and ranks root-cause channels, but its ranking is driven by per-channel error. Anomaly Transformer contrasts each timestep's local and global attention within a series, not across channels.

CGAD is the nearest time-series competitor. It learns lagged causal graphs per window with DYNOTEARS, a continuous-optimization method for dynamic Bayesian networks, and scores structural distance to labeled normal and attack profiles. It shows that lagged structure beats static structure. It needs labeled attack profiles, treats all nodes symmetrically, and does not scale to the thousand-plus channels of transcriptomic data. We keep its key insight that lagged structure matters, and drop the labels and the full per-window graph search.

Finally, Zero-shot forecasting of chaotic systems shows that a univariate foundation model forecasts by copying past context patterns and degrades under parameter drift. This motivates a multivariate forecaster whose cross-channel dependence is inferred from context rather than memorized.

Research Program

The forecaster is a causal-masked mixer in the style of CCM-TAD, with one change: instead of fixed channel clusters, it has attention across channels computed from the current context window. We pretrain it on a large synthetic corpus of coupled stochastic dynamical systems whose coupling graphs, lags, and nonlinearities are randomly drawn and randomly rewired mid-trajectory. The training objective is next-horizon forecasting under random channel masking. The purpose of this pretraining is amortization. The model should learn to read the current couplings from the context, so that masking a channel removes the information that channel actually carries now, not the information it carried during training.

This choice addresses the main weakness of masking-based influence. A forecaster fitted once on normal data would miss newly formed couplings, whereas an in-context coupling reader should not. Whether it succeeds is an empirical question, and we test it directly with the frozen-versus-adaptive ablation below.

On top of the influence maps we build two read-outs. The first is the per-channel dark score defined above. The second is a system-level early-warning score in the spirit of dynamic network biomarkers, computed on the directed map instead of on correlations. We identify candidate modules by clustering the map within each window, then track two quantities: rising influence within a module and falling influence between that module and the rest. For gene-scale data, we first prescreen to a few thousand highly variable genes, as the biological methods also do. Cost is then $C$ forward passes per window, which we will report in wall-clock time against windowed DYNOTEARS.

We will also release a benchmark construction that does not exist today: dark-fault injection into real sensor data. We take normal segments of WADI and SWaT and pick a target channel. We then re-synthesize that channel so that it is driven by a different upstream channel, at a different lag, or with the sign of its coupling flipped. Finally, we rank-remap the result back onto the channel's original empirical distribution, so that its marginal values are statistically unchanged. Ground truth is therefore known: which channel went dark, and when.

Baselines

The naive reference point is per-channel univariate scoring: a rolling z-score of each channel against its reference period, which is what differential-expression analysis amounts to in the time domain. On our injected dark faults it should be near chance by construction. That makes it a sanity check on the injection procedure rather than a competitor, and any real-data dark channel it catches tells us the injection was not truly marginal-preserving.

The strongest simple competitor is a rolling dynamic-network-biomarker score. For each window we compute the Pearson correlation graph and score each channel by its change in degree and its within-module versus between-module correlation, following the recipe of landscape DNB applied to consecutive time windows. This baseline is fast, unsupervised, and designed for exactly the phenomenon we target, so it is the one most likely to match us. We give it the same windows and the same reference period, and we tune its window length and threshold on the same validation split and with the same search budget as our method. Beating it means higher driver-localization accuracy (top-k hit rate and NDCG against the injected or simulated ground truth), specifically on lagged, nonlinear, and common-driver faults. We do not require a win on instantaneous linear faults.

The strongest structural competitor is CGAD made unsupervised: DYNOTEARS run per window, with each node scored by its structural Hamming distance to the reference-period graph, using the same windows. On WADI, CGAD reports F1 = 0.8913 against 0.6679 for the next baseline, SMV-CGAD, in Causal Graph Profiling via Structural Divergence. That result is supervised, detection-level, and not localization, so it is not directly comparable to our task. We will rerun the unsupervised variant ourselves, capped at the same compute budget. On gene-scale inputs it will likely exceed that budget, and we will report where that happens.

For error-based detectors we run CCM-TAD, MTAD-GAT, and Anomaly Transformer with per-channel error used as the attribution. For their standard detection numbers on the original benchmarks, CCM-TAD reports Best-F1 gains of 8.71% on WADI and 8.65% on PSM over the next-best models (CCM-TAD). Anomaly Transformer reports F1 of 92.33% on SMD and 96.69% on SMAP (Anomaly Transformer). MTAD-GAT reports root-cause NDCG@5 of 0.8556 on its proprietary TSA data (MTAD-GAT), which we cannot access. We do not expect to beat these detectors on ordinary value-changing attacks. The claim is that adding the dark score to their scores improves detection on the dark-injected variants without lowering it on the originals.

On the biological side, the reference result is landscape DNB's warning of severe influenza symptoms "at least 8 hours" before onset on GSE30550 (landscape DNB). On synthetic bifurcating gene networks, DNFE is reported to outperform sJSD, SLE, HMM, and KLD, and CCNE is reported to spike near the bifurcation point. No numeric score for either was found on Lacuna, so we will rerun both on our simulations.

Three ablations isolate the key ingredients. The first replaces the in-context pretrained forecaster with one trained only on the reference period and then frozen; this tests whether amortized coupling inference matters. The second restricts masking to lag zero; this tests whether time-lagged influence matters versus instantaneous influence. The third swaps the learned forecaster for a windowed linear vector-autoregression Granger test; this tests whether nonlinearity and amortization matter versus the influence definition itself.

Why not just…?

Why not just run a sliding-window correlation graph and track each gene's degree, as in CSN or landscape DNB? This is the real threat, and we treat it as the primary baseline rather than arguing it away. Correlation within a window is undirected and instantaneous. It therefore cannot separate a channel that drives others from one driven by a shared upstream source, and it can miss couplings that act at a lag or through a nonlinearity with little linear correlation. DNFE reports that undirected co-expression struggles relative to directed flow on noisy data, and scGIR shows that plain degree loses to weighted ranking (ARI 0.53 vs 0.68 on Chu-time). Neither result tests lagged dynamics, though. Our fault-type breakdown settles the question. If rolling correlation matches us on lagged and common-driver faults, the learned influence map is not earning its cost, and we would report the method as a scaling device only.

Why not just use reconstruction or forecast error from a strong multivariate detector and attribute blame to the highest-error channel? Error-based attribution flags the channels whose values became unpredictable. After rewiring, those are typically the downstream channels, not the dark driver whose own values still look normal. This is the core tension identified across CCM-TAD, MTAD-GAT, and the dark-gene work. The injected-fault benchmark measures this directly by comparing the driver's rank to the rank of its downstream targets under each attribution.

Why not just run windowed DYNOTEARS, as CGAD does, and score nodes by structural change? It is the principled version of what we approximate, and on small sensor sets it may be as accurate. Its costs are a full graph search per window and high variance with short windows, which is the typical regime for biological time courses. It also needs a threshold per edge where our map is continuous. We compare on WADI- and SWaT-scale channel counts, where it is feasible, and on 1,000-node simulated gene networks, where we expect it to exceed the budget. If it wins on accuracy at small scale, we will say so and position our method as the scalable approximation.

Why not just forecast each channel with a univariate foundation model like Chronos? A univariate model has no access to the other channels, so it cannot represent a change in who drives whom at all. Zero-shot forecasting of chaotic systems also shows that such models degrade under drift because they copy patterns from their context. We include it only as an error-based baseline.

Proposed Experiments

The first experiment uses simulated gene-regulatory networks at 11, 100, and 1,000 nodes, following the setups used by DNFE and CCNE. We extend the simulations to include explicitly planted dark drivers: nodes whose steady-state expression distribution is held fixed while their outgoing couplings change at a known time before a bifurcation. Metrics are top-k driver recovery, NDCG, and early-warning lead time relative to the bifurcation, with 30 random seeds per setting and false-alarm rates measured on runs with no bifurcation.

The second experiment applies the dark-fault injection to WADI and SWaT. Results are reported per fault type (lag change, driver swap, sign flip, common-driver confound) and per window length, which directly maps the regime where rolling correlation suffices. On the unmodified benchmarks we also report whether combining our score with CCM-TAD's changes its F1.

The third experiment is the real-data check on GSE30550. It tests whether the directed early-warning score separates symptomatic from asymptomatic volunteers, and whether its lead time equals or exceeds the reported 8-hour horizon. It also checks whether the top dark genes overlap those reported by landscape DNB and sJSD more than a random gene set of equal size would.

With 16 time points per subject, per-subject windows are very short. We will therefore pool subjects as a panel and use lags of one or two sampling steps. We treat this experiment as a plausibility check rather than the main evidence, and if the method cannot beat landscape DNB there, we will report that honestly.

The decisive check for the whole proposal is the fault-type breakdown against rolling correlation. The method is worth pursuing only if it recovers dark drivers clearly better on lagged and confounded faults, stays at parity on instantaneous faults, and does so at linear cost in channel count.