Most of the genetic signal behind common diseases sits in thousands of variants too weak to pass genome-wide significance. This is the "missing heritability" problem. One productive response has been to ask where in the gene regulatory network that weak signal accumulates. The clearest answer so far is that it concentrates in densely connected hub genes inside tissue-specific network modules. That answer comes from static maps built from postmortem tissue. Many disease-relevant processes are not static, though. A T cell activating, a stem cell becoming a cardiomyocyte, a placenta failing: each is a transition between states. Control theory gives a precise language for which genes can push a network through such transitions. "Driver" genes are those that must receive an outside input for the whole network to be steerable, and they tend to be peripheral rather than hubs. This proposal asks a simple question with two possible answers. Is disease heritability enriched in the genes that steer a regulatory network through a biological transition, or in the hubs that hold it steady? And does the answer survive once we control for how connected a gene is?
The obstacle so far has been technical. The steering analysis, called structural controllability, needs a directed network. The best tool for separating direct from indirect gene–gene links, partial correlation (the correlation between two genes after removing what all other genes explain), gives only undirected edges. Existing control-theoretic disease pipelines sidestep this by orienting plain Pearson correlations with a generic protein-interaction database. We instead take direction from time. We estimate lagged partial correlations on time-course expression data, so each edge says "gene $j$ now predicts gene $i$ later, given every other gene now." We then compute each gene's controllability role on this directed, context-specific network. Finally, we use stratified LD score regression to test whether those roles carry heritability for matched diseases beyond degree, module hub status, and time-varying expression.
Core Claim
Our hypothesis is testable and falsifiable: in time-resolved networks, genes that are always drivers (unmatched in every maximum matching) or that sit in input-free root components carry heritability enrichment for context-matched traits, and this enrichment remains after conditioning on degree, local hub score, and temporal differential expression. The prediction is that the same annotation built from a static network will show weaker or no conditional enrichment. If controllability roles add nothing once degree and dynamic expression are in the model, that is also a clean and publishable result. It would say the hub-centric picture of The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution is sufficient, and that control-theoretic gene prioritization adds nothing beyond topology proxies.
Where This Sits
Three lines of prior work motivate the question. The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution built bipartite SNP–gene networks for 29 GTEx v8 tissues. It found that roughly 29% of modules were trait-specific and that heritability enrichment peaks in local core hubs. Its maps are static eQTL associations with no notion of direction or steering. Controllability of structural brain networks showed on brain wiring that average controllability tracks degree (r = 0.91), while modal controllability, the ability to reach hard-to-reach states, is almost perfectly anti-correlated with degree (r = −0.99). That result creates the tension we test: hub heritability and transition control point to different genes. It also constrains our design. Because the modal and average metrics are nearly deterministic functions of degree, any enrichment we find for them is uninterpretable unless degree is in the model. For that reason our primary annotation is the matching-based driver set from Emergence of bimodality in controlling complex networks, which is less trivially tied to degree. We use root strongly connected components from A Structured Systems Approach for Optimal Actuator-Sensor Placement in Linear Time-Invariant Systems as a second annotation.
The closest existing pipeline is cPE, from Identifying preeclampsia-associated genes using a control theory method. It builds Pearson correlations from placental microarrays, orients them with a directed protein-interaction network, and ranks critical control nodes by fold change. It reached a 24.9% validation rate against dbPEC, compared with 20.2% for differential expression, 23.9% for betweenness, and 22.8% for hubs. cPE differs from us in three ways. It does not condition out indirect links. It takes direction from a context-free database. It validates against a curated gene list rather than against genome-wide heritability. Validation of gene regulatory network inference based on controllability supplies a principle we reuse: judge an inferred network by whether it gets control right, not only by edge accuracy. That paper found the two can disagree, but it was limited to 10-gene Boolean models.
Our methods come from three sources. The analytic, layer-specific covariance shrinkage of DRAGON stabilizes our precision estimates when genes outnumber samples. Noise cleaning the precision matrix of short time series addresses the short, noisy series typical of human time courses. The polynomial-time minimum input selection of Olshevsky handles driver computation at genome scale. One constraint cuts the other way. geneRNIB found that plain Pearson correlation and GRNBoost2 beat prior-heavy and foundation-model methods, so we do not assume that more elaborate inference helps. Simple builders are baselines throughout.
Research Program
Directed networks from time. For a time course with expression vectors $x(t)$ over $p$ genes, pooled across donors or replicates, we estimate a regularized lagged precision structure. The edge from gene $j$ to gene $i$ is
$$\rho_{j \to i} = \mathrm{corr}\big(x_i(t+1),\, x_j(t) \mid x_{-j}(t)\big),$$
where $x_{-j}(t)$ is every other gene at time $t$. This is the partial regression coefficient of a sparse first-order vector autoregression, rescaled to a correlation. We compute it through a shrinkage estimate of the joint covariance of $[x(t), x(t+1)]$, with separate DRAGON-style shrinkage for the "now" and "later" blocks. The quantity is not new. What is new is using it to give controllability analysis its edge directions. Self-loops are allowed, which matters because they change matching outcomes. For single-cell data we replace clock time with pseudotime bins within a single lineage branch. This avoids the branching "shadow problem" described in scKAN.
Controllability roles. Driver sets from maximum matching are not unique. For each gene we therefore compute its role class: always-driver (unmatched in every maximum matching), never-driver, or sometimes-driver, following Emergence of bimodality in controlling complex networks. We also record whether the gene belongs to a root component. As secondary, degree-entangled quantities we report average and modal controllability. Each role is recomputed over bootstrap resamples of time points, and only roles stable in at least 80% of resamples enter the genetic analysis.
Heritability partitioning. Gene sets become SNP annotations using a standard window around each gene. Stratified LD score regression is then fit jointly with the baseline functional annotations plus four competing gene-level annotations: degree in the same network, local hub core score computed as in the GTEx network study, time-varying differential expression, and an all-expressed-genes control. The test statistic is the conditional per-SNP coefficient for the driver annotation. Marginal enrichment alone is not enough.
Contexts. The primary context is T cell activation. Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes shows that genetic effects there are stage-specific and that immune-disease GWAS offer matched traits. The secondary context is the iPSC-to-cardiomyocyte progression in the 119-individual Yoruba panel used by EGRET, paired with cardiac traits. That panel has only three stages, so lag estimation there is coarse, and we treat it as a replication rather than a primary test. Negative-control traits with no plausible link to the tissue are included in every run.
Baselines
The naive reference point is an annotation of all genes expressed in the context, together with random gene sets matched on degree distribution and expression level drawn from the same network. The driver annotation must exceed the 95th percentile of the matched-random null to count as enriched at all.
The strongest existing competitor on the genetic question is the local-hub annotation of The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution, because it already shows that network position partitions heritability. We recompute core scores on both GTEx eQTL networks for the matched tissue and on our own networks, and fit them jointly with the driver annotation using the same regression, SNP set, and gene windows. The file reports that local hubs carry the highest enrichment, but it gives no enrichment coefficient: no reported result found on Lacuna for a local-hub heritability enrichment value. We therefore report our own replication number before comparing. Beating this baseline means a driver coefficient that is significant after the hub and degree terms are in the model.
For prioritization, the strongest control-theoretic competitor is cPE, from Identifying preeclampsia-associated genes using a control theory method. It reports a 24.9% validation rate against dbPEC on GSE75010. We rerun cPE and our pipeline on GSE75010 with the same gene universe and the same top-k cutoff. That dataset is cross-sectional, so in this comparison our variant uses contemporaneous shrinkage partial correlation oriented by the same directed protein-interaction network. This isolates the effect of conditioning. Beating cPE means a validation rate above 24.9% at matched k. We also use GRNBoost2 as an alternative network builder: its edges are oriented from transcription factor to target and passed through identical matching. geneRNIB found it among the most reliable simple methods, and all builders receive the same hyperparameter search budget.
Two ablations isolate the key ingredients. The first keeps the lagged partial correlation but permutes time labels within donors. This preserves marginal statistics and destroys temporal order, so any enrichment that survives cannot come from directionality. The second keeps the lag but drops conditioning, using lagged pairwise correlation. This isolates the partial-correlation step.
Why not just…?
Why not just use degree or hub score, since controllability is nearly a function of degree? On brain networks, modal and average controllability correlate with degree at −0.99 and 0.91 (Controllability of structural brain networks), so this threat is serious for those metrics. That is why they are secondary. Matching-based driver status depends on in- and out-degree jointly and on path structure, not on total degree. The joint regression settles the question directly. If the driver coefficient vanishes once degree and core score enter, we conclude the reviewer was right.
Why not just annotate genes that change over time? Temporal differential expression is the cheapest dynamic annotation. cPE's own analysis suggests control nodes are often not differentially expressed: 85.2% of its top genes were absent from the differential-expression list. That result used static data, however. Here temporal differential expression is a covariate in every model. If it absorbs the driver signal, the network adds nothing.
Why not just orient a static network with protein-interaction databases, as cPE does? Database direction is not specific to the context, and geneRNIB warns that static priors can discard context-specific links. The permuted-time ablation and the static-network comparison test this head on. Suppose database-oriented static networks give the same conditional enrichment as time-oriented ones. Then time-derived direction is unnecessary, and we would say so.
Why not just use a nonlinear inference method such as GRNBoost2 or scKAN? Both give directed or signed edges and may track nonlinear regulation better. We include GRNBoost2 as a network builder under identical downstream analysis. On BEELINE we also compare against the relative gains scKAN reports, 5.40% to 28.37% in AUROC over signed competitors (scKAN); no absolute AUROC for scKAN is reported on Lacuna. Our claim is not that lagged partial correlation infers edges best. It is that controllability roles computed from any reasonable time-directed network carry heritability, and the builder comparison shows how much the choice of builder matters.
Proposed Experiments
The first experiment is a fidelity check with known answers. On BEELINE synthetic networks (linear, bifurcating, and cyclic, simulated with BoolODE), the true directed graph is known, so the true always-driver set is known too. We measure driver-set Jaccard overlap and root-component recovery for lagged partial correlation, lagged Pearson, GRNBoost2, and contemporaneous partial correlation with random orientation, in addition to AUROC and AUPRC. Following Validation of gene regulatory network inference based on controllability, we expect edge accuracy and driver accuracy to diverge. We select the network builder by driver fidelity, not by AUROC.
The second experiment is the cross-sectional comparison against cPE on GSE75010 at matched top-k, which tests whether conditioning alone improves prioritization.
The third experiment is the decisive one: the heritability test in T cell activation with immune-disease GWAS. The result is positive if the always-driver annotation shows a significant conditional coefficient after degree, core score, and temporal differential expression are in the model; if that coefficient exceeds the one from the static-network version of the same annotation; if it disappears under time permutation; and if it is absent for the negative-control traits. It is negative if the driver coefficient is null conditional on degree and hubs. That outcome would confirm the hub-anchoring view for this context, and we would test it again in the iPSC-to-cardiomyocyte replication before generalizing. A split result is also informative: enrichment in drivers for immune traits but in hubs for cardiac traits would suggest that transition-driving genetics are specific to rapid state changes such as activation, rather than a general property of heritability.
References
- The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution
- Controllability of structural brain networks
- Emergence of bimodality in controlling complex networks
- A structured systems approach for optimal actuator-sensor placement in linear time-invariant systems
- Identifying preeclampsia-associated genes using a control theory method
- Validation of gene regulatory network inference based on controllability
- DRAGON: Determining Regulatory Associations using Graphical models on multi-Omic Networks
- Noise cleaning the precision matrix of short time series
- Minimum input selection for structural controllability
- geneRNIB: a living benchmark for gene regulatory network inference
- Kolmogorov-Arnold Network for Gene Regulatory Network Inference
- Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes
- Predicting genotype-specific gene regulatory networks