Cancer Keeper Genes and Tumor Network Control

• Cancer Keeper Genes (CKGs) are genes essential for tumor maintenance, acting as network control hubs independent of mutational status.
• CKGs underpin critical cellular processes such as cell-cycle progression, stress response, and metabolic homeostasis to sustain malignant states.
• Computational workflows using control theory and sensitivity analysis identify CKGs, offering new targets for therapies that disrupt tumor maintenance networks.

Cancer Keeper Genes (CKGs), also termed cancer-keeping genes in the original bladder-cancer study, are genes—mutated or not—whose activity is indispensable for the maintenance, survival, and homeostasis of established cancer cell states. The concept was introduced to extend the cancer driver gene (CDG) paradigm beyond initiating mutations toward a system-level account of how malignant states are sustained. In the formulation summarized by Zhang et al., CDGs provide the initiating “spark” of tumorigenesis, whereas CKGs constitute the “fuel supply” that keeps a cancer-specific attractor active through continual network support such as cell-cycle progression, stress buffering, and metabolic balance (Zhang et al., 22 Aug 2025).

1. Conceptual distinction from cancer driver genes

Cancer driver genes are defined by their capacity to initiate tumorigenesis via somatic driver mutations that confer a selective growth or survival advantage. Tumors harbor only a handful of such mutated drivers among thousands of passengers, and the CDG paradigm, reinforced by oncogene addiction, has guided targeted-therapy development. The CKG concept departs from this mutation-centric view by focusing on genes required for tumor maintenance irrespective of whether they initiated the malignancy or are themselves mutated (Zhang et al., 22 Aug 2025).

The distinction is not merely semantic. The driver-centric view is described as limited by the broad emergence of frequent therapeutic resistance, the presence of driver mutations in healthy tissues or individuals, and the lack of identifiable drivers in many tumors. The CKG paradigm incorporates non-oncogene addiction by emphasizing reliance on non-mutated pathways crucial for maintaining oncogenic cellular states. A common misconception is therefore that clinically relevant cancer dependencies must coincide with recurrently mutated genes; the CKG framework explicitly rejects that equivalence (Zhang et al., 22 Aug 2025).

Characteristic	CDG	CKG
Mutation status	mutated	mutated or not
Role	Initiation	Maintenance
Pathway scope	Single pathway	Network-wide
Drug resistance	Rapid emergence	Suppressed (theory)
Therapeutic examples	EGFR, BRAF TKIs	PARP inhibitors, HSP90 inhibitors

This comparison suggests that CKGs are intended to capture maintenance vulnerabilities that are convergent across heterogeneous tumor states, even when those vulnerabilities are not evident from mutation frequency alone.

2. Network-controllability formulation

The control-theoretic formalization models a cellular gene-regulatory network as a directed graph and, in linearized form, as

$\dot x = A\,x + B\,u,$

or, in the more detailed synopsis,

$$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$

Here, $x(t)\in\mathbb{R}^n$ is the vector of gene-expression states, $A\in\mathbb{R}^{n\times n}$ is the weighted, directed adjacency matrix of regulatory interactions, $u(t)\in\mathbb{R}^m$ is the vector of external controls, and $B$ specifies which nodes receive inputs (Zhang et al., 22 Aug 2025).

Within structural controllability, minimal driver-node identification reduces to a maximum-matching problem on a bipartite graph derived from the directed network. If $|M|$ is the size of a maximum matching, then the minimum number of control inputs is

$N_D = n - |M|.$

The unmatched nodes under a maximum matching constitute a driver-node set. Zhang et al. emphasize that such control schemes are generally not unique; indeed, the 2022 paper notes that half of the nodes may be driver genes for the cell, which makes standard driver-node analysis impractical as a basis for prioritizing therapeutically actionable dependencies (Zhang et al., 2022).

CKGs are defined as control hubs rather than ordinary driver nodes. Formally, a control hub is a node that lies in the middle of a control path in every possible control scheme, not as a head or tail but as an internal node that all admissible control-signal routes must traverse. The 2022 paper gives an equivalent graph-theoretic characterization: if $H$ is the union of all driver nodes over all schemes in $G$, and $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$0 is the union of all driver nodes over all schemes in the transpose graph $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$1, then the control-hub set is

$$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$2

The 2025 synopsis presents an operational criterion with the same purpose: a node $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$3 is a control hub if, for all driver-node sets derived from maximum matching, removing $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$4 increases $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$5 by at least one (Zhang et al., 22 Aug 2025).

A further refinement is the sensitive control hub, or sCKG. In the 2022 definition, a control hub $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$6 is sensitive if removal of at least one edge causes it to drop out of the control-hub set:

$$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$7

The 2025 review summarizes the same idea as a hub that loses its property under minimal network perturbation, such as deletion of a single edge, and introduces a sensitivity score $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$8 equal to the number of single-edge removals that flip node $$\dot x(t) = A\,x(t) + B\,u(t), \qquad y(t) = C\,x(t).$$9 out of the control-hub set (Zhang et al., 2022).

3. Computational identification workflows

The general workflow described by Zhang et al. begins with network construction from multi-omics data and literature-mined regulatory interactions, including transcription factor to target links and protein-protein regulatory links. The resulting graph is directed and weighted, and edges may be filtered by a confidence threshold, exemplified as interaction score $x(t)\in\mathbb{R}^n$0. Driver-node and control-hub identification then proceeds through maximum matching, for example via the Hopcroft–Karp algorithm on the bipartite transformation, followed by heuristic treatment of alternative matchings because full enumeration is #P-hard. In the 2025 synopsis, control hubs are detected by an efficient edge-removal scan with complexity $x(t)\in\mathbb{R}^n$1, where $x(t)\in\mathbb{R}^n$2 edges and $x(t)\in\mathbb{R}^n$3 is the time for one matching. Sensitivity analysis deletes single edges, recomputes controllability, and ranks candidate hubs by $x(t)\in\mathbb{R}^n$4 and by degree or betweenness. Robustness filtering retains only those control hubs found in at least $x(t)\in\mathbb{R}^n$5 of randomized network variations based on edge-weight permutations (Zhang et al., 22 Aug 2025).

The bladder-cancer implementation in the 2022 paper is more specific. It used 45 known bladder-cancer driver genes from TCGA together with the top 50 most frequently mutated BLCA genes as seed genes. Directed interactions were collected from five sources—NCI Nature Pathway Interaction Database, PhosphoSite, HumanCyc, Reactome, and PANTHER—using ten control-related relation types such as “controls–phosphorylation-of” and “controls–expression-of.” A breadth-first search from the 95 seed nodes yielded a bladder-cancer gene regulatory network, BLCA_GRN, with $x(t)\in\mathbb{R}^n$6 genes and $x(t)\in\mathbb{R}^n$7 directed edges (Zhang et al., 2022).

Control hubs in BLCA_GRN were identified in polynomial time by computing the union of all driver nodes in $x(t)\in\mathbb{R}^n$8, then the corresponding union in the transpose graph $x(t)\in\mathbb{R}^n$9, and finally taking the set difference $A\in\mathbb{R}^{n\times n}$0. The stated complexity for finding all driver-node unions in each graph is $A\in\mathbb{R}^{n\times n}$1, hence total $A\in\mathbb{R}^{n\times n}$2. Sensitive CKG detection by exhaustive single-edge removal has worst-case complexity $A\in\mathbb{R}^{n\times n}$3 (Zhang et al., 2022).

4. Bladder cancer as the reference case study

BLCA_GRN is the canonical empirical demonstration of the CKG framework. In that network, Zhang et al. identified 660 CKGs, corresponding to $A\in\mathbb{R}^{n\times n}$4 of all genes and only $A\in\mathbb{R}^{n\times n}$5 of all driver nodes. This reduction from the larger driver-node universe is central to the rationale for focusing on control hubs: they are intended to isolate indispensable internal nodes rather than any node that can serve as an input in one admissible control scheme (Zhang et al., 2022).

Among the 660 CKGs, 173 were classified as sensitive CKGs, or $A\in\mathbb{R}^{n\times n}$6 of the CKG set. Over $A\in\mathbb{R}^{n\times n}$7 of these 173 had more than one sensitive edge, indicating multiple single-link perturbations capable of disrupting their hub status. To prioritize potentially druggable targets, the 173 sCKGs were intersected with a protein-protein interaction network indispensable-gene set, yielding a core subset of 35 sCKGs. The named members reported in the paper include ACVR1B, ACVR2B, ATF4, BCR, CASP1, CDH2, CFL1, CRK, CRKL, DAG1, DIABLO, EDN1, EP300, ETS2, FN1, GNA12, HMOX2, IL2RG, MAP3K14, NME1, PDGFRA, PML, RPS6KA3, TRAF6, and YWHAZ, plus ten additional genes listed in Supplementary File 2 (Zhang et al., 2022).

Pathway-level enrichment is a prominent feature of the BLCA results. The 2022 study states that $A\in\mathbb{R}^{n\times n}$8 of cell-cycle and p53-signaling pathway genes in BLCA_GRN are CKGs, and that over $A\in\mathbb{R}^{n\times n}$9 of genes in the TGF$u(t)\in\mathbb{R}^m$0 and RTK-RAS pathways are CKGs. The 2025 review similarly describes a bladder-cancer regulatory network with 660 CKGs in red and a subnetwork of 35 sCKGs in green, clustered in the cell-cycle and p53 pathways, underscoring these pathways’ universal importance for malignancy maintenance (Zhang et al., 22 Aug 2025).

These observations support the intended distinction between mutation frequency and network indispensability. The 2022 study states that sCKGs often have moderate mutation frequencies, often below the detection thresholds of frequency-based methods, yet occupy critical network positions upstream of known drivers. This suggests that the CKG concept is designed to recover maintenance-critical loci that would be missed by mutation-centric discovery pipelines (Zhang et al., 2022).

5. Experimental validation and representative non-driver CKG classes

Experimental validation in the BLCA study focused on six sCKGs with no prior extensive characterization in bladder cancer: RPS6KA3, FGFR3, CDH2, EP300, CASP1, and FN1. Transient siRNA transfection used two siRNAs per gene at final $u(t)\in\mathbb{R}^m$1 via Lipofectamine RNAiMAX. Proliferation was assayed by CCK8 at 0, 24, 48, and 72 h with absorbance at 450 nm, and migration was measured by Transwell assay at 24 h with crystal-violet staining and cell counts in five random fields. The reported results were heterogeneous across genes: RPS6KA3 knockdown reduced 72 h proliferation to $u(t)\in\mathbb{R}^m$2 of control and migration to $u(t)\in\mathbb{R}^m$3; FGFR3 knockdown reduced proliferation to $u(t)\in\mathbb{R}^m$4 and migration to $u(t)\in\mathbb{R}^m$5; CDH2 knockdown reduced proliferation to $u(t)\in\mathbb{R}^m$6 and migration to $u(t)\in\mathbb{R}^m$7; EP300 or FN1 knockdown in UMUC3 increased proliferation by $u(t)\in\mathbb{R}^m$8–$u(t)\in\mathbb{R}^m$9; and CASP1 knockdown increased proliferation by $B$0 and migration by $B$1 (Zhang et al., 2022).

In vivo validation was reported for RPS6KA3. BALB/c nude mice, with $B$2 per group, were subcutaneously injected with UMUC3 cells stably expressing RPS6KA3-shRNA or control vector. After 4 weeks, excised tumor weights were $B$3 in controls and $B$4 in the RPS6KA3-knockdown group, with growth curves showing a $B$5 reduction in tumor volume by week 4 (Zhang et al., 2022).

The 2025 review presents the validation landscape more broadly. It states that experimental validations for bladder-cancer sCKGs include six targets—among them FGFR3 and EP300—for which CRISPR-knockout or small-molecule inhibition produced greater than $B$6 reduction in cell viability in vitro and approximately $B$7 tumor-volume reduction in orthotopic mouse models. It also identifies several representative non-driver CKG classes across tumor types: ATR and CHK1 in DNA damage response; HSF1 and HSP90 in proteostasis; NRF2 in redox metabolism; and PARP1 in BRCA1/2-deficient breast and ovarian cancers as a classic non-driver CKG example (Zhang et al., 22 Aug 2025).

These examples are used to illustrate a broader biological principle. CKGs map onto pathways of stress response, DNA damage repair, proteostasis, and metabolism, rather than only onto initiating oncogenic lesions. In the language of the 2025 paper, disabling such genes collapses the cancer attractor without necessarily reversing or removing the original initiating mutations (Zhang et al., 22 Aug 2025).

6. Therapeutic implications, limitations, and prospective extensions

The therapeutic rationale for targeting CKGs is presented as a consequence of network fragility. The 2025 review argues that striking central control hubs disrupts broad swaths of the tumor-maintenance circuitry, making compensatory rewiring more difficult than with single-pathway inhibitors. It further proposes pan-clone targeting, because CKGs may reflect convergent maintenance requirements across heterogeneous subclones, and resistance suppression, because loss of a CKG would force re-architecture of the entire attractor landscape, described as a high-barrier evolutionary step (Zhang et al., 22 Aug 2025).

Several treatment strategies are explicitly proposed. One is combination with driver inhibitors, such as adding an HSP90 or ATR inhibitor to front-line EGFR tyrosine kinase inhibitors in EGFR-positive lung cancer to forestall bypass-pathway activation. Another is maintenance therapy for minimal residual disease, for example administering PARP inhibitors in BRCA-mutants or a CDK4/6 inhibitor in HR-positive breast cancer after cytoreduction. A third is salvage therapy in refractory cancers, exemplified by CHEK1 inhibitors in heavily pretreated ovarian cancer; the review notes that the CHK1 inhibitor prexasertib shows single-agent activity in BRCA-wild-type, platinum-resistant patients (Zhang et al., 22 Aug 2025).

At the same time, the framework has explicit limitations. Current interactomes cover only a fraction of true regulatory edges, so missing or spurious links can misidentify CKGs. CKG profiles vary by tumor type and even by patient, limiting transferability and motivating patient-specific network reconstruction. Static network models may also fail to capture state-dependent rewiring during tumor evolution. These caveats are integral to the paradigm and not external criticisms: the 2025 review presents them as potential challenges for clinical translation (Zhang et al., 22 Aug 2025).

The same review points to future refinements through multi-omics, spatial profiling, and graph-based AI. A plausible implication is that the CKG framework is best understood not as a replacement for mutation-based oncology, but as an additional systems-level layer for identifying maintenance-critical dependencies that are invisible to driver-only analyses. Within that interpretation, CKGs represent a control-theoretic vocabulary for translating network indispensability into therapeutic target selection (Zhang et al., 22 Aug 2025).