When a drug halts cancer cells at a cell-cycle checkpoint, a puzzling thing happens: even in a genetically identical dish of cells given the same dose, a subpopulation keeps dividing. This "fractional resistance" is not genetic mutation — it is non-genetic plasticity, a detour through alternative cell-cycle regulators, demonstrated at single-cell resolution in ER+/HER2- breast cancer cells treated with the CDK4/6 inhibitor palbociclib (Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells). At the same time, we know from an older and under-appreciated result that whether a cell dies under stress depends overwhelmingly on its growth state: subconfluent, actively cycling endothelial cells are killed by amyloid-beta, while confluent, quiescent monolayers of the same cells are completely resistant and even suppress the toxin (Amyloid beta toxicity dependent upon endothelial cell state). Yet the standard tools for finding drug targets — network analyses that rank genes by how centrally connected they are in protein-interaction maps — compute those rankings on static, context-free interactomes, ignoring both the cell-cycle state and the inflammatory stress signaling that reshapes which genes actually matter in a given moment.
This proposal argues that the missing object is the state-conditional hub: a gene whose position as a network bottleneck exists only in a specific cell state, such as "arrested at G1 by CDK4/6 inhibition while NF-κB inflammatory signaling is active." The core claim is that ranking genes by centrality computed on state-specific interactomes, rather than on the usual aggregated maps, will (a) identify the compensatory survival loops that let arrested cells escape, (b) show that the NF-κB pathway — a master inflammatory switch — supplies those loops during checkpoint arrest, and (c) predict in vivo drug efficacy substantially better than existing in-vitro-to-in-vivo extrapolation models that treat every gene in a pathway as equal.
Background and Why the Intersection Is Open
Four literatures meet here, and each stops just short of the connection. The cell-cycle-arrest literature establishes that CDK4/6 inhibitors work clinically in hormone-receptor-positive breast cancer (A phase 1 study of dalpiciclib, a cyclin-dependent kinase 4/6 inhibitor in Chinese patients with advanced breast cancer; A Decade After Approval of the First CDK4/6 Inhibitor: A Look Back at Palbociclib's Journey from Discovery to Approval and What's Next in CDK Inhibition in Breast Cancer) but leaves the dormancy-versus-clearance paradox open: arrested cells that survive become reservoirs of recurrence. The hub-gene literature has powerful machinery — degree and betweenness centrality on protein-interaction networks, which respectively find genes with the most interaction partners and genes that sit on the most shortest paths between other genes — and even a modern reinforcement-learning system, BioMARL, that embeds centrality directly into its reward for selecting genes (Biological Pathway Guided Gene Selection Through Collaborative Reinforcement Learning). But its own practitioners flag the central limitation: the interactomes behind these rankings are static aggregations that fail to reflect how topology rewires across cell-cycle phases or under pharmacological stress. The NF-κB literature shows that inhibiting the TLR4/MyD88/NF-κB cascade — the two-step inflammatory signaling hierarchy running from Toll-like receptor 4, through the adaptor protein MyD88, to the transcription factor NF-κB — quenches destructive inflammatory amplification in heart, lung, spinal cord, and brain models (C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway; Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model; 3-Deoxysappanchalcone attenuates LPS-induced neuroinflammation in microglia cell culture and ameliorates cognitive impairment in traumatic brain injury), but these validations are almost all monoculture assays whose translation is questionable. Finally, the in-vitro-to-in-vivo extrapolation (IVIVE) literature — the study of why culture-dish successes fail in organisms — has produced the strongest existing bridge, Precily, a deep neural network that converts a tumor's gene expression into pathway-activity scores and pairs them with chemical fingerprints of drugs to predict response in xenografts and patients (Gene expression based inference of cancer drug sensitivity).
Precily is the closest prior work and defines what is missing. It averages expression across all members of a pathway, treating a housekeeping peripheral gene and a network bottleneck identically; it uses static snapshot transcriptomics, blind to the dynamic cell-cycle reprogramming that produces fractional resistance; and it models no immune or inflammatory signaling despite evidence that NF-κB-driven stromal alterations help decide in vivo drug failure. Zikry and colleagues supply the single-cell trajectory evidence but only for fourteen hand-picked proteins, with no network topology and no inflammatory readout (Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells). Balcells and colleagues supply the causal demonstration that cell state dictates vulnerability, but with no transcriptomics, no network analysis, and no dissection of which intracellular cascade governs the state dependence (Amyloid beta toxicity dependent upon endothelial cell state). No study in the file computes centrality on a cell-cycle-state-specific network, and none tests whether inflammatory signaling rewires the hub structure of an arrested cell.
Core Claim and Hypothesis
The proposal rests on one falsifiable hypothesis: during CDK4/6 arrest, the surviving fractionally resistant cells reroute their growth-control network through a small set of genes that become betweenness bottlenecks specifically in the arrested state — and the NF-κB pathway is a major supplier of those state-conditional hubs, providing survival signaling that the arrested cell comes to depend on. If true, three things follow. First, combining CDK4/6 inhibition with NF-κB pathway inhibition should collapse the fractionally resistant subpopulation in vitro. Second, centrality computed on state-specific interactomes should predict this dependency, whereas static centrality and unweighted pathway scores should not. Third, a drug-response model whose pathway features are hub-weighted — in place of Precily's unweighted averaging — should predict in vivo efficacy better, closing part of the IVIVE gap for checkpoint inhibitors.
The key computational object is a hub-weighted pathway score. For a pathway $p$ containing genes $g$ with expression $x_g$, Precily-style scoring weights all genes equally; we replace that with
$$S_p(s) = \frac{\sum_{g \in p} C_B^{(s)}(g)\, x_g}{\sum_{g \in p} C_B^{(s)}(g)},$$
where $C_B^{(s)}(g)$ is the betweenness centrality of gene $g$ in an interactome pruned to cell state $s$ — edges kept only where both partner genes are expressed in that state. The form comes from the standard betweenness definition (the fraction of all shortest paths in the network that pass through $g$); the only change from prior practice is that the network, and hence the ranking, is conditional on the measured state.
Research Program
The program has three linked components. The first is a measurement campaign producing state-resolved single-cell profiles of ER+/HER2- breast cancer cells (the T47D model used by Zikry et al.) across a palbociclib time course, crossed with NF-κB pathway perturbation. Multiplexed iterative immunofluorescence imaging — the 4i method, which stains many proteins sequentially in the same single cells — will measure the fourteen cell-cycle regulators from the fractional-resistance study (including CDK2, E2F1, Cyclin B1, Cdt1, and p27) plus NF-κB readouts: p65 phosphorylation and nuclear translocation by immunofluorescence, IκBα levels, and, following the assay logic of the Cordyceps study (Anti-Inflammatory Effects of Cordyceps Cs-HK1 Fungus Exopolysaccharide on Lipopolysaccharide-Stimulated Macrophages via the TLR4/MyD88/NF-κB Pathway), disruption of the TLR4–MyD88 physical interaction by co-immunoprecipitation. Bulk and single-cell RNA sequencing per state provide genome-wide expression for network pruning. Crucially, following Balcells et al., we include the cell-state control that most culture work omits: parallel arms with subconfluent cycling monolayers and confluent quiescent monolayers, so that any claimed state dependence is measured against the cycling-versus-quiescent baseline already known to flip drug vulnerability.
The second component builds the state-conditional interactomes. Starting from a reference protein-interaction database (the STRING/BioGRID type used throughout the hub-gene literature), we prune edges by state-specific expression and compute degree and betweenness centrality per state. The first empirical question is descriptive and sharp: do the cell-cycle detour proteins identified by Zikry et al. — CDK2, E2F1, Cyclin B1 — rise specifically in betweenness in the arrested, fractionally resistant state, and do NF-κB components (p65, MyD88, IKK-complex members) become hubs in the arrested state that were not hubs in the cycling state? If neither happens, the core claim fails early and cheaply.
The third component is predictive and comparative. We retrain a Precily-style deep neural network on the same CCLE and GDSC cell-line drug-screen data, holding the architecture, chemical fingerprints, and tuning budget fixed, and swap only the pathway features: unweighted scores (the Precily baseline), static-interactome-weighted scores, and state-conditional hub-weighted scores. Evaluation follows Precily's own ladder: held-out CTRPv2 screening data, mouse xenografts, and TCGA patient survival stratification. This isolates the single ingredient the proposal adds — state-conditional centrality — against the strongest existing system on identical footing.
Proposed Experiments
Experiment 1 (mechanism, in vitro): In T47D cells, palbociclib alone versus palbociclib plus an NF-κB pathway inhibitor, with the primary endpoint being the fraction of cells still cycling after arrest, measured by EdU incorporation and 4i trajectory inference (Slingshot/TrAGEDy, as in the fractional-resistance study) rather than a 48–72-hour viability endpoint — because static endpoints are exactly what the IVIVE literature identifies as blind to dynamic reprogramming. The hypothesis predicts the combination eliminates or sharply shrinks the cycling fraction that palbociclib alone leaves behind. The inhibitor choice matters for later translation, so we select a small molecule with demonstrated in vivo activity, such as aucubin, which suppressed TLR4/MyD88/NF-κB signaling and improved functional recovery in a spinal cord injury model (Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model), rather than a macromolecule like the Cordyceps exopolysaccharide EPS-LM, whose $6.58 \times 10^5$ Da molecular weight its own authors flag as an unaddressed gastrointestinal-absorption barrier to in vivo use (Anti-Inflammatory Effects of Cordyceps Cs-HK1 Fungus Exopolysaccharide on Lipopolysaccharide-Stimulated Macrophages via the TLR4/MyD88/NF-κB Pathway).
Experiment 2 (topology causality, in vitro): Perturb hub-ranked versus non-hub-ranked NF-κB components. If the state-conditional analysis ranks, say, MyD88 or p65 among the top arrested-state betweenness genes while some canonical-but-peripheral pathway member ranks low, then knockdown of the ranked hub should resensitize fractionally resistant cells to palbociclib and knockdown of the low-ranked member should not. This is the experiment that distinguishes the hub machinery from a generic "NF-κB is pro-survival" story: the ranking, not the pathway label, makes the prediction.
Experiment 3 (prediction, computational): The feature-swap comparison described above, with Precily's evaluation protocol. The decisive metric from the file is Precily's TCGA survival stratification: 72% five-year survival in predicted responders versus 28% in predicted non-responders (Gene expression based inference of cancer drug sensitivity). Beating the baseline means a larger responder-versus-non-responder survival gap on the same cohorts, and higher rank correlation with xenograft tumor-response measurements. For the BioMARL comparison on TCGA breast cancer gene selection, no reported result found on Lacuna for BioMARL's numeric AUC values; the file states only that it outperformed eight baseline feature-selection algorithms, so we will reimplement it under matched compute and tuning and report our own numbers rather than compare against unquoted ones.
Experiment 4 (IVIVE closure, in vivo): ER+/HER2- xenografts treated with palbociclib alone, the NF-κB inhibitor alone, and the combination, with tumor growth inhibition as the endpoint. Before running it, we register the model's prediction from Experiment 3 for each arm. Success is twofold: the combination outperforms palbociclib alone in vivo, and the state-conditional model predicted the magnitude and ordering of the effects better than the Precily-baseline and static-centrality variants. A failure mode worth pre-committing to: if the combination works in vivo but no model variant predicted it, the hub-weighting claim is unsupported even if the pharmacology succeeds.
Baselines
The naive reference point is a static-endpoint IC50 screen: palbociclib dose-response at 72 hours in T47D monoculture, the standard high-throughput readout that the IVIVE literature explicitly faults for blindness to dynamic reprogramming. Any state-conditional method must beat it where it is weakest — predicting the cycling fraction that persists after arrest and the in vivo outcome — not merely reproduce its dish potency. This baseline requires no citation beyond the practice the IVIVE literature critiques.
The strongest existing method is Precily, the deep neural network combining 1,329 pathway-activity scores from Gene Set Variation Analysis with 100-dimensional SMILESVec chemical fingerprints, trained on CCLE and GDSC and evaluated on CTRPv2, single-cell datasets, in-house mouse xenografts, and TCGA (Gene expression based inference of cancer drug sensitivity). It is the strongest competitor because it is the only published system that already makes the in-vitro-to-in-vivo jump for drug response with quantitative clinical results: 72% versus 28% five-year survival stratification in TCGA predicted responders versus non-responders, as reported in the file. The comparison will be run fairly by keeping everything identical — same training data, same chemical embeddings, same evaluation cohorts, same compute and hyperparameter-tuning budget via nested cross-validation — and varying only the pathway-score construction. Beating it means a larger survival separation between predicted responders and non-responders on the same TCGA cohorts and better rank correlation with xenograft response, with the margin assessed across multiple random seeds.
The key ablation isolates the proposal's central ingredient: the same network, the same weighting formula, but with centrality computed on the static aggregated interactome rather than state-pruned interactomes. A second ablation removes weighting entirely, collapsing back to Precily-style unweighted pathway scores. If static centrality performs as well as state-conditional centrality, the "state" in state-conditional hubs is doing no work and the proposal's core claim is refuted even if absolute performance improves over unweighted features.
Why not just…?
Why not just run the palbociclib-plus-NF-κB-inhibitor combination experiment? This is the most serious alternative, because Experiment 1 is cheap and the combination could work without any network analysis. We concede the point partially and deliberately front-load the combination as Experiment 1. But the combination alone cannot tell you which node to hit — the NF-κB cascade has many druggable components, systemic ablation carries the risk the literature flags of impairing host defense and triggering off-target apoptosis, and a blind screen over pathway components is exactly the low-yield approach hub ranking is meant to replace. Experiment 2 is the decisive test of the machinery's value: it makes a component-level prediction that only the state-conditional ranking, and not the generic "inhibit NF-κB" story, generates. If hub-ranked and non-hub-ranked knockdowns resensitize equally well, the computational layer adds nothing and we will say so.
Why not just use Precily as it stands, or add more pathway features to it? Precily already bridges the in-vitro-to-in-vivo gap, so a skeptic could argue the remaining gap is a data problem, not a representation problem. The evidence against comes from Precily's own structure: it weights every gene in a pathway equally, and the single-cell fractional-resistance work shows the deciding biology lives in a handful of detour proteins that unweighted averaging dilutes across hundreds of pathway members (Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells). Adding more pathways would not fix this; reweighting by state-conditional bottleneck position does. The feature-swap ablation settles it empirically: identical architecture and budget, varied only in weighting.
Why not just compute centrality on the standard static interactome, as existing hub-gene studies do? This is standard practice, and it is precisely the ablation we include. The hub-gene literature itself identifies static topology as its central limitation — interactomes like STRING aggregate potential binding events across all contexts and miss rewiring during cell-cycle phases or pharmacological stress. The Balcells result makes the stakes concrete: the same cells in two growth states show opposite vulnerability to the same insult (Amyloid beta toxicity dependent upon endothelial cell state), so a ranking that cannot distinguish the two states is ranking in the wrong space. If the static-centrality ablation matches state-conditional performance, we will report that the state-conditionality is unnecessary and the simpler method should be preferred.
Why not just extend Zikry's fourteen-protein panel with a few NF-κB markers? A panel extension would cheaply test whether NF-κB correlates with fractional resistance, and we adopt a version of it as our pilot measurement. But a hand-picked panel cannot discover unexpected state-conditional hubs outside the cell-cycle and NF-κB gene sets, cannot produce genome-wide rankings that transfer to the Precily-style prediction task across CCLE, GDSC, and TCGA cohorts, and cannot distinguish "NF-κB components are present" from "NF-κB components are topological bottlenecks in the arrested state" — which is the distinction Experiment 2 exploits.
Decisive Checks
Three findings would decide the proposal's fate. First, the descriptive topology result: state-conditional betweenness must rank the known detour proteins and NF-κB components as arrested-state hubs, which the existing Zikry data let us verify partially before any new experiment. Second, the causal result in Experiment 2: hub-ranked knockdowns must resensitize fractionally resistant cells while low-ranked knockdowns must not. Third, the predictive result in Experiment 3: hub-weighted features must beat both unweighted and static-weighted features under Precily's own evaluation, judged against its reported 72% versus 28% survival stratification. If the first two hold but the third fails, the biology stands while the translational claim must be narrowed — and we would report it that way, since a mechanism that explains fractional resistance but does not yet improve prediction is still a real, and previously unmeasured, object.
References
- Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells
- Amyloid beta toxicity dependent upon endothelial cell state
- A phase 1 study of dalpiciclib, a cyclin-dependent kinase 4/6 inhibitor in Chinese patients with advanced breast cancer
- A Decade After Approval of the First CDK4/6 Inhibitor: A Look Back at Palbociclib’s Journey from Discovery to Approval and What’s Next in CDK Inhibition in Breast Cancer
- Biological Pathway Guided Gene Selection Through Collaborative Reinforcement Learning
- C1q/TNF‑related protein 9 decreases cardiomyocyte hypoxia/reoxygenation‑induced inflammation by inhibiting the TLR4/MyD88/NF‑κB signaling pathway
- Aucubin promoted neuron functional recovery by suppressing inflammation and neuronal apoptosis in a spinal cord injury model
- 3-Deoxysappanchalcone attenuates LPS-induced neuroinflammation in microglia cell culture and ameliorates cognitive impairment in traumatic brain injury
- Gene expression based inference of cancer drug sensitivity
- Anti-Inflammatory Effects of Cordyceps Cs-HK1 Fungus Exopolysaccharide on Lipopolysaccharide-Stimulated Macrophages via the TLR4/MyD88/NF-κB Pathway