Live Redox Imaging of Cell Cycle Arrest in Patient Tumor Organoids

When a cancer drug freezes cells at a cell cycle checkpoint, the cells do not just stop — they remodel their metabolism. Some seem to hunker down into a protected, slow-cycling state that shields them from chemotherapy; others appear to sit dangerously close to a lethal oxidative threshold. Which of these is true, for which cells, and whether we can see the difference before a cell either dies or escapes the blockade, is currently unknowable with existing tools: the assays that measure cell cycle arrest require killing the cells, and the organoid models that capture real tumor environments rely on destructive, single-timepoint readouts. This proposal builds a live-imaging platform that watches the same individual tumor cell, inside a patient-derived organoid, as it enters, endures, and either escapes or dies from cell cycle arrest — while simultaneously recording its intracellular redox state, a real-time proxy for metabolic reprogramming.

The core mechanism combines two existing imaging technologies that have never been used together. FUCCIplex, a genetically encoded probe that lights up cells in the G1 phase of the cell cycle in cyan and cells in S/G2/M in near-infrared, deliberately leaves the green and red spectral channels free — but its creators only filled them with structural and calcium markers. Meanwhile, the Oba-Q family of redox sensor proteins reports the balance between oxidized and reduced molecules inside living cells through a quenching mechanism, with a calibrated midpoint redox potential suitable for the cytosol. Putting a redox sensor into the channels FUCCIplex freed, inside patient tumor organoids co-cultured with their native stromal cells, and then applying cell cycle checkpoint inhibitors, would for the first time record the metabolic consequences of arrest in the same cell whose cell cycle state and fate are known.

Core Claim

The claim is that intracellular redox dynamics, measured in real time in individual cells, carry predictive information about cell cycle arrest outcomes that cell cycle state alone does not. Two observations motivate this. First, checkpoint blockade does not arrest cells uniformly: a subpopulation undergoes a non-genetic "detour" and keeps replicating, marked by unexpected early accumulation of Cyclin B1 and CDK2 — but that work was done on fixed 2D cells and could not observe what preceded the escape. Second, stromal co-culture shifts tumor cells into a slow-cycling state that protects them from apoptosis, while reviews of cancer metabolism establish that metabolically stressed cancer cells sit just below a lethal reactive oxygen species (ROS) threshold. These two literatures point in opposite directions about whether arrest is metabolically protective or metabolically fragile, and no study in the supplied material has resolved the tension — because resolving it requires exactly the instrument that does not exist: simultaneous, longitudinal, single-cell measurement of cell cycle state and metabolic redox state in a three-dimensional patient tumor model.

The platform also tests a concrete mechanistic hypothesis: that the cells which escape CDK4/6 blockade are those whose redox state, measured hours before any cell cycle divergence is visible, sits in a specific range — and that local microenvironmental position within the organoid (deep, oxygen-starved core versus oxygenated periphery) sets that range. This would connect the observation that hypoxia gradients drive transcriptional state shifts in pancreatic tumor organoids to the observation that fractional drug resistance is non-genetic, unifying them as one spatially structured, metabolically gated process.

Research Program

Three components must be built. The first is a spectral engineering problem: the published cytosolic redox sensor Oba-Qc is derived from a cyan fluorescent protein, which collides with FUCCIplex's cyan G1 marker. We will re-derive a green-channel variant by transplanting the disulfide-bond redox-switching mechanism from Oba-Qc onto a green fluorescent protein scaffold, using the published construct as the template and validating against its calibrated behavior (its reduced state fluoresces and its oxidized state quenches; the parental CFP-based sensor reports cytosolic glutathione balance with a midpoint potential of −249 mV, and the Grx1-fused version reversibly responds to chemical oxidation and reduction in living HeLa cells). Because quenching-based sensors are not ratiometric and are sensitive to expression level and pH, every experiment includes per-cell sensor expression normalization and a pH-insensitive counterstain control, following the limitations the sensor's own authors document.

The second component is the organoid system. We will use patient-derived pancreatic ductal adenocarcinoma organoids co-cultured with patient-matched cancer-associated fibroblasts in a Matrigel/Collagen I matrix, following the published co-culture protocol that demonstrated stroma-mediated chemoresistance, and estrogen receptor-positive breast cancer organoids, connecting to the system in which fractional resistance was discovered. We adopt the TRACER rolled-scaffold geometry to impose controlled radial oxygen gradients, because that platform demonstrated that gradient position alone drives transcriptional state ($R^2 = 0.78$ between depth/hypoxia and the basal-like subtype), giving us a physical dial for the metabolic microenvironment. We inherit its known practical constraints rather than pretending to solve them: line establishment succeeds in roughly half of attempts and requires a median of 40 days to reach screening biomass, so the study is sized for eight to twelve established lines, not dozens.

The third component is the imaging pipeline. We reuse the CALIPERS smart-imaging logic — low-resolution 2D monitoring that triggers high-resolution 3D confocal stacks only at critical events, which reduced imaging frequency sixfold and preserved viability — extended with two changes: the trigger set includes not only mitotic events but redox-threshold crossings, and the analysis layer segments organoid position radially so that each tracked cell carries its depth coordinate. Cell tracking, lineage assignment, and redox trajectory extraction feed a prediction task: from the first 24 hours of imaging under palbociclib, predict for each single cell whether it will still be cycling, arrested, or dead at 72 hours. We will additionally train a generative trajectory model on the paired redox-plus-cell-cycle state space, adapting Meta Flow Matching's treatment of population dynamics as vector fields, so that short-horizon live measurements can be extrapolated to virtual long-horizon outcomes and compared against the truly observed fates.

Proposed Experiments

The perturbation matrix crosses three axes. The first is cell cycle blockade: palbociclib (CDK4/6, G1 arrest) and the CDC25B degrader WG-391D or fenofibrate (G2/M arrest, the latter having shown radiosensitizing SER of 1.42 in CNE-2 head and neck cells, establishing it as a metabolically active G2/M agent). The second is metabolic pressure: a pyruvate dehydrogenase kinase inhibitor to force pyruvate into mitochondrial oxidation rather than lactate, and a controlled ROS push, chosen because metabolically reprogrammed cancer cells are documented to sit below a lethal ROS threshold. The third is microenvironment: organoid radial position (gradient depth) and CAF presence versus monoculture.

The primary endpoint is a time-lagged prediction benchmark. For each tracked cell, features are (i) cell cycle state from FUCCIplex, (ii) the redox trajectory and its slope from the sensor, and (iii) radial position. The target is the observed 72-hour fate under continuous palbociclib. Prediction quality is measured by area under the receiver operating characteristic curve for escape versus arrest versus death, evaluated per patient line and pooled. The decisive mechanistic readout is the direction of the redox shift upon arrest in CAF-protected versus monoculture organoids: if arrest buffers metabolic stress, arrested cells should drift toward more reduced intracellular state and survive ROS challenge better than cycling neighbors; if arrest exacerbates vulnerability, arrested cells should sit measurably closer to the oxidative threshold and the combination of checkpoint blockade plus metabolic gatekeeper inhibition should kill them preferentially. We will test the therapeutic corollary directly: whether CDK4/6 blockade converts the organoid into a sensitized target for the PDHK inhibitor plus ROS push, measured as a shift in the dose-response curve for cell death — following the logic that Trellis validated, where the YAP inhibitor verteporfin restored SN-38 sensitivity in CAF-protected organoids without altering on-target damage.

A second experiment tests the spatial origin of fractional resistance. If escape from palbociclib is enriched at specific organoid depths and preceded by depth-dependent redox differences, that supports the microenvironmental-metabolic-gating hypothesis over a purely cell-intrinsic stochastic one. The comparison is within-organoid, using each organoid as its own control, which is exactly what the destructive single-timepoint assays could never do.

Baselines

The naive reference point is endpoint-only assessment: fixed-cell imaging of cell death and proliferation at 72 hours, in the style of the DeathPro assay used in the PDAC organoid-fibroblast co-culture work. It is the standard for organoid drug screening and defines the floor of usefulness — if live redox trajectories add nothing over a 72-hour viability readout, the platform's clinical rationale collapses. It will be run on the identical organoid batches and drug plates, scored with the same automated segmentation.

The strongest existing competitor is CALIPERS alone: FUCCIplex cell cycle imaging with smart 3D acquisition, without any redox sensor, under our identical perturbation matrix. This is the fair comparison because it is the best published live-imaging system for cell cycle dynamics in organoids, and its authors explicitly left the green and red channels open for functional sensors; running it here isolates exactly what the redox channel adds. The published CALIPERS result we must at minimum match is its sixfold reduction in imaging frequency relative to continuous automated 3D acquisition while preserving cell viability and event capture — our dual-sensor, redox-triggered variant should not cost more light or more acquisitions than that. On prediction, the claim is beaten only if redox-plus-cell-cycle features yield materially higher escape/death prediction performance than cell cycle features alone at matched tracking density, evaluated with the same cells, the same compute, and the same model class so the comparison isolates the sensor rather than the classifier.

The destructive-profiling baseline is the Trellis/TOBis mass cytometry design: endpoint CyTOF measurement of cell cycle markers, apoptosis, and signaling across the perturbation matrix. Trellis distinguished treatment effects across more than 2,500 organoid conditions and discovered the stromal protection phenotype, so it is the incumbent for phenotype discovery; its own paper reports no numerical accuracy metric on Lacuna for its treatment-discrimination benchmark, so the comparison here is qualitative — whether live trajectories reveal a phenomenon (early redox divergence preceding escape) that endpoint snapshots structurally cannot, namely the ordering of metabolic shift and cell cycle exit in the same cell.

Two ablations isolate the key ingredients. Removing the redox sensor while keeping everything else identical measures the marginal value of metabolic information; removing the CAF co-culture and gradient geometry while keeping both sensors measures the marginal value of the realistic microenvironment. On the analysis side, organoid segmentation and detection will be benchmarked against the MultiOrg dataset, where the strongest reported number on Lacuna is a Single Shot MultiBox Detector achieving 68.09% mean average precision at 0.5 intersection-over-union across 411 images and 63,042 annotated organoids; our pipeline must meet or exceed that on the MultiOrg test split before we trust its organoid-level counts, since inter-rater disagreement and clumping are documented failure modes. For the escape-rate ground truth, no reported numerical fractional resistance percentage on Lacuna exists for the palbociclib study — the 4i paper quantified 14 regulators but reports no escape fraction in its abstract — so we will report our own measured escape fractions per line as a standalone result.

Why not just…?

Why not just use 2D cell lines like T47D? The fractional resistance phenomenon was discovered in T47D cells, so this is a fair first question. But the escape-driving question is upstream of the cell: 2D monocultures expose every cell to identical oxygen and nutrients, while the TRACER work showed radial hypoxia alone drives pancreatic tumor organoids toward the aggressive basal-like state with an $R^2$ of 0.78, and stromal co-culture boosts proliferation 1.3-fold and induces chemoresistance through the CD44-HGF axis. If the hypothesis is that local metabolic microenvironment gates escape, a system with no local microenvironment cannot test it — the 2D experiment would be the ablation, not the main study, and we run it as exactly that.

Why not just do endpoint single-cell profiling (CyTOF, scRNA-seq, 4i immunofluorescence)? These are genuinely strong tools, and Trellis proved they can discover stromal protection at massive scale. The honest answer is that they answer a different question. Reactive intermediates have nanosecond-to-microsecond lifetimes and dissipate before any fixation captures them; more fundamentally, destructive assays cannot observe whether the redox shift came before or after cell cycle exit in a given cell, which is the entire causal question. However, we concede the reviewer's underlying point partially: endpoint profiling has higher per-sample multiplexing than our two optical channels. So we use endpoint scRNA-seq and CyTOF on matched organoid aliquots at 24 and 72 hours as orthogonal validation, and if the live trajectories contradict the endpoint profiles on a line, that contradiction itself is a finding worth reporting, not a reason to bury one method.

Why not just run bulk metabolic assays (Seahorse flux, mass spectrometry metabolomics, FDG-PET)? These are the field's gold standards for measuring reprogramming, and capillary electrophoresis-mass spectrometry has linked FDG uptake to hexosamine pathway activation in lung tumors. But the phenomena we target are subpopulation phenomena: fractional escape and stromally protected slow-cycling cells. Bulk assays average over millions of cells and would dilute a 20% escape subpopulation into noise, and they homogenize the hypoxic core against the proliferative periphery of an organoid, destroying the spatial variable our hypothesis depends on. Bulk metabolomics on whole organoids will still be collected as a sanity check that the optical redox readout tracks global metabolic state — if the sensor and the mass spectrometer disagree, the sensor, not the metabolome, is the suspect.

Why not just run the in-silico generative models (CellFlow, scDEBART, Meta Flow Matching) on existing datasets? This is the most serious objection, because Meta Flow Matching was validated on a 25-million-cell, 10-patient organoid screen and CellFlow cut additive-model error by 66% on drug combinations. But these models are trained on static snapshots and must infer dynamics via optimal transport pairing precisely because single-cell sequencing destroys the measured cell; their predictions of metabolic trajectory during arrest would be extrapolation into a state space they have never seen measured live. The right relationship is symbiotic rather than competitive: our live paired redox-and-cycle trajectories are the ground-truth kinetic dataset that can validate or falsify such models' vector fields, and we will use Meta Flow Matching's released code as the trajectory-modeling layer of the analysis pipeline. If a purely computational model trained on existing static data predicts our observed escape fates as well as the live redox trajectories do, that would be a decisive negative result for this proposal — and we state it in advance as a falsifier.

Decisive Checks

The proposal stands or falls on three outcomes. First, the engineered green redox sensor must reproduce the published Oba-Qc behavior — reversible quenching under chemical oxidation and reduction, with expression-normalized readout — in 2D culture before any organoid work. Second, the time-lagged prediction must show that redox trajectory features improve escape/death prediction beyond the CALIPERS-alone cell cycle baseline on matched cells; if they do not, the sensor's clinical value is unproven and we report the negative result. Third, and most consequential, the arrest-versus-metabolism question must resolve in one direction or the other: CDK4/6-arrested, CAF-protected organoid cells either buffer redox and resist the ROS-plus-PDHK-inhibitor combination, or they sit nearer the oxidative threshold and are preferentially killed by it. Either answer replaces a standing contradiction in the literature with a measured fact, and the combination-therapy direction it selects is directly testable in the same platform.