Delivery Timing as a Hidden Variable in Organoid Drug Response

Every drug screen in a dish shares an unstated assumption: that a drug's effect on a patient's tumor depends only on its identity and concentration, not on the shape of the concentration curve over time. In real patients, timing matters a great deal — a drug dumped into the bloodstream within seconds behaves very differently from the same total dose released slowly over days. Yet the two literatures that know this best never meet. Controlled-release engineering — designing carriers that let a payload escape over hours, days, or weeks — is validated almost exclusively in flat 2D cell sheets or mouse tumors. Meanwhile, patient-derived organoids — miniature 3D tumors grown from a patient's own cells, complete with supporting fibroblasts — are the most realistic human drug-screening platform we have, but every published organoid screen delivers drugs as a single dissolving bolus at constant concentration.

This proposal asks a question no study in the corpus has asked: when the same total dose of a chemotherapy is presented to patient-derived tumor–fibroblast co-cultures as a sharp bolus, as spaced pulses, or as a sustained slow release, does the drug's effective mechanism — which cells die, which molecular circuits are engaged, whether stroma-induced resistance emerges at all — actually change? Our core claim is that stroma-mediated chemoresistance observed in organoids is partly an artifact of bolus dosing, and that delivery timing is a controllable experimental variable that current mechanism-of-action methods ignore. We test this experimentally with matched-kinetics drug exposure and time-resolved gene and protein profiling, and computationally by extending an existing generative model of single-cell drug response to accept the drug's concentration trajectory as an input, so it can predict how a patient's cells respond to any release curve, not just to a fixed dose.

Core Claim

The concentration–time profile $c(t)$ a drug presents to tissue is a first-class determinant of its mechanism of action in patient-derived tumors, on par with dose and target identity. The stromal protection that fibroblasts confer on tumor cells in organoid co-cultures — documented as a shift into slow-cycling, epithelial-to-mesenchymal states — may be entered and exited in response to the sharp concentration transient of a bolus. If sustained, absorption-matched presentation prevents that state transition while preserving on-target damage, then release kinetics are not a delivery detail but a mechanism variable, and organoid screens that only use boluses are systematically mis-measuring drug action.

The motivation comes from three results that sit in tension. A finite-volume transport model validated against porcine coronary arteries showed that bolus release over seconds yields transient, washed-out tissue deposition, while hour-scale release synchronized to tissue absorption maximizes uptake, with a root-mean-squared error below 0.1 against in vivo measurements Intravascular drug release kinetics dictate arterial drug deposition, retention, and distribution; that paper is about arteries, but the principle — matching delivery rate to absorption capacity — should apply to a dense organoid matrix too, and no one has tested it there. On the organoid side, two independent studies showed that cancer-associated fibroblasts shield tumor cells from SN-38, gemcitabine, 5-FU, and paclitaxel by pushing them into protective states: a YAP-dependent slow-cycling "revSC-like" state in colorectal organoids, where the YAP inhibitor verteporfin fully restored sensitivity to SN-38 without weakening on-target DNA damage Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses, and a CD44–HGF-driven epithelial-to-mesenchymal transition in matched pancreatic organoid–fibroblast co-cultures Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system. Both used free-drug boluses. On the delivery side, sustained-release carriers exist and work — an amine-functionalized mesoporous silica nanocarrier with 17 nm pores carried 286.31 ± 8.14 µg of aflibercept per mg (5.28 times more than a small-pore variant) and sustained biologically active release for over 10 days Surface Modification and Pore Size Regulation of MSN as Function Aflibercept Carrier for Anti-Vascular Migration, and an iron-based metal-organic framework sustained pazopanib release and reprogrammed macrophage signaling in breast cancer models Determining M2 macrophages content for the anti-tumor effects of metal-organic framework-encapsulated pazopanib nanoparticles in breast cancer — but neither ever touched a patient-derived 3D culture. The intersection is empty, and that is the gap this proposal fills.

Research Program

The program has three connected parts: a matched-kinetics exposure experiment in patient-derived organoid–fibroblast co-cultures, a time-resolved molecular readout that separates on-target damage from downstream state changes, and a generative model conditioned on the release curve.

The experimental platform builds directly on two published systems rather than inventing a new one. We use patient-derived colorectal cancer organoids co-cultured with matched cancer-associated fibroblasts, the system profiled at single-cell resolution in the Trellis study (over 2,500 co-culture conditions across 10 patients, more than 25 million cells measured by mass cytometry) Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses, and, in a second arm, patient-matched pancreatic organoid–fibroblast pairs in Matrigel/Collagen I Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system. We are explicit about that system's cost: it needed a median of 40 days to expand and succeeded only 50% of the time, which is why the colorectal arm — with its larger established dataset — carries the computational work. Each drug (SN-38 and gemcitabine, chosen because both studies show stromal protection against them) is presented in three arms matched for total exposure, i.e., the same area under the concentration–time curve: a single bolus at the standard screened concentration, a series of spaced small pulses achieved by media exchange, and sustained release from a carrier. For the sustained arm we use mesoporous silica carriers of the type characterized in the aflibercept study, tuned by pore size and surface chemistry Surface Modification and Pore Size Regulation of MSN as Function Aflibercept Carrier for Anti-Vascular Migration, plus empty-carrier and free-drug controls, and we measure the actual $c(t)$ reaching the culture with a fluorescently tagged drug rather than assuming the designed profile. Alongside this, we adapt the finite-volume transport modeling approach validated in the arterial study Intravascular drug release kinetics dictate arterial drug deposition, retention, and distribution to a radially symmetric organoid-in-matrix geometry, so that carrier release rate, matrix diffusion, and cellular uptake can be co-designed rather than guessed; the arterial paper's own limitation — steady laminar flow assumptions — is less severe in a static gel than in a pulsing artery, which makes this a favorable setting for the model.

The measurement strategy addresses a real methodological trap: standard gene and protein assays require grinding up the culture, so they see only one time point per sample, while sustained release unfolds over ten or more days. We solve this with a two-track design. Track one is non-destructive: continuous bright-field time-lapse video of the same living organoids across the whole exposure, scored by the published deep-learning pipeline that predicts endpoint metabolic viability from video with a Pearson correlation of 0.9214 and a mean absolute percentage error of 0.1755 against ATP ground truth — halving the error of the previous best method, which stood at 0.4375 — and which can forecast outcomes from as little as 7.5 hours of pre- and early-treatment footage Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy. Track two is a sacrificial sampling ladder: parallel replicate wells harvested at staggered time points across the release window for mass cytometry (post-translational modification markers, DNA damage, cell cycle, apoptosis — the same panel as the Trellis TOBis assay) and single-cell RNA sequencing. Where automated detection of organoids in crowded bright-field images is needed, we use the MultiOrg benchmark to pick and calibrate the detector, whose best reported model, a Single Shot MultiBox Detector, reached a mean average precision of 68.09% at 0.5 intersection-over-union across 63,042 annotated lung organoids MultiOrg: A Multi-rater Organoid-detection Dataset.

The mechanism-of-action analysis follows the layered logic that the corpus's best mechanistic papers already use: measure the drug's on-target lesion (for SN-38, DNA damage markers), then the intermediate signaling state (YAP activity, epithelial-to-mesenchymal markers such as CD44 and HGF response genes), then the endpoint (apoptosis), and use a chemical rescue — verteporfin, following the Trellis study — as a positive control that we can reproduce. This mirrors how nuclear-versus-cytoplasmic fractionation plus pathway rescue pinned cisplatin's enhancement to nuclear β-catenin depletion Magnesium-Assisted Cisplatin Inhibits Bladder Cancer Cell Survival by Modulating Wnt/β-Catenin Signaling Pathway and how phosphorylation readouts paired with cytokine transcripts mapped dasatinib's anti-inflammatory action Dasatinib regulates LPS-induced microglial and astrocytic neuroinflammatory responses by inhibiting AKT/STAT3 signaling. The new question is whether the layer at which kinetics acts can be located: does sustained presentation preserve per-cell DNA damage while blocking the YAP-state transition, or does it change both?

The computational component extends Meta Flow Matching, a published generative model that takes an initial single-cell population distribution and predicts the post-treatment distribution for unseen patients by learning a vector field over distributions Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold, with public code available Meta Flow Matching. As published, it conditions on a static drug identity and dose. We condition it instead on the discretized concentration trajectory up to the prediction time:

$$ v_t\big(\cdot \mid \phi(p_0),\, c_{0:t}\big), $$

where $p_0$ is the initial single-cell distribution, $\phi(\cdot)$ is the graph-network embedding of that population from the original model, and $c_{0:t}$ is the drug concentration curve sampled at fixed intervals from treatment start to time $t$. The model is trained on the existing Trellis bolus data (which provides static curves — a special case) plus our new matched-kinetics dataset, and evaluated on held-out patients. Trained this way, the model can interpolate to release curves we never ran and, in a second phase, optimize over $c(t)$ to find the presentation profile that maximizes tumor-cell death per unit on-target damage in fibroblast-protected cultures — turning release kinetics from a delivery engineering problem into a searchable variable. We considered the UNAGI approach, which reconstructs temporal gene regulatory networks and perturbs them in silico against a reference compound database to predict therapies, validated when it predicted nifedipine's anti-fibrotic effect in human lung slices A deep generative model for deciphering cellular dynamics and in silico drug discovery in complex diseases; we adopt its time-resolved network view for interpretation but not its method, because it depends on static reference signatures from immortalized cell lines, which the corpus itself flags as blind to patient-specific microenvironment — exactly the context our question lives in.

Baselines

The naive reference point is the status quo of the entire organoid-screening literature: free drug dissolved in media as a single bolus at a fixed concentration over a 72-hour window. This is precisely the design of the ovarian cancer organoid screen that linked homologous recombination deficiency to platinum sensitivity Screening drug effects in patient‐derived cancer cells links organoid responses to genome alterations and of the 2,500-condition Trellis screen Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses. It is the strongest competitor precisely because it is the field's default: if bolus and sustained arms produce indistinguishable single-cell outcomes, the proposal's premise is wrong. We run it at the same total exposure as the other arms, in the same wells, plates, and imaging schedule, so any difference is attributable to timing alone.

The strongest existing experimental competitor is pharmacological rescue: co-treating bolus-exposed, fibroblast-protected organoids with the YAP inhibitor verteporfin, which the Trellis study showed completely restored SN-38-induced apoptosis in protected organoids without altering on-target DNA damage or cell-cycle arrest Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses. We reproduce this arm as the benchmark for "how much protection can be removed." Beating it means sustained presentation alone recovers a comparable fraction of sensitivity — without adding a second drug, and without the off-target concerns a second compound carries.

The strongest computational competitor is Meta Flow Matching as published: the same architecture and public code Meta Flow Matching, conditioned only on drug identity and static dose, which outperformed all baseline generative dynamics algorithms at predicting single-cell distributions of unseen patients on the Trellis data Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold. No reported result found on Lacuna for its numerical Wasserstein distances, so we re-run it ourselves. We give both models the same training data, the same held-out patients, and the same tuning budget. Beating it means measurably lower distributional error (Wasserstein distance between predicted and measured single-cell populations) on held-out patients in the mixed-kinetics evaluation set — conditions where the release curve varies — while matching it on the original static-dose conditions. The key ablation is the same kinetics-conditioned model with the curve input replaced by a constant, which isolates whether any gain comes from the timing signal rather than from extra parameters or fresh data. A secondary experimental benchmark is the video-viability pipeline itself: our re-implementation must reproduce its reported Pearson correlation of 0.9214 and mean absolute percentage error of 0.1755 on its own colorectal organoid video data before we trust its predictions on our new footage Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy.

Why not just…?

Why not just use 2D cancer cell lines? They are cheaper, faster, and the standard venue for exactly this kind of mechanistic work — the magnesium–cisplatin and dasatinib studies did all their expression analysis in flat cultures Magnesium-Assisted Cisplatin Inhibits Bladder Cancer Cell Survival by Modulating Wnt/β-Catenin Signaling Pathway; Dasatinib regulates LPS-induced microglial and astrocytic neuroinflammatory responses by inhibiting AKT/STAT3 signaling. But the corpus is unusually direct about why that fails here: the genetic predictor of platinum response was detectable only in 3D organoids and completely absent in 2D, and 2D monolayers systematically overestimated therapeutic index in the same patients' cells Screening drug effects in patient‐derived cancer cells links organoid responses to genome alterations; fibroblast-mediated protection is structurally invisible in monoculture Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses; and the oxygen-gradient-driven subtype switching that shapes drug sensitivity ($R^2 = 0.78$) simply does not exist in a uniformly oxygenated dish Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes. A question about how tissue absorption dynamics interact with release rate cannot be answered in a system with no tissue dynamics.

Why not just change the media on a schedule instead of engineering carriers? This is the most serious objection, and we do not argue around it — we made it an arm of the experiment. Repeated media exchange with small doses produces a pulsed concentration curve without any nanomaterial, and if the pulsed arm matches the sustained-release arm on every molecular readout, then the carrier adds nothing for mechanism discovery and the simpler protocol wins; that would itself be a useful, publishable finding. What the corpus does establish is that free drug in media is not neutral over long horizons: free aflibercept lost activity quickly and required encapsulation to keep working across 10 days Surface Modification and Pore Size Regulation of MSN as Function Aflibercept Carrier for Anti-Vascular Migration, and the validated transport model shows that how fast drug arrives governs how much the tissue retains Intravascular drug release kinetics dictate arterial drug deposition, retention, and distribution. Media exchange also cannot produce curves slower than the experimenter's schedule or release that is gated by local conditions inside the matrix, and our design separates these: the carrier arm tests genuinely sustained presentation, the pulse arm tests whether cheap approximation suffices, and the fluorescent-drug measurements of actual $c(t)$ make the comparison honest rather than rhetorical.

Why not just use a bulk viability endpoint instead of single-cell expression profiling? Bulk survival readouts would make this a cheap two-variable pharmacology paper. The corpus shows why they cannot answer a mechanism question: in fibroblast co-cultures, bulk viability cannot tell whether tumor cells or fibroblasts are dying, and it was single-cell protein and expression profiling that revealed the protective slow-cycling state and the fact that verteporfin restores sensitivity without weakening on-target DNA damage Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses. Bulk ATP assays also require total lysis, collapsing a ten-day release story into one snapshot Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy. We do include bulk ATP measurements — as the ground truth for the non-destructive video track, following that study's protocol — but the decisive comparisons (protection removed, damage preserved, state transition blocked) are all single-cell quantities by definition.

Decisive Checks

Three outcomes would settle the core claim. First, in the matched-exposure experiment, sustained presentation should reduce the fraction of fibroblast-protected tumor cells in the slow-cycling, YAP-active state while leaving per-cell DNA damage from SN-38 at least as high as under bolus — protection removed without a second drug. If instead all three timing arms produce the same single-cell distributions, the claim is falsified and bolus screening is vindicated for stroma-rich cultures. Second, the kinetics-conditioned model should beat its own static-dose ablation and the published model on held-out patients when release curves vary, and interpolation tests — predicting the pulsed arm from bolus and sustained data only — should land within pre-registered Wasserstein thresholds. Third, the model's optimized release curve, predicted computationally and then run experimentally, should outperform the best human-designed curve on the same total dose; that closes the loop and shows the search over $c(t)$ finds something a delivery engineer would not have guessed. As a final interpretive step, if timing effects are confirmed, we will ask whether they interact with the hypoxia gradients that drive basal-like, chemoresistant identity ($R^2 = 0.78$) Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes — whether sustained delivery penetrates and resensitizes the hypoxic core that a bolus cannot reach — which is the natural second paper this platform enables.