Drug response as a trajectory in perfused organoid arrays

Every preclinical drug screen makes the same two compromises and rarely admits either. First, cells are grown as flat monolayers on rigid plastic, which strips away the polarity, stiffness, and diffusion barriers that govern how real tissues respond to therapy — this is the well-documented "dimensionality gap" (The Dimensionality Gap in In Vitro Disease Modeling). Second, even when researchers move to 3D organoids, they read out drug response as a single number at a single time: a viability assay or an IC50 at 72 hours. Clinical medicine learned the hard way that this snapshot habit misleads. In patients, the trajectory of a biomarker — its velocity, its inflection points, whether it recovers after treatment — predicts outcomes that admission baselines miss entirely: serum albumin trajectories carry an odds ratio of 2.12 for mortality in heart attack patients where static baselines do not (Lei Zhang et al.), and patients whose inflammation is falling still carry a 1.408-fold hazard of metabolic syndrome, a "legacy effect" invisible at any single timepoint (Jianwei Pan et al.).

The core claim of this proposal is that these two failures are one failure. The in vitro snapshot hides exactly the dynamics — drug penetration kinetics, slow-cycling escape states, post-treatment recovery slopes — that determine whether a drug works in vivo, and the field's trajectory-modeling machinery has never been pointed at a living 3D micro-tissue under therapy. I propose to build and validate a perfused organoid array that yields serial, non-destructive biomarker measurements from individual micro-tissues, and to test whether trajectory features (kill velocity, recovery slope, trajectory class) predict in vivo patient-derived xenograft outcomes better than the static endpoint assays that currently gate drug decisions.

Core Claim

Static endpoint assays on 3D cultures systematically misclassify drug efficacy because the decisive biology — transport-limited penetration, mechanotransductive stress resistance, and slow-cycling protected states — expresses itself in the time course of biomarker change, not in its final level. By sampling secreted biomarkers from the perfusate of individual organoids in a microfluidic array over days, and analyzing those series with attrition-aware trajectory statistics, one can (a) detect drug-evasion mechanisms that terminal viability assays miss, (b) call in vivo-matching drug rankings days earlier than endpoint assays, and (c) transfer the clinical insight about recovery slopes and legacy effects into a preclinical setting for the first time. This is not a call for better endpoints; it is a call to change what an assay is.

Research Program

The platform builds on the digital organ-on-a-chip design of Guohua Wu et al., which traps 127 individual 200 µm hydrogel microspheres in microwells and achieves steady-state diffusion 16.8 times faster than bulk-gel chips (with capture efficiency improved from 8.9% to 75.4% by Pluronic F-127 treatment). I extend it in one specific way: the perfusate flowing over each microsphere is fractionated on a schedule (e.g., every 6–12 hours) and assayed for a small panel of secreted biomarkers (cytokines, damage-associated proteins, drug payload itself), so that each individual micro-tissue yields a genuine longitudinal series without ever being fixed or destroyed. Morphological readout runs in parallel via label-free bright-field imaging of nuclear and structure dynamics, following the precedent of Bright2Nuc, which reconstructed virtual nuclear stains from bright-field z-slices with a Pearson correlation of 0.79 and tracked 250,000 nuclei over 72 hours in a 128-chamber chip (Label-free imaging of 3D pluripotent stem cell differentiation dynamics on chip) — though that system tracked only early, unperturbed differentiation and suffered signal-to-noise degradation in live cells, which my drug-perturbed, disease-organoid setting must address head-on (see Decisive Checks).

The analysis layer imports three tools from the clinical trajectory literature, none of which has ever been calibrated against in vitro micro-tissue data. First, latent class trajectory modeling of the kind that separated hsCRP series into "Low-stable", "Increasing", and "Decreasing" groups (Jianwei Pan et al.) will classify organoids by response shape rather than final level. Second, because drug-treated organoids disintegrate or are culled, the attrition statistics developed for mouse clinical trials apply directly: when 88% of dropouts are tumor-burden sacrifices (missing at random), linear mixed-effects models estimate the true growth slope of 0.24 log(mm³)/week while joint shared-random-effects models bias it to −0.02 (On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of PDX Models) — I will use the robust mixed-model formulation, not the joint one, and verify this choice by simulation on my own dropout data. Third, for sparse or missed samples, soft alignment against population-level trajectories reduces prediction error where naive linear imputation inflates it, as shown by a reduction in MMSE prediction RMSE from 0.4188 to 0.2928 on 456 ADNI subjects (Lyujian Lu et al.); the same trick should let a partially sampled organoid borrow strength from the cohort.

The scientific target is the specific class of efficacy failures that endpoints cannot see. The clearest documented case is the Trellis screen, which used mass cytometry on more than 25 million single cells from over 2,500 colorectal cancer organoid–fibroblast co-cultures and found that fibroblasts protect chemosensitive organoids from SN-38 by shifting cancer cells into a slow-cycling state — on-target DNA damage still occurs, so short-term assays look fine, and only a YAP inhibitor (Verteporfin) restores sensitivity (Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses). That screen was destructive and endpoint-only; a protected slow-cycling state is precisely a trajectory phenomenon — decelerating kill velocity, flat recovery slope — and my hypothesis is that perfusate trajectories will flag it without a single cell being fixed.

Proposed Experiments

Experiment 1 (mechanism detection) recreates the colorectal PDO–CAF system on the chip: organoids grown with and without cancer-associated fibroblasts, treated with SN-38, Verteporfin, and the combination. The ground truth is already established by the Trellis study (CAF protection exists; Verteporfin reverses it). The decisive question is whether trajectory features — kill velocity, time-to-plateau, post-treatment recovery slope of secreted damage markers — separate CAF-positive from CAF-negative conditions at a timepoint where a matched terminal viability assay (run on replicate microspheres at the same day) shows no significant difference. Success is detecting the protected state at least 24 hours before any endpoint readout diverges, with the trajectory classifier assigning protected organoids to a "slow-decline" latent class distinct from the "rapid-decline" class of unprotected ones.

Experiment 2 (penetration kinetics versus IC50) uses the gastric cancer organoid setting of Jiale Zou et al., where albumin-bound paclitaxel becomes sequestered in the central organoid lumen while liposomal paclitaxel penetrates the dense outer layers — a 2.3- to 10.8-fold difference in IC50 that correctly forecast a 42-day PDX survival split of 16.7% versus 100%. Their IC50 was a static CellTiter-Glo readout at a single arbitrary timepoint. I will measure the time-resolved perfusate concentration of each formulation and the lag between bath concentration and secreted-damage response, testing whether the trajectory formulation rank-matches the PDX survival split earlier (in days-to-call) and at lower drug exposure than the endpoint assay run in parallel. The secondary outcome is whether trajectory features predict the divergence even in the regime where the endpoint IC50 gap is ambiguous, by deliberately testing dose windows near the overlap of the two formulations' confidence intervals.

Experiment 3 (translational concordance and attrition) connects the chip to the mouse. Using the mixed-effects framework validated on the 225-mouse, 25-model PDX breast cancer cohort (On the Choice of Longitudinal Models...), I will compare chip-derived efficacy slopes against PDX growth slopes for matched drug–organoid pairs, reporting Spearman rank concordance between in vitro trajectory efficacy and in vivo mixed-model slope, with the endpoint IC50 as the competing predictor on identical cultures. A pre-registered simulation on the chip data will also induce artificial MAR dropout at rates up to the 88% observed in the mouse trials and confirm that mixed-model trajectory estimates remain unbiased where joint-model estimates collapse, reproducing the 0.24 versus −0.02 log(mm³)/week failure mode in silico before it can corrupt real data.

Baselines

The "do nothing" reference point is current practice: a static CellTiter-Glo 3D viability endpoint at 72 hours on matched organoids, exactly as run in the gastric PDO study, where this assay produced the 2.3- to 10.8-fold IC50 separation between Albu-PTX and Lipo-PTX (Jiale Zou et al.). It is the strongest competitor for the practical question because it is cheap, standardized, and already correctly ranked that formulation pair; my method must beat it on time-to-call and on the CAF-protection case where it is structurally blind. It will be run on replicate microspheres from the same chip, same passage, same drug lots, so the only difference is what is measured and when.

The strongest existing method for mechanistic drug-response profiling in organoids is the TOBis/Trellis mass-cytometry endpoint screen (Trellis Single-Cell Screening...), which resolved the CAF-protection mechanism at single-cell resolution across more than 25 million cells; no reported result found on Lacuna for a predictive accuracy metric on the Trellis screen, so the comparison will be on the pre-registered detection tasks of Experiment 1 (time-to-detection and classification of the protected state), with Trellis run as a terminal assay on replicate microspheres at matched timepoints and tuning/annotation budgets held equal. The strongest computational competitor is scPPDM, a latent diffusion model that predicts post-perturbation single-cell transcriptomes and outperformed prior models by +36.11% Spearman correlation on unseen drugs in the Tahoe-100M atlas of 50 cancer cell lines and 379 drugs (scPPDM: A Diffusion Model for Single-Cell Drug-Response Prediction); it will be given the same 2D-monolayer and terminal-3D expression data from my cultures as input, and scored on the same outcome (PDX concordance in Experiment 3), which is a task its 2D training never faced. Beating scPPDM means higher rank concordance with in vivo outcome despite its advantage in transcriptomic depth, because it lacks transport, mechanics, and time.

Two ablations isolate the proposal's key ingredient. The first ablation keeps the entire 3D perfused platform but discards all temporal information, using only final-timepoint perfusate and imaging features — if trajectory features do not outperform snapshot features from the identical hardware, the claim collapses to "buy a chip," which is not my claim. The second ablation keeps full trajectory modeling but runs it on 2D monolayer timecourses of the same cells in standard plates, isolating the contribution of dimensionality; the neurotoxicity study of Kuusisto et al. is a cautionary anchor here, since 2D models reached AUC near 1.0 (robust above 0.9 down to 10 genes) versus 0.7–0.8 for 3D models, yet the top predictive genes from 2D and 3D shared zero overlap — clean 2D signal can be artifactual signal, and this ablation tests whether 2D trajectories mislead in the same way. A simple piecewise-linear mixed model on the perfusate series (no latent classes, no soft alignment) is also included, since linear mixed models proved robust under MAR dropout in the mouse trials; if it matches the fuller trajectory analysis, the simpler model wins and I will say so.

Why not just…?

Why not just use 2D cell lines sampled at many timepoints? This is the most serious objection, and the evidence partly supports it: 2D neural progenitor models with 11 timepoints hit AUC near 1.0 while 3D models reached only 0.7–0.8 (Kuusisto et al.). But that same paper found zero overlap in the top predictive genes between 2D and 3D, meaning the clean 2D signal was measuring something other than physiological toxicity. More decisively for drug response, 2D cultures physically cannot exhibit the transport failure at issue: the lumen trap that sank Albu-PTX in organoids and mice (16.7% versus 100% 42-day survival) is invisible on plastic (Jiale Zou et al.), and matrix-stiffness control of radiation damage — 85% versus 12% γ-H2AX foci across stiffnesses — is absent on rigid polystyrene (Biomimetic model for glioblastoma radiosensitivity). My second ablation runs this alternative fairly; if 2D trajectories predict PDX outcomes as well as 3D trajectories, that result would itself be valuable and would overturn my claim.

Why not just measure IC50 at 72 hours like everyone else? Because the clinical literature has already demonstrated, in three independent settings, that snapshots misclassify in exactly the ways that matter here: static albumin baseline missed the S-shaped mortality relationship (OR 2.12 within the inflection window) that the trajectory captured (Lei Zhang et al.); falling-but-elevated hsCRP carried a persistent 1.408 hazard invisible at any single draw (Jianwei Pan et al.); and post-therapy lymphocyte recovery, not baseline counts, predicted survival after chemoradiotherapy (Zhen wei Sun et al.; Yifu Ma et al.). The in vitro analogue is the Trellis finding that fibroblast-protected cells absorb on-target damage without dying — a state that fakes efficacy in a 72-hour viability number. Experiment 1 is designed to settle this empirically: if the endpoint assay detects CAF protection as early and as reliably as the trajectory, the objection stands and the proposal fails.

Why not just use PDX mice, which have real 3D physiology and allow serial tumor measurement? Three concrete reasons from the file. Establishment is too slow for clinical decisions — organoids are ready for screening in roughly three weeks with 90% success, while PDX generation cannot support personalization timelines (Jiale Zou et al.). The in vivo readout is statistically fragile: 88% of mouse dropouts are tumor-burden sacrifices, and the wrong longitudinal model biases the efficacy slope from 0.24 to −0.02 log(mm³)/week (On the Choice of Longitudinal Models...). And mice present every microenvironmental variable at once, so stiffness, hypoxia, and stromal effects cannot be isolated (Biomimetic model for glioblastoma radiosensitivity). PDX cohorts remain the validation standard in Experiment 3 — the proposal needs them — but not the discovery engine.

Decisive Checks

Three results would decide the proposal's fate. First, the CAF-protection experiment: if trajectory features fail to separate protected from unprotected organoids before the endpoint assay does, the central mechanism claim is wrong. Second, the time-to-call analysis against the static IC50 baseline on the gastric PDO panel: the endpoint assay is genuinely strong there (it already produced the 2.3- to 10.8-fold IC50 split), so the honest bar is earlier calls at matched accuracy, not merely equal accuracy. Third, the 2D-trajectory ablation: if 2D monolayer timecourses predict PDX outcome as well as 3D perfused trajectories, then despite the lumen-trap and stiffness evidence, the expensive platform is unjustified and the field should invest in dense 2D timecourses instead — a finding I would report, not bury. A known risk throughout is live-cell signal degradation, which sank label-free tracking fidelity in the Bright2Nuc system (Pearson 0.79 only in fixed cells); the perfusate channel is my hedge, since secreted-biomarker sampling does not depend on optical quality inside thick gels, and the two channels will be reported separately so the reader can see which one carries the result.

References

  1. Association between albumin changes and prognosis in older patients with acute myocardial infarction
  2. Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study
  3. Development of digital organ-on-a-chip to assess hepatotoxicity and extracellular vesicle-based anti-liver cancer immunotherapy
  4. Label-free imaging of 3D pluripotent stem cell differentiation dynamics on chip
  5. On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models
  6. Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments
  7. Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses
  8. Construction of gastric cancer patient-derived organoids and their utilization in a comparative study of clinically used paclitaxel nanoformulations
  9. scPPDM: A Diffusion Model for Single-Cell Drug–Response Prediction
  10. Machine Learning to Predict Developmental Neurotoxicity with High-throughput Data from 2D Bio-engineered Tissues
  11. Biomimetic model for the identification of distinctive microenvironmental factors in glioblastoma radiosensitivity
  12. Effectiveness and mechanisms of lymphocytes at different time points in predicting consolidation immunotherapy following adaptive chemoradiotherapy in locally advanced non-small cell lung cancer
  13. Clinical Considerations of Splenic Dose Constraints to Mitigate Radiation‐Induced Lymphopenia