Spatial SAR: Does chemical structure predict drug penetration and signaling depth in patient tumor organoids?

Every drug candidate is optimized against a flat sheet of cells or a purified protein, yet real tumors are dense, three-dimensional tissues in which a molecule must physically travel through extracellular matrix and stromal cells before it can find its target. Structure-activity relationship (SAR) analysis — the practice of mapping which chemical sub-features of a molecule drive its biological effect — currently has almost nothing to say about this journey: no benchmark in the corpus measures how a structural modification changes a compound's ability to diffuse into, penetrate, and retain potency across the radial depth of a patient-derived organoid (PDO), a 3D culture grown from a patient's own tumor cells. At the same time, theranostic probes — molecules that carry both a therapeutic warhead and an imaging tag on the same chemical scaffold — are built precisely so that the imaging signal reports where the drug itself goes, but they are validated almost entirely in 2D monolayers or mouse xenografts (Dong et al.; Yang et al.). This proposal closes the loop: use a systematic analog grid of theranostic conjugates to measure, inside intact patient tumor organoids co-cultured with their matching stromal fibroblasts, how each structural modification reshapes drug penetration, target occupancy, and downstream signaling — and then learn a spatial SAR model that predicts these depth-resolved outcomes from the chemical graph alone.

The clinical stakes come from three well-documented failures. First, stroma protects tumors: cancer-associated fibroblasts (CAFs) co-cultured with patient organoids push tumor cells into slow-cycling, drug-resistant states through YAP and HGF–CD44 signaling (Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses; Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system), and 3D diffusion barriers alone induce aggressive, treatment-resistant transcriptional subtypes in pancreatic cancer organoids (Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes). Second, translation is leaky: only ~10% of differentially expressed genes overlapped between drug-treated colorectal organoids in vitro and the same organoids grown as xenografts in vivo (Drug repurposing screening and mechanism analysis based on human colorectal cancer organoids), and nobody knows whether that discordance is delivery failure or pathway rewiring. Third, computational SAR generalizes poorly: Therapeutics Data Commons showed drug-target interaction Pearson correlations collapse from ~0.70 on random splits to ~0.40 under realistic temporal splits (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development), plausibly because the labels themselves (flat-culture IC50) misrepresent what structure must accomplish in tissue.

Core Claim

The core claim is that a molecule's structural features — linker length and polarity, conjugation site, fluorophore or chelator bulk, ionizable groups — jointly determine a spatial activity profile consisting of three depth-resolved quantities measurable in the same intact organoid: local drug concentration (from the theranostic imaging signal), local target engagement (from single-cell phosphorylation readouts), and local cell-state response (from cell-cycle, apoptosis, and pathway markers). We claim that these quantities are learnable from the chemical graph, and that a model trained on spatial endpoints will rank analogs in stroma-rich organoids more faithfully than a model trained on conventional flat-culture potency — because the label now contains the physics and biology that actually separate winners from losers in tissue. Crucially, the theranostic design is what makes the measurement possible: because the imaging tag rides on the same scaffold as the drug, the fluorescence gradient is a direct readout of drug distribution, something a co-administered dye cannot provide since two distinct chemical entities diverge in clearance and diffusion (Xiao et al.).

Research Program

The program has four connected parts, each grounded in a method proven in the corpus.

The first part builds the analog grid. Starting from the Osimertinib–Cy7 design of Dong et al., who showed that conjugating a bulky Sulfo-Cy7 fluorophore to the covalent EGFR inhibitor Osimertinib preserved kinase inhibition and gave a 3- to 4-fold fluorescence selectivity for EGFR-mutant over wild-type cells, we synthesize a matched grid of ~24–32 analogs varying one feature at a time: linker length (short to long alkyl/PEG), conjugation site on the scaffold, fluorophore identity and bulk (Cy5, sulfo-Cy7, a chelator-bearing variant), and charge state. This is a controlled theranostic SAR study of exactly the kind Dong et al. called for but did not perform (they tested only parent, one conjugate, and free dye), and it mirrors the modular single-residue tuning logic that Yang et al. used on PSMA radioligands, where swapping one dansylated amino acid changed tumor accumulation from 17.44 %ID/g to 127.36 %ID/g.

The second part establishes the spatial assay. EGFR-mutant non-small cell lung cancer PDOs (driven by the same exon-19-del and L858R/T790M genotypes OSA-Cy7 targets) are grown with and without patient-matched CAFs in a defined 3D matrix, following the co-culture design of Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system. For each analog, we acquire time-lapse fluorescence plus bright-field imaging over 100 hours, quantify the radial drug gradient (surface-to-core fluorescence ratio, penetration half-depth), and — at selected time points — perform multiplexed single-cell readouts of p-EGFR, p-Akt, cell-cycle, apoptosis, and YAP/EMT markers in the manner of the Trellis mass-cytometry framework, which resolved >25 million single-cell signaling events in CRC PDO–CAF co-cultures and pinpointed YAP as the stromal resistance node reversed by Verteporfin (Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses). A key control runs throughout: LC-MS quantification of unmodified parent drug and each conjugate in matched wells, to verify that the fluorophore tag does not itself distort matrix diffusion and that the optical signal is a quantitative proxy for drug mass — the pharmacokinetic identity that defines real theranostics.

The third part addresses the in vitro/in vivo discordance directly. The ~10% gene-overlap finding of Drug repurposing screening and mechanism analysis based on human colorectal cancer organoids is unexplained because that study had no way to see where drug actually went. Our depth-resolved readouts split the explanation into a delivery component (drug fails to reach the core; low fluorescence half-depth with preserved per-cell target engagement where drug is present) versus a signaling component (drug reaches the core but the response pathway is rewired; normal penetration with blunted p-EGFR-to-apoptosis coupling, or YAP/revSC-like state shifts). This decomposition is the mechanism-mapping contribution: it converts a bulk transcriptomic mystery into two separately measurable, separately fixable failure modes.

The fourth part learns the spatial SAR. We train a chirality-aware graph model in the spirit of MolKGNN, whose interpretable convolution kernels recovered known pharmacophores like the trifluoromethyl moiety and achieved an average rank of 2.333 across nine PubChem benchmarks (Interpretable Chirality-Aware Graph Neural Network for Quantitative Structure Activity Relationship Modeling in Drug Discovery), with three outputs per molecule: predicted penetration half-depth, predicted target-occupancy-at-depth, and predicted 100-hour viability outcome. Matched-molecular-pair analysis in the style of the semi-automated QSAR-MMPA workflow, which expanded significant transformation rules from ~8,700 to over 91,000 with R² = 0.986 against experiment (Semi-automated workflow for molecular pair analysis and QSAR-assisted transformation space expansion), then extracts human-readable transformation rules: "lengthening the linker from X to Y improves core penetration by Z while costing affinity." Finally, we test whether the imaging stream improves early outcome prediction over the bright-field-only state of the art, building on the spatio-temporal multiple-instance-learning model of Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy, which predicted 100-hour ATP viability from unlabeled video with a Pearson correlation of 0.9214 and a mean absolute percentage error (MAPE) of 0.1755 using only the first 7.5 hours of footage.

Baselines

The naive reference point is unconjugated parent drug plus a separately administered fluorescent dye, evaluated in the same organoid wells with the same imaging schedule. This is the "do nothing" comparator because it is what most organoid drug screens effectively do: a viability-relevant drug and an independent, non-specific tracker. The expectation, grounded in the pharmacokinetic-identity argument of Xiao et al. — who showed that a co-administered agent's distribution says nothing about another molecule's accumulation — is that the dye gradient will diverge from LC-MS-measured drug distribution; the quantitative check is the correlation between fluorescence intensity and LC-MS drug mass per well, where the conjugate should hold a tight correlation and the co-administered pair should not. If the co-administered dye correlates as well as the conjugate, the theranostic premise of this proposal is wrong and we would report that.

The strongest existing methodological competitor is the combination of OSA-Cy7 itself as the probe and the bright-field video model as the outcome predictor. OSA-Cy7 is the closest prior work in the entire corpus: a theranostic conjugate with preserved EGFR inhibition and 3- to 4-fold mutant-selective fluorescence in PC9 and H1975 lines (Dong et al.). We will run OSA-Cy7 head-to-head against our full analog grid in identical PDO-CAF wells, at equal molar dose, equal imaging parameters, and equal analysis code, so that any advantage of the grid (better core penetration, better occupancy-to-response coupling) is attributable to structure rather than protocol. For outcome prediction, the strongest reported number on this exact task is the video-based model's MAPE of 0.1755 (Pearson 0.9214) on the CRC organoid time-lapse benchmark (Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy); we will re-implement its DINOv2-features-plus-attention architecture on our data and give it the same tuning budget and compute as our fluorescence-augmented variant. Beating it means a statistically significant MAPE reduction across organoid lines, with the pre-registered threshold set at MAPE ≤ 0.15 while retaining Pearson above 0.9 on held-out patient lines — patient-level splits, not random well splits, since MoleculeNet: A Benchmark for Molecular Machine Learning showed random splits flatter model performance.

The key ablation isolates chemical featurization. We train two otherwise-identical models on the identical analog grid and organoid data: one that receives the molecular graph as input (the spatial SAR model) and one that treats each compound as an anonymous categorical label (the setting of the existing video model, which is exactly how Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy handled drugs). The decisive comparison is held-out-analog generalization: leave out several analogs entirely, predict their penetration and efficacy, and test whether the graph model predicts unseen structures while the label model cannot by construction. A second ablation removes the CAFs, quantifying how much of each analog's spatial profile is stroma-dependent rather than geometry-dependent.

Why not just…?

Why not just run the SAR screen in 2D cell lines? Because the outcome we are modeling — penetration through matrix and stroma — does not exist in a monolayer. The corpus is direct about this on both physical and biological grounds: 2D cultures expose every cell to uniform drug concentration and eliminate the radial gradients that drive resistant subtypes in pancreatic organoids (Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes), and stromal protection through EMT and YAP-dependent slow-cycling states is entirely absent without CAFs (Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses; Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system). Still, the objection has real force — 2D is cheaper and higher-throughput — so we will run a parallel 2D arm on the same analogs and report the rank-correlation between 2D potency and 3D spatial efficacy. If the rankings coincide, that is a genuine finding that spatial SAR reduces to flat-culture SAR for this scaffold class, and it would shrink the proposal's claim; we consider this unlikely given the documented stroma effects, but the experiment settles it.

Why not just co-administer a dye with the drug instead of synthesizing conjugates? A separately dosed tracer clears, extravasates, and diffuses differently from the drug, so its signal cannot report the drug's location — the exact failure Xiao et al. documented, and the reason pharmacokinetic identity is the defining requirement of theranostics. The synthesis objection is also empirically weaker than it sounds: Dong et al. showed a bulky Cy7 on Osimertinib preserved Cys797 covalent binding and p-EGFR suppression, and Yang et al. showed single-moiety tuning can multiply tumor exposure eightfold. Our LC-MS-per-well control makes this a measured assumption rather than a defended one: if a conjugate's measured diffusion differs from parent drug, that analog is flagged and the discrepancy itself becomes SAR data.

Why not just predict everything in silico with GNNs or large language models? The corpus gives quantitative grounds for doubt. Out-of-distribution splits cut drug-target interaction correlations roughly in half (~0.70 to ~0.40) in Therapeutics Data Commons, where classical 2D descriptors often matched deep graph models (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development); standard GNNs are at random chance on enantiomer benchmarks (Interpretable Chirality-Aware Graph Neural Network for Quantitative Structure Activity Relationship Modeling in Drug Discovery); and general-purpose LLMs failed to capture single-cell drug responses in the scDrugMap benchmark (scDrugMap: benchmarking large foundation models for drug response prediction). More fundamentally, no training labels for spatial penetration exist anywhere — that is the gap this proposal creates the data for. The in-silico route is not a competitor; it is a downstream consumer of the dataset we build. We do include one honest caveat: with only ~24–32 analogs, our learned spatial SAR model will be data-starved, and its main near-term value will be the extracted matched-pair transformation rules, not a universal predictor.

Why not just use destructive endpoint assays like CellTiter-Glo? They require total organoid lysis, destroying the ability to track drug gradients, sub-clone dynamics, and early response in the same sample — the limitation that motivated the non-destructive video approach in the first place (Spatio-Temporal Analysis of Patient-Derived Organoid Videos Using Deep Learning for the Prediction of Drug Efficacy), and a destructive endpoint gives a single number per well where our design needs depth-resolved, time-resolved signals. The destructive ATP assay retains a role here as the ground truth at 100 hours, exactly as it served the video model, and as the sanity anchor for any co-culture where cell death kinetics must be separated by cell type, following the DeathPro approach of Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system.

Proposed Experiments and Decisive Checks

Experiment 1 (assay validity): in EGFR-mutant PDOs without CAFs, image OSA-Cy7 and the analog grid for 100 hours and quantify radial fluorescence gradients; at matched time points, LC-MS-quantify drug mass per well and per dissected organoid fraction. The decisive check is that fluorescence-to-mass correlation across analogs is high for conjugates and broken for the co-administered-dye control, establishing pharmacokinetic identity — the gap left open by the 2D-only validation of Dong et al. and the xenograft-only validation of Yang et al..

Experiment 2 (does structure control depth?): regress penetration half-depth and surface-to-core fluorescence ratio on the one-variable-at-a-time structural edits, with matched-molecular-pair statistics to extract significant transformation rules (Semi-automated workflow for molecular pair analysis and QSAR-assisted transformation space expansion). The decisive check is that at least one structural edit (linker length, fluorophore bulk, or site) shifts core penetration by a reproducible, patient-line-consistent factor — demonstrating that penetration is an engineerable SAR property, not a fixed property of the scaffold.

Experiment 3 (mechanism decomposition): with patient-matched CAFs added, run Trellis-style multiplexed single-cell readouts of p-EGFR, p-Akt, YAP, and EMT markers stratified by radial depth for the best- and worst-penetrating analogs. The decisive check is the separation of the ~10% in vitro/in vivo discordance of Drug repurposing screening and mechanism analysis based on human colorectal cancer organoids into delivery-limited versus signaling-rewired organoids, and whether a YAP inhibitor co-treatment, in the manner of the Verteporfin rescue in Trellis Single-Cell Screening Reveals Stromal Regulation of Patient-Derived Organoid Drug Responses, restores response only in the signaling-rewired subset.

Experiment 4 (prediction): train the graph-input and label-input models on all but a held-out set of analogs and patient lines; predict 100-hour ATP viability from the first 7.5 hours of combined bright-field-plus-fluorescence video. Beating the re-implemented bright-field-only baseline (reported MAPE 0.1755, Pearson 0.9214; the prior Ins-ATP baseline stood at MAPE 0.4375) on held-out patient lines, plus non-trivial held-out-analog prediction from the graph model alone, would establish spatial SAR as a learnable, generalizing quantity. For the organoid-imaging component we will follow the cohort scale of the corpus's organoid studies (dozens of patient lines, thousands of organoids per condition) rather than the 2D screen scale, accepting lower throughput as the cost of measuring what 2D cannot.

The deliverables are a first spatial-SAR dataset linking chemical graphs to depth-resolved penetration, occupancy, and signaling in patient tissue; a set of validated structural transformation rules for penetration and stromal resistance; and a direct empirical decomposition of organoid-to-xenograft discordance. If the rank-correlation between 2D and 3D efficacy turns out to be near one, or if co-administered dyes track drug distribution as well as conjugates, the core premise fails and we will report it — those are the falsifiable edges of the claim.