From Bath to Gradient: Transferring In Vitro Cytokine-Arrest Maps to In Vivo Lesions

Chronic inflammation often fails to heal tissue not by killing progenitor cells but by freezing them: exposed to inflammatory signals, oligodendrocyte precursors, blood-forming stem cells, and bone progenitors stay alive yet stop maturing, or even switch into immune-active cells that attract T-cell destruction. Three separate literatures each hold one piece of this problem and never meet. Paracrine signaling assays — experiments where donor cells' secreted cocktail is applied to recipient cells without physical contact — prove that soluble factors alone cause the arrest. Inflammatory-microenvironment studies show the arrest is viability-silent, stage-dependent, and resistant to single-cytokine blockade. And the translation-gap literature explains why in vitro findings fail in patients: a culture dish applies a uniform, static "bath" of cytokines, while a living tissue presents each cell with a steep, spatially decaying concentration gradient shaped by diffusion, matrix binding, and clearance. The result is a standing contradiction: bath studies of interferon-gamma at 10 ng/ml report clean, non-toxic maturation arrest (O4+ oligodendrocytes fall from 60–80% to below 20%, iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu), while in vivo studies of the same cytokine report CD8+ T-cell lysis of the same precursors (Oligodendrocyte precursor cells present antigen and are cytotoxic targets in inflammatory demyelination).

The claim of this proposal is that this contradiction is not noise but a missing variable: the local, time-integrated exposure field. We will learn, in vitro, a dose–duration–stage response function mapping cytokine cocktails to progenitor cell states; re-express that function under a spatially resolved diffusion-decay field computed from tissue sections; and use domain adaptation to transfer the combined model from culture to lesion. The deliverables are quantitative predictions no current method makes: the physical radius around an inflammatory cluster within which progenitors arrest (and whether the boundary is sharp or graded), which cytokine combinations must be blocked to restore maturation, and a matched in vitro–in vivo benchmark for progenitor arrest that the field currently lacks entirely.

Core Claim

Maturation arrest is a deterministic function of local cytokine exposure — concentration, duration, and the progenitor's developmental stage — and that function can be learned in vitro where exposure is controlled, then evaluated in vivo once exposure is reconstructed as a spatial field. Specifically, for a ligand $j$ secreted by source cells $i$ at positions $x_i$ in tissue, we model the steady-state exposure at location $x$ with the Green's function of diffusion with decay:

$$E_j(x) \;=\; \sum_i s_{ij}\, \frac{e^{-\|x-x_i\|/\lambda_j}}{4 \pi D_j\, \|x-x_i\|}, \qquad \lambda_j = \sqrt{D_j / \delta_j}$$

where $s_{ij}$ is the secretion rate of ligand $j$ by cell $i$, $D_j$ its diffusivity, and $\delta_j$ its clearance rate, so $\lambda_j$ is the distance over which the signal survives. The arrest response is then a learned function $f_\theta(E(x), a, t) \to$ progenitor state, where $a$ is developmental stage and $t$ is exposure duration. This differs from all three nearest bodies of work. COMMOT reconstructs spatial signaling fluxes with optimal transport but uses mRNA as a proxy for protein, requires hand-set distance thresholds rather than learned decay lengths, and predicts communication, not a functional outcome like arrest. Perturbation predictors such as GEARS and chemCPA predict transcriptional responses to combinatorial perturbations but assume a well-mixed, uniformly exposed population — there is no spatial variable at all. Transfer-learning systems like DeepSADR bridge in vitro to in vivo for drug response in cancer, but their domain shift is distributional; ours is physical, and the exposure field is precisely the correction term that a bath-trained model is missing.

Research Program

The program has three components and one benchmark-building effort.

First, the in vitro arrest-response function. We will train a combinatorial perturbation model in the spirit of GEARS — which predicts single-cell transcriptional outcomes of multi-gene knockouts, reporting a 53% MSE improvement on zero-shot two-gene combinations — but over cytokine cocktails with explicit dose, duration, and developmental-stage inputs, rather than gene knockouts. Training anchors are two existing datasets: GSE147315, bulk and sorted RNA-seq of human iPSC-derived oligodendrocyte precursors under homeostatic versus chronic low-dose IFNγ (over 3,200 differentially expressed genes; iPSCs from people with MS...), and GSE250444, single-cell RNA-seq of 14,248 hematopoietic progenitors exposed to activated CAR-T supernatants (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma). We will extend both with a designed factorial panel: pairwise and triple cytokine combinations (IFNγ, TNFα, IL-1β, IL-6, IL-17) crossed with dose, duration, and initiation day, read out by the MBP-promoter Nanoluciferase reporter line from Reactive Astrocytes Derived From Human iPSCs Suppress Oligodendrocyte Precursor Cell Differentiation plus single-cell RNA-seq, gated by the scPerturb quality thresholds of at least 300 cells per perturbation and 1,000 UMI counts per cell (scPerturb: Harmonized Single-Cell Perturbation Data). The model represents the arrested state continuously — the hybrid MHC+/PSMB8+/TUJ1+ immune-like state documented in both the MS-iPSC and CAR-T studies — rather than as a binary live/dead or mature/immature label, because binary labels are exactly what in vivo data refutes (Macrophages may promote cancer growth via a GM-CSF/HB-EGF paracrine loop).

Second, the exposure-field layer. Given a spatial transcriptomics section of inflamed tissue, we estimate $s_{ij}$ for ligand-producing cells, then solve for $E_j(x)$ with per-ligand decay lengths $\lambda_j$ learned by fitting held-out progenitor states — generalizing COMMOT's fixed maximum-distance transport costs into physically parameterized, learnable kernels, and borrowing geometry-aware message-passing operators from NeuronCtrl, which showed that ignoring true 3D morphology raises safety-violation rates from under 1% to 6–9% in microenvironment simulation. For spatial pretraining and embedding of tissue context we use the SToCorpus-88M corpus (88 million cells across 2,000 slices) behind SToFM, whose codebase is available at PharMolix/SToFM, and volumetric extrapolation via VORTEX, which reconstructs 3D expression from sparse 2D slices (Pearson correlation 0.46 over all genes, 0.57 for the top 50, at depths 171 times the physical section) — necessary because arrest radii live in three dimensions.

Third, the transfer. We adapt chemCPA's architecture — adversarial separation of basal cell state from perturbation-induced shift — so that the in vitro perturbation vector (bath dose) and the in vivo perturbation vector (reconstructed exposure field) share a latent space. The critical test of transfer is whether a model trained where exposure is uniform can predict outcomes where it is graded.

Finally, the benchmark. No matched in vitro–in vivo progenitor-arrest spatial dataset exists; on Lacuna, no reported result exists for one. We will build it: ATCup mice (cuprizone demyelination plus adoptive transfer of myelin-reactive T cells, per Animal models to investigate the effects of inflammation on remyelination in multiple sclerosis and Oligodendrocyte precursor cells present antigen...) with PDGFRα-lineage fate tracing, profiled by spatial transcriptomics, paired with the factorial in vitro panel from the same cytokines. For higher-throughput in vivo perturbation of candidate rescue combinations, the zebrafish mosaic transgenesis pooled-screening method of Library transgenesis in zebrafish... allows in vivo testing of transgene libraries in intact developing tissue.

Baselines

The naive reference point is the uniform-bath assumption applied to tissue: take the mean lesion cytokine level, apply the in vitro dose–response curve at that single concentration, and predict a uniform arrest outcome across the lesion. It is the implicit model behind every bath study and behind single-target blockade trials. It will be run on the same spatial sections with the same learned $f_\theta$, so the only difference is the spatial structure of $E(x)$. Beating it means predicting spatial heterogeneity — arrest near immune clusters and escape at the lesion rim — where the bath model can only predict all-or-nothing, and matching measured arrest gradients significantly better on held-out sections.

The strongest existing competitor is a pretrained perturbation-response model of the chemCPA type, applied directly to in vivo lesional progenitors without any exposure-field layer. chemCPA pre-trains on 1.3 million bulk L1000 profiles and reaches a median $r^2$ of 0.79 on differentially expressed genes in the sci-Plex3 single-cell benchmark; without that pretraining, the same encoder-decoder performs at random-guess level, $r^2 \approx 0.00$ (Predicting single-cell perturbation responses for unseen drugs). We will run this baseline with identical training data, pretraining corpora, and tuning budget as our model, differing only in the absence of the spatial exposure reconstruction. Beating it means higher DEG prediction $r^2$ on held-out in vivo progenitor states, with the margin concentrated in cells far from ligand sources — exactly where a well-mixed model must err. For the retrieval framing (given an arrested population, retrieve the causal cytokine combination), the reference is CellVS-Net's Hits@1 of 0.1250 on DDR-Bench versus 0.0357 for molecular fingerprints and 0.0179 random (Cell-Level Virtual Screening).

The key ablation removes the exposure-field layer alone, feeding the model the lesion-average concentration instead of $E(x)$ while keeping everything else fixed. The precedent for expecting a collapse is direct: in DeepSADR, replacing the adaptive readout with mean pooling dropped average transfer AUC from 0.856 to 0.662; in NeuronCtrl, removing geometry-awareness raised unsafe-field rates from under 1% to 6–9%. If our ablation does not show a comparable drop, the spatial ingredient is not earning its place and the simpler model wins.

Why not just…?

Why not just screen neutralizing antibodies against individual cytokines in culture? The literature contains a clinical refutation: in CAR-T patients with severe marrow suppression and high IL-6, the IL-6 receptor blocker tocilizumab failed to rescue cytopenia (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy...). Secretomes are also non-additive — MSC/endothelial co-culture drives IL-1β to 49.13 ng/ml versus 8.96 and 22.14 ng/ml in the monocultures (Interleukin-1β induces ICAM-1 expression...; Interleukin-1β induces intercellular adhesion molecule-1 expression, thus enhancing the adhesion between mesenchymal stem cells and endothelial progenitor cells via the p38 MAPK signaling pathway) — so monoculture supernatants miss the combinations that matter. And arrest is viability-silent, so simple death-based screens cannot even detect it. Our factorial combination model is designed precisely to predict which multi-target blockades work; the model's rescue predictions will be tested head-to-head against single-cytokine blockade in the in vitro panel and in vivo.

Why not just use combinatorial mouse models (ATCup) directly and skip the computational layer? These models are the right in vivo substrate and we use them — but the review that introduced them states their own limits: they run on young 8–10 week old animals, fail to reproduce decades of smoldering low-level cytokine exposure, and leave the paracrine diffusion thresholds that switch a precursor from repair to arrest completely unquantified (Animal models to investigate the effects of inflammation on remyelination). Without the exposure-field model, ATCup experiments are also low-throughput and cannot tell you which of the hundreds of possible cytokine combinations to block. The model supplies the priors; the mice supply the ground truth; neither substitutes for the other.

Why not just use standard domain adaptation — align global expression vectors between dish and tissue? The benchmarks say this fails at the easier version of the problem. Mean pooling in DeepSADR collapses transfer AUC from 0.856 to 0.662; chemCPA without large-corpus pretraining sits at $r^2 \approx 0.00$; CellVS-Net's authors found gene-network structure carried the therapeutic signal where raw expression snapshots did not (Cell-Level Virtual Screening). Global alignment would average away exactly the spatial dose-decay structure that differs between bath and tissue. Our ablation (removing only the exposure field) is the direct experimental test of whether this concern is real, and we have committed to accepting the simpler model if the drop is negligible.

Why not just multiply ligand expression by receptor expression to infer communication? Unconstrained pairing overestimates signaling badly: over 70% of ligands and 60% of receptors in Fantom5 bind multiple partners, so isolated pairwise products ignore competition (Screening cell–cell communication in spatial transcriptomics via collective optimal transport), and such scores carry no dose, duration, or stage information, so they cannot predict a maturation outcome at all. COMMOT already fixes the physics and competition; we build on it rather than reinvent it, and our baseline suite includes a COMMOT-derived flux field as the input to a simpler outcome model.

Decisive Checks

Four checks decide the proposal. First, the transfer test: DEG prediction $r^2$ on held-out in vivo lesional progenitors, comparing the full model against the chemCPA-type baseline (0.79 on its own sci-Plex3 benchmark is the number to contextualize against; no reported result found on Lacuna for any perturbation model evaluated on in vivo lesional progenitor states). Second, the structure test: whether arrest decays gradedly with distance from immune clusters or switches at a sharp threshold — measured by spatial transcriptomics plus fate tracing in ATCup mice, predicted in advance by the model. Third, the retrospective clinical test: the trained model must predict that IL-6 blockade alone fails to rescue arrest (matching the tocilizumab outcome in Palacios-Berraquero et al.) while correctly ranking combinations that do rescue in the factorial in vitro panel. Fourth, the stage-window test: the model should reproduce the known developmental asymmetry — IFNγ from day 36 drops O4+ cells below 20%, from day 60 much less so (iPSCs from people with MS...) — and extend it to a prediction about which stage distributions in tissue are most vulnerable. If the model passes the ablation-failure condition (spatial layer removable without loss), the honest conclusion is that arrest is bath-like in vivo too, and the field's translation problem lies elsewhere — that would itself be a decisive, publishable answer to the gradient question the literature has left open.

References

  1. iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu
  2. Oligodendrocyte precursor cells present antigen and are cytotoxic targets in inflammatory demyelination
  3. Screening cell–cell communication in spatial transcriptomics via collective optimal transport
  4. Predicting transcriptional outcomes of novel multigene perturbations with GEARS
  5. Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution
  6. DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
  7. Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in multiple myeloma
  8. Reactive Astrocytes Derived From Human Induced Pluripotent Stem Cells Suppress Oligodendrocyte Precursor Cell Differentiation
  9. scPerturb: Harmonized Single-Cell Perturbation Data
  10. Macrophages may promote cancer growth via a GM-CSF/HB-EGF paracrine loop that is enhanced by CXCL12
  11. NeuronCtrl: Geometry-Aware Safe Closed-Loop Generative Control for Neuronal Microenvironment Dynamics
  12. SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics
  13. AI-driven 3D Spatial Transcriptomics
  14. Animal models to investigate the effects of inflammation on remyelination in multiple sclerosis
  15. Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes
  16. Cell-Level Virtual Screening
  17. Interleukin-1β induces intercellular adhesion molecule-1 expression, thus enhancing the adhesion between mesenchymal stem cells and endothelial progenitor cells via the p38 MAPK signaling pathway