A clinical biopsy is a single frozen photograph of a system that is secretly moving: immune cells divide, die, migrate, and get physically trapped by stromal barriers, all on timescales the photograph cannot show. Two literatures each hold half of what is needed. Spatial immune architecture has the representation — multiplexed imaging assays that measure 40-plus proteins per cell at exact physical coordinates, and graph neural networks that read the resulting cell graphs well enough to predict chemotherapy response within a single cohort. But these models are static, and they are trained per cohort, so nobody knows whether they learned universal resistance biology or the cohort's staining idiosyncrasies. Separately, the global temporal consistency literature — developed almost entirely for text timelines — has shown that when local predictions are stitched together without global axioms (nothing precedes itself, ordering is transitive, disjoint states cannot co-occur), systems confidently output logically impossible trajectories, and that baking the axioms into training as differentiable penalties fixes this.
Our claim is that these two halves meet in a place nobody has tested, and that the meeting point is exactly where the third direction — cell line and cohort specificity — lives. The only components of a tumor-immune system that are guaranteed to be identical across cohorts, scanners, and biological contexts are the biophysical axioms: modeled cell mass cannot appear or vanish, populations cannot explode without bound, and a cell cannot simultaneously travel down a phenotype transition and its reverse. We propose that anchoring a dynamical model of spatial immune architecture to these axioms is not merely a plausibility filter — it is a transfer mechanism. The parts of the model that satisfy hard laws are, by construction, the invariant parts; the parts left unconstrained are exactly where cohort-specific shortcuts hide.
Core Claim
No existing work has applied mathematically guaranteed temporal and physical consistency constraints to dynamical models fit on real multiplexed spatial immune data, and no existing work has tested whether such constraints improve transfer across cohorts. The two nearest papers split the gap cleanly between them. Temporal tissue dynamics from a spatial snapshot (OSDR) reconstructs cell-population dynamics from a single imaging mass cytometry biopsy by regressing Ki67 division markers on 80µm neighborhood composition — real data, real clinical payoff (it predicted tumor collapse in the NeoTRIP trial at $P < 10^{-6}$), but zero global constraints: its death rate is a crude spatial average, and nothing stops its forward integration from producing unphysical states. Graph Neural ODEs with Stability and Conservation Guarantees for Tumor Microenvironment Dynamics (GNODE) formulates precisely the missing constraints — exact mass conservation via zero-sum message passing (violations below $10^{-14}$) and Lyapunov stability bounding cell trajectories — but was validated only on a synthetic agent-based/PDE simulator, with no patient data, no immune dynamics, and no cross-domain evaluation. Our proposal is to build and test the model that sits in the empty quadrant: constraint-guaranteed dynamics, fit on real spatial immune graphs, evaluated on whether the constraints close the transfer gap that makes spatial models cohort-specific.
Research Program
The artifact is a dynamical model over cell graphs built from imaging mass cytometry sections, following the conventions already established in the field: nodes are single segmented cells with 40-dimensional protein phenotype vectors, and edges connect cells within an 80µm radius with distance-decay weights, as in OSDR's neighborhood construction and Bi-level Graph Learning Unveils Prognosis-Relevant Tumor Microenvironment Patterns in Breast Multiplexed Digital Pathology. Dynamics are a graph neural ordinary differential equation — a network that predicts each cell's instantaneous rate of state change from its own phenotype and its neighbors' phenotypes, then integrates those rates forward in continuous time:
$$\frac{d x_i}{d t} = f_\theta\!\big(x_i,\, \{x_j\}_{j \in \mathcal{N}(i)}\big), \qquad \sum_{i} \mathbf{1}^{\!\top} \dot{x}_i = 0,$$
where $x_i$ is the state vector of cell $i$, $\mathcal{N}(i)$ its spatial neighborhood, and the second condition — zero-sum message passing borrowed from GNODE — makes total modeled cell mass exactly conserved. To keep trajectories bounded we adopt GNODE's Lyapunov penalty on the learned dynamics. Because a biopsy is an open system (cells enter and leave the section), conservation is enforced on the modeled population with explicit in/out flux variables rather than denied; this is more honest than OSDR, which has no conservation notion at all.
The second ingredient is a differentiable logic layer translating the discrete transition axioms of cell-state biology into training penalties, using the Łukasiewicz t-norm machinery of Clinical Temporal Relation Extraction with Probabilistic Soft Logic Regularization and Global Inference. That paper enforces transitivity of event ordering in clinical text with a continuous loss; we enforce the tissue analogues. If $p_{A \to B}$ is the model's probability that a cell moves from phenotype state $A$ to state $B$ over a predicted interval, then a term like
$$\mathcal{L}_{\mathrm{logic}} = \max\big(0,\; p_{A \to B} + p_{B \to A} - 1\big)$$
penalizes a cell being predicted to transition in both directions at once — the biological twin of a temporal cycle, since text-timeline systems produce exactly this class of self-contradiction when constrained only locally, as Learning Rich Event Representations and Interactions for Temporal Relation Classification demonstrated for pairwise neural classifiers. We add analogous penalties for monotone progression along sensitive-to-persister-to-resistant axes and for mutually exclusive marker-defined states.
Crucially, we do not impose a directed-acyclic view of time. CTRL-PG itself concedes it cannot represent recurring phenomena, and recurrence is the normal mode of tumor-immune biology: OSDR's central discovery was an excitable T-cell/B-cell circuit generating repeated "immune flares." Our consistency layer therefore targets bounded attractors — flares may recur, but counts stay nonnegative and bounded — rather than acyclicity. This is a deliberate departure from the text-timeline tradition, and it is where we think the biological domain should reshape the consistency formalism rather than merely import it.
The third ingredient addresses the specificity axis directly. The model is factorized into a context-specific encoder (absorbing staining batch, scanner, and cohort nuisance, the confounders that Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology showed standard encoders latch onto) and a shared constrained dynamical core. Transfer between cohorts is forced through the constrained core, so anything the model can only express by violating an axiom cannot be transferred — and anything it can express lawfully can. Training signals come from what the cohorts already contain: Ki67-derived division rates per neighborhood (OSDR's supervision), pre/post-treatment biopsy pairs from the TNBC chemotherapy cohort, and outcome labels from NeoTRIP.
Baselines
The naive reference point is the well-mixed bulk ODE: model total cell-population counts per phenotype as ordinary differential equations with no spatial structure, fit to the same Ki67 and outcome data. This is the "do nothing" floor, and the file already quantifies its failure mode — in the GNODE evaluation on the synthetic simulator, the graph-free ODE predicted a resistant cell fraction of 0.000 against a ground truth of 0.242, because pooling cells erases the local niches where resistant populations actually survive. We will run it on the real cohorts to confirm the failure transfers to real data.
The strongest existing competitor on real spatial immune data is the unconstrained graph convolutional network of Identifying tissue states by spatial protein patterns related to chemotherapy response in triple-negative breast cancer, which reports AUROC = 0.71 for predicting chemotherapy response from pre-treatment IMC cell graphs in a 63-patient, 813-region, 4-million-cell cohort. That paper is the strongest competitor because it is the only method in the file that both uses real multiplexed spatial graphs and predicts a clinically central endpoint; its identified hallmarks of chemoresistance (cytotoxic T-cell exclusion, macrophage-fibroblast co-localization) are precisely the spatial structures our dynamics must evolve. We will run it under identical conditions: same segmented cells, same graph construction, same train/test splits, same tuning budget and compute, refit per transfer direction. Beating it means exceeding 0.71 AUROC on its own cohort and degrading less when trained on one cohort and tested on another — the single-number comparison is necessary but not sufficient, since within-cohort superiority alone would not test our claim.
The strongest competitor for the dynamics itself is OSDR Temporal tissue dynamics from a spatial snapshot, which we will refit on every cohort under the same supervision. Beating OSDR means higher neighborhood-rate fidelity on held-out patients, a lower rate of unphysical trajectories under long-horizon forward integration, and — decisively — better rate prediction when transferred across cohorts, which OSDR was never designed or evaluated to do. The file quotes no numeric accuracy figure for OSDR's Danenberg validation or its rate predictions, so we report no reported result found on Lacuna for OSDR's within-cohort rate-prediction accuracy; our comparison numbers will be generated by our own refit under matched conditions.
Finally, the key ablation is our own architecture with the constraint losses removed — conservation, Lyapunov penalty, and t-norm logic all switched off, encoder factorization kept. This isolates the proposal's central ingredient, and its design mirrors the CTRL-PG ablations, where removing the soft-logic regularization dropped F1 by 1.44% and removing the global inference step dropped it by 1.83% on I2B2-2012 — evidence in a different domain that the constraints, not the capacity, carry the gain.
Why not just…?
Why not just apply OSDR to each new cohort? This is the most serious alternative, because OSDR already works on real biopsies and made a validated clinical prediction. But it has three structural defects our proposal targets: its death rate is a global spatial average rather than an inferred quantity, its dynamics are assumed time-invariant even though therapy demonstrably reshapes the microenvironment (the TNBC cohort shows macrophage-fibroblast co-localization intensifying post-treatment), and it has no mechanism — representational or architectural — that would make a mapping learned in one cohort lawful in another. Our experiments settle this head-on: if refit-OSDR matches our model's cross-cohort rate prediction, the constraint hypothesis is wrong and we will say so. What OSDR cannot do by construction is guarantee nonnegative, bounded, mass-conserving trajectories, so the unphysical-trajectory comparison can only be informative in one direction.
Why not just use a bigger unconstrained GNN? The file contains direct evidence against this at two levels. On dynamics, the discrete GNN baseline in the GNODE study predicted a resistant fraction of 0.110 against a 0.242 ground truth — better than the bulk ODE's 0.000, but less than half the constrained model's 0.282 — because discrete message passing cannot represent continuous diffusion and asynchronous division. On prediction, the unconstrained GCN plateaued at AUROC 0.71 within a single cohort with no cross-cohort evaluation at all. Capacity does not supply laws; more parameters give a shortcut-learner more shortcuts to find, which is the cell-line-specificity lesson of the whole third direction. The abation baseline makes this comparison exact rather than rhetorical.
Why not just strip the non-invariant channels, as the nuclear-mask method does? The mask-alignment approach of Are nuclear masks all you need? lifted cross-organ accuracy on Ocelot from 54.5% (standard training) to 70.6% by forcing encoders onto nuclear geometry, and we happily borrow the idea for our encoder. But as an overall strategy it destroys our signal: the clinically decisive quantities in the TNBC cohort — B7H4, FOXP3, CD11b, TIM-3/CD366 — are protein markers with no counterpart in nuclear morphology, and the same paper's own framing ("texture hounds" versus structural invariants) does not extend to functional immune phenotyping. Our claim is orthogonal: we gain invariance by constraining dynamics, not by discarding measurement. The clean test is to run mask-only alignment on our graphs; if it transfers as well while retaining marker resolution, we are wrong.
Why not just do ontology-aware contrastive alignment, as POPPY does? Cross-modal transfer learning for mapping bulk transcriptomes at cellular level prevents dataset shortcuts using Cell Ontology distance and achieved up to 73% silhouette improvement across cohorts — but it is entirely a-spatial and static: it operates on dissociated cells and homogenized bulk profiles, discarding the coordinates, barriers, and neighborhood topology that the entire spatial-architecture literature (from stromal lymphocyte entrapment in Yang et al. to tertiary lymphoid structures in Larsson et al.) identifies as the operative variables. POPPY could serve as a pooled non-spatial baseline, and we will include it as such, but it cannot represent a T-cell being physically excluded by stroma, which is the phenomenon our model must evolve.
Proposed Experiments
The first experiment establishes within-cohort dynamical fidelity and physicality on the Danenberg breast cancer IMC cohort (693 patients, ~859,000 cells). We predict neighborhood division rates from held-out snapshots and compare against refit-OSDR and the bulk ODE, then run long-horizon forward integration and measure the fraction of trajectories that produce negative counts or unbounded growth, and the magnitude of conservation violations. GNODE demonstrated violations below $10^{-14}$ on synthetic data; our question is whether that survives real staining noise and imperfect segmentation. The file reports no numeric accuracy figure for OSDR on this cohort, so this comparison will be generated under matched refits.
The second experiment is the cross-cohort transfer test, which no cited work has run: train on Danenberg, test rate and architecture prediction on the independent IMC breast datasets OSDR itself used for validation (Fischer et al. 2023 and Wang et al. 2023), and test response prediction on the TNBC chemotherapy cohort. The bar for response prediction is the reported AUROC = 0.71 of the TNBC GCN; the bar for transfer is the degradation slope of the unconstrained ablation. Our claim predicts a specific ordering — constrained model degrades least, unconstrained GNN more, per-cohort refits cannot transfer at all — and that ordering, not a single AUROC, is the decisive result.
The third experiment is the computational analog of the in-vitro-to-in-vivo gap, and the sharpest test of the cell-line-specificity claim. GNODE trained only on the homogeneous ABM-PDE simulator — effectively a "cell line" of computational tumor biology, since it contains none of the stromal, batch, or architectural variability of real tissue. We will train our constrained model and the unconstrained ablation on the simulator alone and evaluate zero-shot on real IMC graphs. If constraint anchoring lets simulator-trained dynamics transfer where unconstrained dynamics fail, we have direct evidence that the axioms, not the context, were carrying the signal. This experiment is exactly the gap the GNODE authors left open, and it would either establish or refute the transfer claim cleanly.
The fourth experiment tests recurrence: we require the model to reproduce the excitable T-cell/B-cell flare circuit OSDR discovered, with flares that recur under a bounded attractor while counts remain physical. This is the point where our formalism must outgrow its text-timeline heritage — CTRL-PG's acknowledged inability to handle recurring states is the specific limitation we are repairing — so success is defined as bounded oscillatory dynamics matching the flare statistics, where a DAG-based consistency layer would fail by construction.
The final experiment evaluates clinical utility on NeoTRIP (279 TNBC patients with longitudinal biopsies), predicting pathological response from the pre-treatment snapshot through the dynamics. OSDR achieved $P < 10^{-6}$ for predicting tumor collapse in responders on this cohort; the file reports no AUROC for that analysis, so we will treat the OSDR refit as the head-to-head comparator rather than quoting an unreported number.
We close with the honest limitation: mass conservation and stability are approximations on an open, sampled tissue section, and our flux-variable construction is a modeling choice that could mask real biological influx. If the constrained model wins only within-cohort and not under transfer, the correct conclusion is that constraints buy plausibility but not invariance — a result that would itself discipline a growing habit of importing physical guarantees into biological models without testing what they actually purchase.
References
- Temporal tissue dynamics from a spatial snapshot
- Graph Neural ODEs with Stability and Conservation Guarantees for Tumor Microenvironment Dynamics (Student Abstract)
- Bi-level graph learning unveils prognosis-relevant tumor microenvironment patterns in breast multiplexed digital pathology
- Clinical Temporal Relation Extraction with Probabilistic Soft Logic Regularization and Global Inference
- Learning Rich Event Representations and Interactions for Temporal Relation Classification
- Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology
- Identifying tissue states by spatial protein patterns related to chemotherapy response in triple-negative breast cancer
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level