Command–Response Fields for Immune Exclusion in Cold Tumors

In speech research, the Fujisaki model explains a speaker's pitch contour as the sum of a constant baseline, a handful of slow "phrase" resets, and short local "accent" steps — each command passed through a damped second-order filter so the resulting curve is always smooth and physically realizable, because vocal-fold muscles cannot change tension instantaneously (Hiroya Fujisaki). Cold tumors — tumors where cytotoxic T cells are either absent entirely or physically trapped in the surrounding stroma (Ming Yi et al.; Mengke Niu et al.) — have immune landscapes with exactly this two-tier shape: slow tissue-wide gradients of hypoxia, perfusion decay, and TGF-β-driven fibrosis (Aijun Yang et al.; Alexandro E. Trevino et al.), onto which sharp local structures are superimposed: the fibrotic "interface niche" that walls lymphoid aggregates off from malignant cells, and focal chemokine or necroptotic bursts that try to pull T cells back in (Alexandro E. Trevino et al.; Zitong Wang et al.; Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation).

This proposal imports the command–response mathematics as a compact physics of immune exclusion, and pairs it with a second import from a distant field: distant supervision. Expert annotation is the bottleneck in both domains — discourse treebanks stalled at 385 gold documents (Gregor Weiss; Shafiq Joty et al.) until Patrick Huber et al. showed that back-propagating coarse document-level star ratings through a multiple-instance-learning parser could manufacture 250,000 silver discourse trees; meanwhile, the spatial studies that best explain immunotherapy resistance rest on as few as 13 patients (Alexandro E. Trevino et al.). A model whose entire description of a tumor is a few dozen physically meaningful commands counters both problems at once: it cannot memorize 13 patients, and its parameters are directly readable as candidate interventions.

Core Claim

A spatial immune or stromal protein field measured by multiplexed imaging can be decomposed, accurately and with very few parameters, into a baseline, a small number of smooth global gradient commands, and a small number of sharp region-bound step commands, each convolved with a critically damped radial kernel whose decay length encodes how far that signal penetrates fibrotic tissue. Three consequences follow, and each is testable: the fitted step commands should localize the interface niche and separate responders from non-responders; the command features should predict response across cohorts better than composition biomarkers and comparably to deep graph models while transferring better across cancer types; and counterfactual interventions designed in command space should be physically feasible in a way that unconstrained pixel edits are not.

Research Program

The forward model is a direct spatial transplant of the Fujisaki decomposition. For a scalar field $S(\mathbf{x})$ over tissue coordinates $\mathbf{x}$ (for example smoothed TGFB1 or SPP1 expression, collagen density, or a chemokine):

$$S(\mathbf{x}) \;=\; B \;+\; \sum_{i=1}^{I} A_i\, G_\alpha\!\big(\|\mathbf{x}-\mathbf{c}_i\|\big) \;+\; \sum_{j=1}^{J} a_j \!\int_{R_j} G_\beta\!\big(\|\mathbf{x}-\mathbf{u}\|\big)\, d\mathbf{u}$$

Here $B$ is the tissue baseline; the first sum holds point-source commands at centers $\mathbf{c}_i$ with amplitudes $A_i$, modeling global gradients (hypoxia fields, perfusion decay); the second sum holds step commands on regions $R_j$ with amplitudes $a_j$, modeling localized structures such as the interface-niche ring around a lymphoid aggregate or a focal CXCL16 burst. $G_\alpha(r) = \alpha^2 r e^{-\alpha r}$ is the radial analog of Fujisaki's critically damped impulse response, and the decay lengths $1/\alpha$ and $1/\beta$ are the physical content of the model: they say how far a signal travels through stroma before damping out. This decoupling of global scale from local identity mirrors what Jinfu Ni et al. demonstrated for tone and intonation in speech — global modulation acts as a latent scale transformation over local commands — and it is precisely the structure that dual TGF-β × PD-L1 blockade lacks clinically: shifting the global fibrotic baseline without adding local recruitment steps failed in consecutive trials, and efficacy returned only when a focal STING-driven CXCL16 command was added (Triple targeting of STING, TGF-β, and PD-L1 boosts CXCL16–CXCR6 signaling for potent antitumor response; Xinrun Wang et al.).

Three components must be built. First, an inversion procedure that recovers commands from an observed field. This is an ill-posed inverse problem with many local minima, exactly as automatic Fujisaki parameter extraction is for Cantonese pitch (Towards automatic parameter extraction of command-response model for Cantonese); we will use differentiable optimization with a sparsity penalty on command count, initialized by peak and edge detection, and validated by reconstruction error. The methodological precedent is the electrolaryngeal speech work of Physically Constrained Statistical F0 Prediction for Electrolaryngeal Speech Enhancement, which showed that coupling a statistical decoder to the physically constrained generative model — rather than regressing the raw contour — raised the $F_0$ correlation coefficient from 0.40 with a plain Gaussian mixture model to 0.50 with their hidden-Markov-model decoder. That 0.40-to-0.50 gain is the central evidence that the constraint does real work, and it is the pattern we aim to reproduce in tissue. Second, a distant-supervision training scheme: patient-level outcome labels (response, recurrence, survival) are the analog of star ratings, and a multiple-instance-learning attention layer over candidate commands — the tissue analog of Huber's attention over discourse units — learns which command configurations matter without any spot-level expert annotation. Third, a counterfactual module: given fitted commands, we search for the minimal command addition or amplitude change that flips an excluded phenotype to an infiltrated one, subject to damping feasibility — the required gradient must penetrate the measured stromal width at a deliverable amplitude.

Baselines

The naive reference point is aggregate composition: cell-type proportions, total TIL density, CD8/Treg ratio, and tumor mutational burden. This is the standard clinical comparator, and it is the floor to clear: on the Stanford-CRC cohort, SPACE-GM reports that composition-based methods performed near random chance where their graph model reached an ROC-AUC of 0.739, and Alexandro E. Trevino et al. found cell composition alone could not distinguish responders from non-responders in tumor-dense regions. Any model that fails to beat composition has no spatial content worth publishing.

The strongest existing competitor is SPACE-GM, a three-layer graph isomorphism network over 3-hop cellular neighborhoods within 75 microns (SPACE-GM: geometric deep learning of disease-associated microenvironments from multiplex spatial protein profiles), together with the multiplexed-immunofluorescence GNN of Alexandro E. Trevino et al., which reports an AUROC of 0.93 on pre-treatment HCC biopsies, and SPACE-GM itself at a ROC-AUC of 0.98 on the 13-patient HCC cohort of Alexandro E. Trevino et al.; on UPMC-HNC, SPACE-GM reports ROC-AUC above 0.86 and a C-index around 0.799. We will run SPACE-GM and the MIF-GNN on identical input cell tables, identical patient splits (leave-one-patient-out on the small HCC cohorts, and cross-cancer transfer such as training on the head-and-neck cohorts and testing on Stanford-CRC), with the same nested tuning budget and compute. Beating this baseline does not mean a higher within-cohort AUROC on 13 patients — a few-dozen-parameter model will likely lose that race. It means matching within the confidence interval under leave-one-patient-out while winning the cross-cancer transfer comparison by a clear margin, and doing so with parameters that name interventions.

For the counterfactual arm, the competitor is Morpheus, which trains CNNs on multiplexed tissue images and inverts them by unconstrained pixel optimization; on its melanoma dataset it reports that a counterfactual perturbation raised predicted CD8+ T-cell infiltration from 3% to 69% (Zitong Wang et al.). We will fit our commands to the same images and ask for perturbations achieving a comparable predicted shift. Beating Morpheus here means the recommended perturbations achieve similar predicted infiltration while remaining physically feasible — the induced gradient's decay length covers the measured stromal width and the required amplitude falls within deliverable ranges — whereas unconstrained pixel edits carry no such guarantee.

The key ablation is the same architecture with the physics removed: an equally-sized set of free-form region parameters with no damped kernel and no superposition structure, trained identically. A second ablation drops distant supervision and trains only on the gold-labeled HCC cohort, and a third drops the local step commands, keeping only global gradients — this last one isolates the decoupling claim, since it emulates a pure baseline-shifting therapy of the kind that failed in the TGF-β × PD-L1 trials (Triple targeting of STING, TGF-β, and PD-L1).

Why not just…?

Why not just count cells? Because the evidence says composition does not carry the signal. Composition baselines performed near random chance on Stanford-CRC where the graph model reached 0.739 (SPACE-GM), and high tumor mutational burden is insufficient on its own in immune-excluded tumors where a collagen wall physically blocks T-cell contact (Xinrun Wang et al.; Christian Hörner et al.). The relevant question is whether spatial structure matters, and that question is already settled in favor of structure.

Why not just fine-tune SPACE-GM on each new cohort? This is the real threat, and we will not argue around it. A 0.98 ROC-AUC from 13 patients (Alexandro E. Trevino et al.) is either a remarkably clean biological law or a memorized cohort, and the existing literature cannot tell us which because cross-cohort transfer was not the headline evaluation. Our decisive check is designed to settle this: train each model on the pooled head-and-neck and colorectal cohorts and test, without fine-tuning, on the HCC trial cohorts and on the Morpheus melanoma and colorectal-liver-metastasis images. If SPACE-GM transfers as well as the command model, the honest conclusion is that the parametric bottleneck buys interpretability and intervention design, not predictive accuracy — and the counterfactual feasibility experiments would then carry the proposal. We state this plainly because it is a possible outcome.

Why not just write a reaction–diffusion PDE? A generic PDE gives forward simulation but no decoder from measured images to parameters, and no discrete vocabulary of interventions — it cannot tell a clinician "add one step command here." The hard part, demonstrated in speech, is coupling statistical estimation to the physically constrained generative model: the electrolaryngeal work only succeeded once the Fujisaki structure was embedded in the decoder, improving correlation from 0.40 to 0.50 (Physically Constrained Statistical F0 Prediction for Electrolaryngeal Speech Enhancement), and inversion remains ill-posed even in one dimension (Towards automatic parameter extraction of command-response model for Cantonese). Our contribution is precisely that coupling, in two dimensions, with distant supervision supplying the training signal.

Why not just give patients the drug combination and skip the modeling? Because the empirical combination search has already produced expensive failures: dual TGF-β × PD-L1 blockade failed in consecutive trials, and the triple-targeting work had to rediscover by drug engineering what a decoupled command model would have predicted — that removing the global baseline without adding local recruitment steps does nothing (Triple targeting of STING, TGF-β, and PD-L1). A model that quantifies decay lengths and required amplitudes turns that search into arithmetic.

Proposed Experiments and Decisive Checks

The first experiment fits commands to the HCC spatial cohorts of Alexandro E. Trevino et al. and Alexandro E. Trevino et al. and tests descriptive validity: does the inversion place a step command on the fibroblast-rich, SPP1/TGFB1-expressing ring around lymphoid aggregates, with decay length consistent with the measured barrier width, and does the interface command amplitude separate the "exposed" responders from the "interfaced" non-responders? The second experiment is the scaling study: pool UPMC-HNC (308 samples), Stanford-CRC (292), DFCI-HNC (58), and the Morpheus melanoma (69 patients) and colorectal-liver-metastasis (30 patients) images under coarse outcome labels only — no spatial expert annotation — train the multiple-instance command parser, and trace prediction quality against silver-corpus size, the tissue analog of Huber's demonstration that 250,000 noisy trees beat 385 gold ones; DFCI-HNC has no reported result found on Lacuna for any method, which makes it a clean held-out transfer target. The third experiment is the counterfactual comparison against Morpheus on the melanoma dataset, judged on feasibility-adjusted infiltration gains rather than raw predicted percentages. A secondary bulk-level check fits only the global command parameters to TISMO's 2,123 RNA-seq samples across 68 syngeneic mouse models, including 832 checkpoint-blockade samples (TISMO: syngeneic mouse tumor database to model tumor immunity and immunotherapy response), to test whether global-gradient parameters alone predict immunotherapy response — they should partially, and the residual should belong to the local commands that bulk data cannot see.

The decisive checks are three. The damping ablation must hurt: if free-form parameters do as well, the Fujisaki structure is decoration and the proposal fails. The transfer gap must be real: the command model must degrade less than the graph baseline when moved across cancer types. And the counterfactuals must be feasible: at least some Morpheus-style pixel edits must be shown, within the damped model, to require physically impossible gradients given the measured stromal width. If all three hold, the field has a compact, trainable physics of immune exclusion — and a bridge from static spatial atlases to dosable interventions. What this proposal deliberately does not attempt is in vivo validation; the in vitro transwell assays used by Morpheus are the natural next step, and the damped counterfactuals are designed to be directly testable in exactly that setup.

References

  1. In search of models in speech communication research
  2. Exploiting innate immunity for cancer immunotherapy
  3. The enhanced antitumor activity of bispecific antibody targeting PD-1/PD-L1 signaling
  4. CD8 + T cell infiltration is associated with improved survival and negatively correlates with hypoxia in clear cell ovarian cancer
  5. 1415 Multimodal spatial signatures of immunotherapy response in hepatocellular carcinoma
  6. 478 Spatial omics and deep learning define an immunotherapy resistance niche in hepatocellular carcinoma
  7. 1131 Spatially-guided therapy predictions with morpheus
  8. Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation
  9. Learning Representations for Text-level Discourse Parsing
  10. Discourse Analysis and Its Applications
  11. MEGA RST Discourse Treebanks with Structure and Nuclearity from Scalable Distant Sentiment Supervision
  12. Estimation of intonation variation with constrained tone transformations
  13. Triple targeting of STING, TGF-β, and PD-L1 boosts CXCL16–CXCR6 signaling for potent antitumor response
  14. Bispecific antibody targeting TGF-β and PD-L1 for synergistic cancer immunotherapy
  15. Towards automatic parameter extraction of command-response model for Cantonese
  16. Physically Constrained Statistical F0 Prediction for Electrolaryngeal Speech Enhancement
  17. SPACE-GM: geometric deep learning of disease-associated microenvironments from multiplex spatial protein profiles
  18. TISMO: syngeneic mouse tumor database to model tumor immunity and immunotherapy response