Ranking EMT Inhibitors by Collapse of Spatial Signaling Loops

When epithelial cells turn cancerous or a tissue turns fibrotic, the deadliest change is often a change of shape and behavior, not of growth rate. In the epithelial-mesenchymal transition (EMT), tightly adherent lining cells dismantle their junctions, lose the adhesion protein E-cadherin, gain mesenchymal markers like Vimentin, and become motile and drug-resistant. Multi-target compounds — including botanical formulations such as Biejiajian Pill, which cut lung metastasis incidence from 80% to 20% in mouse liver-cancer xenografts (Biejiajian Pill Inhibits Carcinogenesis and Metastasis via the Akt/GSK-3β/Snail Signaling Pathway in Hepatocellular Carcinoma) — can suppress this program in vivo. But we cannot rank candidate compounds or explain failures, because the standard evaluation pipeline, network pharmacology, treats the tissue as a well-mixed pool of proteins and asks only which intracellular cascades a compound touches.

This matters because EMT is not purely a cell-internal decision. It is continuously induced and maintained by signaling between neighboring cells: cancer-associated fibroblasts broadcasting Wnt5a to stem-cell niches (Cancer associated fibroblasts serve as an ovarian cancer stem cell niche through noncanonical Wnt5a signaling), and, as Bocci et al. showed computationally, hybrid epithelial/mesenchymal intermediates that carry the highest stemness and metastatic potential also exhibit the highest cell-cell signaling capacity (Single-Cell RNA-Seq Analysis Reveals the Acquisition of Cancer Stem Cell Traits and Increase of Cell–Cell Signaling during EMT Progression). My proposal is to rank EMT inhibitors by a single, spatially resolved quantity: how completely their combined target set collapses the intercellular signaling loops that sustain hybrid EMT states, with the answer readable at scale from ordinary stained tissue sections.

Core Claim

A compound's anti-EMT efficacy in situ can be predicted by a communication-perturbation score computed on spatial ligand-receptor networks, and this score can be estimated both from spatial transcriptomics and, prospectively, from hematoxylin-and-eosin (H&E) histology images alone. Concretely, three pieces exist separately and have never been joined. Spatial communication inference via collective optimal transport — a method that assigns signaling flux between physically proximate cells while accounting for the fact that over 70% of ligands and 60% of receptors in the Fantom5 database compete for multiple partners (Screening cell–cell communication in spatial transcriptomics via collective optimal transport) — supplies the network. Per-cell EMT state scoring, following Bocci et al.'s epithelial/mesenchymal signature approach, identifies which cells sit in dangerous hybrid states. Network-pharmacology target maps, of the kind the morusin aortic-dissection study built by docking a flavonoid against seven hub proteins with binding energies of -9.0 to -7.0 kcal/mol (Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection), supply the perturbation.

The joining works as follows. For a compound $c$ with target set $T_c$, let $a_{c,i} \in [0,1]$ be an affinity prior for target protein $i$ (a normalized docking score). We attenuate the expression of every targeted ligand or receptor, $\tilde{g}_i = (1 - a_{c,i})\, g_i$, recompute the spatial communication flux $F$, and compare it to the unperturbed flux. The ranking score is the fraction of total flux collapse that lands on EMT-sustaining edges:

$$S_c \;=\; \frac{\Delta F_{\mathrm{EMT}}}{\Delta F_{\mathrm{tot}} + \epsilon},$$

where $\Delta F_{\mathrm{EMT}}$ is the reduction in flux delivered to cells in hybrid EMT states, $\Delta F_{\mathrm{tot}}$ is total flux reduction, and $\epsilon$ avoids division by zero. A high $S_c$ means the compound preferentially silences the paracrine loops feeding plastic cells rather than indiscriminately jamming tissue-wide signaling. This is deliberately a ratio: it penalizes the non-specific toxicity that bulk network pharmacology cannot see, because a compound that shuts down every pathway in every cell also "treats" EMT on paper.

Research Program

The first work package tests whether Bocci et al.'s coupling between EMT state and signaling capacity actually holds in intact tissue. Their result came from dissociated single-cell suspensions and 2D cultures; no one has checked it against physical distance and tissue architecture. Using the 1,229 paired spatial-transcriptomics-plus-histology samples in HEST-1k, spanning 26 organs and 367 cancer samples with over 76 million segmented nuclei (HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis), plus STimage-1K4M's 4,293,195 image-spot pairs across 50 tissue types (STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics), I will compute per-spot EMT signatures and local incoming/outgoing communication flux, then measure whether hybrid-EMT locations carry elevated local flux in situ. If the coupling dissolves under spatial constraints — plausible, since dissociation can create false-positive communication partners — the whole proposal should be abandoned, and this experiment is designed to falsify it early.

The second package builds the perturbation model and ranks a fixed compound panel: the ten BJJP constituents (including resveratrol, wogonin, ursolic acid), the Qixue Huazheng flavonoids quercetin, kaempferol, and isorhamnetin that target ESR1 and RAF1 (Integrated serum metabolomics and network pharmacology reveal molecular mechanism of Qixue Huazheng formula on peritoneal fibrosis), morusin, curcumin (which raises E-cadherin and lowers N-cadherin in fibrotic lung fibroblasts; Curcumin ameliorates pulmonary fibrosis progression by inhibiting the fibrotic process of lung fibroblasts), and the single-target SNAIL disruptor GN-25 as a control arm (Systems analysis reveals a transcriptional reversal of the mesenchymal phenotype induced by SNAIL-inhibitor GN-25). The key mechanistic prediction is that multi-component formulations score high because their constituents hit the ligand on the sending cell and the receptor on the receiving cell of the same paracrine loop — the complementary two-sided blockade that no existing method models — while GN-25, which acts only downstream inside each cell, leaves residual paracrine flux that the network can reroute around.

The third package adds a histology-only readout. I will extend a whole-slide image-to-gene-expression model of the STFlow type — which avoids the quadratic memory cost of global attention using local spatial attention and already achieves an 18% average relative improvement over the UNI and Gigapath pathology foundation models on the HEST-Benchmark (Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching) — with a communication-flux head, so that EMT-weighted flux can be predicted from H&E alone on held-out slides. This makes the score deployable on routine clinical sections, where spatial transcriptomics is rarely available.

Baselines

The naive reference point is the zero-perturbation expectation: untreated tissue should show no communication collapse beyond natural slide-to-slide drift. I will measure that drift band across untreated HEST-1k replicate slides of the same tumor type, and any predicted treatment effect must exceed it to count as signal at all. The strongest existing competitor is the standard network-pharmacology ranking used in the morusin study — docking affinities plus degree and betweenness centrality on a protein-protein interaction network, which there yielded 84 overlapping targets out of 281 compound targets and 1,741 disease genes (Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection). I will run that exact pipeline on my compound panel with the same affinity priors and the same compute budget, so the only difference is that my score is computed on spatial intercellular edges rather than a well-mixed protein graph. Beating it means a materially higher rank correlation with measured in situ EMT reversal (spatial E-cadherin/Vimentin shifts and, in the retrospective arm, metastasis outcomes), with a pre-registered Spearman threshold of 0.5 versus the baseline's correlation. A third competitor is the Bocci-style non-spatial analysis: CellChat-style communication propensity on dissociated single-cell data from the same tissues, which ignores distance and competitive binding; it is beaten if spatially constrained predictions track measured flux from paired spatial transcriptomics while the dissociated version does not. Finally, an ablation isolates the proposal's key ingredient by removing the communication-perturbation coupling and keeping only the image-to-gene model — effectively STFlow plus EMT signatures, no ligand-receptor layer. On the gene-prediction subtask the reference numbers are STFlow's reported 18% average relative improvement over UNI and Gigapath, and the finding that H-Optimus-0, a 1.1-billion-parameter vision transformer, achieved the best average performance among foundation models with a logarithmic parameter-count-to-performance correlation of R=0.81 (HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis); my added layer must not degrade gene prediction below that level while substantially improving flux prediction, for which no reported result exists on Lacuna for numerical transport error in COMMOT, so this work establishes the first reference values.

Why not just…?

Why not just run bulk RNA sequencing after drug treatment? It is cheaper and standardized, and it does capture canonical marker shifts. But it averages over cell types: the morusin study found its hub targets upregulated specifically in fibroblasts and immune subsets (Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection), and bulk homogenization would dilute exactly the sender-receiver structure my score depends on. More decisively, bulk expression cannot distinguish a compound that silences a fibroblast-to-tumor Wnt5a loop from one that merely lowers global Wnt5a transcript while the loop persists; the first experiment in the program, correlating flux with marker reversal, will show whether that distinction carries outcome information bulk measures miss.

Why not just run CellChat on dissociated single-cell data, as Bocci et al. did? That is precisely the third baseline, and it is a real threat because it is established and simple. The file's own evidence says it falls short twice: pair-at-a-time scoring overestimates interactions when over 70% of ligands and 60% of receptors have multiple partners, and dropping spatial distance creates false-positive communication between cells that never sit within signaling range (Screening cell–cell communication in spatial transcriptomics via collective optimal transport). Still, if the non-spatial score correlates with in situ flux collapse as well as mine does, the spatial machinery is not earning its cost — and that head-to-head is built into the evaluation rather than argued around.

Why not just fine-tune a pathology foundation model to classify responders from H&E directly? This would skip all the molecular modeling. But Chen et al. document severe cross-study batch effects for exactly this approach (STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics), and a black-box responder classifier gives no mechanism: it cannot say which ligand-receptor pairs to hit next, cannot rank untested compounds, and was outperformed by spatially aware whole-slide modeling by 18% on average (Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching). I will nevertheless report a fine-tuned foundation-model classifier as a reference on the responder-prediction task, because if it matches my pipeline's accuracy with none of the machinery, the mechanistic story is a luxury, not a necessity.

Decisive Checks and Retrospective Validation

The first decisive check is the spatial confirmation of the EMT-communication coupling on HEST-1k cancer samples; failure there kills the project cheaply. The second is a retrospective test against the Biejiajian Pill study: given only its ten identified constituents and their targets, the score must predict (a) strong convergence of the formulation on paracrine edges feeding hybrid-EMT cells, consistent with its observed suppression of lung metastasis from 80% to 20%, and (b) partial, bypass-prone reversal for GN-25, visible as residual post-perturbation flux rerouting through parallel routes (Biejiajian Pill Inhibits Carcinogenesis and Metastasis via the Akt/GSK-3β/Snail Signaling Pathway in Hepatocellular Carcinoma). The third check is generalization: the histology-only flux head must predict measured spatial communication on held-out HEST-1k slides across organs, quantified by Spearman correlation between predicted and measured EMT-weighted flux, reported per tissue type. Because no pre- and post-treatment paired spatial-transcriptomics-plus-histology dataset for EMT exists in the retrieved literature, the final package generates one: a small BJJP-versus-vehicle xenograft arm with paired spatial transcriptomics and H&E at treatment onset and endpoint, testing whether communication collapse, predicted from baseline slides, precedes and predicts morphometric reversal — cell elongation, nuclear aspect ratio, and collagen area measured with the quantitative histomorphometry used in fibrosis studies (All-trans retinoic acid regulating angiopoietins-1 and alleviating extracellular matrix accumulation in interstitial fibrosis rats). If flux collapse at baseline predicts which animals later histologically reverse, the communication-centric view of EMT inhibition earns its place; if marker shifts alone predict equally well, the simpler cell-autonomous picture stands.