Epithelial-to-mesenchymal transition (EMT) — the process by which adherent epithelial cells loosen junctions and become migratory — is no longer viewed as a binary switch but as a spectrum of stable hybrid epithelial/mesenchymal states. These intermediate states are not passive waypoints: across melanoma cohorts, circulating tumor cells, and breast cancer models, hybrid E/M cells combine invasive and cohesive survival traits and concentrate tumor-initiating capacity (Intermediate cell states in epithelial-to-mesenchymal transition; Hybrid Stem Cell States: Insights Into the Relationship Between Mammary Development and Breast Cancer Using Single-Cell Transcriptomics). Crucially, they are also the tissue's heaviest signalers: single-cell analyses of four independent cancer systems show that cells in intermediate EMT states send and receive more intercellular signals than either endpoint state (Single-Cell RNA-Seq Analysis Reveals the Acquisition of Cancer Stem Cell Traits and Increase of Cell–Cell Signaling during EMT Progression).
Wireless network theory has a mature and uncomfortable result about exactly this situation: when many uncoordinated transmitters share a physical medium, adding transmitters eventually destroys the network. Because every new transmitter adds interference to every existing link, aggregate throughput peaks at a finite density and then collapses — the waterfilling game-theoretic analysis found total network rate peaked at approximately 14 transmitters for 10 receivers under low noise (The Waterfilling Game-Theoretical Framework for Distributed Wireless Network Information Flow), and percolation analysis proved that network connectivity survives only inside a bounded density interval $[\lambda_{\min}, \lambda_{\max}]$ (Percolation and Connectivity on the Signal to Interference Ratio Graph). No intercellular signaling model in computational biology has this property. CellChat (Inference and analysis of cell-cell communication using CellChat), COMMOT, and HoloNet all assume communication scales monotonically or merely saturates with signaling-cell density. This proposal tests whether dense hybrid-cell niches in tumors are interference-limited communication systems — and whether self-interference, not just intracellular circuitry, is what stabilizes the intermediate states that dominate metastasis.
Core Claim
The claim is that paracrine signaling among hybrid E/M cells operates in an interference-limited regime. Three specific, falsifiable predictions follow. First, the functional signaling graph of a tissue — links that actually change downstream gene expression — percolates only within a bounded density of signaling-competent hybrid cells; below $\lambda_{\min}$ cells are too isolated to coordinate, and above $\lambda_{\max}$ promiscuous ligand crosstalk drowns cognate signals, fragmenting the network into isolated signaling islands. Second, this interference creates a biochemical analog of the interference floor documented in mixed RF relaying, where boosting transmit power produced zero further reliability improvement (Multi-User Mixed FSO-RF Systems With Aperture Selection Under Poisson Field Interference): upregulating ligand expression in an already-dense niche should plateau target-gene response rather than increase it. Third, the interference floor acts on EMT dynamics: by injecting a persistent, spatially structured noise term into the saddle-node bifurcation that governs EMT commitment (Tipping points in epithelial-mesenchymal lineages from single-cell transcriptomics data), dense paracrine fields should delay or prevent commitment, trapping cells in the high-entropy intermediate state that trajectory analysis already identifies as an obligatory routing hub (Spatial transition tensor of single cells). If correct, this supplies a density-dependent mechanism for why the EMT spectrum persists in crowded tumors rather than resolving to clean epithelial or mesenchymal endpoints.
The Bio-SIR Model
The proposal builds a signal-to-interference formalism for cell-cell communication. For a receiver cell $j$ and a candidate sender $i$, the received signaling power is $P_{ij} = \ell_i \, r_j \, g(d_{ij})$, where $\ell_i$ is sender ligand output, $r_j$ is cognate receptor abundance at the receiver, and $g$ is a path-loss kernel over distance $d_{ij}$. The key quantity is the ratio of wanted power to everything else arriving at the same receptors:
$$\mathrm{SIR}_{ij} = \frac{\ell_i \, r_j \, g(d_{ij})}{\sum_{k \neq i} \ell_k \, r_j \, g(d_{kj}) + N},$$
where the sum runs over all other ligand sources in the niche — including non-cognate ligands that share the receptor, which is the biological counterpart of non-orthogonal channels, and justified by the Fantom5 statistics that over 70% of ligands and 60% of receptors bind multiple partners (Screening cell–cell communication in spatial transcriptomics via collective optimal transport). $N$ is baseline expression noise. A functional signaling link exists only when $\mathrm{SIR}_{ij}$ exceeds a threshold $T$, exactly the edge rule of the SIR-graph percolation model (Percolation and Connectivity on the Signal to Interference Ratio Graph). Because every added sender appears in the denominator of every other pair's ratio, the model has infinite-range dependencies and is structurally non-monotonic in density — the property that distinguishes it from every existing biological communication model.
Two refinements come directly from the wireless literature. The path-loss kernel need not be isotropic: in obstacle-rich environments, distance-based association catastrophically misidentifies effective links while power-based thresholding achieved a gain factor up to 7.83 (Downlink Small-cell Base Station Cooperation Strategy in Fractal Small-cell Networks); tumor extracellular matrix is precisely such an anisotropic, barrier-rich medium, so $g$ will be fit per-tissue rather than fixed as a Euclidean Gaussian. And the percolation result that $n$ nodes in a shared medium need orthogonal resources scaling as $O(\log n)$ to stay connected (Percolation and Connectivity on the Signal to Interference Ratio Graph) suggests a reading of the colorectal finding that partial EMT decomposes into three distinct spatial programs — Inflammatory, Intermediate, and Mesenchymal pEMT — each using different paracrine axes (Single-cell and spatial transcriptomic analysis reveals tumor cell heterogeneity and underlying molecular program in colorectal cancer): program diversification may be the tissue's version of channel splitting, and this will be tested rather than assumed.
Research Program
The work proceeds in four parts. First, implementation: the bio-SIR graph will be computed on spatial transcriptomic data using the CellChat database of 2,021 curated interactions (Inference and analysis of cell-cell communication using CellChat) for ligand-receptor pairs, with per-tissue path-loss kernels fit by holding out known functional edges. Second, numerical percolation studies: synthetic point-process tissues with tunable hybrid-cell density will be used to locate the equivalent of $\lambda_{\min}$ and $\lambda_{\max}$, mirroring the lattice-discretization approach of the SIR-graph paper, and testing whether the interference alignment intuition from femtocell networks (Interference Alignment in a Poisson Field of MIMO Femtocells) has a biological counterpart in cells segregating ligand outputs to reduce shared-channel contention. Third, dynamics: the tipping-point ODE framework of RNA velocity coupled to local Jacobian spectra (Tipping points in epithelial-mesenchymal lineages from single-cell transcriptomics data) will be extended with a paracrine interference term drawn from the local bio-SIR field, testing whether interference shifts the timing of the bifurcation or stabilizes the saddle. Fourth, target validation: predicted functional edges will be checked against downstream gene-expression change, using the HoloNet decomposition of observed expression into a cell-type baseline plus a communication-induced change (Decoding functional cell–cell communication events by multi-view graph learning on spatial transcriptomics) as the readout of whether a link actually did something.
Proposed Experiments
The MCF10A 2D colony system is the primary testbed because it carries a natural density and EMT gradient in one preparation: outer colony cells are more mesenchymal with altered communication while interior cells stay epithelial (Single-Cell RNA-Seq Analysis Reveals the Acquisition of Cancer Stem Cell Traits and Increase of Cell–Cell Signaling during EMT Progression), and the line natively sits in the $I_1$ intermediate state with bidirectional potential (An Ovol2-Zeb1 Mutual Inhibitory Circuit Governs Bidirectional and Multi-step Transition between Epithelial and Mesenchymal States). The decisive experiment is a density manipulation — plating the same cells at graded initial densities — asking whether measured functional signaling throughput (communication-induced expression change, aggregated over the colony) is non-monotonic in hybrid-cell density. The OVCA420 TGF-$\beta$ time course serves the dynamics question; the largest Jacobian eigenvalue peaks around day 3 post-induction in that data (Tipping points in epithelial-mesenchymal lineages from single-cell transcriptomics data), and the test is whether an interference term computed from concurrent cell-cell signaling explains additional variance in when individual cells cross the tipping point beyond what the cell-autonomous Jacobian model captures. The colorectal Visium HD samples provide the orthogonalization test, checking whether the three pEMT programs occupy sufficiently distinct receptor-channel space that crosstalk between them stays below threshold, and whether the Intermediate pEMT program's TGFB1-to-fibroblast axis (Single-cell and spatial transcriptomic analysis reveals tumor cell heterogeneity and underlying molecular program in colorectal cancer) is recovered by the bio-SIR graph at spatial resolutions the mass-action model blurs; the constraint is honest — only three spatial samples exist, so this is a validation site, not a discovery site. Finally, to verify that adding the interference field does not corrupt transition inference, the extended spatial random walk will be checked against the synthetic ground-truth-velocity benchmark used by the spatial transition tensor work, which reported higher cosine similarity to true velocities than scVelo and CellDancer, though no reported result found on Lacuna for the numerical similarity value.
For clinical relevance, any density-derived risk feature extracted from the bio-SIR graph will be benchmarked against the colorectal 15-gene survival signature, which achieved an AUC greater than 0.8 for 1-to-10-year survival prediction (Single-cell and spatial transcriptomic analysis reveals tumor cell heterogeneity and underlying molecular program in colorectal cancer).
Baselines
The "do nothing" reference point is standard mass-action communication scoring: CellChat's ligand-receptor interaction strengths computed per cell pair, which treats each link independently and predicts communication grows monotonically with the number and expression of signaling cells, bounded only by receptor saturation (Inference and analysis of cell-cell communication using CellChat). This is the right null because it is what nearly every published EMT signaling analysis actually computes, including the study that motivates this proposal (Single-Cell RNA-Seq Analysis Reveals the Acquisition of Cancer Stem Cell Traits and Increase of Cell–Cell Signaling during EMT Progression); beating it means demonstrating a regime where measured functional signaling declines with added signaling cells, which mass-action cannot predict in principle. It will be run on the identical datasets, ligand-receptor database, and expression inputs, with its own threshold parameter tuned on the same held-out functional edges.
The strongest existing competitor is COMMOT's collective optimal transport (Screening cell–cell communication in spatial transcriptomics via collective optimal transport), because it is the only prior method that models competition — multiple ligands tug-of-war over shared receptors — with explicit distance costs and mass preservation, making it the closest formal relative of the bio-SIR model. It will be run on the same spatial data with the same database and given an equal tuning budget over its transport-cost parameters; no reported result found on Lacuna for a numerical accuracy figure from COMMOT on the EMT datasets used here, so the comparison will be established de novo on the shared functional-edge prediction task. Beating COMMOT means predicting which communication-induced expression changes actually occur better than the transport solution does — specifically in the dense-niche regimes where allocation efficiency (COMMOT's strength) and interference collapse (its blind spot) diverge.
The key ablation removes exactly one ingredient: the non-cognate terms in the SIR denominator. A "cognate-only" variant computes the same ratio but sums interference only from ligands of the same pathway family, keeping the threshold rule, the path-loss kernel, and everything else identical. If the full model outperforms this ablation specifically at high hybrid-cell density, the cross-talk interference term — the element absent from every prior method — is doing the work. A second ablation replaces the SIR edge rule with a Boolean distance rule (link if within radius), matching the connectivity construction in the spatial transition tensor framework (Spatial transition tensor of single cells), to isolate the contribution of power-thresholded rather than distance-thresholded association.
Why not just…?
Why not just use Michaelis-Menten or Hill saturation to explain density plateaus? Receptor saturation and ligand depletion are the standard biochemical answer, and they are simpler. But they predict a plateau, not a collapse: saturating kinetics bound throughput from above while never driving total functional signaling down as more signaling cells are added. The wireless results are explicit that the decline comes from interference power entering the denominator, not transmitter saturation — throughput peaked at roughly 14 transmitters and then fell (The Waterfilling Game-Theoretical Framework for Distributed Wireless Network Information Flow). The MCF10A density experiment is designed precisely to separate these: if functional throughput merely plateaus, the simple view wins and this proposal is wrong; if it declines, no saturating kinetics explains it.
Why not just use COMMOT, which already handles ligand competition? COMMOT is a genuine threat and the closest prior work; it resolves the tug-of-war of promiscuous ligands as a global transport-optimization problem. But it treats the tissue as a coordinated allocator — one optimizer assigning signals efficiently — whereas the wireless results show the failure mode of interest arises when transmitters are uncoordinated and each pair's link quality is degraded by everyone else, with a "Price of Anarchy" that widens with density (The Waterfilling Game-Theoretical Framework for Distributed Wireless Network Information Flow). COMMOT contains no threshold, no SIR, and no mechanism by which adding signaling cells breaks existing links, so it cannot produce or predict the non-monotonic regime. The head-to-head on shared data will settle whether that difference matters empirically, and I will report it either way.
Why not just add distance-decay weights to existing spatial models? This is what current spatial methods already do — Gaussian kernels in HoloNet and STT's spatially constrained random walk (Spatial transition tensor of single cells). But the fractal small-cell result shows distance is actively deceptive in anisotropic, obstacle-rich media: distance-based association degraded performance as density grew, while power-based thresholding gained up to a factor of 7.83 (Downlink Small-cell Base Station Cooperation Strategy in Fractal Small-cell Networks). Tumor stroma is that kind of medium, so the Euclidean-kernel approach risks misattributing which cells actually communicate. The Boolean-versus-SIR ablation above directly measures how much this matters for EMT data.
Decisive Checks
Three outcomes would decide the proposal's fate. The non-monotonicity test — functional signaling throughput declining at high hybrid-cell density in manipulated MCF10A colonies — is the single most important result, and it is one the mass-action null and COMMOT both predict will not occur. The interference-floor test asks whether experimentally boosting ligand output (mimicking, for example, the elevated signaling of intermediate states) in an already-dense niche yields zero additional target-gene response, the direct analog of the outage floor where power scaling produced no reliability gain (Multi-User Mixed FSO-RF Systems With Aperture Selection Under Poisson Field Interference). And the trapping test asks whether the interference-augmented tipping model explains the persistent occupancy of intermediate states in the OVCA420 time course better than the cell-autonomous Jacobian model alone. If the first test yields a plateau rather than a decline, the honest conclusion is that tumor signaling is saturation-limited, not interference-limited, and the density-stabilization mechanism for the EMT spectrum would need to be sought elsewhere.
References
- Intermediate cell states in epithelial-to-mesenchymal transition
- Hybrid Stem Cell States: Insights Into the Relationship Between Mammary Development and Breast Cancer Using Single-Cell Transcriptomics
- Single-Cell RNA-Seq Analysis Reveals the Acquisition of Cancer Stem Cell Traits and Increase of Cell–Cell Signaling during EMT Progression
- The Waterfilling Game-Theoretical Framework for Distributed Wireless Network Information Flow
- Percolation and connectivity on the signal to interference ratio graph
- Inference and analysis of cell-cell communication using CellChat
- Multi-User Mixed FSO-RF Systems With Aperture Selection Under Poisson Field Interference
- Tipping points in epithelial-mesenchymal lineages from single-cell transcriptomics data
- Spatial transition tensor of single cells
- Screening cell–cell communication in spatial transcriptomics via collective optimal transport
- Downlink Small-Cell Base Station Cooperation Strategy in Fractal Small-Cell Networks
- Single-cell and spatial transcriptomic analysis reveals tumor cell heterogeneity and underlying molecular program in colorectal cancer
- Interference Alignment in a Poisson Field of MIMO Femtocells
- Decoding functional cell-cell communication events by multi-view graph learning on spatial transcriptomics
- An Ovol2-Zeb1 Mutual Inhibitory Circuit Governs Bidirectional and Multi-step Transition between Epithelial and Mesenchymal States