Enzyme-Screened, Saturating Signaling Kernels for Spatial Cell–Cell Communication

Cells in a tissue talk by secreting signaling molecules (ligands) that drift through the space between cells and bind receptors on neighbors. Tools that infer this "cell–cell communication" from spatial transcriptomics, which measures gene expression while keeping each cell's position, mostly score a sender–receiver pair by how much ligand and receptor messenger RNA (mRNA) each expresses and how far apart they are. The best current evidence says signal strength should fall off as the reciprocal of distance, $1/d$. In engineered molecular communication, the field that designs signaling between nano-machines over diffusing molecules, this form is a familiar special case. It is what diffusion from a continuously secreting source gives you when nothing in the medium destroys the molecule. Engineers also know two ways that picture breaks. Degrading enzymes near the receiver shorten how long molecules linger. Receptors saturate and stay bound, so molecules left over from earlier pulses, called inter-symbol interference, crowd out new ones.

We propose to bring these communication-channel effects into spatial cell–cell communication inference as one physically derived kernel. A ligand's effective reach shrinks wherever neighboring cells express enzymes that degrade it. The receiver's response saturates instead of growing linearly. The timing structure of secretion, meaning how coherently neighboring senders pulse relative to how fast the space clears, decides whether leftover molecules smooth the signal into a steady background or leave it spiky. The model reduces exactly to the $1/d$ kernel when degradation is absent. It makes a falsifiable prediction: gains over $1/d$ should be concentrated in tissue neighborhoods rich in ligand-degrading enzymes, and should vanish when cell coordinates are scrambled to mimic the spatial context lost when tissue is dissociated into single cells.

Core Claim

stMLnet links ligand–receptor pairs to downstream target genes through a distance model grounded in diffusion and mass action. It found that a reciprocal-distance kernel predicts target-gene activation better than constant or radial-basis kernels. We read that result as the zero-degradation limit of the steady-state diffusion field. With a local first-order degradation rate $k$, a sender $i$ secreting at rate $s_i$ produces at receiver $j$ the concentration

$$\bar c_j = \sum_i \frac{s_i}{4\pi D\, d_{ij}} \exp\!\left(-d_{ij}/\lambda_{ij}\right), \qquad \lambda_{ij} = \sqrt{D/k_{ij}},$$

where $D$ is the ligand's diffusion coefficient, $d_{ij}$ is the sender–receiver distance, and $\lambda_{ij}$ is a screening length. The degradation rate $k_{ij} = k_0 + \beta \sum_{m} e_m w(d_{im}, d_{mj})$ rises with the expression $e_m$ of a curated degrading-enzyme gene set in cells $m$ lying along or near the path. Here $w$ is a fixed geometric weight and $\beta$ and $k_0$ are fitted. This is the tissue analogue of the result in Effective inter-symbol interference mitigation with a limited amount of enzymes, which showed in idealized 3D simulations that where a fixed supply of enzymes sits, especially clustered around the receiver, strongly changes how much residual signal arrives.

The receiver's response is taken as the time-averaged occupancy of a saturating receptor, following the receptor-clogging Markov model of Capacity of Diffusion-based Molecular Communication with Ligand Receptors:

$$r_j \approx \frac{\bar c_j}{\bar c_j + K} - \frac{K\,\sigma_j^2}{(\bar c_j + K)^3}.$$

$K$ is an effective dissociation constant, set per receptor and scaled by receptor mRNA. $\sigma_j^2$ is the temporal variance of concentration at $j$. The second term is the second-order Jensen correction: a saturating receptor exposed to a fluctuating field responds less than it would to the mean field.

Temporal synchronization enters through $\sigma_j^2$. If senders pulse incoherently, or pulse faster than the local clearance time $\lambda^2/D$, their overlapping leftover tails, which are the inter-symbol interference, smooth the field and $\sigma_j^2$ is small. If senders pulse coherently and the space clears quickly, which is exactly the enzyme-rich case, the field is spiky and the response drops below what the mRNA levels suggest. This is the tissue counterpart of the "power paradox" in Capacity of Discrete Molecular Diffusion Channels: more output does not monotonically mean more transmitted signal. We parameterize $\sigma_j^2$ with a single per-pathway coherence parameter $\rho \in [0,1]$ and the ratio of pulse period to clearance time.

What is new is the combination of three physical ingredients, each motivated by molecular-communication results, inside cell–cell communication inference: spatially varying screening driven by enzyme expression, receptor saturation, and the effect of sender timing on the time-averaged field. It also comes with a stratified test that can show whether the physics explains the gains or whether extra flexibility does.

Research Program

The work has three parts.

The first is a ground-truth simulator. It extends the 3D Brownian, enzyme-kinetics setup of Cho et al. and the clustered-node geometry of Fundamentals of Clustered Molecular Nanonetworks to tissue-like layouts taken from real spatial coordinates. Senders secrete pulse trains with controllable timing jitter and coherence. Enzyme-secreting cells act as local sinks. Receivers carry finite receptor pools with binding and release kinetics, and downstream "target genes" read out time-averaged occupancy plus noise. A final snapshot gives the static data an inference method would see.

The Clustered Nanonetworks paper assumes perfect synchronization in free space. The synchronization-error models in Variance-constrained capacity of the molecular timing channel with synchronization error and On the Impact of Time-Synchronization in Molecular Timing Channels explicitly leave out multi-symbol interference. The simulator's main conceptual contribution is to model jitter, leftover-molecule tails and spatially heterogeneous sinks together. Its practical use is to tell us when the kernel's parameters can be recovered from a single snapshot at all.

The second part is the inference model itself. It is a drop-in replacement for the distance kernel inside the stMLnet ligand → receptor → transcription factor → target pipeline, fitted by maximum likelihood on held-out target-gene expression. Competition among ligands for shared receptors, which COMMOT shows affects over 70% of Fantom5 ligands and over 60% of receptors, is kept by summing all competing ligand fields before the saturating step.

The third part is an identifiability analysis of the timing term. A static snapshot confounds diffusion-driven decay with changes in transcription over time. We therefore treat $\rho$ as the claim most at risk, and we report honestly whether it can be identified from snapshots at all.

Baselines

The naive reference is an expression-only score: the product of ligand and receptor mRNA, with neighbors chosen by a fixed radius and no distance weighting. We pair it with a coordinate-shuffled version of every method, which permutes cell positions within each cell type. This is our in-silico stand-in for tissue dissociation. It keeps the cell-type composition that single-cell RNA sequencing would capture and destroys the spatial arrangement. The gap between real and shuffled coordinates measures how much predictive power spatial context carries at all. Every proposed gain must exceed that gap's noise.

The strongest existing method is stMLnet with its $1/d$ kernel. It is the most physically grounded competitor, it targets the same task of predicting downstream target genes, and our model contains it as the special case $\beta = k_0 = 0$, $K \to \infty$, $\rho = 0$. stMLnet reports that it significantly outperformed NicheNet, CytoTalk and MISTy at target-gene prediction on its Visium datasets, but no reported result found on Lacuna for its numerical AUROC or error on Visium, seqFISH+, MERFISH, Slide-seq, or its 100 synthetic distance-weighting sets. We will therefore rerun it ourselves. The only adjacent number in the file is from CellChat: modeling heteromeric complexes reduced the false-positive rate by roughly 10–14%. We use this as a rough scale for what a meaningful structural improvement looks like in this literature.

COMMOT is run as the strongest competitor that handles competition, and FlowSig as a directional-flow comparison where target genes can be matched.

A flexible learned baseline is essential. It is a graph neural network in the spirit of STELLAR, which learns from a cell-proximity graph and reached 93% annotation accuracy on Barrett's esophagus CODEX, 11% above expression alone. We repurpose it to predict target genes from the same neighbor inputs, including enzyme gene expression.

Ablations isolate each ingredient: a single global screening length with no enzyme term; enzyme screening without saturation; saturation without screening; the full model without the timing term; and, most importantly, the full model with the enzyme gene set swapped for a random expression-matched gene set.

All methods share the same ligand–receptor and transcription-factor databases, the same target genes, and the same spatially blocked train/test splits, so neighboring cells cannot leak across folds. Each gets an equal hyperparameter search budget. Beating stMLnet means a consistent improvement in held-out target-gene AUROC across datasets, larger than its variation across splits. In addition, that improvement must be significantly larger in the top enzyme-expression quartile of neighborhoods than in the bottom quartile, and it must vanish under the random-gene-set ablation.

Why not just…?

Why not just learn a flexible distance kernel or a graph neural network instead of deriving one? This is the most serious threat. A graph network given neighbor enzyme expression could in principle learn context-dependent decay. The proposal answers it empirically rather than by argument. If the learned model matches the physical kernel on held-out tissues, the physics adds interpretability but no accuracy, and we will say so. Where we expect the physical kernel to win is transfer: fitting on one tissue (Visium breast cancer) and testing on another (glioma, COVID-19 lung) with few parameters. We also expect it to win on the synthetic benchmark, where the true kernel is known and the graph network's fitted map can be checked against it.

Why not just add exponential decay with one global length scale to stMLnet? That is our first ablation. If a single length scale captures all the gain, the enzyme-dependence claim is false. The stratified test separates the two hypotheses. A global decay length predicts gains spread uniformly across neighborhoods. Enzyme screening predicts gains concentrated where degrading enzymes are expressed.

Why not just put a Hill-type saturation on COMMOT or CellChat scores? Saturation alone probably captures part of the effect. The receptor-clogging analysis already implies that response is bounded. But saturation without screening cannot explain why the same ligand reaches farther in some niches than in others, and the factorial ablation measures each ingredient's share. We will report it even if saturation dominates.

Why not just drop the timing term, since cells have no shared clock and snapshots cannot see dynamics? This may be right. The timing-channel literature shows synchronization strongly affects pulsed signaling, but tissue snapshots may simply not contain this information. The simulator tells us in advance whether $\rho$ can be recovered from a snapshot at realistic noise levels. If it cannot, we fix $\rho = 0$ and report the negative result, which still tells the field that inferring pulse coherence needs time-resolved or perturbation data.

Proposed Experiments

On the simulator, we vary enzyme density, pulse coherence, receptor count and timing jitter. We measure how well each method recovers the true sender–receiver influence matrix and the true screening lengths, and at what noise level $\rho$ stops being identifiable.

On real data, we use the stMLnet benchmark datasets: Visium breast cancer, glioma and COVID-19 lung; seqFISH+ mouse cortex; Slide-seq v2 hippocampus; and MERFISH hypothalamus. The primary metric is held-out target-gene AUROC under spatially blocked splits. The decisive check is the enzyme-quartile stratification combined with the random-gene-set ablation and the coordinate-shuffle control. Any genuine physical effect must disappear when the spatial arrangement is destroyed.

A further check uses the eight breast-cancer perturbation studies (TGFB1/IL6 stimulation, NRP1 knockout, CXCR4 mutation) from stMLnet. They test whether pathways whose fitted screening lengths are short respond more weakly to stimulation at a distance than the $1/d$ model predicts.

The result that would sink the proposal is gains that are uniform across enzyme quartiles and survive the random-gene-set swap. That pattern would mean the extra parameters bought flexibility, not physics.