When a spatial transcriptomics array pools the mRNA of up to 30 cells into one measurement spot, a rare cell type does not disappear. It leaves a fingerprint in how spots differ from their neighbors. If a minority population sits only in one tissue layer and is scattered sparsely within it, the counts of its marker genes will jump from spot to spot more than counting noise allows, and only inside that layer. Neuroscience has a name for this pattern: cross-frequency coupling, where the phase of a slow, broad oscillation controls the amplitude of fast, local bursts. We propose to treat spot averaging as a low-pass filter over cells, and to detect and quantify the populations it hides by measuring how coarse tissue structure gates the fine-scale, above-noise variability of gene counts. This needs no single-cell reference.
The opportunity is practical. The largest benchmark of spatial deconvolution, Li et al., which compared 18 methods on 50 datasets, found that reference-based methods degrade when the single-cell reference and the spatial assay come from different platforms. Reference-free methods, in turn, struggle to separate subtypes. Current spatial methods either smooth over neighbors or throw away high spatial frequencies as noise. GraphST does single-scale graph smoothing. waveST shrinks high-frequency wavelet coefficients independently of the coarse bands. In both, the band where masked minorities live is the band being discarded.
Core Claim
Our claim has three parts. First, spot-to-spot fluctuation in excess of counting noise carries recoverable information about cell populations inside a spot. Second, when that excess fluctuation is modulated by coarse tissue position, the modulation is a specific signature of a spatially confined minority. Third, a simple moment identity lets us separate the minority's density from its per-cell marker expression, which spot means alone confound.
Suppose rare cells land in each spot roughly as a Poisson process with local density $\lambda$, and each carries a marker whose counts are Poisson with mean $m$. The spot count $x$ is then compound Poisson, with mean $\lambda m$ and variance $\lambda m + \lambda m^2$. It follows that
$$\frac{\mathrm{Var}(x) - \mathbb{E}(x)}{\mathbb{E}(x)} = m,$$
so the normalized excess variance recovers the per-cell load $m$. Dividing the local mean by $m$ then gives the density $\lambda$. Estimated locally and read out as a function of coarse tissue coordinate, these two quantities form a reference-free map of where a hidden population lives and how dense it is.
This is new relative to the work in the file. Beyond Power Spectra showed, for image diffusion and flow models, that per-frequency power can look identical across systems whose cross-frequency interactions differ. That result motivates our view that marginal summaries hide structure, but it was only tested on continuous pixel grids and used purely as a diagnostic. We move that lens to integer counts on irregular tissue graphs and use it for inference.
Research Program
Representation on the native spot graph. We build a spot-adjacency graph like GraphST's and take the low-frequency eigenvectors of its graph Laplacian as the coarse tissue coordinate. This replaces the "slow phase" of neural cross-frequency coupling. It avoids waveST's regridding, which averages spots and loses resolution. In laminar tissues such as cortex, the first nontrivial eigenvector typically tracks depth, which gives a one-dimensional phase-like axis. In other tissues we use a two- or three-dimensional coarse embedding and measure dependence with mutual information instead of phase bins.
Count-exact residuals. For each gene we fit a low-pass mean $\mu_s$ at each spot $s$ from library-size-scaled counts, restricted to the coarse eigenvectors. We then form Pearson residuals $r_s = (x_s - \mu_s)/\sqrt{\mu_s}$. Under pure Poisson sampling these have unit variance whatever $\mu_s$ is. This step matters because the obvious confound is that counting noise already grows with the mean. That would produce spurious "coupling" between tissue layer and fluctuation size, much as Untangling cross-frequency coupling in neuroscience warns that sharp waveforms create spurious harmonic coupling in electrophysiology.
The coupling statistic. The local amplitude $A_s$ is the neighborhood average of $r^2$ over the high-pass graph band. The coupling statistic is the dependence of $A_s$ on the coarse coordinate. Significance comes from two nulls. The first resamples integer counts from Poisson$(\mu_s)$ with library sizes preserved. This stays on the integer lattice, respecting Count Bridges' argument that continuous relaxations distort count geometry. The second applies spatial block shifts that keep the fine structure but break its alignment with coarse position.
An internal control for technical overdispersion. Technical sources such as capture efficiency and ambient RNA also inflate variance. We estimate a shared technical overdispersion floor from spatially uniform genes. These are the "context genes" that LGDiST found useful for stabilizing spatial completion. We subtract that floor before applying the moment identity. The technical floor should be common across genes, whereas minority-driven excess should scale with each marker's per-cell load. That difference is testable.
From genes to populations. Genes whose excess amplitude is gated by the same coarse region, and whose fine-scale residuals co-fluctuate across spots, are grouped into candidate hidden programs. For each group we report its $(m, \lambda)$ map.
Baselines
Naive reference points. Two naive references anchor the comparison. The first is plain spot-level mean expression of each marker, which is what a masked readout gives. The second is a global overdispersion test that ignores position. Beating these shows that the spatial gating, not overdispersion alone, carries the signal.
Strongest existing methods. The strongest competitors for recovering rare populations are the reference-based tools that Li et al. rank as overall top performers: CARD, Cell2location, and Tangram. We also include Count Bridges, which enforces exact integer sums during deconvolution, and the reference-free STDeconvolve that Count Bridges reports surpassing.
For numbers, Tangram reports that on MERFISH leave-one-out tests 75% of genes were predicted with spatial correlation above 40%. For per-spot proportion accuracy of CARD and Cell2location, no reported result was found on Lacuna. Li et al. give rankings without figures in our file, and Count Bridges reports only qualitative wins over CIBERSORTx, MuSiC and STDeconvolve.
All methods run on identical pseudo-spot data. Every method gets the same gene panel. Reference-based methods get the same held-out single-cell reference, in two versions: platform-matched and deliberately mismatched. We give each baseline the tuning budget its authors recommend and cap wall-clock per dataset equally.
Our method beats them if it recovers the planted minority's density map at least as well under the matched reference, clearly better under the mismatched or missing reference, and uses no reference at all. Recovery is scored by Spearman correlation with true $\lambda$ and by detection AUROC.
Ablations. Three ablations isolate the key ingredients. The first drops the coarse gating and uses only local excess variance. The second replaces Pearson residuals with raw residuals, which should produce false coupling. The third replaces graph eigenvectors with gridded wavelet bands as in waveST.
Why not just…?
Why not just run Cell2location or CARD and read off the rare cell type's proportion? When the reference contains the state and matches the platform, this may well win. We will report exactly that. The failure cases are the ones Li et al. document: platform mismatch, and states absent from the reference, which no reference-based alignment can infer. Our experiments deliberately degrade and delete reference entries to measure where the crossover occurs. If reference methods remain better even under mismatch, our contribution shrinks to a reference-free fallback. That is still useful, but we would state it as such.
Why not just test overdispersion separately within annotated tissue layers? This is the most serious threat. In DLPFC, with its manual layer labels, a layer-stratified dispersion test is nearly our statistic with discrete bins. We will run it as a baseline wherever labels exist. We expect parity there. Our method should help only where labels are absent, where minorities straddle label boundaries, or where the gating follows a continuous gradient rather than a layer. If parity holds everywhere, the honest conclusion is that the moment identity matters and the cross-frequency framing does not.
Why not just use a multi-scale graph network with cross-band attention, as Ada-FCN does for fMRI? Ada-FCN learns frequency bands and asymmetric cross-band attention for disease classification, reaching 79.68% accuracy on ADNI. A supervised network of that kind needs labels that do not exist for hidden populations. It also gives no interpretable split into density and per-cell load. We include a GraphST embedding followed by clustering as the learned-representation comparator. If its clusters isolate the planted minority as well as our maps do, that argues against our approach.
Why not just denoise with wavelets as waveST does? Its thresholding removes exactly the high-frequency band where the minority signature lives. The waveST ablation arm will measure how much signal that removal destroys.
Proposed Experiments
Controlled experiment on MERFISH. The primary testbed is MERFISH mouse motor and visual cortex, with 254 genes at single-cell resolution. There we know every cell's type and position. We bin cells into 55 µm pseudo-spots, matching Visium pooling. We treat each rare cell type as the hidden target and vary its abundance down to under one cell per spot.
Decisive check. The decisive test keeps each spot's rare-cell count distribution per layer fixed and permutes rare-cell positions within versus across layers. The first manipulation leaves marginal per-layer proportions unchanged and preserves gating. The second destroys gating while keeping global abundance. Per-spot proportion error and mean maps should barely distinguish the two, which mirrors the Mehrpanah finding that marginal spectra miss cross-frequency structure. Our coupling statistic should separate them sharply. If it does not, the core claim fails.
Moment identity. On the same data we test whether estimated $m$ matches true per-cell marker load, and whether $\lambda$ matches true density. We also check whether the context-gene floor correctly removes injected technical overdispersion of known size.
Evaluation of existing methods. Using the 50-dataset collection of Li et al., we ask whether ranking deconvolution methods by preserved coupling reorders them relative to their proportion-error rankings. A reordering would show that the standard metric is blind to how well methods preserve hidden minorities.
Real data. Finally, on the 12 DLPFC Visium slices, we check that detected gated programs do not simply reproduce layer markers. We also check that their $\lambda$ maps agree in location with sparse cell types mapped by Tangram from a matched reference. Here we use Tangram as a weak external check, not as ground truth.
References
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST
- Spatial Transcriptomics Dimensionality Reduction using Wavelet Bases
- Beyond Power Spectra: Cross-Frequency Interactions in Generative Dynamics
- Untangling cross-frequency coupling in neuroscience
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Latent Gene Diffusion for Spatial Transcriptomics Completion
- Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram
- Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based Brain Disorder Classification