Focusing on What the Reference Misses: Time-Reversal Detection of Unreferenced Rare Cells in Spatial Transcriptomics

Spatial transcriptomics platforms such as 10X Visium measure gene expression in small tissue "spots", but each spot holds several cells, so the readout is an average. Standard tools then "deconvolve" each spot, estimating what fraction of it comes from each cell type in a single-cell reference atlas. This works for common, well-catalogued cell types. It fails exactly where biology cares most. Rare stem-like or drug-resistant populations, and aberrant tumor states, are often absent from the reference, so the method has no name to assign their signal to. Their signal is either forced onto a neighbouring type or left in the unexplained remainder. We propose to recover these populations from that remainder, which we call the residual. The approach borrows an idea from wave physics. In time reversal, a strongly scattering medium is turned from a nuisance into a lens: recorded echoes are played backwards and pile up on a hidden scatterer. A known variant first suppresses the bright scatterers it already knows about, so that weak ones become visible.

The concrete bet has three parts. The residual matrix (spots by genes) left after reference deconvolution plays the role of the scattered field. The spatial maps of the known cell types play the role of the bright scatterers to be suppressed. Variation in which cells sit next to which across the tissue plays the role of the random medium. Wave theory says that medium randomness widens the effective aperture of a small time-reversal mirror. The biological analogue is a testable and somewhat counterintuitive prediction. Mixing inside a spot hides a rare population. Diversity across spots in what it co-occurs with makes it easier to find. A population that always sits in the same niche next to the same neighbour should be nearly undetectable from residuals, however abundant it is.

Core Claim

We claim that a time-reversal-style detector can find and spatially localize cell populations that are missing from the reference. The detector has three steps. It builds a spot-by-spot operator from count-whitened deconvolution residuals. It projects out the spatial directions already explained by known cell types. It then reads off outlying eigenvalues against a random-noise baseline. We further claim that its detection limit follows a law with an explicit dependence on how diverse the population's neighbourhoods are.

Let $\tilde E$ be the residual matrix after Pearson-type whitening under a Poisson or negative-binomial count model, with $S$ spots and $G$ genes. Let $P_\perp$ be the projector in spot space onto the complement of the known types' proportion maps. The operator is

$$W = P_\perp \tilde E \tilde E^\top P_\perp .$$

This is the analogue of the time-reversal operator $R R^\dagger$ used in optics, with already-focused scatterers removed. Each eigenvector of $W$ is a candidate "focal spot", meaning the spatial footprint of an unexplained population. Back-projecting it, $\tilde E^\top v$, gives that population's gene signature.

Suppose a rare type occupies $n$ spots at mean proportion $p$, with whitened signature energy $\|\Delta\|^2$ (growing with sequencing depth and with the number of informative markers). Let $\rho_\perp^2$ be the fraction of its footprint lying outside the span of the known types' maps. Spiked random-matrix theory gives a sharp threshold, from the phase transition at which a planted rank-one component separates from noise. Applying it here, the population is detectable roughly when

$$n\, p^2\, \|\Delta\|^2\, \rho_\perp^2 \;\gtrsim\; \sqrt{S\,G}.$$

The factor $\rho_\perp^2$ is where across-spot heterogeneity acts as an aperture. It is close to one when the population co-occurs with varied neighbours, and close to zero when it is locked to a single niche.

We are candid about the algebra. Without whitening, nulling and the diversity analysis, this reduces to principal component analysis of residuals. The contribution therefore does not rest on the eigen-decomposition being new. It rests on three things. The first is a count-aware construction under which the random-noise baseline is actually valid. The second is a spatial-nulling step taken from selective time-reversal focusing. The third, and most important, is a falsifiable detection-limit law plus a benchmark for unreferenced rare types, which current deconvolution benchmarks do not measure.

Relation to Prior Work

The wave-physics side supplies the method and the predictions. The standard three-stage time-reversal process is described in Time Reversal for 6G Spatiotemporal Focusing: probe the channel, reverse the recorded response, then pre-filter transmissions. We reuse only its logic of focusing by correlation with the measured response rather than by inverting the channel.

Two papers by Fouque, Garnier and Sølna motivate the diversity prediction. Time-Reversal Aperture Enhancement shows that randomness in the medium makes the refocused field fall off far faster than in a uniform medium. Time reversal super resolution in randomly layered media shows that multiply scattered coda waves can refocus onto buried passive scatterers. Our claim that varied neighbourhoods enlarge $\rho_\perp$ is the biological translation of this result. It could fail, and we test it directly. Statistical stability in time reversal supplies the expectation that a detector averaged over many decorrelated "frequencies" becomes stable across realizations of disorder. Here genes, grouped into weakly correlated modules, play the role of frequencies, and tissue sections or donors play the role of disorder realizations.

The closest existing work in tissue is CTR-CLASS, High-throughput volumetric adaptive optical imaging using compressed time-reversal matrix. It builds a time-reversal operator $W = R_s R_s^\dagger$ from a reflection matrix to correct optical aberrations in mouse brain. It is a physical imaging method that sees structure, not molecular identity, and it never touches expression counts. We keep its operator construction and move it to the residual of a statistical model. The learned detector in Simultaneous Localization and Recognition of Subwavelength Noncooperative Entities Based on SISO Time Reversal and Neural Networks recognizes and localizes small objects in a metal cavity, but only on pre-calibrated grids and for objects whose classes were seen in training. Our setting is the opposite: the target is by construction absent from the reference.

On the biology side, several papers motivate the gap. Rare stem-like populations that drive treatment resistance are invisible in bulk measurements according to A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers and Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance. Pro-tumour macrophage subsets surfaced only with single-cell profiling in Global immune characterization of HBV/HCV-related hepatocellular carcinoma.

Other papers constrain the design. POPPY, Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, aligns bulk profiles to single-cell atlases through the Cell Ontology. Its authors acknowledge that attribution fails when the atlas lacks a state; that failure mode is our target. Count Bridges enable Modeling and Deconvolving Transcriptomic Data shows that treating counts as continuous misspecifies the problem. This is why we whiten under a count model rather than using raw or log residuals; otherwise the noise baseline is wrong and false detections follow. The HESCAPE benchmark, A Large-Scale Benchmark of Cross-Modal Learning for Histology and Gene Expression in Spatial Transcriptomics, found that batch effects degrade expression prediction. We therefore treat cross-platform gene-level distortion as a first-class stress test rather than an afterthought.

Research Program

The first component is the detector itself. We run a strong reference deconvolution method, then compute count-model residuals. Genes are grouped into modules with low within-reference correlation, which serve as the "frequencies". We form $W$ with spatial nulling and compare its eigenvalues against the noise-only edge predicted for a matrix of the same shape. We validate that edge by permuting spots within each module.

Detection is followed by iterative time reversal. This means alternating between the spot footprint and the gene signature, with non-negativity on both and light spatial smoothing on the footprint. We stop when the strongest remaining component converges, add its footprint to the nulled set, and repeat until no eigenvalue exceeds the edge. The output is a set of footprints and signatures. We then cluster each signature against the held-out truth or annotate it by marker enrichment.

The second component is the detection-limit law. We fit the left-hand side of the threshold inequality on simulated data and check whether the empirical 50%-detection boundary collapses onto a single curve across spot count, depth, marker count and co-occurrence pattern. Deviations would tell us which ingredient of the analogy breaks. For example, a curve that depends on raw gene count rather than on decorrelated module count would mean the "frequency diversity" reading is wrong.

Baselines

The naive reference point is the deconvolution method's own output. If a rare type is missing from the reference, the only "detection" available is to flag spots with large residual norm. We implement that threshold directly, giving the spot-level score of a method that does nothing beyond standard deconvolution. Beating it shows that structure in the residual, and not just its size, carries the signal.

The strongest existing deconvolution methods are CARD, Cell2location and Tangram. A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics identified them as the top performers across 50 datasets. That benchmark reports them categorically, and the file has no numeric score for them, so for these methods no reported result was found on Lacuna for proportion accuracy, and none for rare-type detection either. We run each with its recommended settings on identical simulated spots and identical references. Each method serves both as the upstream model for our detector and as a competitor. As a competitor, it is given the most generous possible version of the task: the missing type is added to its reference under an "unknown" label built from the mean of all other types. We also run it with an oracle reference that contains the true held-out signature, as an upper bound.

Count Bridges, Count Bridges enable Modeling and Deconvolving Transcriptomic Data, is reported to outperform CIBERSORTx, MuSiC and STdeconvolve. The file gives no numbers, so no reported result was found on Lacuna for its spot deconvolution accuracy. If code is available, we include it as an upstream model.

Reference-free factorization is the other serious competitor, because it never needed the reference in the first place. We represent it by non-negative matrix factorization of the raw spots and by STdeconvolve, swept over the number of factors with the same tuning budget as our method's single hyperparameter, the smoothing strength.

Three ablations isolate the key ingredients. Plain principal component analysis on unwhitened log residuals removes both the count model and the nulling. Our detector without spot-space nulling, keeping only gene-space whitening, tests whether the "suppress bright scatterers" step matters. Our detector with nulling but with genes treated individually rather than as decorrelated modules tests the frequency-diversity reading.

Beating the baselines means two things. First, a higher area under the precision-recall curve for spot-level presence of the held-out type at abundances where the oracle-reference method still succeeds. Second, a lower abundance at which detection reaches 50% power, at a matched false-detection rate on null tissues containing no held-out type.

Why not just…?

Why not just run principal component analysis on the residuals? This is the real threat, because the core algebra is the same. Our view is that unwhitened residuals mix count noise of very different sizes across genes. That breaks the noise baseline, so outlying eigenvalues will be dominated by highly expressed genes rather than by a coherent population. Without nulling, the leading components will also mostly re-describe mis-estimated proportions of known types. The ablation settles this. If plain residual PCA matches the full detector, we will report that the detector is unnecessary. What remains is the detection-limit law and the unreferenced-type benchmark, which we would then present as the contribution.

Why not just use reference-free factorization such as STdeconvolve? Reference-free methods must spend their capacity explaining the dominant types, so a population at a few percent of a few spots competes poorly for a factor. The benchmark in A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics found reference-free factorization weaker at fine-grained discrimination. Our residual approach explains the known types first and then searches only what is left. The abundance sweep shows directly whether this matters. If factorization detects the held-out type at the same abundance, the residual route offers no advantage.

Why not just add an "unknown" slot or a better atlas, as with POPPY? An "unknown" component absorbs everything the reference misfits: platform distortion, mis-specified types and the real rare population together. It gives no spatial footprint or signature that separates these. POPPY's acknowledged failure on missing states, Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, suggests that better atlases shrink the problem but do not remove it, since tumours keep producing new states. The "unknown"-slot baseline and the batch-distortion test show whether our nulled operator separates true populations from technical misfit better than a catch-all component does.

Proposed Experiments

The main testbed reuses the simulation design of the deconvolution benchmark. Pseudo-Visium spots are built from seqFISH+ and MERFISH tissues with known single-cell positions, and Xenium sections from the HESCAPE collection across 54 donors are aggregated into spots in the same way. In each run, one cell type is held out of the reference, and its abundance is thinned by subsampling its cells. The independent sweeps are spot count, sequencing depth, number of informative markers, and co-occurrence pattern. For the co-occurrence pattern, the held-out cells are relocated either into a single fixed niche or dispersed among varied neighbours, while abundance is held fixed.

The decisive checks are three. First, the empirical detection boundary should collapse onto the threshold law when plotted against $n p^2 \|\Delta\|^2 \rho_\perp^2 / \sqrt{SG}$. Second, at matched abundance, niche-locked populations should be markedly harder to detect than dispersed ones, and this gap should vanish when nulling is turned off. That pattern would confirm the aperture-from-diversity mechanism and also expose its cost. Third, under the statistical-stability prediction, the variance of the detection score across donors and sections should shrink as the number of decorrelated gene modules grows, not as raw gene count grows.

For batch robustness, the reference is drawn from a different platform than the spots. We also impose per-gene multiplicative distortions, and we require the false-detection rate on null tissues to stay controlled. For a real-data check, we deconvolve 10X Visium mouse brain against a MERFISH-derived reference with one annotated subtype removed, as in the zero-shot transfer setting used by Count Bridges. We then ask whether the recovered footprint and signature match that subtype's MERFISH locations and markers. Signature recovery is scored by correlation with the held-out profile.

A negative outcome on the second check would be informative in its own right. It would mean the wave analogy says nothing useful about heterogeneity, and that the method's value, if any, is ordinary residual analysis done carefully.