What Should a Photodetector Under Each Capture Spot Measure?

Many tissue-profiling assays read RNA from small groups of cells at once. Bulk sequencing pools a whole sample, and spatial platforms such as 10X Visium pool roughly 10 to 50 cells per capture spot. The pooled signal hides rare or transitional cell states. Computational "deconvolution" tries to undo this by unmixing each pooled measurement into cell types. It works best when a reference atlas of single-cell profiles already contains every type in the tissue, and it fails exactly where heterogeneity matters most: cell types the atlas lacks. Electronics fabricated in the same process as the capture surface, such as photodiodes, filters, and flat meta-optics (patterned thin films that shape light), could record a few physical facts about each spot as its RNA is captured. Those facts include how many cells sit there, how large they are, and a coarse color signature. Because the detector would be built in the same lithographic step as the spot, these measurements would be registered to the spot with no alignment error.

This note proposes to find out, before anyone builds such a chip, whether those measurements are worth having and which ones matter. The core claim is that a measured cell count per spot does something a reference atlas cannot. Once you know how many cells are present, any part of the RNA signal that the reference types cannot account for becomes evidence of an unseen cell population, not noise to be absorbed into the nearest known type. We will build a count-constrained version of an integer-count deconvolution model and drive it with simulated detector outputs that are degraded to realistic on-chip specifications. The result is a quantitative specification: the pixels per spot, spectral bands, and gain stability needed before on-chip detection beats purely computational unmixing.

Core Claim

A spot's RNA counts $x_{sg}$ for spot $s$ and gene $g$ can be modeled as Poisson draws from the summed expression of the cells present:

$$ x_{sg} \sim \text{Poisson}\Big(\sum_{k=1}^{K} n_{sk}\,\mu_{kg} + m_s\,\nu_{g}\Big), \qquad \sum_{k} n_{sk} + m_s = \hat N_s . $$

Here $n_{sk}$ is the number of cells of reference type $k$ in spot $s$, $\mu_{kg}$ is that type's mean expression, $m_s$ is the number of cells of an unreferenced type with a profile $\nu_g$ learned jointly across spots, and $\hat N_s$ is the detector's noisy cell-count estimate. Without $\hat N_s$, the model can trade $m_s$ against inflated $n_{sk}$ freely. With it, every unexplained cell has to show up as $m_s > 0$. The claim we will test is that this constraint, even when the count is approximate, sharply improves recovery of held-out cell types. It should also improve known-type proportions in spots that contain unseen types. Coarse size and spectral readings should add further gains by separating types whose expression overlaps.

The inference machinery comes from Count Bridges. That work models integer transcript counts with a Poisson birth-death process and deconvolves multi-cell mixtures by expectation-maximization. On simulated spots it reports Jensen-Shannon divergence 0.231 and RMSE 0.110, against 0.288 and 0.177 for the reference-free STDeconvolve. The same paper notes that integer models become expensive or degenerate at high counts. Small, sparse capture spots are the regime where it is strongest, so the proposal stays there.

Why this, and how it differs from existing work

Two problems documented in the file motivate the work. First, POPPY maps bulk transcriptomes into a single-cell foundation-model space, using an ontology-aware distance and cross-attention to reference atlases. It states that its cell-type attribution fails when the reference misses a cell state. Second, the 18-method spatial deconvolution benchmark ranked CARD, Cell2location, and Tangram highest and identified the mismatch between reference and target platforms as a main source of error. Both papers treat missing information as something to be inferred. We instead ask which physical side-measurement removes the most ambiguity.

The closest work that joins detectors to single cells does not touch deconvolution. Neuromorphic cytometry counts and sizes cells in a microfluidic channel with an external event camera, reporting 98.1% gated bead purity versus 95.6% on a commercial cytometer, and performs no molecular readout. Optoelectronic tweezers in NanoPen arrays isolate thousands of single cells, but with external projectors and microscopes. A review of on-chip dielectrophoretic manipulation handles cells by their electrical properties and notes that these overlap across cell types. Closed-loop microfluidic microscopy automates focus and segmentation with external optics. None of these feeds a physical measurement into RNA unmixing, and none asks what detector specification such a measurement would need.

The device papers set the realistic detector constraints. Flat-band meta-optic photodiodes give about 10.3-fold responsivity enhancement, a filter about 10 nm wide, and an acceptance angle of about ±18° in a 230 nm thick device. Their enhancement degrades at higher numerical aperture, so a per-spot detector should be assumed to collect light from only a narrow cone. Spiking cone photoreceptors discriminate three wavelengths at under 400 pW per spike. They reached about 83.2% on a color-digit task versus 75.5% for a color-blind control, which is a modest spectral margin. Passive kinematic coupling of disposable biosensors reports a 5.1 dB standard deviation in coupling loss across 99 units. We use that figure as the pessimistic chip-to-chip gain variation. Low-temperature nanoimprinted photonics shows that optics can be added on top of finished electronics, which makes a detector under the capture surface plausible in principle.

Research Program

The work has three parts.

Simulator. We start from imaging-based spatial datasets with single-cell ground truth, MERFISH and seqFISH+, as used in the deconvolution benchmark and the Count Bridges paper. Cells are binned into Visium-sized pseudo-spots. Each pseudo-spot has known transcript counts, known cell identities, segmented cell areas, and stain-image intensities. A detector model then turns each spot's image into what an on-chip sensor would plausibly report, with five adjustable settings:

  1. Pixels per spot: 1, 2×2, or 4×4.
  2. Spectral bands: zero to three bands about 10 nm wide, following the meta-optic filter.
  3. Angular blur from the limited acceptance cone, which mixes in light from neighboring spots.
  4. Shot and read noise.
  5. Chip-level gain variation, swept from zero up to 5.1 dB standard deviation.

Cell count is estimated from this degraded signal by a small calibrated regressor. The estimate therefore inherits realistic errors and is not the true count.

Model. We extend Count Bridges' expectation-maximization with the count constraint above, an unreferenced-type component, and optional per-cell size and spectral likelihood terms tied to type. Chip gain is treated as a per-chip latent scale so that coupling variation is absorbed in a principled way rather than ignored. If the constrained EM proves too slow, we fall back to a Poisson factorization with the same constraint. The scientific question does not depend on Count Bridges specifically.

Specification map. We sweep detector settings against deconvolution error and report the smallest detector that reaches the accuracy obtained from full-resolution microscope-derived counts. The output is a curve that a device group could design against, and it may show that no feasible on-chip detector is worth building.

Baselines

The naive reference point is reference-based deconvolution with no physical information. It assigns all signal to known types and imposes no count. Its error on spots containing held-out types shows how much masking the problem actually involves.

The strongest existing methods are CARD, Cell2location, and Tangram, the top performers in the 18-method benchmark, along with unmodified Count Bridges. The best number available in the file is Count Bridges' JSD 0.231 and RMSE 0.110 on simulated spots. The benchmark paper reports no single numeric score for CARD, Cell2location, or Tangram that we can quote, so we will rerun all of them on our identical pseudo-spots. Every method receives the same reference with the same held-out types removed and the same hyperparameter search budget. Wherever a method accepts a per-spot cell-abundance prior, it gets the same detector-estimated counts our model uses. This ensures that any gain comes from how the count is used and not merely from having it.

The microscope ceiling feeds counts and sizes from full-resolution segmentation of the registered stain image. We also add image-to-spot misregistration offsets to quantify what lithographic registration buys.

The key ablations follow. Count-only versus count plus size versus count plus spectral isolates the contribution of each measurement. A shuffled-count control, which permutes $\hat N_s$ across spots, tests whether gains come from information or from regularization. A latent-count variant infers $N_s$ from library size alone, without a measurement.

Beating the field means two things on spots containing held-out types. First, a clear reduction in JSD and RMSE of known-type proportions relative to the best baseline, judged across five held-out-type splits with bootstrap intervals. Second, higher area under the ROC curve (AUROC) for detecting which spots contain the unseen type. Both must hold while performance on fully-referenced spots stays no worse.

Why not just…?

Why not just image the tissue with a microscope, as Visium workflows already do? This is the most serious threat. If a registered microscope image is available, it probably supplies everything an on-chip detector would, and better. We treat it as the ceiling, not as a competitor to defeat. The proposal's value is conditional on two cases: settings without a microscope, such as disposable or point-of-care capture chips, and settings where image-to-spot misregistration matters. The misregistration sweep and the minimal-detector curve settle this. If microscope priors with realistic registration error already match zero-error priors, and degraded on-chip detectors fall far short of them, the honest conclusion is that on-chip integration is not worth pursuing for this purpose. We will report that.

Why not just treat cell count as a latent variable inferred from total RNA? Library size depends on cell type, cell size, and capture efficiency, so it confounds count with composition. That confound is exactly what lets missing types hide inside known ones. The latent-count ablation tests this directly. If it performs as well as the measured count, the core claim is false.

Why not just use a more complete atlas, such as 3CA with POPPY? Bigger atlases shrink the gap but cannot cover aberrant tumor clones or transitional states, which POPPY's own limitations section names. Our held-out-type splits simulate that case. We will also report how quickly the advantage of physical priors shrinks as held-out types are restored, which shows when a better atlas suffices.

Why not just dissociate cells and sequence them one at a time? Dissociation removes the spatial context in which layered-tissue questions are posed, such as ectoderm versus mesenchyme in facial prominences. The proposal addresses spatial capture, where single-cell isolation is not an option.

Decisive Checks

We consider the idea supported if three results hold:

  1. Detector-estimated counts, under realistic noise and up to 5.1 dB gain variation, beat every reference-only baseline on held-out-type spots.
  2. The shuffled-count and latent-count controls lose most of that gain.
  3. A detector of at most 2×2 pixels and three bands reaches a substantial fraction of the microscope ceiling.

The idea is refuted, and the device direction should be abandoned, if any of the following holds:

  1. The latent-count variant matches measured counts.
  2. Gains vanish at realistic gain variation.
  3. Only microscope-grade detectors help.

A physical prototype, such as a flat-band photodiode array beneath microwells, is justified only if the first outcome holds. Its target specification would come directly from the curve produced here.