Imaging spatial transcriptomics is now routine on archived clinical tissue. In this family of in situ hybridization methods, fluorescent probes stick to individual RNA molecules in a tissue slice and a microscope records where each one sits. Researchers increasingly want to go one step further and infer the proteins those cells make, because proteins, not transcripts, carry out function, and RNA predicts protein levels only 40–60% of the time (A high-stringency blueprint of the human proteome). The natural pipeline has three steps: draw cell boundaries in the image, count the RNA inside each boundary, and feed that count vector to an RNA-to-protein model trained on dissociated cells. The first step is known to be unreliable. Boundaries drawn in dense or irregular tissue are wrong often enough that transcripts from one cell land in its neighbor, and this holds on every commercial platform (Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues). A B cell that has absorbed T-cell transcripts will then be predicted to carry T-cell proteins. When this prediction is compared against measured protein, the error looks like biology, a "discordance" between transcript and protein, when it is really an image-analysis artifact.
We propose to make that distinction measurable and to remove the artifact before translation. The method treats each segmented cell's RNA counts as an integer mixture of the cell's own transcripts plus contributions from its physical neighbors. It unmixes that mixture at the level of whole-number counts, and only then passes the cleaned own-cell counts to a translator that predicts the full distribution of protein levels rather than a single average. The central claim is testable: a substantial share of the apparent RNA–protein discordance in imaging in situ data is segmentation spillover. It can be removed by unmixing that knows about neighbors. Any remaining protein signal that depends on a cell's surroundings should be evaluated only after this cleaning, because otherwise spatial models will learn to reproduce the spillover.
Why this is the right cut through the three directions
The three directions play different roles here. Heterogeneity masking is the mechanism we attack. It shows up not in bulk samples or array spots but inside the "single cells" produced by segmentation. Transcript–protein discordance is the quantity we want to estimate honestly. In situ hybridization is the regime: sensitivity, error rates, and panel size come from the platforms themselves.
The closest existing work falls into three groups, and each misses a different piece. Dissociated-cell translators are the first group. totalVI jointly models RNA and surface-protein counts with a variational autoencoder, a neural network that compresses data into a probabilistic latent code (Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells). scTranslator is a pretrained transformer that maps any gene panel to thousands of proteins (A pre-trained large generative model for translating single-cell transcriptome to proteome). scChord first learns a smooth protein latent space with a count-aware autoencoder, then links RNA to it with flow matching, a generator that learns straight paths from noise to data. This lets it recover bimodal and heavy-tailed markers such as CD4 and CD45RA that regression averages away (scChord). All three assume the input RNA vector belongs to one cell. In segmented in situ data that assumption fails. We reuse scChord's distributional translator as a component and change its input.
Spatial protein predictors are the second group. STProtein predicts spot-level protein from spot-level RNA with a graph attention autoencoder over similar spots (STProtein). CANDIES denoises one modality using another in co-profiled spatial multi-omics (CANDIES). Both motivate the problem, but they constrain the design in a specific way. Pooling features from neighbors is itself a form of spillover. A graph model applied to uncleaned in situ cells could fit measured protein better while making the mixing problem worse. CANDIES also requires both modalities to be measured together, and imaging RNA panels on archived tissue do not have that.
Count Bridges supplies the unmixing method. It models integer counts with Poisson birth–death dynamics and deconvolves aggregates with an EM procedure. EM, or expectation–maximization, alternates between guessing the hidden per-cell counts and refitting the model. Count Bridges forces the guessed per-cell counts to sum exactly to the observed total, and it beats STDeconvolve on multicellular MERFISH spots (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). It has never been used on segmentation spillover, and it has never been connected to protein. Our unmixing step differs from Count Bridges and from spot deconvolution tools such as CARD, Cell2location, and Tangram (benchmark) in one structural way. The mixture components are not abstract cell-type proportions from a reference atlas. They are the specific adjacent cells, and each neighbor's weight depends on shared boundary length and local density.
Core Mechanism
For a segmented cell $i$ with observed integer count vector $x_i$, we write
$$x_i = z_{i\to i} + \sum_{j \in N(i)} z_{j\to i},$$
Here $z_{j\to i}$ is the vector of transcripts truly originating in cell $j$ but assigned to $i$, and $N(i)$ is the set of physical neighbors. Each cell's true profile $y_j = \sum_k z_{j\to k}$ is drawn from a count generative model $p_\theta(y)$ pretrained on a dissociated reference; we use a Count Bridge model. The split of $y_j$ into assigned pieces follows a multinomial whose probabilities $\pi_{j\to k}$ are predicted from image geometry: boundary contact, distance from the boundary, and nuclear versus cytoplasmic position of each molecule when available. EM alternates between two steps. The first samples the per-cell latent counts, with the sampling projected so that the observed totals are matched exactly. The second refits the spillover probabilities $\pi$.
The cleaned own-cell estimates $\hat y_i$ go into a distributional translator, a retrained scChord, which outputs samples of the cell's protein vector. We train it on dissociated CITE-seq data, which pairs RNA and protein in the same cells. Before training we downsample the RNA to match each in situ platform's per-gene capture efficiency, so the translator sees inputs as sparse as real Xenium or CosMx cells. The optional third piece conditions the translator on neighbor composition. It is added only after unmixing, so any benefit it shows cannot come from re-absorbing spillover.
Research Program
The program has three parts.
A controlled benchmark. No public dataset in our file provides single-cell imaging in situ RNA with ground-truth protein, so we construct one. We take real cell centroids, cell-type labels, and segmentation masks from the FFPE tissue-microarray benchmarking corpus (Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues). We fill each position with a CITE-seq cell of matching type from the human PBMC atlas (GSE164378, 228 proteins) or the murine spleen and lymph node data. We then simulate molecule placement and platform capture. Capture efficiency is set to Xenium-like and CosMx-like levels, reflecting the reported roughly 2-fold per-gene sensitivity gap. Spillover is controlled by density and boundary error.
Because the true protein and true own-cell RNA are known, observed discordance splits exactly into a spillover part and a true translational part. We also plant a known neighbor-dependent translation effect in a subset of proteins. For example, protein $p$ is scaled up in cells adjacent to type $T$ while its RNA is left unchanged. This tests whether neighbor conditioning recovers real context effects and not artifacts.
Real-tissue checks at spot level. On SPOTS mouse spleen and Stereo-CITE-seq thymus, where measured protein exists only for aggregated spots, we sum our per-cell protein predictions within each spot footprint and compare them with STProtein on its own benchmark.
A small paired validation. The decisive real-data test needs a section with an imaging RNA panel plus multiplexed protein immunofluorescence on the same cells, in lymphoid tissue where lineage-exclusive markers exist. We will generate this dataset. There is no reported result found on Lacuna for single-cell in situ RNA-to-protein prediction against same-section protein.
Baselines
The naive reference applies no model. The protein is read off as its own gene's RNA count in the segmented cell. The file indicates this recovers well under full concordance, and it gives the floor for every metric.
The strongest existing translator is scChord applied directly to raw segmented cells, trained with the same platform-matched downsampling as our model. It is the strongest competitor because it already solves the smoothing problem that plain regression suffers from, so any gain we show must come from the cleaning step. The file reports no scChord numbers, only that it beats totalVI, sciPENN, and scTranslator on RMSE and Pearson correlation on GSE100866. We therefore re-run it on identical splits. scTranslator and totalVI are run under the same protocol. The one reported comparison for scTranslator is a 10.7% cosine-similarity gain over sciPENN and cTP-net (scTranslator).
At spot level, STProtein is the reference. It reports RMSE 0.95 on SPOTS mouse spleen, 0.04 better than Dengkw (STProtein). We will reproduce that number under its preprocessing before comparing. Beating it means lower spot RMSE with our aggregated per-cell predictions, without training on spot protein.
Two ablations isolate the key ingredient. The first replaces Count Bridge unmixing with non-negative least squares, which fits each cell as a non-negative weighted sum of neighbor mean profiles. This tests whether integer-count modeling matters or whether knowing the neighbors is enough. The second replaces neighbor-specific components with reference cell-type proportions (CARD-style). This tests whether neighbor identity matters.
All methods share data splits, platform simulation, the translator architecture where relevant, and an equal hyperparameter-search budget of GPU hours. Beating scChord-on-raw means two things. The first is a statistically reliable reduction in the spurious lineage co-positivity rate, defined as the fraction of cells predicted positive for a protein exclusive to another lineage, such as CD3 in B cells. The second is no loss in per-protein Wasserstein distance or in correlation-matrix distance.
Why not just…?
Why not just run scChord or scTranslator on segmented cells? This is the main threat and the central comparison. If spillover is small at realistic density, raw translation will match ours, and the benchmark's spillover dial will show where that happens. Our prediction is that the gap grows with density and boundary error, which the benchmarking study identifies as highest in stroma and tumor (iST benchmark).
Why not just count nuclear transcripts only, or use better segmentation? Nuclear-only counts avoid most spillover but throw away cytoplasmic molecules. This hurts most on lower-sensitivity platforms, which already detect 2- to 10-fold fewer transcripts per gene. We include nuclear-only and dilated-nucleus inputs as baselines on both Xenium-like and CosMx-like simulations. If nuclear-only matches us on the high-sensitivity setting, we will report that honestly and restrict the claim to lower-sensitivity data.
Why not just use a spatial graph model like STProtein at single-cell resolution? Averaging neighbor features can raise fit to measured protein while encoding spillover. The planted-effect experiment separates the two cases. A graph model on raw cells should "find" context effects in proteins where none were planted, and our conditioned model should not.
Why not simple non-negative least squares instead of count bridges? This is plausible, and it is run as an ablation. If it performs equally, we adopt it, and the contribution becomes the neighbor-component formulation rather than the count machinery.
Proposed Experiments
The first decisive check runs on the synthetic benchmark. It asks how much of the observed discordance, measured as the gap between true protein and naive RNA-based prediction, is attributable to spillover, how this share scales with density, and what fraction our unmixing removes compared with scChord-on-raw and nuclear-only inputs. Metrics are per-protein Pearson correlation and RMSE, per-protein Wasserstein distance with CD4 bimodality recovery, correlation-matrix distance across the full 228-protein panel, and the lineage co-positivity rate.
The second check asks whether neighbor conditioning after cleaning recovers the planted context effects at a low false-positive rate on unplanted proteins. The same conditioning is also applied before cleaning, and we expect it to produce false context effects there.
The third check aggregates per-cell predictions to spots on SPOTS spleen and Stereo-CITE-seq thymus and compares them against the reproduced STProtein RMSE of 0.95.
The last check, on the paired immunofluorescence section, applies the same co-positivity and correlation metrics. If unmixing does not lower co-positivity there relative to scChord-on-raw, the central claim fails on real tissue.
References
- A high-stringency blueprint of the human proteome
- Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- STProtein: predicting spatial protein expression from multi-omics data
- Cross-modal Denoising and Integration of Spatial Multi-omics data with CANDIES
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics