Spatial transcriptomics machines such as 10x Visium measure gene counts in small tissue spots, and each spot holds roughly five to thirty cells. The cells' individual identities are averaged away. Most tools that recover which cell types sit in each spot need a complete single-cell reference atlas, so they misassign or drop cell states the atlas lacks. We propose a reference-free alternative built on a result from theoretical neuroscience. When a network's units cannot fire negatively and must pay for every spike, each unit is pushed to represent exactly one independent cause of the data, and this yields readable "cell types" in the network itself. Our addition is a fix for the condition under which that result breaks. The guarantee needs the hidden causes to be independent, and cell types in tissue are not: they co-occur by region. We argue that conditioning each spot's spike budget on its physical neighbors absorbs this regional co-occurrence, which restores near-independence and with it the push toward one-unit-one-cell-type.
Put in two sentences: we build an autoencoder whose hidden units emit nonnegative spike counts and are charged a metabolic cost relative to what their spatial neighbors already express, so a unit pays only for local surprise. If anatomy removes the correlations between cell types and energy cost then separates what remains, the units should become individual cell-type programs that can be read directly, including rare and atlas-missing ones, without post-hoc dissection.
Core Claim
Disentanglement with Biological Constraints: A Theory of Functional Cell Types shows that nonnegative firing combined with an energy penalty creates an "energy tax" on multiplexing. A unit that encodes two factors must keep a high baseline rate so it can modulate downward, and this is costly, so units specialize. The paper also shows that correlated factors produce warped, mixed representations. That is the regime of tissue, where tumor cells, fibroblasts and immune cells co-vary with anatomical region. Spatio-relational inductive biases in spatial cell-type deconvolution supplies the other half of the argument. It reports that compositional dissimilarity between spots grows with physical distance, which means much of the co-variation among cell types is explained by location. Our claim is that the factors which are correlated across a whole slide become approximately independent once we condition on the neighborhood. Making the metabolic prior neighborhood-conditioned should therefore recover the specialization that the theory predicts for independent factors.
The mechanism comes from Metabolic cost of information processing in Poisson variational autoencoders, which shows that a variational autoencoder (an encoder–decoder trained to compress and reconstruct data under a probabilistic prior) with Poisson spike-count latents pays a cost that grows with firing rate. A Gaussian version can make silence free by inflating prior variance. For a latent unit $k$ in spot $s$ with posterior rate $\lambda_{sk}$ and prior rate $\lambda^0_{sk}$, the per-unit cost is
$$ \mathrm{KL}\big(\mathrm{Pois}(\lambda_{sk})\,\|\,\mathrm{Pois}(\lambda^0_{sk})\big) = \lambda_{sk}\log\frac{\lambda_{sk}}{\lambda^0_{sk}} - \lambda_{sk} + \lambda^0_{sk}. $$
In that paper $\lambda^0$ is a global constant. Here we set $\lambda^0_{sk}$ to a stop-gradient average of the posterior rates of unit $k$ over the spatial neighbors of $s$. Two consequences follow. A program that is present throughout a region is cheap inside it. A unit that fires where its neighbors are silent, which is the signature of a regionally rare cell type, pays the full tax, so it will fire only when the counts genuinely demand it. The decoder maps spikes to expected gene counts through a nonnegative weight matrix, so each unit's column of weights is a gene program. Each unit's posterior rate is then that program's abundance in the spot.
Positioning Against Prior Work
The ingredients have been tested separately, and each test motivates or constrains this design. Reference-based spatial deconvolution with a neighbor graph, the GAT-C2L model of Spatio-relational inductive biases, improves rare-cell detection over spatially blind Cell2Location. It still requires reference signatures and outputs proportions, not discovered programs. We keep its spatial prior and drop the reference. Count Bridges models integer counts exactly and resolves MERFISH and Visium spots into single-cell profiles. Its latent process is a statistical bridge with no unit-level meaning and no spatial prior; we borrow its insistence on discrete count likelihoods. POPPY aligns bulk data to single-cell foundation-model embeddings using the Cell Ontology. It depends on atlas completeness, which is exactly the failure we target. Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models extracts features from dense models after training and finds about two-thirds of audited features interpretable, with the rest uncharacterized or polysemantic. This sets the bar our built-in interpretability must clear. Pathway-masked autoencoders such as NetActivity give readable units, but they work on homogenized bulk tissue with fixed gene-set scaffolds. We deliberately do not hard-code pathways, because the cell programs we want to discover may cut across curated sets.
One prior result constrains us directly. The open-problems summary for metabolic coding notes that L1 and L2 activity penalties in continuous deep networks often produce dead units instead of clean specialization. We take this seriously. It is one reason we use Poisson latents, where silence is a natural state, and we will report the dead-unit fraction explicitly.
Research Program
The model is small. A graph encoder reads a spot's counts and those of its neighbors and outputs Poisson rates for $K$ latent units, with $K$ set above the expected number of cell types. A nonnegative linear decoder, scaled by library size, outputs Poisson rates for genes. The loss is the negative Poisson log-likelihood plus $\beta$ times the neighbor-conditioned KL above. We train on one tissue and analyze the learned units. Units are matched to ground-truth cell types by maximal correlation of abundance maps, and we score how exclusively each unit tracks one type using a mutual-information gap of the kind Whittington et al. report.
The scientific core is a controlled test of the decorrelation hypothesis. Starting from MERFISH mouse brain data, where single-cell identities and positions are known, we bin cells into Visium-sized pseudo-spots. We then build variants of the same tissue by reassigning cells across regions. This lets us dial the marginal correlation between cell-type abundances while independently controlling the correlation that remains once the neighborhood is known. The theory makes a sharp prediction. Without the spatial prior, disentanglement should degrade as marginal correlation rises. With it, disentanglement should track only the conditional correlation. A shuffled-adjacency control, analogous to the weight shuffling that destroyed attractor dynamics in From Synapses to Dynamics, tests whether real geometry is doing the work.
Baselines
The naive reference point is spot-wise nonnegative factorization with a Poisson likelihood and sparsity, the classical reference-free approach in the same family as STDeconvolve. It shares our nonnegativity and count likelihood but has neither spatial conditioning nor an amortized encoder. It tells us how much of any gain is just "nonnegative Poisson factorization." Count Bridges reports beating STDeconvolve on MERFISH, but no reported result found on Lacuna for the numeric margin. We run it with the same $K$, the same genes and a matched tuning grid. We also run a spatially smoothed variant, factorization with a graph-Laplacian penalty on abundances. This is the strongest simple competitor, and it is discussed below.
The strongest existing reference-based method is GAT-C2L from Spatio-relational inductive biases. It reports consistent gains over Cell2Location in Pearson correlation, Jensen–Shannon divergence and AUPRC on simulated MPOA and Xenium breast data, but no reported result found on Lacuna for the actual values. We will rerun it on identical pseudo-spots and give it the full single-cell reference, which is an advantage our method does not get. We also run it a second time with one or two cell types deleted from the reference, to measure the atlas-dependency gap. We count it as beaten if we match it within noise on proportion recovery with the full reference, and clearly exceed it on recovering the deleted types. Count Bridges is run as the strongest count-exact deconvolver, with no reported result found on Lacuna for its MERFISH numbers.
Three ablations isolate the key ingredient. The first replaces Poisson latents with Gaussian-ReLU latents plus an L1 penalty, while keeping the neighbor-conditioned prior mean. Vafaii and Yates report that Gaussian-ReLU sparsity stays near 50% as $\beta$ varies, while Poisson sparsity moves from 7% to 96% on natural image patches; we test whether that contrast survives on count data. The second uses Poisson latents with a global prior rate, which removes anatomy. The third uses the same prior but feeds a shuffled adjacency graph. As an interpretability baseline, we train a TopK sparse autoencoder after the fact on the hidden layer of an unconstrained count autoencoder, following Pedrocchi et al.. We then compare the fraction of units that map cleanly to one cell type. All methods get the same data splits, the same $K$ grid and an equal number of hyperparameter trials.
Why not just…?
Why not just run sparse nonnegative factorization with spatial smoothing? This is the real threat. Sparse nonnegative factorization already embodies nonnegativity plus an energy-like cost, and a Laplacian penalty already uses anatomy. The difference is where anatomy enters. Smoothing pulls abundances toward neighbors in the output, which blurs boundaries, the known oversmoothing failure of spatial graphs. Our prior instead changes what it costs to deviate, so a sharp boundary is expensive only once and then cheap on its own side. If the smoothed factorization matches us on the correlation sweep and on boundary-adjacent rare types, the honest conclusion is that the neural machinery is unnecessary. We would then report the decorrelation result as a property of the objective, not the architecture.
Why not just use a reference-based deconvolver such as Cell2Location or GAT-C2L? Where the atlas is complete, these may well match or beat us, and we expect that. Their weakness, documented for atlas-aligned tools like POPPY and catalogued in the breast-cancer deconvolution benchmark, is states missing from the reference. The deleted-reference experiment measures this directly.
Why not just apply sparse autoencoders to a foundation model? Post-hoc dissection leaves about a third of features uncharacterized and inherits superposition from the base model (Pedrocchi et al.). Those features also carry no abundance meaning in a mixed spot. Our post-hoc baseline on a matched count autoencoder tests whether the training-time constraints produce cleaner units than after-the-fact decomposition.
Why not just use a Gaussian VAE with L1 sparsity? Vafaii and Yates argue that Gaussian latents make silence cheap through variance inflation, so sparsity does not respond to the regularizer. The first ablation tests whether this costs disentanglement on transcript counts, not just on images.
Proposed Experiments
The primary benchmark is MERFISH mouse brain binned into pseudo-spots with exact ground truth, followed by the MPOA and Xenium breast cancer settings used for GAT-C2L. Metrics are Pearson correlation and Jensen–Shannon divergence of recovered proportions, AUPRC for regionally rare types, the mutual-information gap between units and cell types, the dead-unit fraction, and recovery of reference-deleted types. On Xenium breast cancer we additionally bin high-tumor-purity regions to probe the "Purity Trap" from the benchmark study. There, normal epithelium is a regionally rare population that the neighbor-conditioned cost should keep from being absorbed into the dominant tumor unit. Finally, we train on MERFISH and apply the model to real Visium sections, following the zero-shot protocol of Count Bridges. We check unit maps against known anatomical layers.
Two outcomes decide the project. The first is the correlation sweep. Disentanglement should follow conditional rather than marginal correlation only when the true adjacency graph is used. If the shuffled graph does as well, or the global-prior model is unaffected by marginal correlation, the hypothesis is false. The second is the comparison with spatially smoothed factorization on boundary-adjacent rare cell types and deleted-reference types. Winning there is the evidence that pricing deviations from the neighborhood, instead of smoothing toward it, is the contribution.
References
- Disentanglement with Biological Constraints: A Theory of Functional Cell Types
- Spatio-relational inductive biases in spatial cell-type deconvolution
- Metabolic cost of information processing in Poisson variational autoencoders
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models
- NetActivity enhances transcriptional signals by combining gene expression into robust gene set activity scores through interpretable autoencoders
- From Synapses to Dynamics: Obtaining Function from Structure in a Connectome Constrained Model of the Head Direction Circuit
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures