Neurons of the same type are not identical. The receptors that set how strongly a synapse lets calcium into a cell differ from spine to spine and from neuron to neuron, and some of these differences track where the neuron sits in tissue. Cell DNA carries chemical tags, methyl groups added to cytosine bases, that can quietly turn genes down without changing the sequence. This gene silencing is a plausible candidate for a slow, stable "dimmer" that sets how much of each receptor subunit a neuron can make. Nobody can currently test this in intact brain tissue. Spatial transcriptomics, which measures RNA while keeping each cell's position, does not read DNA methylation. Single-cell methylation sequencing reads methylation, but only after the tissue has been dissociated and every position discarded.
We propose to rebuild the missing map computationally and then try hard to break it. We will place individual cortical methylomes from the BICCN motor cortex atlas onto spatially resolved cells. We will then ask a narrow question: within a single neuron type, do the methylation levels at regulatory regions of NMDA receptor subunit genes (Grin2a, Grin2b) vary continuously with cortical depth, and does that variation predict the depth gradient of those genes' RNA and of the receptor current? The claim is not "methylation can be projected into space." The claim is that a projection preserving within-type continuous variation shows a depth-graded methylation dial at synaptic receptor genes that cell-type averages cannot show. The test is built so that the simple explanation, cell-type composition changing with depth, can win.
Why this, and why now
Three findings motivate the question. The first is about synapses. Sobczyk et al. showed with two-photon glutamate uncaging that spines on one hippocampal neuron vary 10-fold in calcium entry per unit current. NR2B (GluN2B) enrichment in small spines drives this. Spine geometry explains only about half the variance, and blocking NR2B with ifenprodil abolishes the excess variability. That work locates the "missing half" in subunit composition but says nothing about what sets a neuron's subunit supply. The second is about tissue position. The BICCN census (A multimodal cell census and atlas of the mammalian primary motor cortex) found that intratelencephalic (IT) neurons form continuous molecular and physiological gradients aligned with cortical depth. In the striatum, Muñoz‑Manchado et al. showed that Pvalb is a continuous, spatially graded axis inside a broader Pthlh population rather than a discrete type. Neither study asked what holds such gradients in place. The third is about mechanism. Rao et al. showed for CAV1 that silencing is controlled by methylation of CpG island shores, the lower-density flanks within about 2 kb of the island, with r < −0.7 against mRNA, while the island itself stays unmethylated. scMET showed that single-neuron methylation variability is not noise and concentrates at distal enhancers near identity genes such as Gad1 and Neurod6. Together these suggest graded shore or enhancer methylation as a candidate mechanism for within-type receptor gradients. Shore regulation has been shown in cancer cell lines, not neurons, so the transfer is a hypothesis to test, not an assumption.
Core claim and what is new
The closest existing work is Tangram's "spatial multi-omics" (Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram). Tangram aligns the RNA half of paired SHARE-seq cells to a spatial scaffold and carries their chromatin accessibility along. It differs from our setting in three ways. It relies on RNA and chromatin being measured in the same cell; BICCN snmC-seq2 nuclei have no RNA. It projects accessibility, not methylation. It was never evaluated on whether within-type variation survives the projection; its validation is held-out gene correlation. BICCN integrated methylation with RNA using SingleCellFusion and LIGER, but only to agree on a taxonomy, not to assign positions. SpaOTsc uses optimal transport (a way of computing the cheapest soft matching between two point clouds) to place dissociated cells in space, but carries only transcript abundance. Standard deconvolution methods (stereoscope, CARD, Cell2location; see the benchmark) model each spot as a mixture of fixed type profiles. By construction they erase within-type drift. Those methods supply our machinery. Our contribution is an unpaired, within-type-preserving methylome-to-space projection with circularity controls, and a falsifiable biological test applied to synaptic receptor genes.
Research Program
The pipeline has two legs. The first leg links each methylome nucleus to RNA-profiled nuclei. It uses only "bridge" features: gene-body methylation summaries of the kind BICCN integration already relied on. All synaptic receptor genes and their flanking regions are removed from the bridge. The second leg places RNA nuclei onto the 254-gene MERFISH cells with Tangram-style alignment. Composing the two legs gives each methylome nucleus a soft distribution over spatial cells.
Run naively, both legs tend to collapse cells onto their type centroid. That collapse would make any within-type gradient an artifact or erase it. We therefore restrict matches to the same consensus type and add a term that rewards keeping each cell's position along its type's main continuous axis. Concretely, for type $k$ we solve for a coupling matrix $P$ between methylome cells $i$ and spatial cells $s$:
$$\min_{P \ge 0} \sum_{i,s} P_{is}\, C_{is} \;+\; \lambda \sum_{i,s} P_{is}\,\big(z_i - z_s\big)^2 \;+\; \varepsilon\, H(P)$$
The symbols are defined as follows. $C_{is}$ is the bridge-feature distance through the RNA leg. $z_i$ and $z_s$ are the cells' coordinates on a within-type latent axis, learned separately in each modality and aligned by rank. $H(P)$ is an entropy regularizer. The marginals are left unbalanced as in SpaOTsc. Setting $\lambda = 0$ recovers plain type-restricted matching, and that is our key ablation.
After placement, regulatory regions are scored for each projected cell. These are the promoter island, the 5′ and 3′ shores (≤ 500 bp and ≤ 2 kb), and candidate enhancers for Grin2a, Grin2b, Gria1 and a set of non-synaptic control genes. Because 80–95% of CpGs are unobserved per cell (scMET), region methylation is estimated with scMET's Beta-Binomial model. Uncertainty is then propagated through the coupling instead of using raw ratios. The biological analysis fits, within each type, projected region methylation against cortical depth. It then asks whether shore methylation, rather than island methylation, tracks the depth profile of the gene's RNA. If the CAV1 logic transfers to neurons, we predict that Grin2b shore methylation rises with depth where Grin2b/Grin2a falls, while the island stays flat.
Baselines
The naive reference is cluster-mean label transfer. Every MERFISH cell receives the average methylation of its consensus type, so any depth trend comes only from type composition changing with depth. This is the most important comparator because it encodes the skeptical hypothesis directly. It uses the same snmC-seq2 data, the same type labels, and the same region scoring. Our method beats it only if within-type depth trends in held-out ground truth, described below, are reproduced by our projection and not by the cluster mean.
The strongest existing method is Tangram's projection route applied to the unpaired setting. A methylome nucleus is linked to RNA nuclei by LIGER/SingleCellFusion-style integration, and Tangram then carries the attached methylation to space with $\lambda = 0$. No within-type term and no type restriction beyond Tangram's defaults are used. Both arms get identical bridge features, the same GPU budget, and the same hyperparameter search over the entropy weight and learning rate, tuned on a validation split of RNA genes. The RNA leg must first match Tangram's reported quality on this exact data: in leave-one-out tests on MERFISH, 75% of genes were predicted with spatial correlation above 0.4 (Tangram). SpaOTsc reports held-out gene AUCs of 0.95 (zebrafish) and 0.88 (Drosophila) (SpaOTsc), but on embryos, not cortex. We will run SpaOTsc as a second transport baseline under the same budget. No reported result exists on Lacuna for spatial projection of single-cell DNA methylation by any method, so methylation-level comparisons are new and will be measured, not quoted.
The key ablation is $\lambda = 0$ versus tuned $\lambda$, with everything else fixed. It isolates whether the within-type term adds information or simply manufactures a gradient. A permutation control shuffles depth labels within type before fitting, which sets the null distribution for "gradient strength" for every method.
Why not just…?
Why not just assign each spatial cell its type's average methylation? This is a serious threat. If within-type methylation at receptor genes does not vary with depth, the type average is the correct answer and our extra machinery adds nothing. We do not argue around this. The cluster-mean baseline, scored against layer-resolved ground truth, settles it. If it matches our projection within the permutation null, we report that the receptor gradients are compositional or post-epigenetic.
Why not just run smFISH or immunostaining for NR2A, NR2B and their mRNAs across depth? Imaging would measure the RNA or protein gradient directly and better than imputation. We will use it as validation of the RNA side. But no in situ assay on a standard section reads shore or enhancer methylation, so imaging alone cannot tell graded silencing from, for example, graded translation. The question here is the regulatory one.
Why not just microdissect layers, sort a cell type, and run bulk bisulfite sequencing? For the coarse question, this is the cheapest and most direct design, and it may answer it outright. We therefore adopt it as our ground truth instead of competing with it. It has two limits. First, it gives a handful of depth bins per type, not continuous positions. Second, it targets the regions chosen in advance, while the projection works genome-wide and can nominate unexpected shores or enhancers. If layer-sorted bulk data and the projection agree, the projection is credible where bulk cannot reach. If they disagree, the projection is wrong.
Why not just project chromatin accessibility, as Tangram already does? Accessibility and methylation are related but not the same mark. The shore-versus-island dissociation in Rao et al. is specifically a methylation result. A positive accessibility gradient would support our methylation finding but would not substitute for it.
Proposed Experiments and decisive checks
The data are the BICCN mouse MOp snRNA-seq, snmC-seq2 and 254-gene MERFISH datasets, plus Patch-seq cells with recorded soma depth (BICCN). We use IT subtypes as the primary test bed because their depth continuum is documented. Ground truth is new layer-microdissected, type-sorted bulk bisulfite sequencing at the target regions in four depth bins. Where available, we also use the dissection metadata of the snmC-seq2 nuclei themselves, held out from fitting.
The first check concerns geometry. For each method, we compute the within-type correlation between projected depth and true dissection bin for methylome nuclei. Shuffling depth labels within type must destroy this correlation. The second check concerns methylation. We compare projected region methylation per depth bin with bulk ground truth on regions excluded from the bridge, scoring with a rank correlation across bins after subtracting the type mean. This metric isolates within-type signal, and the cluster-mean baseline scores zero on it by design. The third check is the shore-versus-island prediction. Across the receptor and control genes, shore methylation should anticorrelate with the within-type depth profile of the gene's RNA, while island methylation should not. The RNA profile is measured by MERFISH where the gene is on the panel and by validated imputation or smFISH otherwise. Control genes without a depth gradient guard against a generic depth artifact.
The fourth check is a small physiological test that ties the map back to synaptic heterogeneity. In acute slices, we will record the ifenprodil-sensitive fraction of NMDA current in IT neurons at known soma depths. This is the somatic counterpart of the NR2B contribution measured by Sobczyk et al.. We then ask whether depth-binned projected Grin2b shore methylation predicts this fraction better than depth alone. This measures a soma-level supply ceiling, not single-spine stoichiometry, and we state that limit plainly.
Any one of three outcomes counts as a clean negative result: the projection fails the geometry check, the within-type methylation signal does not beat the cluster-mean baseline, or shores behave no differently from islands. Each is informative about whether DNA methylation is a lever on synaptic receptor gradients at all.
References
- NMDA Receptor Subunit-Dependent [Ca2+] Signaling in Individual Hippocampal Dendritic Spines
- A multimodal cell census and atlas of the mammalian primary motor cortex
- Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq
- CpG island shore methylation regulates caveolin-1 expression in breast cancer
- scMET: Bayesian modeling of DNA methylation heterogeneity at single-cell resolution
- Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram
- Inferring spatial and signaling relationships between cells from single cell transcriptomic data
- Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics