When a mouse learns that a tone predicts reward, dopamine floods small patches of its motor cortex within milliseconds — and, remarkably, different patches tens of micrometers apart carry different dopamine signals during movement, reward anticipation, and reward consumption, as shown by two-photon imaging with the genetically encoded sensor dLight1 (Ultrafast neuronal imaging of dopamine dynamics with designed genetically encoded sensors). But the optical recording is chemically anonymous: it cannot tell you which cells sit inside each active patch, what receptors they express, or whether the functional differences between patches trace back to differences in local cellular composition. Spatial transcriptomics — sequencing mRNA while keeping each measurement anchored to a physical position in the tissue — and single-cell RNA sequencing, which profiles the full transcriptome of individual dissociated cells, know the molecular identities but only after the tissue is fixed and dead. The two measurement families have never been registered to each other, so a basic question is unanswered: is micro-scale functional heterogeneity in neuromodulator signaling a signature of local transcriptomic architecture, or is it noise?
The proposal here is to build that bridge and exploit it in both directions. Concretely, I will co-register, within the same animal and the same block of tissue, in vivo dLight1 dopamine imaging with post-hoc spatial transcriptomics, and develop a deconvolution model — the computational step that unmixes each 55-micrometer capture spot (which pools 10–50 cells) into its constituent cell types — that is guided by the live dopamine recordings. The model addresses a second, independent problem: existing deconvolution assumes tissues are made of discrete cell-type categories, while single-cell data from the striatum shows that key neuromodulator-target populations form continuous molecular gradients, not bins (Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq). Live dynamics offer a functional readout that can constrain where along such continua a spot's cells sit.
Core Claim
The central claim is twofold. First, experimentally: with a same-animal protocol (chronic window two-photon dLight1 imaging during a Pavlovian task, followed by immediate spatial transcriptomics of the imaged block, registered by vascular and fiducial landmarks), the dopamine dynamics of individual imaged regions can be linked to the transcriptomic composition of the very spots that contain them. Second, computationally: a deconvolution model in which spot composition is parameterized on a continuous cell-state manifold, and in which spots with similar dopamine dynamics are pushed toward similar predicted compositions, will both (a) unmask cell mixtures more accurately than discrete-category deconvolution in tissue dominated by gradients, and (b) predict held-out dopamine dynamics from composition alone — something no static method can do. If micro-scale dopamine heterogeneity is functional rather than noise, predicted local composition should carry real information about local signaling.
This is deliberately different from existing intersections. Tangram aligns dissociated single-nucleus profiles onto static spatial scaffolds and expands targeted gene panels to the full transcriptome (Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram), but its objective is purely static transcript abundance. Stereoscope and Cell2location unmix spots probabilistically against fixed, discrete single-cell references (Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography; A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics), with no notion of time or function. SpaOTsc and FlowSig infer intercellular signaling from ligand–receptor mRNA co-expression (Inferring spatial and signaling relationships between cells from single cell transcriptomic data; Inferring pattern-driving intercellular flows from single-cell and spatial transcriptomics) — an assumption that breaks completely for dopamine and norepinephrine, which are enzymatic metabolites, not transcripts. None of these frameworks consumes a millisecond-resolution optical signal; none can ask whether composition predicts dynamics.
Research Program
The experimental arm adapts the dLight1 preparation directly: AAV-expressed dLight1 in mouse motor cortex and dorsal striatum, chronic cranial window, two-photon imaging across multi-day Pavlovian conditioning, exploiting the sensor's ~9.5 ms rise and ~90 ms decay kinetics and its ~70-fold selectivity for dopamine over norepinephrine (Ultrafast neuronal imaging of dopamine dynamics with designed genetically encoded sensors). Immediately after the final behavioral session, the imaged tissue block is harvested, fiducial landmarks (surface vasculature, window-induced marks) are used to align the imaging field-of-view to the histological slide, and 10x Visium spatial transcriptomics plus H&E histology are collected from the same block. A single-nucleus RNA-seq reference is generated from the contralateral homotopic region of the same animals, minimizing the "platform effect" — systematic mismatch between reference and tissue — that the large deconvolution benchmark identified as a major error source (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics).
The computational arm is a function-guided, continuum-aware deconvolution model. Standard deconvolution writes the observed gene counts at a spot as a weighted sum of discrete cell-type profiles, $Y_s \approx \sum_k w_{sk} X_k$, where $X_k$ is the expression signature of type $k$ and $w_{sk}$ its proportion at spot $s$. I replace the discrete weights with interpolation weights over a continuous cell-state manifold learned from the single-nucleus reference (the manifold on which, for example, the Pvalb/Pthlh interneuron gradient of the striatum lives as a trajectory rather than two clusters; Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq), and I add a coupling term that ties composition to function. With $d_s$ denoting a vector of dopamine-dynamics features extracted at spot $s$ (reward-locked transient amplitude, movement-locked response, decay slope), and $c_s$ the predicted composition embedding, the loss is
$$\mathcal{L} \;=\; \mathcal{L}_{\text{NB}} \;+\; \lambda \sum_{s,t} \big( \operatorname{sim}(d_s, d_t) - \operatorname{sim}(c_s, c_t) \big)^2 ,$$
where $\mathcal{L}_{\text{NB}}$ is the standard negative-binomial count likelihood used by stereoscope and Cell2location (Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography) and $\operatorname{sim}$ is a normalized similarity (correlation for dynamics, cosine for composition). The coupling is symmetric in principle but asymmetric in practice: it can be used at inference time on spots with recorded dynamics, and, once trained, the model predicts dynamics features for spots (or animals) never imaged — the predictive test that makes the claim falsifiable. Setting $\lambda = 0$ recovers a continuum-only deconvolution; collapsing the manifold to cluster centroids recovers a discrete deconvolution — these two knobs define the ablations.
A secondary application tests generality on the norepinephrine system: the finding that microglial P2Y$_{12}$ signaling suppresses norepinephrine transmission to promote sleep (Microglia regulate sleep through calcium-dependent modulation of norepinephrine transmission) relied on Cre drivers targeting all microglia, masking transcriptomic sub-states; the same co-registration protocol with a norepinephrine sensor can ask whether predicted microglial sub-state composition predicts where norepinephrine suppression is strongest. This arm is secondary and would follow the dopamine demonstration.
Baselines
The naive reference point is spot-level raw counts with no deconvolution, plus a morphology-only predictor in the style of DeepSpot2Cell, which uses pathology foundation models and a permutation-invariant network to predict virtual single-cell expression from H&E tiles under spot-level supervision (DeepSpot2Cell: Predicting Virtual Single-Cell Spatial Transcriptomics from H&E images using Spot-Level Supervision). This baseline matters because it is the strongest "static visual features only" competitor: if H&E morphology alone predicts local dopamine dynamics as well as my dynamics-guided model, the functional-modality ingredient adds nothing. It will be run on the identical Visium-plus-H&E slides with matched compute and tuning, using the reported evaluation protocol of cell-expression correlation for the top 100 genes; for context, DeepSpot2Cell reports a 46% improvement in out-of-sample cell-expression correlation for the top 100 genes over its best competitor (scstGCN) in lung cancer, and over 50% improvement in the out-of-distribution breast cancer setting. Beating this baseline means achieving higher held-out ROI dynamics prediction from predicted composition than from predicted morphology features alone.
The strongest existing deconvolution competitor is Cell2location, which the comprehensive benchmark of 18 methods across 50 datasets identified among the top performers (together with CARD and Tangram), and which is particularly effective on large spot arrays (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics); the benchmark reports rankings and practical guidelines but no numeric accuracy values, so no reported result found on Lacuna for Cell2location's numeric accuracy on that benchmark suite. Cell2location will be run on my Visium slices with the same single-nucleus reference, same spot set, and same compute budget, and compared on two axes: deconvolution accuracy against ground truth on simulated data generated with scCube, which can vary spot size and tissue geometry and which showed that deconvolution accuracy degrades as spot size grows (Simulating multiple variability in spatially resolved transcriptomics with scCube); and prediction of held-out dopamine dynamics from its estimated compositions. Beating Cell2location requires matching or exceeding its deconvolution accuracy on scCube ground truth while significantly exceeding its dynamics-prediction correlation — a standard discrete model should capture coarse cell-type effects but miss gradient-driven variation. Tangram serves as an auxiliary baseline for the transcriptome-expansion step, where it reports that 75% of genes were predicted with a spatial correlation greater than 40% in leave-one-out tests on MERFISH data (Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram); Starfysh, the reference-free histology-integrated alternative that reports deconvolution RMSE ≈ 0.09, comparable to reference-based tools (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs), is included because it avoids single-cell reference batch effects, though its archetypal anchors presuppose transcriptionally pure spots that dense neural tissue lacks.
The key ablation isolates each ingredient of my model: the full model versus $\lambda = 0$ (no dynamics coupling) versus discrete cluster references (no continuum), all on identical data and compute. If the full model does not beat both ablations on held-out dynamics prediction, the specific ingredients — not merely extra data — are not earning their place, and the claim should be scaled back to the experimental co-registration contribution alone.
Why not just…?
Why not just run subcellular in situ platforms (MERFISH, Xenium) on the imaged tissue and skip deconvolution entirely? This is the most serious alternative, and the proposal does not argue around it — it incorporates it. In situ panels are targeted: MERFISH scaffolds in the Tangram work measured 254 genes and STARmap 1,020 (Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram), so transcriptome-wide coverage still requires computational expansion, and automated segmentation of intermingled cells in dense neuropil remains a bottleneck. More importantly, the in situ arm answers "which cells are here" but not, by itself, "does composition predict dynamics" — that requires the predictive model and held-out evaluation this proposal builds. The honest resolution is empirical: a MERFISH validation arm with a receptor- and gradient-focused panel runs on a subset of animals, and if a few-hundred-gene panel plus simple regression predicts dynamics as well as the full continuum model, that result will be reported as the simpler sufficient method, with the full model retained only for untargeted discovery settings.
Why not just dissociate the imaged region and do deep snRNA-seq or PatchSeq? Dissociation destroys exactly what the question needs: the physical registry between an imaged location and its cellular neighborhood, a limitation both Tangram and SpaOTsc were built to partially repair after the fact (Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram; Inferring spatial and signaling relationships between cells from single cell transcriptomic data). PatchSeq couples physiology and transcriptomes at single-cell resolution but is low-throughput, ex vivo, and restricted to intrinsic membrane properties of the patched cell (Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq) — it cannot measure extracellular dopamine in a behaving animal or the composition of the surrounding spot. SnRNA-seq is still essential here, but as the reference that parameterizes the continuum, not as the measurement of the imaged tissue itself.
Why not just image one cohort of mice and build atlases in a separate cohort? Parallel atlasing avoids the co-registration engineering but forfeits the within-animal link: expression levels and dopamine dynamics both vary across animals, so cross-cohort correlations cannot distinguish local functional architecture from inter-animal covariation. The decisive metric in this proposal is trained-on-some-animals, tested-on-held-out-animals prediction of ROI-level dynamics from composition — a design that directly subsumes the parallel-atlasing question. If composition-to-dynamics mappings generalize across animals, then atlas-level inference is validated; if they do not generalize, that is itself an important negative result showing that functional–transcriptomic coupling is animal-specific, which no parallel design could have detected.
Proposed Experiments and Decisive Checks
The experimental sequence is as follows. First, the co-registration protocol is validated in a small pilot (dLight1 motor cortex, Visium post-hoc) by quantifying registration error between imaging fields and histological landmarks; the study proceeds only if spot-level registration error is below one spot diameter. Second, the continuum deconvolution model is benchmarked against Cell2location, CARD, Tangram, and Starfysh on scCube-simulated ground truth with matched spot sizes and cell-mixture geometries (Simulating multiple variability in spatially resolved transcriptomics with scCube), and on the benchmark's seqFISH+/MERFISH-derived ground-truth datasets (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics). Third, the core test: using the same-animal dataset, predict held-out ROI dopamine-dynamics features from predicted spot composition, reporting the correlation against the DeepSpot2Cell-style morphology-only predictor, the $\lambda=0$ ablation, the discrete-reference ablation, and Cell2location compositions. Fourth, a negative control shuffles dynamics features across spots before coupling; if the full model still outperforms ablations under shuffling, the coupling is fitting an artifact, and the claim fails. Fifth, the MERFISH validation arm checks predicted continuum positions and receptor densities against direct in situ measurement on a subset of animals.
The decisive numbers are these: the full model must beat both ablations and all baselines on held-out dynamics prediction (correlation, with cross-animal held-out prediction as the strictest setting), without degrading deconvolution accuracy relative to Cell2location on scCube ground truth. If morphology-only prediction matches the full model, the functional modality is redundant to static structure and the proposal collapses to a measurement paper; if discrete-reference deconvolution matches the continuum model on gradient-rich simulated tissue, the continuum ingredient is unnecessary; if neither holds, the result establishes that live neuromodulator signaling carries information about tissue molecular architecture that no static method — sequencing, histology, or their combination — can recover, and opens a route to testing, in the norepinephrine system (Microglia regulate sleep through calcium-dependent modulation of norepinephrine transmission), whether masked cellular sub-states are the hidden organizers of neuromodulatory microenvironments.
References
- Ultrafast neuronal imaging of dopamine dynamics with designed genetically encoded sensors
- Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq
- Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram
- Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Inferring spatial and signaling relationships between cells from single cell transcriptomic data
- Inferring pattern-driving intercellular flows from single-cell and spatial transcriptomics
- Microglia regulate sleep through calcium-dependent modulation of norepinephrine transmission
- DeepSpot2Cell: Predicting Virtual Single-Cell Spatial Transcriptomics from H&E images using Spot-Level Supervision
- Simulating multiple variability in spatially resolved transcriptomics with scCube
- Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs