Whether a neurotransmitter excites or inhibits a neuron is not fixed. GABA, the brain's main inhibitory signal, becomes excitatory whenever the intracellular chloride concentration rises enough to flip the direction of chloride flow through GABA-A receptors. This flip underlies newborn-brain development, circadian resetting, chronic pain, and epileptogenesis, and it is executed by a handful of chloride-handling genes — SLC12A5 (the extruder KCC2), SLC12A2 (the importer NKCC1), ANO1/TMEM16A, and CLCN2/CLC-2. The problem is that these flips are intensely local: they occur in specific anatomical subregions (subicular sharp-wave-ripple studies), in restricted cortical columns during bursts (sensorimotor cortex work), and only where reactive astrocytes sit next to vulnerable neurons (Ju et al.). Yet the tools we use to interpret the non-coding variants that regulate these genes — the more than 90% of disease-associated genetic variants that live in regulatory switches rather than protein-coding exons — assume a static, spatially uniform regulatory logic computed from DNA sequence alone.
This proposal builds a model that predicts how a non-coding variant changes the expression of chloride-transporter genes conditioned on where the cell sits in tissue: its cell type and the identity of its physical neighbors. The motivation is a three-way mismatch documented across the literature. Spatial transcriptomics of the epileptogenic hippocampus shows neurons acquire sparse, spatially specific synaptic alterations while glia show diffuse inflammatory activation, but it profiles only 247 pre-selected genes and touches no genetics (the KA-TLE Xenium study). Bulk epigenomics links non-coding variants to active enhancers across tissues, but explicitly aggregates cell types and has no spatial coordinates (GTEx H3K27ac profiling). And sequence models like AlphaGenome predict variant effects from DNA alone, while their own authors flag condition- and microenvironment-specific variant effects, and regulatory influence beyond 100 kilobases, as open failures (AlphaGenome). Nobody has asked the question that sits at the junction: does a non-coding variant near SLC12A5 do something different in a subicular neuron surrounded by reactive astrocytes than in an otherwise identical neuron two hundred microns away?
Core Claim
The claim is that non-coding variant effects on chloride-homeostasis genes are a function of the tissue niche, not just the DNA sequence and the cell type, and that this niche dependence can be modeled and measured. Concretely, we will train a predictor whose input is the local DNA sequence around a variant, the variant allele, a cell-type embedding learned from single-nucleus data, and a spatial-niche embedding learned from physical neighbor graphs in intact tissue, and whose output is the predicted expression change (a log fold-change, following GRAM's "logSkew" convention) of the transporter genes in that cell at that location:
$$\hat{\Delta}_{g} = f\big(E_{\text{seq}}(x \pm v),\; h_{\text{cell}},\; h_{\text{niche}}\big),$$
where $x \pm v$ is the sequence window carrying reference versus alternate allele, $h_{\text{cell}}$ embeds the cell type from nuclear RNA profiles, $h_{\text{niche}}$ embeds the spatial neighborhood from an in-situ tissue graph, and $g$ ranges over the four transporter genes. The falsifiable prediction is that the niche term changes predicted effect sizes and sometimes signs for a substantial fraction of variants — something no current model represents — and that this improves prediction of observed, spatially resolved transporter expression over cell-type-only conditioning.
Background and the Open Gap
Three lines of work bracket this idea without occupying it. GRAM predicts cell-type-specific non-coding variant effects by coupling 515 universal transcription-factor binding models with cell-type TF expression levels, and shows that conservation scores contribute almost nothing to immediate transcriptional impact (GRAM). Its cell types, however, come from dissociated or cultured profiles, so a neuron's neighbors are invisible to it. The KA-TLE study resolves both cellular heterogeneity (31,390 glial cells by single-cell RNA sequencing, 48,221 neuronal nuclei by single-nucleus sequencing) and spatial context (Xenium in-situ imaging of 247 genes across 27 cell types in intact hippocampal sections), and shows that neuronal alterations cluster in specific subregions while glial inflammation is everywhere (the Xenium TLE paper) — but it is purely observational, with no genetic variation, chromatin, or sequence-to-function modeling. GTEx H3K27ac profiling maps 282,000 active regulatory elements and 130,000 histone-acetylation QTL variants, and demonstrates that many disease variants alter enhancer chromatin even when bulk expression QTLs are diluted to invisibility (Multitissue H3K27ac profiling) — but its own authors state that single-cell epigenomics is required for true cell-type resolution, and its samples are homogenates with no coordinates. Our proposal is the intersection none of them covers: genetics plus cell type plus physical space, aimed at the specific gene family whose function is known to invert in a spatially restricted way.
The tension that makes this worth testing rather than merely plausible: functional chloride reversals are strictly localized phenomena, while every deployed variant-effect model assumes spatially invariant regulatory rules. AlphaGenome matched or exceeded prior systems on 22 of 24 track-prediction tasks and improved eQTL sign prediction from 75% to 80% auROC (AlphaGenome), and Enigma reaches 90.4–97.3% of AlphaGenome's zero-shot variant-effect performance at a fraction of the compute (Enigma) — but both are trained on bulk reference tracks. Meanwhile TraitGym shows sequence models collapse exactly where we need them: on complex-trait distal non-coding enhancer variants, where alignment-based CADD reaches 0.85 AUPRC on Mendelian traits but the complex-trait gap is large (TraitGym benchmark). Whether transporter-gene regulation follows GRAM's immediate-TF-biophysics story or TraitGym's deep-constraint story, within specific neural cell types and niches, is unresolved — our experiments are designed to answer that too.
Research Program
The program has three components. First, we assemble a niche atlas of the epileptogenic hippocampus by reanalyzing the KA-TLE Xenium and snRNA-seq data (the Xenium TLE paper) and generating one new dataset: an extended custom Xenium panel on kainic-acid and control hippocampi that adds Slc12a5, Slc12a2, Ano1, Clcn2, and a set of activity- and inflammation-responsive transcription factors to the 247-gene panel. The file does not report the original panel's gene list, so this extension is a prerequisite, not an embellishment — if the transporter genes were already on the panel the first atlas reanalysis alone would partially serve. Second, we build the embeddings: cell-type embeddings from the 48,221-nucleus RNA data using a cross-modal encoder in the spirit of ChromFound, which learned single-cell chromatin representations across 1.97 million cells and 30 tissues (ChromFound); and spatial-niche embeddings from a cell-proximity graph in the spirit of STELLAR, which models cells as nodes and physical adjacency as edges and used that graph to annotate 2.6 million cells across 64 tissue sections (STELLAR). Third, we bolt these onto a GRAM-style predictor: universal TF binding scores for the sequence window, but with TF activity levels read out per cell type per niche from the spatial data rather than from dissociated averages. Training signal comes from pseudo-bulk transporter expression computed per (cell type × anatomical subregion × condition), with variants drawn from brain haQTLs and GWAS loci for epilepsy and chronic pain (GTEx H3K27ac supplies the haQTL variant catalog).
The honest framing of the bet is this: STELLAR showed that adding physical spatial organization to molecular profiles improved cell-state annotation accuracy by 11% and reached 93% accuracy on CODEX imaging of Barrett's esophagus (STELLAR). We are betting that an analogous gain exists one level up — not for annotating cells, but for predicting what a genetic variant does to a cell. If the bet fails, the ablation will show it cleanly.
Baselines
The naive reference point is nearest-gene assignment: map each non-coding variant to the closest transcription start site, as standard genetic epidemiology does. This is the "do nothing" comparator because it uses no model at all, and it is known to be biologically wrong at complex loci — the GTEx H3K27ac study showed causal variants frequently skip adjacent genes to contact distal promoters, and its gLink scoring framework was built precisely because distance rules fail (Multitissue H3K27ac profiling); chromatin looping work reaches the same conclusion across megabases (HiView). No reported performance number for nearest-gene assignment exists on Lacuna for transporter loci, so we will characterize it on our own held-out data. It will be run on the identical variant set as our model, and beating it means simply recovering known distal links that it misses by construction.
The strongest existing competitor is GRAM, the only published model that predicts non-coding variant effects conditioned on cell type: it couples universal DeepBind transcription-factor binding models with cell-type TF expression and predicts expression fold-change, reporting an AUROC of 0.938 for its universal activity score, 0.72 on MPRA variant prediction in GM12878, and 0.66 across multiple cell lines (GRAM). It is the strongest competitor because it already solves the cell-type half of our problem, and if niche conditioning adds nothing beyond cell type, GRAM is what our model collapses to. We will run GRAM on the same transporter-locus variants with its own cell-type TF profiles derived from our dissociated snRNA-seq, with the same sequence windows, the same tuning budget, and matched compute; its MPRA AUROC of 0.72 on GM12878 is the pre-registered number to beat on the shared MPRA-based evaluation, and we additionally require a niche-stratified analysis in which GRAM is flat across niches while our model tracks observed spatial expression. Alongside GRAM we run AlphaGenome zero-shot on the same variants as the genome-wide reference for sequence-only machinery; its reported eQTL sign-prediction improvement from 75% to 80% auROC (AlphaGenome) sets the scale of what a purely sequence-based ceiling looks like, and Enigma's 90.4–97.3% of AlphaGenome performance at ~10.9% of the compute (Enigma) lets us include an efficient sequence-only comparator at equal cost to our own inference.
The key ablation is our own model with the spatial-niche term removed — identical architecture, identical training data, cell-type conditioning only. This isolates exactly the ingredient the proposal is about, and it is the direct analog of STELLAR's demonstrated 11% accuracy improvement from adding spatial context to molecular profiles (STELLAR). Success requires the full model to beat the ablation on niche-stratified transporter-expression prediction; if it does not, the niche hypothesis is wrong and we will say so.
Why not just…?
Why not just run bulk eQTL and haQTL analysis on brain samples, like GTEx? This is the cheapest path, and it is documented to fail in exactly the way our question requires. The GTEx H3K27ac study itself found many disease variants that alter enhancer acetylation with no detectable bulk eQTL, because the expression effect is diluted across unresponsive cell types (Multitissue H3K27ac profiling); its authors explicitly state that single-cell resolution is needed. In neural disease, bulk analysis cannot even tell whether a transporter-gene change is a primary neuronal defect or a secondary astrocyte response (Ying Sun et al.), and the subiculum-scale localization of functional inversions (status epilepticus subiculum work) is far below any bulk sample's resolution. The bulk path can supply the variant catalog — we use it for that — but it cannot supply the measurement.
Why not just use dissociated single-cell or single-nucleus RNA sequencing, without spatial data? This resolves heterogeneity but destroys the very thing our claim depends on. Dissociation severs contact- and proximity-dependent signaling (Xin Shao et al.; Jovan Tanevski et al.), and chloride flips are confined to microdomains: the sensorimotor cortex study showed depolarizing GABA confined to restricted columns during bursts (2012 sensorimotor cortex work). Empirically, adding spatial coordinates to molecular profiles improved cell-state annotation by 11% (STELLAR) — spatial information is not redundant with molecular profiles. This is also the honest caveat for snRNA-seq specifically: nuclei under-sample cytosolic and dendritic mRNA, which matters for transporter isoforms localized to dendrites.
Why not just use off-the-shelf sequence models like AlphaGenome or Enigma directly? These are real threats, and part of our evaluation plan is to run them seriously rather than dismiss them. But two documented failure modes bite our question. On TraitGym, sequence models struggle specifically with complex-trait distal non-coding enhancer variants — the exact class of variant we target (TraitGym), and AlphaGenome's authors name both beyond-100-kb regulation and condition- and microenvironment-specific effects as open limitations (AlphaGenome). A sequence-only model cannot, even in principle, represent "variant X is damaging only when the neuron is surrounded by reactive astrocytes." Our decisive experiment — the co-culture MPRA below — tests whether such niche dependence actually occurs; if variants behave identically across niches in that assay, the off-the-shelf models win and our premise fails.
Why not just assign variants to the nearest transporter gene? Covered under baselines: the GTEx H3K27ac gLink framework exists precisely because causal variants skip neighbors (Multitissue H3K27ac profiling), and 3D chromatin folding brings distal variants into contact across megabases (HiView). Nearest-gene assignment will still be run as the naive floor.
Why not just extend the Xenium TLE study with more genes and no modeling? Descriptive spatial profiling, however large, cannot distinguish which of many candidate variants drives an observed spatial expression pattern. The atlas extension is a component of this proposal; the proposal itself is the conditional predictor and its validation.
Proposed Experiments and Decisive Checks
Experiment one is benchmark-internal: validate the predictor machinery on human MPRA data before touching neural tissue, using the same benchmarks GRAM used. GRAM reports an AUROC of 0.72 on MPRA variant prediction in GM12878 and 0.66 across multiple cell lines (GRAM); our reimplementation must reach parity without the niche term before the niche term is evaluated, so that any later gain is attributable to space and not to engineering. We also evaluate on the TraitGym complex-trait set of 1,140 variants across 83 traits (TraitGym); no reported AUPRC for the complex-trait subset was found on Lacuna, so we will establish that number ourselves for CADD, AlphaGenome, Enigma, GRAM, and our model under identical evaluation code.
Experiment two is the spatial hold-out on the mouse atlas. We train on all anatomical subregions except one, then predict pseudo-bulk transporter expression per cell type in the held-out subregion, comparing the full model, the cell-type-only ablation, GRAM with dissociated profiles, and AlphaGenome zero-shot. The decisive metric is Spearman correlation between predicted and observed Slc12a5/Slc12a2 expression across (cell type × subregion × condition) triples, with the pre-registered success condition being that the full model beats the ablation by a margin at least as large as the 11% spatial-context gain STELLAR demonstrated in cell annotation (STELLAR). A secondary check asks whether predicted variant effects localize to the subiculum and dorsal-horn-like regions where functional inversion is documented (subiculum status epilepticus work; astrocytic GABA tonic excitation studies).
Experiment three is the decisive wet-lab test of niche dependence, the load-bearing claim. We run a massively parallel reporter assay — a method that measures the regulatory activity of thousands of DNA sequences and their variant alleles simultaneously in a pooled experiment — carrying candidate regulatory elements from the four transporter loci, in human iPSC-derived excitatory neurons cultured alone versus co-cultured with activated astrocyte-like cells, and in each case with and without NKCC1 blockade or KCC2 enhancement, which are the pharmacological levers documented in the circadian and pain literature (Choi et al.; Chen et al.). The pre-registered outcome: if a measurable fraction of variants change activity, magnitude, or sign between the alone and co-culture conditions, niche-dependent variant effects are real and sequence-only models are structurally blind to them; if not, the core claim fails and the honest conclusion is that cell-type conditioning, already available through GRAM, is sufficient.
The final check interprets human data: among epilepsy and neuropathic-pain GWAS loci, does our model nominate transporter-regulating variants and effector links that bulk haQTL colocalization misses or misattributes? The comparison set is the 1,070 GWAS–haQTL colocalizations across 46 traits from the GTEx H3K27ac resource (Multitissue H3K27ac profiling); we ask specifically whether our niche-conditioned predictions recover neuronal links that bulk analysis attributed to dominant glial signals — the masking pattern the KA-TLE study demonstrates directly, with diffuse glial inflammation over sparse neuron-specific changes (the Xenium TLE paper). Quantifying that recovery rate against an independent measurement is the study's exit criterion: it is the number that would tell a geneticist whether spatially conditioned interpretation changes which variants they act on.
References
- Status epilepticus induces chronic silencing of burster and dominance of regular firing neurons during sharp wave-ripples in the mouse subiculum
- Columnar distribution of activity dependent gabaergic depolarization in sensorimotor cortical neurons
- Tonic excitation by astrocytic GABA causes neuropathic pain by augmenting neuronal activity and glucose metabolism
- Single-cell, single-nucleus and xenium-based spatial transcriptomics analyses reveal inflammatory activation and altered cell interactions in the hippocampus in mice with temporal lobe epilepsy
- Multitissue H3K27ac profiling of GTEx samples links epigenomic variation to disease
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- GRAM: A GeneRAlized Model to predict the molecular effect of a non-coding variant in a cell-type specific manner
- Enigma: An Efficient Model for Deciphering Regulatory Genomics
- Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics
- ChromFound: Towards A Universal Foundation Model for Single-Cell Chromatin Accessibiltiy Data
- Annotation of spatially resolved single-cell data with STELLAR
- HiView: an integrative genome browser to leverage Hi-C results for the interpretation of GWAS variants
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- New avenues for systematically inferring cell-cell communication: through single-cell transcriptomics data
- Gene selection for optimal prediction of cell position in tissues from single-cell transcriptomics data
- Excitatory Actions of GABA in the Suprachiasmatic Nucleus
- The Role of TMEM16A/ERK/NK-1 Signaling in Dorsal Root Ganglia Neurons in the Development of Neuropathic Pain Induced by Spared Nerve Injury (SNI)