Enzyme × Access: Predicting Whether Glutamate-Clearance Boosters Survive the Jump from Dish to Brain

Every excitatory signal in the vertebrate brain ends with a clean-up job. Glial cells sweep the neurotransmitter glutamate out of the synapse through surface transporters and then neutralize it internally by converting it to glutamine with the enzyme glutamine synthetase (GS), a loop sustained by further enzymes such as glutamate dehydrogenase (GDH) and the mitochondrial sirtuin SIRT4 (Glial Regulation of Glutamate Homeostasis Against Excitotoxicity). When this clearance fails, glutamate spills onto extrasynaptic receptors that trigger neuronal death cascades. Hundreds of compounds that stabilize these enzymes in cultured cells have nonetheless failed in human stroke and neurodegeneration trials (The Role of Glutamate Receptors in Ischemic Stroke). I argue the field is blind for two specific, fixable reasons: the assays that measure enzymatic capacity average over millions of cells and so cannot see which astrocyte subpopulations are failing, and the culture systems used for screening lack the physical contact between glia and synapses that gates whether an enzyme can act at all. The first problem is cellular heterogeneity masking; the second is the in vitro–in vivo translation gap.

The proposal in one sentence: build and validate a model in which a compound's in vivo effect on glutamate clearance factorizes into its effect on enzymatic capacity per astrocyte subpopulation, times the fraction of synapses those subpopulations still physically cover — and test whether this two-factor model predicts in vivo outcomes that in vitro potency alone gets wrong. The key motivation comes from a contradiction in the literature: bulk homogenates show global GS suppression after excitotoxic insult (Rapid neonatal hearing screening using maximum length sequence auditory brainstem response, chirps and automated signal detection), while nanoscale imaging in Huntington's model mice shows astrocytes selectively retracting from mature synapses while staying engaged with immature ones (Astrocytic engagement of the corticostriatal synaptic cleft is disrupted in a mouse model of Huntington's disease). Bulk measurements cannot distinguish "every astrocyte is a bit worse" from "a subpopulation has disconnected entirely," and the two scenarios predict opposite outcomes for an enzyme-boosting drug: in the first, boosting GS helps everywhere; in the second, it helps only where glia still touch the synapse.

Core Claim

Define effective regional clearance as a sum over astrocyte functional states $k$:

$$C(d, r) \;=\; \sum_k \pi_k(d, r)\; a_k(r)\; E_k(d)$$

where $\pi_k(d, r)$ is the fraction of astrocytes in state $k$ at brain region $r$ under drug $d$, $a_k(r)$ is that state's synaptic access fraction (the share of synapses its processes envelop), and $E_k(d)$ is its enzymatic clearance capacity under the drug. The claim is that standard screening predicts only $E$, that $E$ alone systematically overestimates in vivo efficacy in structurally decoupled disease tissue, and that $\pi$, $a$, and $E$ can each be estimated with existing computational machinery — count-based deconvolution for $\pi$, imaging-derived structural priors for $a$, and perturbation-transcriptomics transfer learning for $E$ — and then validated against direct optical measurement of glutamate clearance kinetics in vivo.

No existing work assembles these pieces. DeepSADR bridges cell-line screens to patient response with adaptive, permutation-invariant readouts, but only in cancer pharmacogenomics and without any notion of cell-state composition or physical access (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). SCAD similarly transfers bulk drug-sensitivity models to single cells to find resistant subpopulations, again only in cancer (Enabling Single‐Cell Drug Response Annotations from Bulk RNA‐Seq Using SCAD). POPPY and Count Bridges deconvolve bulk and spatial profiles into cell-level states with ontology alignment and discrete count modeling, respectively, but have been evaluated on cancer and immune atlases, never on glial metabolic states (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, Count Bridges enable Modeling and Deconvolving Transcriptomic Data). The glial-biology papers that do connect enzymes to in vivo outcomes — the Drosophila sulfite oxidase shopper work, the retinal SIRT4 study, and the astrocytic Rac1 review — all measure GS, GDH, or clearance by bulk Western blot, homogenate activity, or gross phenotype, and explicitly leave the subpopulation structure unresolved (The sulfite oxidase Shopper controls neuronal activity by regulating glutamate homeostasis in Drosophila ensheathing glia, SIRT4 Is Highly Expressed in Retinal Müller Glial Cells, Mechanistic Role of Astrocytic Rac1 Protein in Alzheimer's Disease). The proposal's role assignment is deliberate: enzymatic neurotransmitter regulation is the mechanism, heterogeneity masking is the estimation problem, and the translation gap is both the motivation and the evaluation regime.

Research Program

The first component builds an astrocyte enzymatic-state atlas. We define a "clearance module" gene set (GS/GLUL, GDH/GLUD1 and GLUD2, SIRT4, the transporter SLC1A2, and the Rac1/NLRP3 inflammatory axis) and apply count-based deconvolution, in the style of Count Bridges' Poisson birth-death modeling on integer-valued reads, to bulk and spatial transcriptomic profiles of mouse brain under excitotoxic and neurodegenerative conditions (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). The output is $\pi_k(r)$: the regional mixture of astrocyte states that are enzymatically competent versus uncoupled, estimated where bulk assays currently report only a single averaged GS level. A crucial design constraint, taken directly from the deconvolution benchmarking study of 18 tools across 50 spatial datasets, is that reference-free methods struggle to separate closely related subtypes, so we use a reference-guided design and guard against the atlas-dependency failure mode — if a diseased glial state is absent from the reference, transcripts get misallocated to normal states (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics, Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). We therefore seed the reference with disease-state profiles rather than healthy atlases alone.

The second component trains the transfer model. The model takes a compound's structural fragments and an in vitro perturbation signature (from a newly generated astrocyte perturbation compendium, supplemented by LINCS L1000, which profiles 978 landmark genes across thousands of chemical perturbations) and predicts two things separately: the per-state enzymatic response $E_k(d)$, via an adaptive, permutation-invariant readout in the spirit of DeepSADR's Set Transformer module, and the state-composition shift $\pi_k(d, r)$. We deliberately split the readout into these two heads rather than predicting a single efficacy score, because the single-score formulation is exactly what conflates capacity with access. The structural access prior $a_k(r)$ comes from published imaging: the R6/2 serial block-face electron microscopy data quantify astrocytic coverage of mature versus immature spines in striatum, giving a region-specific access reduction that flat cultures cannot represent (Astrocytic engagement of the corticostriatal synaptic cleft is disrupted in a mouse model of Huntington's disease). The biophysical bridging study that connected in vitro conductance models to in vivo Neuropixels recordings — separating inhibitory subclasses with 82.5% spiny/aspiny accuracy — is our template for coupling reductionist measurements to in vivo observables through an explicit intermediate model rather than direct correlation (Associations between in vitro, in vivo and in silico cell classes in mouse primary visual cortex).

The third component is validation, which the glutamate-sensor work makes genuinely quantitative: iGluSnFR4f resolves sensory-evoked cortical glutamate transients up to 20 Hz in vivo with a ~26 ms decay time, and the slow variant iGluSnFR4s (decay ~153 ms, with midbrain fiber-photometry signals ~3.6-fold larger than its predecessor) tracks slower tonic clearance (Glutamate indicators with increased sensitivity and tailored deactivation rates). This gives us a measured in vivo clearance rate to compare against the predicted $C(d, r)$ — the sensor paper also supplies a cautionary methodological note, since its proteins showed reduced dynamic range as purified soluble assays despite high in vivo performance, which is precisely the in vitro–in vivo discordance we are modeling (Glutamate indicators with increased sensitivity and tailored deactivation rates).

Proposed Experiments

The first experiment benchmarks enzymatic-state deconvolution on real spatial data. Using the MERFISH mouse brain dataset and the 50-dataset spatial benchmarking suite, we synthesize mixed bulk profiles from known single cells and measure how well each method recovers the astrocyte state mixture, with a targeted test in which a rare uncoupled state (5–15% abundance, clearance-module-suppressed) is spiked into a background of normal astrocytes — the exact scenario the R6/2 structural data motivate. Success is recovering the spiked state's abundance with higher correlation and lower bias than the strongest reference-guided baselines.

The second experiment tests the coupling-gate hypothesis with a region-contrast design. The model predicts that resveratrol-like GS activators — which upregulate GS and reduce apoptosis in the retina, where Müller glia remain structurally engaged (SIRT4 Is Highly Expressed in Retinal Müller Glial Cells) — should show sharply attenuated benefit in R6/2 striatum, where astrocytes have retracted from mature spines (Astrocytic engagement of the corticostriatal synaptic cleft is disrupted in a mouse model of Huntington's disease). The decisive measurement is iGluSnFR4-based clearance kinetics in both regions before and after treatment: the enzyme-only model predicts equal improvement; the coupling-gated model predicts clearance improvement proportional to residual coverage. A third, cheaper in vivo intermediate is the Drosophila shopper system, where glial-specific sulfite oxidase knockdown slows peristalsis from ~1.2 Hz to 1.12 Hz and stretches muscle burst cycles from 12.4 s to 16.1 s, with rescue by GDH overexpression — a genetic, not pharmacological, perturbation that cleanly isolates the enzymatic branch of the model (The sulfite oxidase Shopper controls neuronal activity by regulating glutamate homeostasis in Drosophila ensheathing glia). We acknowledge honestly that the file documents no existing astrocyte-specific chemical-perturbation transcriptomic compendium ("no reported result found on Lacuna for a CNS glial perturbation screen comparable to LINCS L1000"), so generating one (primary astrocytes, ~50 compounds including resveratrol analogs and propolis constituents, glutamate-challenged and resting conditions) is a required, and substantial, data-generation component rather than a download.

The third experiment is the translational ranking test. We rank a held-out compound set by in vitro GS/GDH activation alone and by the full coupling-gated model, and evaluate both rankings against measured in vivo clearance rescue across the three animal systems. The headline metric is ranking correlation with in vivo outcome and AUC for predicting in vivo success/failure; the model wins only if it specifically corrects the false positives that in vitro potency produces in decoupled tissue.

Baselines

The naive reference point is current practice: rank compounds by in vitro enzymatic activation, measured as GS or GDH activity in cultured glia, and assume the top in vitro hit is the best in vivo candidate. This is the implicit default that produced the clinical failure record reviewed in the ischemic stroke paper, where broadly effective in vitro neuroprotectants failed uniformly in human trials (The Role of Glutamate Receptors in Ischemic Stroke). Any model that does not beat this reference on in vivo outcome prediction — not on in vitro metrics — has added nothing.

The strongest deconvolution competitor is the reference-guided family that the 50-dataset benchmarking study found most accurate overall — CARD, Cell2location, and Tangram — rather than the linear tools a reviewer might first name (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics). The file reports no single headline accuracy number for these tools on a common metric ("no reported result found on Lacuna for a reported accuracy value for CARD, Cell2location, or Tangram on the 50-dataset suite"), so we run them ourselves on our MERFISH-derived synthetic mixtures with the same reference atlas, the same gene set, and the same compute budget as our count-based pipeline. Beating them means two things: lower error on the spiked uncoupled-state recovery task, and, more importantly, calibrated estimates of state proportions rather than only cell-type identities, since the downstream transfer model consumes proportions.

The strongest transfer-learning competitor is the DeepSADR architecture itself, retrained end-to-end on our glial perturbation data with its bipartite fragment–pathway graph and adaptive readout intact. This is the right strongest baseline because DeepSADR is the best-documented bridge from in vitro screens to in vivo response in the file: it reached an AUC of 0.957 on Sorafenib in TCGA patient samples versus 0.811 for the next-best baseline, and removing its adaptive readout collapsed average transfer AUC from 0.856 to 0.662, showing that pooling-based readouts fail under translational distribution shift (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). We give it identical training data, tuning budget, and fragment vocabulary; the only architectural difference is our two-headed output (per-state enzymatic response and composition shift) and the external access prior. It beats us if a single pooled response score predicts in vivo clearance rescue as well as the factorized model; we beat it only if the factorization specifically improves prediction in decoupled regions while matching it elsewhere. The key ablation removes the access term $a_k(r)$ from our own model (setting it to 1 everywhere), isolating whether the structural gate — the proposal's central ingredient — carries real predictive weight or is decorative.

Why not just…?

Why not just use standard linear deconvolution like CIBERSORTx or MuSiC? The count-modeling work showed that linear and continuous baselines mishandle the integer geometry of RNA-seq counts, with Count Bridges outperforming CIBERSORTx and MuSiC on synthetic mixtures and real single-cell deconvolution (Count Bridges enable Modeling and Deconvolving Transcriptomic Data), and the benchmarking study showed reference-free tools fail on closely related subtypes (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics). But the honest answer is that reference-guided tools like CARD and Cell2location are a genuine threat, which is why they — not the linear methods — are our strongest baselines, run on identical data. Our differentiator is not raw deconvolution accuracy but the downstream consumption of state proportions by the transfer model; if the benchmarking tools match our state recovery, we will use them and the proposal's novelty rests entirely on the coupling-gated transfer component.

Why not just measure marker genes (GLUL, SLC1A2, SIRT4) by RT-qPCR or bulk differential expression? This is cheap and standard, and it fails for the exact reason motivating the proposal: a bulk GS measurement cannot distinguish uniform mild downregulation from the loss of a small uncoupled subpopulation, and these two worlds predict opposite drug outcomes. There is also direct evidence that static expression carries less signal than higher-order structure: CellVS-Net predicted drug-disease indications from gene-gene network rewiring better than even experimentally observed "oracle" expression snapshots on DDR-Bench, with Hits@1 of 0.1250 versus 0.0357 for molecular fingerprints (Cell-Level Virtual Screening). Our first experiment makes this falsifiable: if bulk marker levels predict in vivo clearance rescue as well as the deconvolved mixture, the heterogeneity component of the claim is wrong.

Why not just run 2D astrocyte–neuron co-culture screens? Faster and cheaper, but structurally incapable of representing the failure mode at issue: astrocytic retraction from mature spines is a three-dimensional tissue phenomenon documented by electron microscopy in R6/2 striatum and cannot occur on flat plastic without intact dendritic architecture (Astrocytic engagement of the corticostriatal synaptic cleft is disrupted in a mouse model of Huntington's disease). The sensor paper adds a sharp mechanistic precedent: proteins that looked poor as purified soluble assays performed well in their native membrane context in vivo (Glutamate indicators with increased sensitivity and tailored deactivation rates). This alternative is not a threat to the proposal — it is the proposal's control condition, and we run it as the in vitro arm whose rankings we show are miscalibrated.

Why not just align in vitro and in vivo transcriptomes by global correlation or PCA? DeepSADR's own ablation answers this: replacing its permutation-invariant adaptive readout with ordinary pooling dropped transfer AUC from 0.856 to 0.662, because whole-profile alignment washes out the modular fragment–pathway interactions that actually drive response (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). SCAD found the same for bulk-to-single-cell transfer, where adversarial domain adaptation was necessary to reach AUCs up to 0.937 for resistant-subpopulation prediction (Enabling Single‐Cell Drug Response Annotations from Bulk RNA‐Seq Using SCAD). We include a global-alignment variant as an ablation head anyway, because if it matches the factorized model on our (smaller, noisier) glial data, the added structure is not earning its cost.

Decisive Checks and Closure

Three results would make or break the proposal. First, the deconvolution spike-in test: if no method, including ours, can recover a 5–15% uncoupled astrocyte state from realistic bulk mixtures, the whole estimation layer is not yet ready and the project should stop at the spatial-data stage. Second, the region contrast: if resveratrol-class compounds improve striatal clearance in R6/2 mice as much as retinal clearance despite documented spine retraction, the access gate is empirically negligible and the two-factor model collapses to the enzyme-only baseline. Third, the ranking test: the coupling-gated model must beat in vitro potency at predicting in vivo outcomes specifically by fixing false positives in decoupled tissue — if it wins only on average, driven by regions where access is intact, it has not demonstrated the mechanism it claims. The broader payoff, if the checks pass, is a reusable pattern: deconvolve the target tissue into functional states before transferring in vitro pharmacology, and gate enzymatic efficacy by physical access. That pattern is what the cancer transfer models lack and what neurodegenerative drug screening, after decades of failed translation, most needs.