Astrocytes are not one cell type. The same glial cell manages metabolic sensing in the hypothalamus, clears glutamate in the cortex, and releases GABA in the striatum, with each region running a different molecular program — Lee et al. showed hypothalamic astrocytes synthesize GABA within 30 minutes of fatty-acid exposure while cortical astrocytes do not, and Park et al. showed a high-fat diet switches on a NAD-salvage pathway in hypothalamic astrocytes that is entirely absent in cortex and hippocampus. Yet nearly every pathway-level conclusion about these cells is drawn from one measurement tier: RNA sequencing, run through functional enrichment tools that map long gene lists onto curated pathway labels. Two results say this is unsafe. In human cortex, protein diversity is about 1.6-fold higher than transcript diversity, meaning regional specialization is dominated by post-transcriptional mechanisms (Zheng et al.); and in inter-regional brain proteomics, energy-metabolism pathways diverge sharply at the protein level while synaptogenesis signaling looks conserved (Davidson et al.).
The problem is that no one has measured, for a single purified cell type taken from multiple brain regions, how often pathway calls made from RNA are wrong at the protein level. Bulk proteomic atlases blend astrocytes with neurons (Zheng et al.; Davidson et al.), astrocyte-only nuclear sequencing projects measure no protein at all (Serrano-Pozo et al.), and the one study of astrocyte local translation stayed in cortex only (Boulay et al.). This proposal builds the missing matched measurement and uses it to answer a concrete mechanistic question that RNA data alone cannot.
Core Claim
If we isolate astrocytes from four mouse brain regions (hypothalamus, hippocampal dentate gyrus, striatum, cortex), profile each sample simultaneously with translating-ribosome RNA sequencing and deep quantitative proteomics, and run the same functional enrichment analysis on both tiers, then (a) a measurable fraction of RNA-derived pathway calls will fail at the protein level in a region-dependent way, and (b) a region-specific RNA-to-protein transfer map will predict which calls fail. The payoff is a correction layer for functional enrichment: before trusting an enriched pathway term from astrocyte RNA-seq, check whether its genes are in the region-specific set known to break the RNA-protein link. Two concrete biological questions ride on the same data: the unexplained paradox that losing the BEST1 channel makes GABA accumulate in hippocampal and thalamic astrocytes but decrease in striatal astrocytes (Joo et al.), and whether the hypothalamic NAD-salvage response to a high-fat diet (Park et al.) is fully visible from RNA alone.
Research Program
Three coordinated components, all in mice to allow perturbations impossible in human tissue.
First, the matched atlas. For each region we collect astrocyte nuclei by fluorescence-activated nuclei sorting, as used to resolve nine astrocyte subclusters in the Alzheimer's progression atlas (Serrano-Pozo et al.). In parallel, from littermates, we collect astrocyte translating mRNA via TRAP — pulling down epitope-tagged ribosomes from astrocytes — because TRAP captures cytoplasmic and distal-process mRNA pools that nuclear sequencing structurally discards; Boulay et al. showed cortical astrocytes localize 337 transcripts to their vascular endfeet with 650 more bound to translating ribosomes there, none of which survive nuclei-only isolation. The protein tier is deep DIA mass spectrometry on sorted astrocyte material using pressure-cycling sample preparation, the method behind 11,772 proteins across 2,367 human brain sites (Zheng et al.); our acceptance threshold for technical reproducibility is a replicate Pearson correlation at or above the 0.868–0.929 range reported for inter-regional brain proteomics (Davidson et al.). We profile both sexes: the NAMPT phenotype was characterized only in males (Park et al.) and the fasting atlas only in females (Al-Nusaif et al.), so the field has never had both sexes in one design.
Second, the two-tier enrichment protocol and the discordance measurement. From the matched data we derive differentially expressed gene lists (RNA tier) and differentially abundant protein lists (protein tier) for each pairwise regional contrast, then run standard overrepresentation analysis on Gene Ontology and KEGG terms on both, with matched background sets so the comparison is not confounded by detection differences. A pathway call is an enriched term at 5% false-discovery rate. The headline statistic is the concordance rate: the fraction of RNA-tier calls whose member proteins move coherently in the protein tier, reported per region contrast. We then fit a simple region-specific transfer function, predicting log protein abundance from log RNA abundance plus a region effect per gene, and test — under cross-validation on held-out samples — whether enriching on predicted protein abundance recovers the measured protein-tier calls better than enriching on raw RNA does. This is one regression per gene; the point is calibration, not a large model.
Third, the mechanistic anchors. For the BEST1 paradox, we measure mRNA, protein, and enzymatic activity of the GABA-handling enzymes — 4-aminobutyrate aminotransferase (ABAT, the degradation enzyme), monoamine oxidase B, and glutamate decarboxylase — in striatal, dentate, and hypothalamic astrocytes. Joo et al. hypothesized regionally high ABAT catabolism to explain why striatal astrocytes lose GABA when its exit channel is removed while hippocampal astrocytes accumulate it; no one has measured ABAT protein across these regions. For the metabolic perturbation, a 12-week high-fat-diet cohort (Park et al.) lets us ask whether RNA-only enrichment in hypothalamic astrocytes would have found the NAMPT/CD38 pathway and whether the protein tier reveals compensatory routes the RNA tier missed — the unbiased, genome-wide follow-up that Park et al. explicitly did not perform.
Baselines
The naive reference point is what most of the field does today: bulk microdissected tissue RNA-seq per brain region, followed by standard Gene Ontology overrepresentation analysis, with astrocyte signal inferred informatically. We will run this pipeline on aliquots of the same dissected regions before astrocyte isolation. It is the strongest "do nothing" competitor because it is cheap and dominant in practice, and the comparison is fair by construction — same tissue, same sequencing budget per region, same enrichment tooling. It beats us if sorting astrocytes changes no pathway call; we beat it if it misses a pathway call we recover, which is what the dilution record predicts: bulk sequencing hid the VGLUT1-expressing astrocyte subpopulation entirely until single-cell resolution found it (de Ceglia et al.). No reported result found on Lacuna for a bulk-versus-sorted pathway-call error rate, which is precisely why this measurement is worth making.
The strongest existing method as a competitor is the astrocyte snRNA-seq atlas strategy of Serrano-Pozo et al., who profiled 628,943 astrocyte nuclei across five human brain regions and resolved nine subclusters. We will run their design (FANS, nuclei-only, same subclustering and enrichment tooling) on our mouse material so the only intended difference is the added tiers — TRAP translatome and proteomics. Beating it means two things: identifying specific pathway calls that the nuclear-transcriptome tier makes and the protein tier contradicts in at least two regions, and recovering distal-process transcripts (against the cortical endfeetome gene list of Boulay et al.) that nuclei-only isolation cannot contain.
The strongest enrichment-software competitor is DeepSyn, which replaces static Gene Ontology curation with a machine-learned knowledge graph of 85,000 entities mined from 16.7 million PubMed abstracts and reported 0.74–0.96 AUC across functional tasks, outperforming co-occurrence baselines by at least 18% in recovering Gene Ontology categories (Wang et al.). We will run DeepSyn on the identical differentially expressed gene lists. It is the strongest competitor because it is the best-documented attempt to fix enrichment's context blindness — but it still operates on RNA-only input, so it cannot know whether a pathway's proteins actually changed. If DeepSyn's RNA-only calls match the measured protein tier as well as our calibrated pipeline does, our calibration layer adds nothing and the proposal's methodological claim fails; that is the falsification condition, stated plainly. Finally, the key ablation withholds the protein tier from our own pipeline: run the TRAP-RNA tier plus enrichment, then ask how many calls the protein tier would have revised. That number, by itself, is the paper's central result.
Why not just…?
Why not just do bulk RNA-seq per region and deconvolve astrocytes computationally? This is the cheapest design, and it fails for a documented reason: bulk tissue sequencing diluted the transcripts of the rare VGLUT1-positive astrocyte subpopulation below detection until single-cell methods recovered it (de Ceglia et al.), and bulk homogenates confound glial proportion shifts with genuine expression shifts (Serrano-Pozo et al.). Since our target statistics — per-region RNA-protein discordance — require astrocyte-purified protein, bulk data cannot supply the denominator at any price. We still run it as the naive baseline so the cost-benefit is measured, not asserted.
Why not just run RNA-seq and standard GO/KEGG enrichment, and skip the proteomics? Because the two best-available numbers say the tiers diverge where it matters most: protein diversity exceeds transcript diversity roughly 1.6-fold across human cortex (Zheng et al.), and energy-metabolism pathways — exactly the pathways regional astrocyte biology is about — diverge at the protein level while other pathways look flat (Davidson et al.). Post-transcriptional repression, where mRNA holds steady while protein drops, is a well-attested mechanism (Guan et al.; Hu et al.). If our matched atlas turns out to show near-perfect concordance, we will report that RNA-only enrichment is validated for astrocytes and the field can stop paying for mass spectrometry; either outcome is informative, but only the experiment can tell you which.
Why not just run targeted RT-qPCR and Western blots for the candidate enzymes (ABAT, MAO-B, NAMPT)? Honestly, this would likely resolve the BEST1 paradox on its own, and it is what the candidate-gene studies actually did (Park et al.; Joo et al.). But it cannot answer the general question — which RNA-derived pathway calls to trust in a new region or condition — because it requires guessing the candidates in advance, and the field's own record shows the important players are often not the guessed ones (the NAMPT pathway surfaced only after targeted probing, and Cluster 7 only after unbiased single-cell work). We therefore include the targeted enzymatic panel as the validation arm, not the whole study, and the high-fat-diet arm is exactly the unbiased version of what Park et al. left open.
Decisive Checks
Three outcomes decide the proposal. First, the concordance rate and its regional variation, with the ablation-derived revision count as the headline figure; the ABAT panel is the mechanistic proof that discordance has consequences, if striatal astrocytes show higher ABAT protein and activity than hippocampal ones, resolving the Joo et al. paradox, and mRNA alone fails to predict the direction. Second, the calibrated transfer function must beat raw RNA at predicting held-out protein-tier pathway calls; if it does not, the useful residue is still the discordance map itself as a public caution-list. Third, in the high-fat-diet cohort, the hypothalamic protein tier must confirm the NAMPT/CD38 response that Park et al. established, and the two tiers must be compared on whether compensatory metabolic routes are visible in one and not the other. All raw data, the region-specific transfer coefficients, and the per-term concordance table are released as a reusable resource, so any future astrocyte RNA-seq study can look up which of its enriched terms survive translation in that region.
References
- Fatty Acid Increases cAMP-dependent Lactate and MAO-B-dependent GABA Production in Mouse Astrocytes by Activating a GαsProtein-coupled Receptor
- Hypothalamic astrocyte NAD+ salvage pathway mediates the coupling of dietary fat overconsumption in a mouse model of obesity
- A spatially-resolved human brain proteome atlas for understanding function and disease
- Inter-Regional Proteomic Profiling of the Human Brain Using an Optimized Protein Extraction Method from Formalin-Fixed Tissue to Identify Signaling Pathways
- Astrocyte transcriptomic changes along the spatiotemporal progression of Alzheimer’s disease
- Translation in astrocyte distal processes sets molecular heterogeneity at the gliovascular interface
- Generation of Astrocyte-specific BEST1 Conditional Knockout Mouse with Reduced Tonic GABA Inhibition in the Brain
- Elucidating brain metabolic changes during short-term fasting: a comprehensive atlas
- Specialized astrocytes mediate glutamatergic gliotransmission in the CNS
- Deep functional synthesis: a machine learning approach to gene functional enrichment
- Up‐regulation of miR‐10b‐3p promotes the progression of hepatocellular carcinoma cells via targeting<scp>CMTM</scp>5
- MiR-361-3p regulates ERK1/2-induced EMT via DUSP2 mRNA degradation in pancreatic ductal adenocarcinoma