Is It More Glia or Angrier Glia? Separating Cell Abundance from Cell State in Neuroinflammation Omics

A standard design in neuroinflammation research grinds up brain tissue from a diseased or treated animal, measures thousands of transcripts (RNA sequencing) and small molecules (metabolomic profiling), and asks which biological pathways are "enriched" among the molecules that changed. This test is called functional enrichment analysis. It checks whether genes from a known pathway, such as inflammatory signaling, are overrepresented among the changed genes. The answer that comes back is almost always some version of "inflammation went up." There is a confound in this answer. In neurodegeneration the immune cells of the brain, microglia and reactive astrocytes, multiply and move in. A tissue sample with twice as many microglia will show more inflammatory transcripts and different metabolites even if no individual cell has changed its behavior. Bulk enrichment cannot tell "more glia" apart from "each glial cell has switched to a toxic state." Yet the second claim is the mechanistic one that papers usually make.

This note proposes to build and validate a method that pulls the two apart for transcripts and metabolites together. First, we measure what each glial cell type contributes to the transcript and metabolite pool. We do this in an engineered human brain culture, removing or stimulating one cell type at a time. Second, we use these reference profiles to split tissue-level changes into an abundance part and a per-cell state part. Enrichment analysis is then run only on the state part. The deliverable is a re-analysis tool, together with a direct test of how many published "glial inflammation" enrichments survive once cell numbers are accounted for.

Core Claim

Our claim is that a substantial share of inflammatory pathway enrichment in bulk neuroinflammation omics reflects shifts in glial abundance rather than changes in per-cell state. We further claim that metabolites are where this confound is least addressed and most damaging. For transcripts, researchers can at least check marker genes or run single-cell sequencing. A metabolite in a bulk extract carries no label saying which cell made or consumed it. We propose to attach that label experimentally and then use it statistically.

The motivating biology comes from the review Elucidating the Interactive Roles of Glia in Alzheimer's Disease Using Established and Newly Developed Experimental Models. It describes a feed-forward loop between astrocytes and microglia, the astrogliosis–microgliosis axis. In this loop, microglial TNF, IL-1α and C1q push astrocytes into a neurotoxic state. The review also reports that the extent of glial activation tracks cognitive decline better than plaque burden. Both statements are about abundance and state together, and current bulk designs cannot separate them. The imaging study Cyclic multiplex fluorescent immunohistochemistry and machine learning reveal distinct states of astrocytes and microglia in normal aging and Alzheimer's disease resolves glial states in tissue with up to 16 protein markers, which shows the distinction is real at the cellular level. It does not connect those states to transcriptome-wide or metabolome-wide pathway claims. We use it as independent ground truth for cell composition, not as a competitor.

In the model can be stated in one line. For a bulk measurement $y_g$ of gene or metabolite $g$ in a sample,

$$y_g = \sum_k f_k\,(\mu_{kg} + \delta_{kg}) + \varepsilon_g,$$

where $f_k$ is the fraction of cell type $k$ in the sample, $\mu_{kg}$ is the contribution of a baseline cell of type $k$ to $g$, $\delta_{kg}$ is that cell type's state-induced change, and $\varepsilon_g$ is noise. For metabolites, $\mu_{kg}$ can be negative, because cells consume as well as secrete. Naive enrichment tests changes in $y_g$. We test changes in $\delta_{kg}$.

Research Program

The program has three parts.

The first part is the reference atlas. We use the engineered immuno-glial-neurovascular culture from Engineered 3D immuno-glial-neurovascular human miBrain model. It assembles separately differentiated human stem-cell-derived neurons, astrocytes, microglia, oligodendrocytes, pericytes and vascular cells in a brain-like gel. Because each cell type is added independently, we can build cultures with a type left out, or with a titrated amount of it. Differentiation protocols for the glial components are reviewed in Differentiation of Glial Cells From hiPSCs: Potential Applications in Neurological Diseases and Cell Replacement Therapy, and patient-line considerations in Phenotyping Neurodegeneration in Human iPSCs.

For each culture we collect paired bulk RNA sequencing, untargeted metabolomics of cells and conditioned medium, and single-nucleus RNA sequencing on a subset. The culture conditions cross two factors. One is the dose of microglia and astrocytes, which changes abundance only. The other is stimulation, which changes state. Stimulation uses LPS, which triggers an immune response, or the TNF/IL-1α/C1q triad that the astrogliosis–microgliosis axis names as the driver of the toxic astrocyte state. From these conditions we estimate $\mu_{kg}$ and $\delta_{kg}$ per cell type. The key output is metabolite attribution: which cell types produce or consume each metabolite at baseline and under stimulation.

The second part is joint deconvolution. We fit the equation above to bulk samples, sharing the fractions $f_k$ across the transcript and metabolite layers of the same sample. This sharing is the real use of pairing the two measurements. Transcripts are rich in cell-type markers and pin down $f_k$. Metabolites, which lack clean markers, then inherit those fractions, and their residual changes can be attributed to state. Pathway enrichment then runs on the estimated $\delta$ terms. It uses pathway definitions that link each enzyme's gene with the metabolites it acts on, so a pathway counts as altered only if its genes and metabolites move consistently in the same cell type.

The third part is re-analysis of published claims. The mouse literature in the file offers concrete test cases. Alternate day fasting alleviates neuroinflammation in diabetic mice by regulating δ-valerobetaine-carnitine-microglia axis via enrichment of Akkermansia muciniphila links a metabolite to microglial inflammation in diabetic mice. Our method asks whether the carnitine-pathway change is attributable to a per-cell microglial state change or to changed microglial numbers. The same question applies to the many "compound X suppresses NF-κB/iNOS/COX-2" studies, including:

These papers motivate the proposal. They establish the inflammatory pathways we test, and their tissue-level conclusions are exactly the kind that abundance could explain. We will generate matched tissue data in LPS-treated and diabetic mice, using mouse-orthologous reference profiles where the human atlas transfers. We will not assert that any specific published result is wrong until the test is run.

Baselines

The naive reference point is standard enrichment on raw bulk fold changes. Transcripts and metabolites are tested separately, and there is no adjustment for cell composition. This is the analysis most papers in the file appear to rest on. We run it on exactly the same samples and pathway definitions as our method. Beating it means lower false-positive enrichment under pure abundance shifts, with no loss of recall under true state changes.

The strongest simple competitor is marker-score regression. Before enrichment, we regress each bulk feature on marker-gene scores for microglia and astrocytes and test the residuals. It is cheap, widely understood, and handles the transcript side directly. It gets the same marker panels and the same samples as our method, and we tune it on the same held-out mixture set. Our method beats it only if it recovers the known state-driven pathways in metabolites with clearly higher precision at matched recall. On transcripts alone we expect the two to be close. Parity there would not count as a win.

The strongest existing method for cell-resolved transcripts is single-nucleus sequencing of the same samples. It is ground truth for transcript state, not a competitor, but it defines the ceiling for the transcript half. No reported result was found on Lacuna for any composition-adjusted enrichment baseline, and none for joint transcript–metabolite deconvolution in glial tissue. We therefore report every baseline number from our own runs.

An ablation isolates the key ingredient. We run our deconvolution with the transcript and metabolite layers fit independently, with no shared $f_k$. If shared fractions do not improve metabolite attribution over independent fitting, the claim that paired omics helps is refuted.

Why not just…?

Why not just run single-cell RNA sequencing and skip deconvolution? For transcripts this works, and we use it as ground truth. It says nothing about metabolites, which is the layer the published axis-style claims increasingly rest on. The delta-valerobetaine–carnitine claim in the fasting study is an example. Single-cell data can only serve as a proxy for metabolites through enzyme expression, and that proxy is itself tested in our ablations. We measure how often enzyme-expression attribution agrees with the attribution from the removal and titration experiments.

Why not just regress out marker-gene scores? This is the real threat. If marker regression matches our method on metabolite attribution, the atlas is unnecessary and we will say so. We expect it to fall short for two reasons. Metabolites have cell-type-specific signs, since one type secretes what another consumes, and a single fraction covariate cannot capture that. Also, stimulated glia change their own marker expression, which contaminates the covariate. The mixture experiment below settles the question directly.

Why not just correlate metabolites with transcripts across samples in a joint network? Correlation networks inherit the abundance confound wholesale. When microglia increase, all of their products rise together, so microglial transcripts and metabolites correlate whether or not any cell changed state. The pure-abundance arm of our mixture design is a null in which such networks should still report couplings, and we will quantify how many they do.

Proposed Experiments

The decisive check is a mixture design with known answers. We combine unstimulated and stimulated miBrain cultures, and cultures with and without microglia, at defined ratios. This yields samples whose true abundance shifts and true state changes are known.

In the pure-abundance arm, only unstimulated microglia are added. Correct behavior is no inflammatory state enrichment. We record the false-positive rate for the naive analysis, for marker regression, for the correlation network, and for our method.

In the pure-state arm, the cell numbers are fixed and the TNF/IL-1α/C1q triad or LPS is applied. Correct behavior is recovery of the pathways that single-nucleus sequencing and targeted metabolite checks confirm. We record precision and recall per cell type.

The two effects are then crossed at graded ratios. We plot pathway recovery against composition error, measured against the known mixing fractions.

In tissue, we compare deconvolved microglial and astrocyte fractions against imaging-based counts obtained with the cyclic multiplex approach. For mouse tissue we use standard glial marker staining.

The headline result is the fraction of naive inflammatory enrichments in the LPS and diabetic-mouse tissue that vanish after composition adjustment. If that fraction is near zero, the confound is small in practice, which is itself a useful result. If it is large, published mechanistic claims resting on bulk enrichment need revisiting.