aiMOOC

Aging – Pooling, pseudobulk and non-identifiable interactions

QR-Code

Introduction

When you see thousands of nuclei on a single-nucleus RNA-sequencing plot, how many independently treated animals do you really have? This advanced course teaches you to audit Gene Expression Omnibus (GEO) records before drawing conclusions about aging, microglia, TIMP2, gene-expression changes, and interactions. You will reconstruct sample provenance, separate mice from pools and sequencing libraries, make a defensible Pseudobulk analysis plan, and test whether proposed regression coefficients are identifiable.

Central principle: The unit of independent treatment assignment and sampling determines biological replication. A nucleus, read, barcode, gene, or repeated plot point cannot replace an independent experimental unit. Pooling samples can preserve some comparisons, but it does not create individual-mouse observations. A full factorial model requires observations in the corresponding factorial cells.

The worked audit concerns two distinct public mouse datasets accompanying Hemmer and colleagues' 2026 study of tissue inhibitor of metalloproteinases 2 (TIMP2): GSE297916 and GSE297917. One tests treatment in 20-month-old wild-type male hippocampus by single-nucleus RNA-seq; the other tests neonatal knockout versus wild type using bulk RNA-seq from pooled primary microglia. They are not four cells of a shared genotype-by-treatment experiment. This audit reads the deposited series and GSM metadata and research methods; it does not claim to have recomputed expression estimates from all raw reads.

Illustrative microglial cell, not an image from either GEO experiment. Microsome, Wikimedia Commons, CC BY 3.0.

How to Audit Experimental Provenance

GEO series, samples and underlying units

A GSE is a study-level record, whereas a GSM documents a deposited sample or library. A sample labeled “biological replicate” is not automatically an individual animal. Read the GSM title, characteristics, source, extraction and library protocols, platform, supplementary file names, and SRA links. Then reconstruct this hierarchy:

Animal or independently created pool → nuclei or isolated cells → library or GSM → sequencing reads → computational observations.

Compare the hierarchy with the actual assignment of genotype or experimental treatment. A technical sequencing rerun of the same library adds read depth, not an independent animal. Multiple nuclei from the same pool share treatment, preparation, and possible batch effects. In a mixed-sex pool, you cannot recover the sex of each contributing cell unless donor-identification information exists.

An audit trail you can reproduce

Open each series accession and at least one GSM from each group. Capture the organism, age, sex, genotype, treatment, tissue/material, whether cultures were prepared, pooling text, number of contributing animals, distinct pool identity, library name, assay, count-file provenance, and analytical unit. Record reported and inferred separately: the deposits report pooling but do not provide nucleus-to-animal assignments. Do not infer mice from barcode numbers or assign treatment to a dataset where it was never administered.

The series were submitted in 2025 and made public in June 2026; the linked peer-reviewed article is Hemmer et al., Nature Communications 17, 8173 (2026), DOI:10.1038/s41467-026-74906-z. For source-level traceability, follow GSM record URLs rather than relying exclusively on a figure legend.

Reconstructing GSE297916: Aged Hippocampal Nuclei

Four libraries built from pooled animals

GSE297916 contains four GSM records. Their characteristics report 20-month-old, wild-type C57BL/6 male mice, hippocampus, with systemic TIMP2 or PBS treatment. The extracted material is nuclear RNA; the assay is 10x Genomics Single Cell 3' v3 on an Illumina NovaSeq 6000. Every relevant GSM extraction protocol says that two biological samples of the same treatment were pooled per library, loading 20,000 nuclei in total, described as 10,000 from each source. There are two pooled libraries per treatment, for an implied four contributing mice per arm if those source samples represent distinct animals. The released data do not recover which individual mouse produced which nucleus.

GSM	Source label	Treatment	Library	Defensible independent unit
GSM9001952	biol rep1	TIMP2	TD006476-1	pooled library A
GSM9001953	biol rep1	PBS	TD006476-2	pooled library A
GSM9001954	biol rep2	TIMP2	TD006476-3	pooled library B
GSM9001955	biol rep2	PBS	TD006476-4	pooled library B

Accounting: Four GSMs = four 10x pooled libraries = two independently prepared pools for TIMP2 plus two for PBS. Based on the stated two source animals per pool, there were four animal contributions per arm, but the analysis-ready n is two pools per treatment, conditional on the pools' independence. Neither 20,000 loaded nuclei per library nor the 2,831 microglial nuclei described in the paper supplies additional experimental replication. The sample names rep1 and rep2 label pools, not individually estimable mouse effects. All four records show the same stated batch date, 4.18.23; there is no within-series batch contrast to estimate from that field.

The GSM H5 files contain barcoded nucleus-expression information. Cell Ranger, SoupX, DoubletFinder, STACAS and Seurat are reported in processing metadata. These can identify and organize nuclei, but none reconstructs missing mouse labels after unbarcoded pooling.

What the aged data can answer

Within these four comparably prepared pooled libraries, a TIMP2 versus PBS contrast for aged WT hippocampal cell types is the prospective analysis. It is not a knockout effect, an age effect, or a treatment-by-genotype interaction. Sampling uncertainty must reflect the two independent pools per arm and the very small residual degrees of freedom, not the number of nuclei. The paper's microglial subclusters are biologically interesting, but cell-level p-values that ignore shared pooled-library origin risk Pseudoreplication.

Reconstructing GSE297917: Neonatal Bulk Microglia

Fourteen separate libraries from pooled pups

GSE297917 is bulk RNA-seq, not single-cell or single-nucleus sequencing. Neonatal P6–P7 mouse forebrains were pooled by TIMP2 knockout versus wild-type genotype. Primary microglia were isolated and cultured before total RNA extraction. The extraction protocol specifies seven samples per genotype, each a pool of five pups. GEO lists sex as “pooled male and female” and also uses the term “sex-matched”; the precise sex composition for every pool is not recoverable from these brief fields. The instrument is Illumina HiSeq 4000, with raw gene counts supplied as GSE297917_raw_counts.csv.gz.

Knockout pools (seven GSMs): GSM9001956, GSM9001958, GSM9001960, GSM9001962, GSM9001964, GSM9001966, GSM9001968.

Wild-type pools (seven GSMs): GSM9001957, GSM9001959, GSM9001961, GSM9001963, GSM9001965, GSM9001967, GSM9001969.

These alternating GSM titles are labeled biological replicates 1 through 7 within each genotype. Each deposited sample is one bulk library from one pool of five pups. If pools contain distinct pups and are independently processed, n = 7 pools per genotype, not 35 independent expression profiles per genotype. Five pups in a pool are contributors, not five separately observed RNA-seq outcomes. The total recorded animal contributions are nominally 35 pups per genotype; the metadata alone do not verify unique pup identities or exclude litter-level dependence. Culturing may introduce shared conditions and batch effects; missing litter and per-pool sex details limit adjustment.

Crucially, GSE297917 has no experimental TIMP2-versus-PBS treatment factor. “KO” is genotype, not a drug-treatment indicator. Its supportable contrast is neonatal KO versus WT expression in cultured pooled primary microglia. A conventional bulk count model at the pool-library level is appropriate: filter low-count genes, normalize library sizes, use an appropriate negative-binomial or precision-weighted model, examine sample-level QC, and describe pooling and unmeasured confounders. Do not create artificial cell labels or pseudobulk replicates from this bulk matrix.

Mechanistic context only: public-domain CC0 schematic from Wikimedia Commons/WikiPathways; it is not a dataset-specific pathway result.

Pseudobulk Without Pseudoreplication

The defensible aggregation rule

For GSE297916 alone, start from raw integer UMI counts and a justified nucleus-level cell-type assignment. After QC, background correction, doublet review, consistent cell annotation, and examination of each library's cell-type yield, sum counts for each gene × cell-type or subcluster × GSM pooled library. A separate pseudobulk count profile is then obtained for every observed library–cell-type pair, never for a random split of nuclei. Keep the four GSM identifiers as the sample IDs, retain numbers of nuclei used, and exclude/report cell-type strata lacking adequate nuclei across pools.

For each gene and cell type, the aggregation is:

Pseudobulk count for gene g in library s and cell type c = sum of raw counts for gene g over eligible nuclei assigned to s and c.

For example, if two TIMP2 libraries contain 240 and 390 quality-controlled microglial nuclei, summing expression creates two microglia pseudobulk profiles, not 630 biological replicates. These numbers are a fictional teaching example, not observed GEO counts. Use DESeq2, edgeR, limma-voom, or muscat with the pool-library as independent sample; model treatment for each sufficiently populated cell type. Do not use log-normalized matrices as though they were raw counts, and do not infer mouse-specific effects from pooled 10x barcodes.

A credible reporting boundary

For aged microglia, the simple design is ~ treatment over four pooled library observations. The intercept-plus-treatment matrix has rank 2 and only 2 residual sample degrees of freedom. With n = 2 pools per arm, dispersion estimation, effect-size uncertainty, low-abundance subclusters, and cell-composition analyses are fragile. Report effect sizes, pool-to-pool variation, uncertainty, and sensitivity to filtering rather than equating thousands of cells with high replication. If a cell type disappears from a group, an expression-state comparison may be undefined; a change in cell-type abundance is a separate question requiring sample-level compositional methods.

For neonatal GSE297917 the design is ~ genotype over fourteen independently prepared pooled bulk libraries, with matrix rank 2 and nominal 12 residual library-level degrees of freedom, subject to hidden dependencies. A pseudobulk step is unnecessary because each GSM already measures bulk RNA from pooled microglia. Do not combine the single-nucleus pseudobulks and neonatal bulk libraries in one differential-expression fit just because their gene IDs can be harmonized.

Rank, Confounding and Missing Factorial Cells

The design you might wish to estimate

Suppose the scientific question is whether genotype modifies the effect of TIMP2 treatment. A genuinely comparable 2 × 2 experiment requires four conditions in the same age/material context: WT+vehicle, WT+TIMP2, KO+vehicle and KO+TIMP2, with independent units in each cell. A model can use columns intercept, KO genotype, TIMP2 treatment, genotype × treatment. Its interaction coefficient is the difference between the treatment effect in KO and the treatment effect in WT.

The actual GEO evidence is different:

Context	Genotype	Experimental treatment	Independent outcomes	Tested question
Aged hippocampal nuclei, 20 months	WT	PBS	two pooled libraries	baseline for aged treatment contrast
Aged hippocampal nuclei, 20 months	WT	TIMP2	two pooled libraries	effect of TIMP2 within aged WT
Neonatal cultured microglia, P6–P7	WT	none specified	seven pooled bulk libraries	baseline for neonatal genotype contrast
Neonatal cultured microglia, P6–P7	KO	none specified	seven pooled bulk libraries	KO effect within neonatal cultured microglia

There are no aged KO+PBS or aged KO+TIMP2 nuclei libraries, and no TIMP2- or PBS-treated neonatal WT or KO bulk groups. Furthermore, “no treatment specified” for the neonatal cultures is not automatically equivalent to an aged-mouse PBS-injection control. The studies differ simultaneously in age, tissue/material, isolation, culture, assay, platform, sex specification and genetic manipulation.

An explicit algebraic rank check

Define G as KO genotype (WT is zero) and T as TIMP2 treatment (PBS is zero). If you improperly coerce the neonatal untreated samples to T = 0 and stack all eighteen libraries, the attempted four-column matrix consists of [1, G, T, G×T]. For all observed rows, G×T = 0, since no KO animal has a TIMP2-treatment observation. Under this invalid cross-context coding the mathematical rank is only 3 rather than the required 4. The interaction column is identically zero. Software may drop the column, return a non-estimable coefficient, or show a singular-design warning. This rank demonstration is a diagnosis of missing cells, not approval to fit the stacked analysis.

Within aged nuclei, G is constant WT, so ~ genotype * treatment collapses to ~ treatment (rank 2 on four libraries). Within neonatal bulk, no treatment factor varies or is defined, so ~ genotype is the possible rank-2 model on fourteen libraries. Adding age, material, culture, or platform to the pooled combined data cannot conjure unobserved interactions; context and design factors are inseparable for many proposed effects. The study cannot support a causal age comparison between matched samples because neonatal and aged experiments were not performed with the same factorial treatment and assay.

Permissible and non-identifiable conclusions

Permissible within-study contrasts: aged WT TIMP2 against aged WT PBS at pool-library resolution, and neonatal KO against neonatal WT at pool-library resolution. Use the corresponding platforms and count-generation pipelines independently; these outcomes can motivate hypotheses for later work.

Non-identifiable from the combined two-series design: genotype × TIMP2 interaction, KO response to TIMP2, a genotype effect in aged hippocampal nuclei, genotype-by-age or treatment-by-age effects, and a clean age main effect. An apparent old-versus-neonatal expression difference is confounded by many simultaneous protocol changes. A cell-type-specific treatment effect from the aged series cannot be declared a universal bulk-microglial genotype effect. Confidence intervals and thousands of nuclei do not solve missing design cells.

Minimum Study Design for the Next Test

Restore all conditions within a common context

To estimate genotype × treatment, recruit or generate independent biological units for WT+vehicle, WT+TIMP2, KO+vehicle and KO+TIMP2 at a prespecified common age, tissue, sex strategy, and assay protocol. Randomize treatment within genotype, balance processing batches and acquisition order, preregister the main comparison, and track mouse IDs, litters, library IDs and cell-type labels. Use at least three independent mice or genuinely independent pools per factorial cell as a conceptual minimum for replication; a real study should generally plan more, according to biological variation and an effect-size-based power calculation. One library per animal and separate mouse barcodes are preferable for single-nucleus studies.

When individual mice must be pooled, use independently constructed, non-overlapping pools as the experimental replicates and be explicit that inference targets pool-level means. Multiple nuclei within a pool improve characterization of that pool; they do not add treatment-independent degrees of freedom. If animal-level inference is needed, collect separate animal-level libraries or validated donor multiplexing.

To identify age × genotype × treatment as well, repeat all four cells at each of at least two prespecified ages, holding isolation and sequencing approaches comparable. This yields eight factorial condition cells, each requiring its own biological replication. Include both sexes when appropriate, define the estimand, handle litter and cage structure, and do not confuse repeated measurements with independent assignments. The arithmetic size of a factorial matrix is less important than verified nonempty, independent, balanced experimental groups.

Illustrative microglia; Explore Biology, Wikimedia Commons, CC BY 4.0. Images illustrate cell biology, not replicate counts or experimental findings.

Interactive Tasks

Quiz: Test Your Knowledge

Which experiment is GSE297916? (Single-nucleus RNA sequencing of aged WT hippocampus after TIMP2 or PBS treatment) (!Bulk RNA sequencing of neonatal KO and WT microglia) (!A factorial KO by treatment experiment in aged mice) (!A randomized human aging clinical trial)

What is the defensible sample number per GSE297916 treatment arm? (Two independently prepared pooled libraries) (!Twenty thousand independent mice) (!Every captured nucleus) (!Fourteen bulk RNA libraries)

What does one GSM in GSE297917 represent? (One bulk library made from a pool of five neonatal pups) (!One nucleus from an aged hippocampus) (!Five separately sequenced and labeled mice) (!One TIMP2 injection)

Which contrast is available in GSE297917? (Knockout versus wild type in neonatal pooled primary microglia) (!Aged knockout TIMP2 versus PBS) (!An age by genotype interaction) (!Drug treatment within aged knockout microglia)

What is the appropriate grouping for snRNA pseudobulk counts? (Cell type and independent pooled library identifier) (!Each nucleus as a separate animal) (!Randomly assigned groups of sequencing reads) (!Genotype without any library identity)

What should the pseudobulk procedure sum? (Raw gene counts within each library and cell type) (!Log normalized expression values across both experiments) (!Cell type percentages over unrelated assays) (!P values across cell types)

What is the rank of the intercept plus treatment design for four aged pooled libraries? (Two) (!Four) (!Twenty thousand) (!Zero)

Why is the joint genotype by treatment interaction not estimable? (No knockout sample has an observed TIMP2 treatment) (!The interaction is known to equal zero biologically) (!Nuclei are larger than RNA molecules) (!There are too many independently randomized conditions)

Why must age effects not be estimated by directly comparing these two series? (Age is confounded with material preparation and sequencing platform) (!Both series use identical animals and protocols) (!Neither series contains any experimental contrast) (!GEO prohibits all aging research)

What minimal new comparison makes a genotype by treatment interaction estimable? (Replicated wild type and knockout groups under both vehicle and TIMP2 in one comparable context) (!Extra sequencing reads from an existing pooled library) (!More microglial nuclei from only treated WT mice) (!A renamed dataset without new biological conditions)

Memory Game

GSM	Deposited experimental sample record
Independent pool	Distinct mixture of contributing animals represented by one observed sample
Pseudobulk	Summed counts over cells of one type within one sample
Library	Prepared collection of molecules used for sequencing
Confounding	Different explanatory factors change together and cannot be isolated
Matrix rank	Number of linearly independent design columns

Drag and Drop

Match the correct terms.	Meaning
Aged treatment comparison	TIMP2 versus PBS within wild-type hippocampal nuclei pools
Neonatal genotype comparison	Knockout versus wild type in pooled primary microglia
Donor provenance	Trace an animal to its pool and sequencing library
Pseudobulk profile	Raw gene counts aggregated per cell type and pooled library
Missing interaction	Absent crossed genotype and treatment observations

Crossword Puzzle

Pooling	Combining material from several animals before measurement
Nuclei	Subcellular objects profiled in the aged hippocampus assay
Genotype	Genetic status distinguishing knockout from wild type
Library	Sequencing preparation associated with a deposited sample
Confounding	Mixing two explanations so that a unique effect cannot be isolated
Pseudobulk	Count aggregation by sample and cell identity

LearningApps

Use this topic-based LearningApps discovery page to find or adapt exercises; it is a search, not a verified bespoke exercise.

Cloze Text

The GEO series GSE297916 profiles aged hippocampal

rather than neonatal cultured microglia.

Each aged sequencing library contains a

of two biological sources.

There are two independent library-level replicates per aged

group.

The GSE297917 assay is bulk

sequencing of neonatal primary microglia.

Each neonatal bulk sample combines material from five

.

The seven observations in each neonatal genotype arm are seven separately prepared

.

For single-nucleus differential expression, sum raw counts within each cell type and

.

Testing every nucleus as an independent mouse would cause

.

A factorial model requires an identifiable

column for a genotype-dependent treatment effect.

When the required crossed groups are absent, the coefficient is not

.

Open Tasks

Easy

GEO metadata: Create a provenance worksheet with all four GSM identifiers, treatment groups and library names from GSE297916, attaching links to the source records.
Experimental unit: Draw a visual hierarchy from aged mouse to pooled nuclei to barcode and explain which objects may count as independent replicates.
Microglia: Write a short illustrated explanation of how the aged nucleus assay differs from neonatal cultured microglia bulk sequencing.
Pooling: Make a diagram showing a library built from two adult mice and another from five neonatal pups, with accurate labels and a warning about n.

Standard

Bioinformatics: Make a fourteen-row GSM-to-genotype table for GSE297917 and identify its seven independent sample pools per genotype.
Single-cell sequencing: Draft an annotated pseudobulk workflow that sums raw counts only within nucleus subtype and pooled library, with filtering and QC decision points.
Research methods: Interview a sequencing-core scientist or use published core documentation to describe how multiplexing might preserve individual mouse origin.
Design matrix: Write the two within-series model formulas and calculate matrix ranks and nominal residual sample degrees of freedom.

Advanced

Statistical identifiability: Construct the naive joint four-column genotype-by-treatment matrix and prove that the interaction column is all zero without treating an undefined neonatal drug factor as truly controlled.
Differential gene expression: Reanalyse the neonatal raw count table using a sample-level bulk workflow and communicate effects and uncertainty without inflating n.
Reproducible research: Design a sensitivity and transparency report for the aged nucleus pseudobulk comparison, explicitly stating why n is two pools per arm and why results are exploratory.
Factorial experiment: Produce a preregistered, budget-aware plan for replicated WT and KO mice assigned to vehicle or TIMP2 at one age, then extend it to two ages for an age interaction.

Learning Check

Source evaluation: Using GSM extraction protocol text, distinguish reported numbers of source animals, physical pools, libraries, and captured nuclei, and explain every change of denominator.
Hypothesis testing: Explain why a very large number of nuclei can coexist with uncertain sample-level treatment effect estimates when there are only two independent pools per arm.
Model matrix: Demonstrate the within-series rank calculations and the singularity of a naive joint genotype-by-treatment design, identifying the structurally missing groups.
Statistical inference: Defend the separate neonatal genotype and aged treatment contrasts, specifying the population, source material and analytic limitations for each.
Experimental design: Specify the minimum crossed and replicated groups needed for genotype-by-treatment and age-by-genotype-by-treatment tests, including randomization, pooling and library mapping.
Scientific communication: Rewrite an exaggerated cross-study claim into a precise conclusion that distinguishes association, within-study comparison and non-identifiable effects.

Learning evidence

You can document your competence with a reproducible GEO metadata audit containing every GSE/GSM reference and an explicit animals-to-pools-to-libraries map; an annotated count-table analysis plan that distinguishes bulk libraries from true single-nucleus pseudobulk profiles; a checked design matrix with rank and estimability notes; a concise contrasts table that separates supported conclusions from absent factorial cells; and an experimentally feasible follow-up design with independent biological units. Include a clear evidence ledger: metadata observed, model assumption, deduction and not identifiable. Where sex/litter composition, nucleus-to-mouse assignments or raw-data reprocessing are unavailable, state that limitation rather than fabricating details.