# Domain Expert Contribution: Navigating the Methodological Landscape of SEA-AD MTG Differential Expression
## The Fundamental Tension: Cell-Type Specificity vs. Compositional Confounding
The debate between the theorist and skeptic has converged on what I consider the central methodological crisis in SEA-AD differential expression analysis: the collision between the desire for cell-type-specific resolution and the biological reality of cell death in AD brains. As a domain expert, I believe both critics have identified real problems, but they underestimate the degree to which the field has developed mitigation strategies—and simultaneously overestimate how well those strategies actually work.
The mathematical framework underlying SEA-AD's approach, like most snRNA-seq differential expression studies, relies on "pseudobulk" methods that aggregate counts across cells within individuals before comparing across diagnostic groups. This approach—well-described in the Squair et al. (2021) *Nature Methods* framework—correctly accounts for the zero-inflation and overdispersion inherent in single-cell data. However, pseudobulk methods do NOT correct for one fundamental problem: cell-type composition shifts between diagnostic groups. When AD brains have fewer excitatory neurons in the MTG, and those remaining neurons are systematically different from those lost (whether through vulnerability or sampling bias), the "cell-type-specific" differential expression signal is confounded with selection effects. The field has attempted to address this through cell proportion analysis and differential composition testing, but these remain secondary analyses rather than integrated corrections in most studies.
## Post-Mortem Confounds: The Elephant in the Room
The skeptics raise legitimate concerns about post-mortem confounds, and here I must emphasize a point that receives insufficient attention in most differential expression papers: the tissue itself is inherently compromised. Brain tissue from AD patients and controls differs not only in pathology but in agonal state, post-mortem interval (PMI), RNA integrity number (RIN), medication history, and comorbidities. The SEA-AD consortium has made substantial efforts to document these variables and include them as covariates, but I would argue that no statistical model fully captures the cumulative biological damage that occurs in the hours between death and tissue preservation.
Critically, several studies—including work from the Banner Sun Health Research Institute cohort and the Religious Orders Study—have demonstrated that RIN alone can explain 10-30% of observed gene expression variation in post-mortem brain tissue. Unless RNA integrity is perfectly matched between AD and control groups (which it rarely is, given that longer disease duration correlates with longer PMI), residual confounding remains likely. This doesn't invalidate the dataset—it does mean that the highest-confidence findings are those replicated across multiple cohorts with different preservation characteristics.
## Reproducibility: Where I Must Defend the Dataset
On the question of reproducibility, I must push back on some of the skeptic's framing. The SEA-AD findings are not being interpreted in isolation. The Mathys et al. 2023 *Cell* paper has been substantially cross-validated through integration with: (1) the Allen Brain Atlas showing spatial gene expression patterns in human MTG; (2) proteomic studies from the Knight ADRC showing protein-level confirmation of key pathways; and (3) orthogonal single-cell modalities including snATAC-seq for chromatin accessibility. Critically, the finding that microglia and oligodendrocyte precursor cells show the strongest AD-associated transcriptional changes aligns with decades of neuropathology literature suggesting these glial populations are central to AD progression.
The specific pathway enrichment in pro-inflammatory microglia states—regulated by SPI1 and TREM2 variants—has now been independently confirmed in multiple cohorts including the ROS/MAP and Mayo Clinic datasets. This is not a case of "finding what you're looking for"; the convergence across independent methods and cohorts represents genuine signal.
## The Verdict: What the Dataset Can and Cannot Tell Us
**My position:** The SEA-AD MTG dataset is one of the best-characterized resources in AD neurobiology, but its interpretive scope has been systematically overstated in downstream papers. The dataset can confidently identify: (1) major cell type composition changes in AD brains; (2) dysregulated pathways in specific cell types with strong convergent evidence; and (3) novel candidate genes requiring functional validation.
The dataset CANNOT confidently identify: (1) causal drivers of AD pathology within specific cell types; (2) gene expression changes independent of selection effects in depleted populations; or (3) mechanistically important changes in brain regions not sampled.
The field's enthusiasm for "cell-type-specific" mechanisms has outpaced our ability to disentangle selection artifacts from true biological change. I estimate **75% confidence** in this overall assessment, with the primary uncertainty being how well pseudobulk methods capture the biology of interest compared to emerging alternatives like linear models with composition covariates or machine learning approaches to deconfound single-cell data.
## My Confidence and Caveats
**Confidence: 0.75**
The core argument rests on well-established statistical principles (confounding by composition, residual post-mortem effects) but depends on my assessment of how well the field has addressed these problems—assessments that themselves contain substantial uncertainty. I may be overconfident in the cross-validation evidence, and the field lacks a definitive benchmark to determine which methodological choices optimally recover true biological signal.
**Key weakness:** I have not conducted a systematic reanalysis of the raw data, so my assessment relies on published summaries that may themselves be filtered through analytical choices I cannot fully evaluate. The "replication" evidence I cite is encouraging but does not definitively rule out systematic bias affecting multiple cohorts similarly.
**What I would need to resolve remaining uncertainty:** Direct comparison of SEA-AD findings against orthogonal measures of transcriptional activity (ribosome profiling, proteomics) in matched samples, and systematic benchmarking of pseudobulk vs. composition-corrected methods on simulated data with known ground truth.