# Domain Expert Contribution: SEA-AD MTG Differential Expression Dataset
## Experimental Design Strengths and Limitations
The SEA-AD dataset represents one of the most comprehensive single-nucleus RNA sequencing (snRNA-seq) resources for Alzheimer's disease research, with the Mathys et al. 2023 *Cell* publication reporting data from over 1.2 million cells across hundreds of donors. The study's core strength lies in its scale: examining the middle temporal gyrus (MTG) provides access to a region implicated in both early AD pathology and language-related circuits, offering insight into vulnerability patterns distinct from the prefrontal cortex or hippocampus. The inclusion of subjects spanning cognitive resilience to frank dementia enables pseudo-longitudinal analysis of disease progression that few datasets can match.
However, several design considerations warrant scrutiny. First, the cross-sectional nature of post-mortem tissue collection introduces confounding variables that differ substantially between donors—PMI (post-mortem interval), RNA quality, agonal state, and medication history can profoundly influence transcriptomic profiles. The Mathys et al. analysis addressed batch effects using Harmony integration and accounted for demographic covariates, but residual confounding remains possible, particularly for rare cell types where power is limited. Second, the cell-type composition problem is acute in AD: if disease causes selective neuronal loss in excitatory neuron clusters, comparing expression profiles between groups conflates *per-cell* transcriptional changes with *population-level* compositional shifts. The appropriate remedy—cell-by-cell rather than cell-type-level analysis—is technically possible but statistically challenging.
## Statistical Methodology Evaluation
The differential expression framework employed in SEA-AD uses standard negative binomial models (likely MAST or similar) with appropriate handling of expression as counts. The inclusion of covariates (age, sex, PMI, cognitive diagnosis) is essential and generally well-executed. However, a critical limitation persists in how the study addresses the multiple testing burden: with ~20+ cell types/subtypes × thousands of genes, stringent FDR correction may sacrifice power to detect true signals in smaller clusters. The meta-analysis approach across discovery and validation cohorts represents best practices, but I note that both cohorts derive from the same geographic/biospecimen provenance, limiting true external generalizability. A more rigorous validation would involve orthogonal datasets from different institutions (e.g., Mayo Clinic RNA-seq, ROSMAP) to confirm cell-type specificity of findings.
## Reproducibility and Translational Concerns
From a drug development perspective, the most actionable findings from SEA-AD involve pathway-level insights—upregulation of innate immune pathways in microglia, dysregulation of mitochondrial/ribosomal genes in specific excitatory neuron subtypes—rather than single-gene targets. The reproducibility of cell-type-specific DEGs across independent cohorts remains uncertain; published comparisons suggest moderate overlap (~40-60% for top hits), which is actually reasonable given the noise inherent in snRNA-seq. A deeper concern involves the biological interpretation challenge: does the observed *in vivo* transcriptional dysregulation reflect causal disease drivers or downstream epiphenomena? For target validation, functional follow-up (iPSC models, CRISPR perturbations) is essential before assuming therapeutic relevance. The dataset's greatest value may lie in hypothesis generation rather than immediate target identification.
---
**My Confidence Score: 0.75**
The Mathys et al. 2023 paper provides substantial methodological detail, but I cannot verify specific parameter choices without deeper access. The concerns I raise are well-documented in the broader snRNA-seq literature for neurodegenerative disease, though specific claims about SEA-AD would benefit from independent evaluation. I recommend cross-referencing with the original publication's supplementary methods and considering the ROSMAP snRNA-seq dataset as an independent comparator.
n implicated in both early AD pathology and language-related circuits, offering insight into vulnerability patterns distinct from the prefrontal cortex or hippocampus. The inclusion of subjects spanning cognitive resilience to frank dementia enables pseudo-longitudinal analysis of disease progression that few datasets can match.
However, several design considerations warrant scrutiny. First, the cross-sectional nature of post-mortem tissue collection introduces confounding variables that differ substantially between donors—PMI (post-mortem interval), RNA quality, agonal state, and medication history can profoundly influence transcriptomic profiles. The Mathys et al. analysis addressed batch effects using Harmony integration and accounted for demographic covariates, but residual confounding remains possible, particularly for rare cell types where power is limited. Second, the cell-type composition problem is acute in AD: if disease causes selective neuronal loss in excitatory neuron cl