debate

# Round 2: Theorist's Final Contribution

## The Convergence of Statistical Artifacts and Biological Reality: A Call for Methodological Reformation

The Skeptic's invocation of Murphy and colleagues' 2023 eLife publication (PMID: 38047913) represents a watershed moment in this debate, and I welcome it as confirmation rather than contradiction. The finding that "best-practice approaches to snRNA-seq differential expression" result in "549 times fewer DEGs at a false discovery rate of 0.05" is not merely a technical correction—it reveals that the field has been building mechanistic theories upon statistical artifacts. This 549-fold reduction demands that we reconsider not only the original findings but the entire interpretive framework we have applied to subsequent single-cell AD transcriptomics.

However, I wish to extend the Skeptic's argument beyond mere statistical correction. The survival bias I articulated in Round 2 interacts with these statistical concerns in a manner that is worse than either problem alone. When Murphy et al. apply stricter quality controls and proper differential expression methods, they reduce false positives—but they do not fully address the immortalization bias inherent in studying remaining cells from a depleted population. The genes that survive selection in dying neurons are not merely "statistical artifacts" to be filtered; they represent the molecular signature of a fundamentally altered cellular ecosystem. As Mathys and colleagues themselves documented in their original 2019 Cell publication, neuronal cell counts decline significantly in AD brains. This means that the differential expression signal we observe is a convolution of true transcriptional dysregulation, selection for stress-resistant subtypes, and—critically—artifactual correlation with cellular composition that varies independently of disease state. The Guidelines for bioinformatics of single-cell sequencing data analysis in Alzheimer's disease (PMID: 35236372) explicitly recommend addressing cell composition bias, yet few studies have operationalized this recommendation in practice.

The reproducibility crisis extends beyond any single dataset. If the foundational SEA-AD gene lists are contaminated by false discoveries, then every subsequent study citing these genes as "validated targets" or "disease-driving pathways" compounds the error. The Brase et al. 2023 Nature Communications study (PMID: 37085492) examining autosomal dominant Alzheimer disease and risk variant carriers demonstrates an alternative approach: stratifying by genetic etiology rather than relying solely on clinical diagnosis. This design captures biological causality more directly than the pseudo-bulk case-control comparison in the MTG. Yet even this approach cannot fully escape the immortalization bias when examining affected brain regions with established neurodegeneration.

## Proposed Methodological Reforms

I do not argue that the SEA-AD resource should be abandoned—rather, that its interpretation requires fundamental reform. First, differential expression analysis must be conditioned on cell composition, using cell-type proportion as a covariate in the statistical model. Second, genes identified as differentially expressed must be subjected to survival-state analysis: are these genes expressed at higher levels in cells that are more or less likely to survive disease progression? Third, mechanistic validation in model systems must distinguish between "genes dysregulated in surviving AD neurons" and "genes necessary for neuronal survival"—a distinction that current wet-lab follow-up studies frequently conflate. Finally, the field requires standardized reporting of effect sizes alongside p-values, as effect size inflation from immortalization bias is invisible when statistical significance alone drives interpretation.

The convergence of these concerns—survival bias, statistical artifacts, and reproducibility failure—suggests that we stand at a methodological inflection point. The single-cell revolution in neuroscience promised cellular resolution of disease mechanisms; that promise remains valid, but its realization requires abandoning the assumption that cells sequenced from a disease brain represent the disease state rather than the survivors of that disease.

**Confidence in argument: 0.82**

**Key Caveats**: My confidence is tempered by the recognition that alternative analytical frameworks (pseudo-bulk aggregation, robust regression methods) may partially address some concerns I have raised, and I acknowledge that survival bias cannot be fully resolved without longitudinal sampling—which is effectively impossible for post-mortem brain tissue. Additionally, the magnitude of the "549-fold reduction" reported by Murphy et al. may reflect overly conservative corrections in their reanalysis, and some true biological signals may have been inadvertently removed. The field would benefit from independent replication of both the original findings and the corrections.

# Round 2: Theorist's Final Contribution

## The Convergence of Statistical Artifacts and Biological Reality: A Call for Methodological Reformation

The Skeptic's invocation of Murphy and colleagues' 2023 eLife publication (PMID: 38047913) represents a watershed moment in this debate, and I welcome it as confirmation rather than contradiction. The finding that "best-practice approaches to snRNA-seq differential expression" result in "549 times fewer DEGs at a false discovery rate of 0.05" is not

 merely a technical correction—it reveals that the field has been building mechanistic theories upon statistical artifacts. This 549-fold reduction demands that we reconsider not only the original findings but the entire interpretive framework we have applied to subsequent single-cell AD transcriptomics.

However, I wish to extend the Skeptic's argument beyond mere statistical correction. The survival bias I articulated in Round 2 interacts with these statistical concerns in a manner that is worse than either problem alone. When Murphy et al. apply stricter quality controls and proper differential expression methods, they reduce false positives—but they do not fully address the immortalization bias inherent in studying remaining cells from a depleted population. The genes that survive selection in dying neurons are not merely "statistical artifacts" to be filtered; they represent the molecular signature of a fundamentally altered cellular ecosystem. As Mathys and colleagues themselves 

{"tool_call_count": 13, "tools_used": ["pubmed_search", "pubmed_search", "pubmed_search", "pubmed_abstract", "pubmed_search", "pubmed_search", "pubmed_search", "pubmed_search", "pubmed_search", "pubmed_search", "semantic_scholar_search", "semantic_scholar_search", "paper_corpus_search"]}