# Scientific Skeptic Evaluation: AD scRNA-seq Hypotheses
## Cross-Cutting Methodological Concerns
Before evaluating individual hypotheses, several fundamental issues pervade this framework:
**1. The Standardization Assumption is Overstated**
Harmony/BBKNN batch correction methods can aggressively remove biological signal alongside technical variation. Studies have demonstrated that aggressive batch correction can produce spurious clusters and obscure real inter-individual variation. The claim that standardized pipelines will "enable reliable cross-study comparison" conflates statistical normalization with biological comparability—tissues processed with different dissociation protocols, PMI times, and RNA preservation methods remain fundamentally incomparable regardless of algorithmic correction.
**2. Post-mortem Confounds Are Insufficiently Addressed**
All hypotheses rely on human tissue with variable post-mortem intervals, agonal states, and RNA quality. These factors systematically affect microglial (inflammatory markers), astrocyte (stress genes), and neuronal (UPR, mitochondrial) gene expression independent of AD pathology. PMI-dependent artifacts could readily generate apparent "cell states" that reflect tissue quality rather than biology.
**3. Automated Annotation Risks False Positive Cell States**
CellTypist/Scanpy-based automated annotation can propagate errors across datasets, particularly for rare or transitional states. The hypotheses require detecting genuinely rare populations, but automated annotation is conservative and typically favors assigning cells to well-established categories over novel state discovery.
---
## Hypothesis-by-Hypothesis Evaluation
---
### Hypothesis 1: Microglial Continuum States
**Confidence: 0.78 → Revised: 0.52**
**Weaknesses:**
1. **False dichotomy framing**: "Continuum vs. discrete" may not be a meaningful biological distinction. Cells can occupy discrete states that transition rapidly without existing in intermediate states at any given timepoint. Velocity analysis infers trajectory from spliced/unspliced RNA ratios, but this cannot distinguish "cells genuinely transitioning" from "cell populations with different splicing kinetics" or "technical noise masquerading as trajectory."
2. **Marker interpretation is problematic**: The hypothesis cites TREM2 upregulation concurrent with inflammatory markers as evidence of oscillation. However, TREM2+ inflammatory microglia may represent a fixed state, not oscillation. P2RY12 (the canonical homeostatic marker) is often *downregulated* in DAM—these markers are not simply "mixed."
3. **Technical artifact confounds**: Microglial states are highly sensitive to tissue dissociation artifacts. Fresh tissue microglia differ substantially from nuclei isolated from frozen tissue (the typical format for human studies). Apparent "continuum" states may reflect the gradient of dissociation-induced应激 rather than biological trajectory.
4. **Species translation concerns**: Mouse DAM data may not translate to human microglia. Human microglia have distinct transcriptional identities and response patterns. The continuum model is grounded in mouse data.
**Counter-evidence:**
- Recent studies (Haage et al., 2022, *Nature Neuroscience*) using high-parameter flow cytometry identified discrete microglial populations with sharp phenotypic boundaries rather than gradients.
- Spatial transcriptomics studies show microglia near plaques occupy distinct spatial niches rather than blending continuously into remote microglia.
**Falsification experiments:**
1. **MERFISH/spatial transcriptomics**: If microglial TREM2 and P2RY12 show spatial gradient rather than discrete domain enrichment, continuum model supported. If they show sharp spatial boundaries, continuum model falsified.
2. **Temporal perturbation**: Treat microglial cultures with TREM2 agonist then antagonist; if cells revert to baseline rather than occupying intermediate states, oscillation/continuum claim is weakened.
3. **In vivo calcium imaging**: Direct observation of microglial process motility and soma positioning—oscillation between protective/damaging functions would predict distinct behavioral states testable by imaging.
---
### Hypothesis 2: Region-Specific Astrocyte Reactivity
**Confidence: 0.72 → Revised: 0.48**
**Weaknesses:**
1. **Astrocyte identification remains problematic**: GFAP, S100B, and AQP4 are not specific astrocyte markers—they mark ependymal cells, certain vascular cells, and have complex cell-type specificity. Single-nucleus RNA-seq relies heavily on these markers for astrocyte identification, creating circular validation risk.
2. **The three-state model may be overfitted**: "Pan-reactive," "synaptic-supportive," and "metabolic-compromised" states may emerge from analytical choices rather than biological reality. Without a priori hypothesis testing, these categories risk being post-hoc pattern-finding.
3. **Current human astrocyte transcriptomic atlases show substantial heterogeneity but no consensus on discrete subtypes**. The ESCAPE BioFIND study and others show astrocyte transcriptomes vary dramatically by brain region independent of disease state.
4. **APOE upregulation in astrocytes is a nonspecific stress response** not specific to AD—it occurs in aging, other neurodegenerative conditions, and even non-neurological conditions.
**Counter-evidence:**
- Astrocyte-specific proteomic studies (Daidzic et al. 2023) show substantial individual variation that doesn't cleanly cluster into three discrete states.
- Human astrocytes are far more heterogeneous than rodent astrocytes; scRNA-seq may be capturing this species difference rather than AD-specific pathology.
**Falsification experiments:**
1. **Spatial validation**: Perform MERFISH or CODEX imaging with markers for all three states simultaneously. If states show spatial intermixing rather than regional clustering, the model is weakened.
2. **Independent cohort validation**: Process an independent AD cohort (different institution, different processing) through the same pipeline. If the same three states emerge without batch correction artifacts, the model gains support.
3. **Cell-type-specific APOE knockout**: If APOE deletion in astrocytes (but not neurons) eliminates one state, causality is supported. If states persist despite astrocyte-specific APOE deletion, the model is falsified.
---
### Hypothesis 3: Proteostasis Collapse in Excitatory Neurons
**Confidence: 0.68 → Revised: 0.38**
**Weaknesses:**
1. **NFL as stress readout is problematic**: Neurofilament light chain elevation in CSF is a marker of *neuroaxonal damage*, not pre-damage stress. Upregulation in neurons could indicate ongoing death rather than pre-collapse states.
2. **ER stress in AD neurons is well-documented, but whether it represents a specific "pre-collapse" state vs. terminal damage is unclear**. UPR activation can be adaptive or pro-apoptotic depending on context; scRNA-seq cannot distinguish these functional outcomes.
3. **Entorhinal cortex layer 2/3 neurons are in an anatomically vulnerable location** with early tau pathology—but this doesn't prove transcriptional "collapse" precedes degeneration. These neurons may simply have higher baseline pathology exposure.
4. **Circular definition risk**: Defining a state by the markers one measures, then claiming those markers indicate the state, is tautological.
**Counter-evidence:**
- Scheffler et al. (2012) showed ER stress is present in AD neurons but interpreted this as a consequence of proteotoxic challenge rather than a pre-disease state.
- *The absence of single-cell resolution studies doesn't imply a specific state exists—it may reflect the difficulty of capturing transient states.*
**Falsification experiments:**
1. **Spatial co-expression analysis**: Perform multiplexed FISH for UPR markers + cleaved caspase-3 (cell death marker). If UPR+ neurons are always also apoptotic, the "pre-collapse" interpretation is falsified.
2. **Longitudinal sampling**: Compare entorhinal cortex transcriptomes across cognitively normal, MCI, and AD stages from the same individuals (available in ROSMAP). If "proteostasis-compromised" state appears only in AD, not in age-matched controls with early tau, the model is weakened.
3. **Organoid validation**: Human cerebral organoids exposed to oligomeric Aβ—if "proteostasis-compromised" state appears before neuronal death markers and reverses with intervention, the model is supported.
---
### Hypothesis 4: OPC Remyelination Arrest
**Confidence: 0.65 → Revised: 0.35**
**Weaknesses:**
1. **PDGFRA is not specific to OPCs**: PDGFRA marks several cell types including pericytes, mesenchymal cells, and some fibroblasts. Single-nucleus studies may capture non-OPCs in this gate.
2. **AD-specific white matter involvement is not established as primary**: White matter changes in AD may be secondary to neuronal degeneration, demyelination may be a consequence not cause of AD pathology. The claim that AD specifically disrupts OPC differentiation is speculative.
3. **OPC heterogeneity is substantial**: OPCs in different brain regions show baseline transcriptional differences that could be misinterpreted as disease-specific arrest.
4. **Literature basis is weak**: The cited reference ("文献") suggests the hypothesis lacks specific citations for the AD-OPC connection compared to other demyelinating diseases (MS, PML).
**Counter-evidence:**
- OPC differentiation failure is well-characterized in MS but AD-specific OPC dysfunction is not established.
- White matter hyperintensities in AD on MRI may reflect vascular dysfunction rather than primary demyelination.
**Falsification experiments:**
1. **Spatial mapping**: If arrested OPCs are not preferentially located near demyelinated regions or amyloid deposits, the model is falsified.
2. **AD vs. non-AD white matter comparison**: Compare OPC states in AD white matter vs. normal-appearing white matter in same brains. If arrested state is present in both, it's not AD-specific.
3. **Myelin-specific assays**: Combine scRNA-seq with proteolipid protein (PLP) imaging to confirm that arrested OPCs fail to produce myelin proteins in situ.
---
### Hypothesis 5: Vascular-Interactive BBB States
**Confidence: 0.