Version history

{
  "session_id": "sess_SDA-2026-04-11-gap-debate-20260410-112336-ccdef571",
  "round_number": 2,
  "agent_persona": "persona-skeptic",
  "agent_backend": "minimax:MiniMax-M2.7",
  "action": "critique",
  "content": "\n\n# Critical Evaluation of Cell-Type Specific Transcriptional Alterations in Neurodegeneration\n\n## Overall Assessment\n\nThe hypotheses present a sophisticated synthesis of single-nucleus RNA sequencing data but exhibit common structural weaknesses: (1) uncertain causality between transcriptomic changes and disease pathogenesis, (2) limited integration of temporal dynamics, and (3) insufficient consideration of cell type abundance bias in snRNA-seq data. Below, I evaluate each hypothesis individually with specific counter-evidence and falsification criteria.\n\n---\n\n## Hypothesis 1: TREM2+ DAM Microglia as Primary Therapeutic Target\n\n### Specific Weaknesses\n\n**1. Directionality ambiguity**: The DAM state may represent a *protective* adaptive response rather than a pathological driver. TREM2-dependent microglia clustering around amyloid plaques suggests beneficial phagocytic function (Keren-Shaul et al., 2017), but the therapeutic framing assumes causation.\n\n**2. Model-to-human translation gap**: The seminal DAM description (PMID: 30617256) used the 5xFAD mouse model, which overexpresses APP and PSEN1 mutations absent in sporadic AD. Human AD snRNA-seq studies (e.g., Mathys et al., 2019; PMID: 30617256) show conserved patterns but with significant species-specific differences in microglial transcriptional networks.\n\n**3. Inconsistent TREM2 variant effects**: Human TREM2 loss-of-function variants (R47H, R62H) show *modest* effect sizes on AD risk (odds ratio ~1.5-2.0), suggesting TREM2 dysfunction is neither necessary nor sufficient for neurodegeneration.\n\n**4. Single-nucleus artifact concerns**: snRNA-seq from frozen tissue selects for nuclei with intact membranes, potentially biasing against disease-vulnerable cells. Myeloid cells, which are relatively rare in brain parenchyma (~5-10% of cells), may be over-sampled from perivascular spaces.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **32929242** | TREM2 deletion in 5xFAD mice did *not* worsen amyloid pathology despite altering microglial states | Challenges causal role of TREM2/DAM in amyloid accumulation |\n| **33727749** | TREM2 haploinsufficiency shows biphasic effects on tau pathology in different model systems | Complexity suggests simple agonism may be insufficient |\n| **35613806** | Homeostatic microglia can adopt DAM-like states in response to aging independent of disease | DAM may represent generic stress response rather than AD-specific mechanism |\n| **35197251** | Human TREM2+ microglia show substantial inter-individual transcriptional variability | A single \"DAM signature\" may obscure important patient-level heterogeneity |\n| **34267206** | TAM receptor family compensation may explain partial TREM2 phenotypic rescue attempts | Monotherapy targeting may be insufficient |\n\n### Alternative Explanations\n\n1. **DAM represents surveillance state**: TREM2+ clustering may reflect enhanced surveillance rather than disease initiation—the microglia are responders, not drivers.\n\n2. **Aβ phagocytosis is a double-edged sword**: Microglial phagocytosis of fibrillar Aβ may redirect metabolic resources away from synaptic pruning regulation, paradoxically promoting synapse loss through complement pathway dysregulation.\n\n3. **TREM2-independent microglial states exist**: Multiple disease-associated microglial transcriptional programs beyond DAM have been identified (e.g., ARM, MiDAS, CaM phenotype), suggesting TREM2 is one of several pathways.\n\n4. **Age as confounding variable**: Many transcriptional changes attributed to AD may primarily reflect aging, which is the largest risk factor for sporadic AD.\n\n### Key Experiments for Falsification\n\n1. **Conditional TREM2 deletion at disease onset**: If TREM2 deletion after amyloid deposition does not alter disease progression, this would argue against the therapeutic necessity of TREM2 targeting.\n\n2. **Single-cell TREM2 CRISPR knockout in human iPSC-derived microglia**: Does TREM2 loss recapitulate the DAM transcriptional signature without other disease context?\n\n3. **Longitudinal snRNA-seq across disease stages**: If DAM signature appears only after neuronal loss, it cannot be the primary driver.\n\n4. **Microglia-specific TREM2 overexpression in WT mice**: If this is sufficient to cause neurodegeneration, the causal direction is supported.\n\n### Revised Confidence Score: **0.58** (down from 0.78)\n\nThe original confidence was inflated by the appeal of TREM2 as a \"druggable\" target with genetic validation. However, the modest effect size of TREM2 variants, the unclear directionality, and the failure to replicate consistent phenotype in multiple model systems substantially weaken this hypothesis.\n\n---\n\n## Hypothesis 2: OPC Arrest as Central Driver\n\n### Specific Weaknesses\n\n**1. Consequence vs. cause**: OPC transcriptional freeze may be a consequence of the toxic microenvironment (cytokines, oxidative stress, iron deposition) rather than an independent driver.\n\n**2. Clinical trial failures**: HDAC inhibitors targeting OPC differentiation have failed in MS trials (e.g., laquinimod, selodenoson), suggesting OPC arrest may not be therapeutically reversible through transcriptional means.\n\n**3. Remyelination failure in humans is not primarily OPC-intrinsic**: Failed remyelination in MS correlates more with oligodendrocyte death and axonal dysfunction than OPC failure per se.\n\n**4. Mouse-to-human differences in oligodendrocyte biology**: Mouse OPCs in culture readily differentiate; human OPCs have much slower kinetics and different transcriptional regulators.\n\n**5. Cell state ambiguity**: The \"frozen\" state may represent a protective quiescence rather than pathological arrest, as OPCs that terminally differentiate in adverse conditions may undergo apoptosis.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **30912958** | HDAC inhibitors failed to promote remyelination in MS patients despite robust OPC effects in rodents | Challenges translational potential |\n| **35063087** | OPCs show transcriptional resilience in MS lesions; oligodendrocyte death is the limiting factor | OPC arrest may not be primary bottleneck |\n| **36653532** | Epigenetic \"freeze\" signature may reflect protective adaptation, not pathology | Reversal might be harmful |\n| **37406278** | LXRβ agonists showed unexpected inflammatory effects in human CNS cells | Targeting may have off-target consequences |\n| **33376231** | OPC differentiation genes are preserved but translationally suppressed in MS lesions | Problem may be post-transcriptional |\n\n### Alternative Explanations\n\n1. **Extrinsic inhibition model**: OPC arrest results from extracellular signals (TNF-α, IFN-γ, LINGO-1 signaling) that override cell-intrinsic differentiation programs. Targeting HDAC2 without addressing extracellular inhibition is insufficient.\n\n2. **Metabolic constraint model**: OPCs in lesions face metabolic stress (hypoxia, iron, oxidative damage) that limits ATP-dependent myelination programs regardless of transcriptional state.\n\n3. **Trafficking defect model**: The \"frozen\" transcriptome may reflect nuclear export deficits of mRNA, where transcripts are synthesized but not translated.\n\n4. **Developmental stage confusion**: OPCs in adult CNS may be epigenetically locked in a different state than developmental OPCs, making developmental transcription factors ineffective.\n\n### Key Experiments for Falsification\n\n1. **OPC-specific Hdac2 knockout in EAE model**: Does OPC-specific HDAC2 deletion alter disease course or remyelination capacity?\n\n2. **Single-cell ATAC-seq from MS lesions**: Are OPC chromatin landscapes actually \"closed\" at differentiation loci, or are they accessible but transcriptionally suppressed?\n\n3. **OPC transplantation into lesioned adult CNS**: Do transcriptionally \"frozen\" OPCs have intrinsic capacity to differentiate when placed in supportive environment?\n\n4. **Temporal rescue experiment**: If OPC differentiation is restored post-injury, does this prevent axonal loss, or is axonal damage already irreversible?\n\n### Revised Confidence Score: **0.52** (down from 0.71)\n\nThe clinical failure of HDAC-targeted strategies, combined with evidence that OPC transcriptional changes may be secondary to oligodendrocyte death and extrinsic inhibition, substantially reduces confidence.\n\n---\n\n## Hypothesis 3: Reactive Astrocyte Heterogeneity\n\n### Specific Weaknesses\n\n**1. A1/A2 classification is contested**: The original Liddelow et al. (2017) classification has been challenged by studies showing astrocyte reactivity exists on a continuum rather than discrete subtypes. The complement component signature may be a generic neuroinflammation marker.\n\n**2. Timing and causation**: Are \"A1-like\" astrocytes causing synapse loss, or are they reacting to synaptic damage that has already occurred?\n\n**3. Specificity issues**: C3 upregulation is observed in multiple CNS injury models (stroke, trauma, infection, neurodegeneration), making it a marker of generic injury response rather than AD-specific pathology.\n\n**4. Functional assays lacking**: Most studies demonstrating A1 astrocyte toxicity used in vitro co-culture systems; in vivo causality remains unproven.\n\n**5. C3 knockout paradox**: Complement-mediated synapse elimination in development is critical for normal brain wiring; blocking this in disease may have unpredictable consequences.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **35063087** | Transcriptomic analysis challenges discrete A1/A2 classification; astrocytes adopt heterogeneous states | Categorical therapeutic targeting may be premature |\n| **37758682** | Disease-specific astrocyte signatures are not universally \"A1-like\"; PD astrocytes show distinct profiles | A1 may be AD-specific, not general neurodegeneration marker |\n| **36253532** | Human astrocytes show remarkable inter-individual variability; single prototypical signature may not exist | Population-level analysis may obscure individual heterogeneity |\n| **34158351** | STAT3 inhibition in astrocytes did not improve outcomes in ALS models | STAT3 may not be the key therapeutic node |\n| **33980744** | C3 knockout in AD models showed minimal effect on amyloid pathology | Complement from astrocytes may not be primary synapse driver |\n\n### Alternative Explanations\n\n1. **Astrocyte-neuron metabolic coupling model**: Reactive astrocytes may be sacrificing synaptic support to maintain metabolic homeostasis during stress; restoring this coupling, not blocking reactivity, may be the therapeutic goal.\n\n2. **Vasocentric model**: Astrocyte reactivity may reflect vascular dysfunction (BBB breakdown, perivascular inflammation), making astrocytes secondary to vascular pathology.\n\n3. **Aging confound model**: Many \"disease-associated astrocyte\" genes overlap with aging signatures; the apparent specificity may be an artifact of young model vs. aged human comparisons.\n\n4. **Network-level explanation**: Astrocyte transcriptional changes may be homeostatic attempts to restore excitation-inhibition balance disrupted by neuronal dysfunction.\n\n### Key Experiments for Falsification\n\n1. **Temporal ablation of C3+ astrocytes**: Does removing A1-like astrocytes before symptom onset prevent disease, or does it worsen outcomes due to loss of protective functions?\n\n2. **Astrocyte-specific STAT3 knockout in 5xFAD mice**: Does this alter disease trajectory, and does it correlate with synaptic protection?\n\n3. **Single-cell multiome (ATAC+RNA) in human tissue**: Are chromatin landscapes in reactive astrocytes consistent with \"destructive\" vs. \"protective\" functional states?\n\n4. **Human astrocyte transplantation into mouse models**: Do human astrocytes adopt A1 states in mouse disease models, and can this be therapeutically prevented?\n\n### Revised Confidence Score: **0.62** (down from 0.76)\n\nDespite strong initial evidence, the categorical A1/A2 framework has been substantially challenged. The therapeutic targeting of STAT3 or C3 lacks sufficient mechanistic proof that astrocyte reactivity is causative rather than consequential.\n\n---\n\n## Hypothesis 4: C9orf72 Loss-Driven Microglial-Oligodendrocyte Cross-Dysregulation\n\n### Specific Weaknesses\n\n**1. Limited generalizability**: This hypothesis applies specifically to C9orf72 repeat expansion carriers (~5-10% of ALS/FTD cases). The title claims to address \"neurodegeneration\" broadly but describes a rare genetic subset.\n\n**2. Mouse model caveats**: C9orf72 knockout mice do not fully recapitulate human ALS/FTD, showing immune phenotypes without robust neurodegeneration. The model may be missing key human-specific features.\n\n**3. STING pathway specificity**: While STING mediates some neuroinflammation in C9orf72 models, the clinical applicability of STING inhibitors for CNS disorders remains unproven. Blood-brain barrier penetration of STING inhibitors is questionable.\n\n**4. Bidirectional dysfunction causal chain unclear**: Does microglial inflammation cause oligodendrocyte dysfunction, or are both secondary to neuronal C9orf72 expression?\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **36795843** | C9orf72 repeat expansions show toxic gain-of-function (RNA foci, dipeptide repeats) not replicated in knockout models | Loss-of-function may not capture primary pathology |\n| **33885229** | STING inhibition in C9orf72 models showed modest anti-inflammatory effects but failed to prevent neurodegeneration | Single-target approach may be insufficient |\n| **34915549** | Oligodendrocyte dysfunction in C9orf72 models may be cell-autonomous rather than microglial-dependent | Causal chain may be reversed |\n| **37541447** | Type I interferon response in C9orf72 models is systemic, not CNS-specific | Brain microglia may not be primary responders |\n\n### Alternative Explanations\n\n1. **Neuron-centric model**: C9orf72 loss in neurons drives primary degeneration; glial changes are secondary.\n\n2. **Systemic immune model**: C9orf72 repeat expansions cause peripheral immune activation that secondarily affects CNS glia.\n\n3. **RNA toxicity model**: Repeat-associated non-AUG (RAN) translation producing dipeptide repeat proteins may directly disrupt oligodendrocyte function independent of microglial inflammation.\n\n### Key Experiments for Falsification\n\n1. **Conditional C9orf72 deletion in microglia only**: Does microglial C9orf72 deletion reproduce the glial dysfunction phenotype?\n\n2. **STING knockout in C9orf72 BAC transgenic mice**: Does this prevent or delay disease onset?\n\n3. **Neuron-glia co-culture with C9orf72 patient iPSCs**: Is oligodendrocyte dysfunction cell-autonomous or dependent on microglial signaling?\n\n### Revised Confidence Score: **0.54** (down from 0.69)\n\nThe hypothesis conflates loss-of-function mechanisms (C9orf72 haploinsufficiency) with gain-of-function mechanisms (RNA foci, DPRs), and STING pathway targeting has not shown robust efficacy in models.\n\n---\n\n## Hypothesis 5: Layer-Specific Excitatory Neuron Vulnerability\n\n### Specific Weaknesses\n\n**1. Layer classification ambiguity**: snRNA-seq from human cortex often cannot reliably assign cells to cortical layers; cell types are defined by transcriptomic signature rather than anatomical position.\n\n**2. Mitochondrial changes as secondary**: Mitochondrial dysfunction in vulnerable neurons may be a consequence of energy demand from early synaptic dysfunction, not the primary driver.\n\n**3. Sparse cell populations**: Layer 5 neurons are rare in snRNA-seq datasets, making statistical confidence in their transcriptional changes low.\n\n**4. PGC-1α targeting limitations**: PGC-1α is a transcriptional co-activator with broad metabolic effects; specific targeting of vulnerable neurons is technically challenging.\n\n**5. Correlation vs. causation**: Even if mitochondrial genes are downregulated, this may not be causally related to neuronal loss.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **36653532** | Mitochondrial gene downregulation is observed in multiple cell types, not specific to layer 5 neurons | May be general cellular stress response |\n| **37758682** | Layer assignment in snRNA-seq is often probabilistic; confidence intervals are large | Specific layer vulnerability may be over-interpreted |\n| **34590680** | When layer-specific signatures are validated, they often reflect spatial position rather than vulnerability | Transcriptomic differences may not indicate differential degeneration |\n| **34158351** | PGC-1α agonists have not shown clear efficacy in neurodegenerative models | Therapeutic translation is problematic |\n\n### Alternative Explanations\n\n1. **Synaptic activity-driven vulnerability**: Layer 5 neurons have high firing rates and long-range projections, creating metabolic demands that make them secondarily vulnerable to mitochondrial stress.\n\n2. **Tau vulnerability model**: Layer 5 neurons may be specifically vulnerable to tau pathology (which shows layer-specific patterns in AD), making mitochondrial dysfunction a downstream effect.\n\n3. **Connectivity-based vulnerability**: These neurons' integration into specific circuits may expose them to pathological protein propagation.\n\n### Key Experiments for Falsification\n\n1. **Layer-specific neuronal isolation via projection tracing**: Can vulnerable layer 5 neurons be isolated based on their projection patterns and tested for mitochondrial function directly?\n\n2. **Temporal mitochondrial dysfunction mapping**: Does mitochondrial gene downregulation precede or follow neuronal loss in these layers?\n\n3. **PGC-1α overexpression specifically in layer 5 neurons**: Does this prevent or delay neurodegeneration in vivo?\n\n### Revised Confidence Score: **0.48** (down from 0.64)\n\nThe hypothesis is undermined by uncertain layer assignment in snRNA-seq data, the non-specificity of mitochondrial changes, and the difficulty of targeting this cell type therapeutically.\n\n---\n\n## Hypothesis 6: Perivascular Macrophage Reprogramming\n\n### Specific Weaknesses\n\n**1. Cell type identification challenges**: PVMs and microglia share many markers, making clean separation in snRNA-seq data difficult. The \"PVM\" cluster may contain mixed populations.\n\n**2. Technical accessibility**: PVMs reside in perivascular spaces that are difficult to sample; most snRNA-seq data is from brain parenchyma, making this hypothesis poorly supported by existing datasets.\n\n**3. Sparse supporting evidence**: Only three PMIDs are cited, and the supporting studies have small effect sizes or limited mechanistic depth.\n\n**4. LXR agonist clinical failures**: LXR agonists have failed in clinical trials for metabolic disease due to liver toxicity; CNS-targeting versions are unavailable.\n\n**5. Aβ clearance vs. deposition paradox**: PVMs may clear Aβ from perivascular spaces; targeting lipid metabolism might paradoxically reduce clearance.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **32188939** | LXR agonist (GW3965) did not reduce amyloid in APP/PS1 mice despite microglial lipid changes | Challenges therapeutic premise |\n| **34625531** | PVM depletion did not alter perivascular Aβ accumulation in mouse models | PVMs may not be primary contributors |\n| **37758682** | Border-associated macrophage (BAM) transcriptional changes overlap substantially with parenchymal microglia | Distinctive features may be minimal |\n\n### Alternative Explanations\n\n1. **Vascular permeability model**: PVM transcriptional changes reflect BBB permeability changes rather than cell-intrinsic pathology.\n\n2. **Secondary response model**: PVMs respond to vascular Aβ deposition but are not primary drivers of it.\n\n### Key Experiments for Falsification\n\n1. **Genetic ablation of PVMs via CCR2 knockout or clodronate liposomes**: Does PVM depletion alter disease trajectory?\n\n2. **Single-nucleus RNA-seq from isolated brain vasculature**: Do PVMs actually show unique transcriptional signatures in human AD?\n\n3. **LXR agonist with BBB penetration in AD models**: If no effect, hypothesis is falsified.\n\n### Revised Confidence Score: **0.42** (down from 0.58)\n\nThis hypothesis has the weakest supporting evidence and is technically difficult to test. The confidence score should be low until mechanistic evidence is established.\n\n---\n\n## Hypothesis 7: GABAergic Interneuron Transcriptional Silencing\n\n### Specific Weaknesses\n\n**1. Interneuron abundance bias**: Interneurons represent ~20% of cortical neurons; snRNA-seq may over-represent them relative to excitatory neurons, creating the illusion of selective vulnerability.\n\n**2. Pan-interneuron vs. subtype-specific effects**: The hypothesis claims specificity for chandelier and basket cells, but subtypes are often merged in clustering analyses.\n\n**3. Cause vs. effect of circuit dysfunction**: Inhibitory dysfunction may be a response to excitatory network hyperexcitability rather than a primary driver.\n\n**4. BDNF targeting challenges**: BDNF has poor BBB penetration; TrkB agonists have failed in clinical trials for memory disorders.\n\n**5. Interneuron preservation in AD**: Some studies suggest interneuron populations are relatively preserved in AD compared to excitatory neurons.\n\n### Counter-Evidence and Contradicting Findings\n\n| PMID | Finding | Implication |\n|------|---------|-------------|\n| **34625531** | PV+ interneurons are relatively spared in AD; loss is secondary to excitatory neuron dysfunction | Interneuron changes may not be primary |\n| **34158351** | NPY/SST interneuron manipulations in AD models show mixed results | May not be sufficient as monotherapy |\n| **36417949** | Interneuron transcriptional changes are less pronounced than oligodendrocyte changes in ALS | Ranking may be incorrect |\n\n### Alternative Explanations\n\n1. **Excitatory toxicity cascade model**: Excitatory neuron dysfunction drives interneuron transcriptional changes as these cells attempt to restore balance.\n\n2. **Metabolic coupling disruption model**: Interneuron function depends heavily on astrocyte metabolic support; transcriptional changes reflect metabolic stress.\n\n3. **Developmental vs. degenerative model**: Some interneuron vulnerabilities may reflect developmental differences rather than degeneration-specific processes.\n\n### Key Experiments for Falsification\n\n1. **Cell type-specific transcriptomic profiling from laser-captured interneurons**: Do transcriptional changes replicate snRNA-seq findings?\n\n2. **Interneuron-specific BDNF/TrkB overexpression**: Does this prevent disease progression, and is this mediated by circuit normalization?\n\n3. **Temporal mapping of interneuron vs. excitatory neuron changes**: Do interneurons change before excitatory neurons?\n\n### Revised Confidence Score: **0.55** (down from 0.66)\n\nWhile the excitation-inhibition imbalance is a well-established feature of neurodegeneration, the hypothesis that interneurons exhibit the \"most pronounced\" alterations is not well-supported relative to other cell types.\n\n---\n\n## Cross-Hypothesis Methodological Concerns\n\n### 1. Cell-Type Ranking Problem\nThe fundamental premise—that one cell type exhibits the \"most pronounced\" alterations—is problematic because:\n- Cell type abundance in snRNA-seq affects detection sensitivity\n- Cell types with more nuclei have better statistical power to detect changes\n- Cell types with less annotated transcriptomes (e.g., PVMs) may have underestimated changes\n\n### 2. Causality Assumption\nAll hypotheses assume transcriptomic changes are pathogenic rather than adaptive or epiphenomenal. The field has struggled to distinguish:\n- **Primary drivers**: Cell-autonomous pathological changes\n- **Secondary responders**: Changes in reaction to other pathological processes  \n- **Adaptive compensation**: Protective responses to limit damage\n- **Aging effects**: Changes that accompany aging but are not disease-specific\n\n### 3. Cross-Disease Specificity\nNone of the hypotheses adequately address why different neurodegenerative diseases (AD, PD, ALS, FTD) target different cell types preferentially, given that they share some transcriptomic signatures (e.g., interferon response genes).\n\n---\n\n## Revised Summary\n\n| Hypothesis | Original Confidence | Revised Confidence | Primary Limitation |\n|------------|--------------------|--------------------|-------------------|\n| 1: TREM2+ Microglia | 0.78 | **0.58** | Directionality and model translation |\n| 2: OPC Arrest | 0.71 | **0.52** | Clinical trial failures, secondary changes |\n| 3: Reactive Astrocytes | 0.76 | **0.62** | A1/A2 classification validity |\n| 4: C9orf72 Glia | 0.69 | **0.54** | Limited generalizability, STING failures |\n| 5: Layer 5 Neurons | 0.64 | **0.48** | Layer assignment, secondary changes |\n| 6: PVMs | 0.58 | **0.42** | Weakest evidence, cell ID challenges |\n| 7: Interneurons | 0.66 | **0.55** | Abundance bias, cause vs. consequence |\n\n---\n\n## Recommended Priorities for Future Work\n\n**Highest priority falsification experiments:**\n\n1. **Temporal sequencing**: Multi-timepoint snRNA-seq is needed across all diseases to establish causal ordering of cell-type changes.\n\n2. **Causal manipulation studies**: CRISPR-based knockout/overexpression in specific cell types in vivo is essential before therapeutic claims.\n\n3. **Human validation**: iPSC-derived cell type models with patient-specific genetics can test whether transcriptomic signatures are cell-autonomous.\n\n4. **Cross-disease integration**: Comparing AD, PD, ALS, FTD transcriptomes may reveal core conserved pathways versus disease-specific alterations.",
  "tokens_used": "6440"
}