# Critical Evaluation of AD Fine-Mapping Hypotheses
## Hypothesis 1: Strong LD Enables Narrow Credible Sets
### Weaknesses
**The fundamental problem with this hypothesis is a category error**: strong LD is a *hindrance*, not a help, for variant-level resolution. When variants are highly correlated, posterior probability diffuses across the LD block, making pinpointing the causal variant statistically harder, not easier. The hypothesis conflates "high statistical power to detect association" with "narrow credible sets."
The APOE/TOMM40 region is particularly problematic as an exemplar. Despite the strong effect size, fine-mapping has been notoriously contentious:
- The extended LD block spans APOE promoter regions, TOMM40, and downstream elements
- The "APOE ε4 tag SNP" problem has been debated for years—some argue rs429358/rs7412 (the functional APOE4 missense variants) should be the causal anchor, while others identify regulatory proxies
- Karch et al.'s posterior probability of >0.85 should be scrutinized: is this for the missense variants themselves or for tagging SNPs? These are fundamentally different claims.
The phrase "natural statistical bins" is circular—the challenge is precisely identifying which bin contains the causal variant.
### Counter-Evidence
- Fine-mapping studies of the APOE region (e.g., Ridge et al., 2013; Corradini et al., 2020) have repeatedly failed to achieve single-digit credible sets despite decades of study
- The region defies easy resolution because LD spans both regulatory and coding variation
- Even the CASZ1 locus adjacent to APOE demonstrates how extended LD can confuse inference
### Falsification Experiments
1. **Allele-specific expression in human brain tissue**: Test whether the highest-posterior variant shows allele-specific expression in microglia or neurons. If the top SNP has no allele-specific effects while a lower-probability SNP does, the posterior is misallocated.
2. **CRISPR base editing of all top 10 credible set variants in iPSC-derived microglia**: Quantify APOE expression changes. The causal variant must show functional effects; others should not.
3. **ATAC-seq allelic imbalance in brain tissue**: Functional validation of regulatory variant candidacy.
### Revised Confidence: **0.58** (down from 0.82)
The confidence of 0.82 is substantially inflated. The stated mechanism is misunderstood—strong LD complicates rather than simplifies fine-mapping. The confidence should reflect the known difficulty of the APOE region despite its large effect size.
---
## Hypothesis 2: Multi-Omics Integration Sharpens Credible Sets 40-60%
### Weaknesses
**1. The claimed magnitude (40-60%) lacks theoretical justification or empirical precedent.**
The reduction from incorporating chromatin annotations is bounded by how much posterior mass currently concentrates on non-coding regulatory regions versus tagging SNPs. If 70% of posterior probability already falls on a variant in LD with the functional causal variant, annotation integration can only recover 30% at maximum—even this assumes perfect annotation calibration.
**2. Informative priors from ATAC-seq and ChIP-seq carry substantial assumptions:**
- Tissue specificity: ATAC-seq from bulk microglia contains mixed cell types; accessible chromatin may reflect multiple lineages
- Temporal specificity: AD-relevant chromatin states may differ from young adult tissue donors
- The assumption that "accessible = causal regulatory variant" conflates accessibility with actual regulatory function
**3. Model dependence:** The 40-60% reduction is highly sensitive to:
- How functional annotations are weighted relative to statistical LD
- Prior specification choices (Laplace versus normal mixtures)
- Calibration of annotation weights against truth-known benchmarks
**4. INPP5D/PLCG2 as the target gene is problematic:**
- INPP5D shows complex splicing patterns with multiple isoforms
- PLCG2 is a signaling enzyme with limited clear eQTL patterns compared to surface receptors
- These are not the canonical GWAS signal genes for AD—they appear secondary
### Counter-Evidence
- The GARFIELD algorithm (Iotchkova et al., 2019) showed more modest improvements (~20-30%) from regulatory annotations in fine-mapping
- Enrichment of GWAS variants in regulatory elements doesn't translate linearly to posterior probability allocation
- Many "functional" regulatory variants in ATAC-seq peaks are not causal for trait differences
### Falsification Experiments
1. **Simulations with known causal variants**: Generate GWAS summary statistics under realistic LD with known causal variants embedded in regulatory elements. Apply multi-omics Bayesian integration. Measure calibration: are 90% of credible sets actually covered? If coverage is <85%, the priors are miscalibrated.
2. **Holdout validation**: Train annotation-informed priors on 20 AD loci, test on held-out loci with known causal variants (from Mendelian mutations or functional studies). This tests generalizability.
3. **Null experiments**: Apply the same multi-omics framework to null phenotypes (e.g., hair color in AD cohorts). If credible set reduction occurs similarly, annotations are spurious.
### Revised Confidence: **0.52** (down from 0.76)
The mechanism is plausible in principle but the claimed effect size (40-60%) is unsupported by theory or existing benchmarks. The target gene selection is also suboptimal. A confidence in the 0.50-0.55 range better reflects the uncertainty.
---
## Hypothesis 3: Allelic Heterogeneity Confounds 3-5 Loci
### Weaknesses
**1. The "3-5 loci" estimate may be conservative.**
Conditional analyses from Bellenguez et al. (2022) identified secondary signals in BIN1, CLU, PTK2B, and others. How many of the remaining 21 loci would show allelic heterogeneity with larger sample sizes? The number could approach 8-10, not 3-5.
**2. Standard fine-mapping tools can accommodate multiple signals—but with caveats:**
- SuSiE and DAP can model multiple causal variants, but they require:
- Larger sample sizes (current AD GWAS may be underpowered for detecting modest secondary signals)
- Proper LD reference panels
- Computational burden increases substantially
- The hypothesis states these tools "can accommodate" but doesn't address whether current sample sizes provide sufficient power for stable multi-signal estimation
**3. The phrase "Standard fine-mapping assumes a single causal variant" reveals a strawman.**
Most current fine-mapping methods (FINEMAP, CAVIAR, SuSiE) explicitly allow multiple causal variants. The real issue is whether secondary signals are adequately powered.
### Counter-Evidence
- The number of secondary signals detected has consistently increased with sample size in other complex diseases (T2D, schizophrenia GWAS), suggesting the current "3-5" may be an artifact of limited power
- BIN1's primary signal is in strong LD with a splicing QTL that may be the causal mechanism; secondary signals may be tagging distinct regulatory elements
### Falsification Experiments
1. **Bootstrap confidence intervals**: For each locus, compute credible sets with 100 bootstrap resamples of the summary statistics. If credible set size varies by >2-fold across bootstraps, the locus is underpowered for stable multi-signal estimation.
2. **Leave-one-population-out meta-analysis**: Remove one ancestral group; if secondary signals disappear, they may be artifacts
3. **Power calculation**: For each locus, compute the expected power to detect a secondary signal at OR=1.05 with current sample sizes. Loci with power <80% cannot be reliably classified.
### Revised Confidence: **0.74** (up from 0.68)
This hypothesis is actually better supported than the others. Allelic heterogeneity is a known complication in complex trait genetics, and the proposed range (3-5 loci) is likely conservative. The main uncertainty is whether this is actually 3-5 or higher.
---
## Hypothesis 4: Brain eQTL Colocalization Doubles Confidence (Factor of 2.5)
### Weaknesses
**1. The factor of 2.5 has no theoretical basis.**
Colocalization posterior probability is bounded by the prior probability of colocalization, which depends on:
- eQTL effect size
- GWAS effect size
- LD structure
- Prior probability that a variant affects both traits
Doubling or tripling posterior probability requires specific parameter combinations—the general claim that brain eQTL colocalization universally increases confidence 2.5-fold is unjustified.
**2. eQTL hotspots cause false colocalizations.**
The MS4A locus is particularly susceptible: MS4A6A, MS4A4A, MS4A2, MS4A3 form a tight cluster. A variant affecting overall chromatin accessibility in the region can appear to colocalize with multiple genes' expression without being causal for any specific one.
**3. Statistical limitations of coloc:**
The coloc method (Giambartolomei et al., 2014) computes P(H4 | data), the posterior probability that a single causal variant explains both GWAS and eQTL signals. However:
- It assumes a single causal variant per signal—problematic given allelic heterogeneity
- The H4 posterior is sensitive to prior specifications (π₁, π₂ parameters)
- eQTL sharing may reflect regulatory networks, not causal relationships
**4. Pleiotropy complicates interpretation.**
A variant could colocalize with MS4A expression without being causal for AD—it could affect both through independent pathways.
### Counter-Evidence
- Systematic benchmarks of colocalization methods (e.g., Wallace et al., 2021) show high false positive rates when eQTL and GWAS signals are close but not shared
- Many AD colocalization findings haven't translated to functional validation
- The MS4A locus is unusual—most AD loci don't have such strong brain eQTL signals
### Falsification Experiments
1. **Test colocalization on negative control gene pairs**: Use expression of genes with no biological relationship to AD (e.g., liver-specific genes). Measure colocalization posterior probabilities. If >0.5 of negatives show colocalization, method is overcalibrated.
2. **Permutation test**: Permute eQTL sample labels to destroy real eQTL associations. Measure residual colocalization—this is the false positive rate.
3. **Functional follow-up**: Take the top 10 colocalized variants (post-hoc) and test in cellular models. If <2 show functional effects, colocalization is not capturing causality.
### Revised Confidence: **0.48** (down from 0.74)
The claimed 2.5-fold confidence increase is too specific to be credible given the complexities of colocalization analysis. The MS4A locus is a reasonable target but the methodology has known limitations that are underweighted in this hypothesis.
---
## Hypothesis 5: Low-LD Loci Yield Credible Sets >50 Variants
### Weaknesses
**1. The hypothesis conflates "large credible set" with "interpretation infeasible."**
A credible set of 50 variants is not inherently infeasible—it depends on:
- Whether these variants cluster in accessible regulatory regions
- Whether functional annotation can prioritize them
- Whether experimental follow-up can test them in parallel
The problem is "large" but not necessarily "unsolvable."
**2. CASS4 is the correct exemplar, but the effect size cited (OR ~1.1) suggests it's underpowered rather than simply "low LD."**
The poor resolution is driven more by weak statistical signal than by LD architecture alone. A locus with OR=1.1 will have wide confidence intervals regardless of LD structure.
**3. The "centromere/chromosomal arm" localization is overly specific.**
Recombination hotspots create low-LD regions throughout the genome, not only near centromeres.
**4. "Unacceptably large" is a value judgment, not a quantitative threshold.**
What makes 50 variants unacceptable? CRISPR screens can handle this scale.
### Counter-Evidence
- PTK2B (also a top AD locus) has moderate LD structure but manageable credible sets
- CASS4's small effect size may simply reflect that current GWAS sample sizes lack power, not that the locus is inherently unresolvable
- With larger GWAS (future meta-analyses), credible set sizes will shrink regardless of LD structure
### Falsification Experiments
1. **Simulate GWAS with OR=1.1 for true causal variant; compute credible set size**: If <50 variants in simulations with current LD architecture, the hypothesis overstates the problem
2. **Cross-population LD**: Test whether CASS4 credible sets in African ancestry populations (with different LD patterns) are smaller. If yes, the problem is ancestry-specific, not inherent to the locus.
3. **Compare to matched OR loci**: Compute credible set sizes for other OR~1.1 loci with better LD. If those are similarly large, the problem is effect size, not LD structure.
### Revised Confidence: **0.77** (up from 0.71)
This hypothesis is well-supported and may even underestimate the problem (more than 1-2 loci may have this issue). The framing is slightly problematic but the core claim—that sparse LD combined with weak effects yields large credible sets—is correct.
---
## Hypothesis 6: Cell-Type Specific Regulatory Architecture Identifies Microglia-Predominant Causal Variants
### Weaknesses
**1. MEF2C is a reasonable target but the hypothesis oversimplifies regulatory biology.**
MEF2C has multiple isoforms and complex regulatory architecture. "