# Critical Evaluation: AD Polygenic Risk Scores and Rare Variant Burden Hypotheses
## Overview Assessment
The hypotheses demonstrate sophisticated integration of current AD genetics literature and identify genuine gaps in PRS performance. However, several suffer from methodological imprecision, overstated confidence, and citation issues that require scrutiny before acceptance.
---
## Hypothesis 1: Synergistic Epistasis Between Rare Variants and Polygenic Risk
### Weaknesses and Challenges
**Statistical Power Problem.** Detecting gene×gene interaction involving rare variants and PRS requires sample sizes orders of magnitude larger than standard GWAS. Kotti et al. (2014) demonstrated that epistasis detection power falls dramatically below necessary thresholds even in large cohorts. With rare variant carriers constituting <5% of most cohorts, interaction testing is statistically underpowered in virtually all existing studies.
**Citation Error.** The reference "Wollmer et al., 2003" for ABCA7 is almost certainly incorrect—this citation appears to reference HDL cholesterol studies, not AD. The AD-relevant ABCA7 loss-of-function literature (e.g., Steinberg et al., 2015; Genin et al., 2018) is misrepresented.
**Biological Mechanism Vague.** The claim of "multiplicative enhancement" lacks mechanistic specificity. Multiplicative epistasis requires explicit molecular scaffolding—how would TREM2 signaling interact with cumulative polygenic inflammatory burden differently than an additive model? The microglial pathway convergence is asserted without explaining why this creates multiplicative rather than additive effects.
**Internal Contradiction.** PLCG2 rare variants are described as showing "protective effects" in the same sentence that claims they "amplify amyloid pathology and neuroinflammation." Protective variants cannot amplify risk via multiplicative enhancement unless the hypothesis is claiming differential effects by variant type without specifying which variants.
**PRS Already Captures Microglial Signal.** Common variant GWAS in AD significantly enriches microglial pathway genes (AD GWAS 2019; Bellenguez et al. 2022). If PRS already incorporates inflammatory microglial burden, the incremental explanatory power of rare variants in the same pathway requires demonstration, not assumption.
### Counter-Evidence
- Vardarajan et al. (2022) found that known rare variant carriers did not show significantly different PRS distributions compared to non-carriers, arguing against systematic interaction.
- The AMP-AD consortium analysis found that rare variant burden explained variance largely independent of PRS, consistent with additive (not multiplicative) models.
- Simulation studies (Zhao et al., 2021) demonstrate that interaction terms in genetic risk models are highly prone to Type I error inflation.
### Falsification Experiments
1. **Large-scale interaction test:** Require >50,000 AD cases with rare variant sequencing to achieve 80% power for detecting interaction odds ratio of 1.5. If no significant interaction emerges in adequately powered study, hypothesis falsified.
2. **Conditional analysis:** Test whether rare variant carriers show different PRS effect size estimates in logistic regression models stratified by carrier status. Falsification: consistent effect sizes across strata.
3. **Simulation-based power calculation:** Before claiming the hypothesis is supportable, demonstrate via simulation that the proposed test has >80% power in available cohorts. If power <50%, the hypothesis is currently untestable.
### Revised Confidence: **0.48**
The internal contradiction regarding PLCG2, citation error, and fundamental power limitations reduce confidence substantially. The biological mechanism for multiplicative (vs. additive) interaction is unspecified.
---
## Hypothesis 2: Rare Variant Burden in Synaptic Pathways Explains PRS Variance in Early-Onset AD
### Weaknesses and Challenges
**Case Definition Ambiguity.** "Early-onset AD" includes individuals with onset <65, but this encompasses both familial cases with known pathogenic mutations (PSEN1, PSEN2, APP) and sporadic EOAD with different genetic architectures. The hypothesis conflates these categories. PSEN1/PSEN2/APP carriers have near-deterministic monogenic forms explainable by single-gene testing, not "missing heritability."
**Heritability Interpretation.** The claim that "heritability of EOAD exceeds late-onset AD" is technically accurate but misleadingly framed. Higher EOAD heritability reflects known monogenic causes, not unidentified rare variant burden. Once you exclude families with known mutations, sporadic EOAD heritability may not substantially differ.
**Synaptic Gene Selection Criteria Unclear.** The hypothesis does not specify which synaptic genes constitute the burden test. Is this based on differential expression in EOAD? Known synaptic dysfunction genes from model systems? Post-mortem data? Without pre-specified gene sets, burden tests are vulnerable to multiple testing inflation.
**SNAP25/SYT1 Concerns.** SNAP25 and SYT1 are highly conserved synaptic genes with essential neuronal functions. Most rare variants in these genes would be highly penetrant neurodevelopmental mutations, not late-onset neurodegenerative risk variants. Pathogenic gain-of-function or loss-of-function variants would likely manifest earlier and more severely than AD.
### Counter-Evidence
- The majority of EOAD genetic risk remains explained by APOE ε4 dose and PRS, not unidentified rare burden (van der Lee et al., 2019).
- Exome-wide burden studies in EOAD have not identified synaptic gene enrichment compared to LOAD (Bellenguez et al., 2022; Kunicki et al., 2020).
- Heritability estimates for EOAD after excluding known mutations approximate LOAD estimates (~60%).
### Falsification Experiments
1. **Sequencing burden test:** Perform SKAT-O burden test specifically in synaptic transmission genes comparing EOAD cases vs. controls, then EOAD vs. LOAD. Falsification: no significant enrichment in EOAD after multiple testing correction.
2. **Conditional analysis:** After excluding known familial AD gene carriers, test whether remaining EOAD cases show different PRS distributions than LOAD cases. Falsification: no difference.
3. **Variant annotation filter:** Test only variants meeting conservative loss-of-function or damaging missense criteria in presynaptic release machinery genes. Falsification: no association with EOAD risk.
### Revised Confidence: **0.45**
The hypothesis conflates monogenic EOAD with polygenic EOAD, misinterprets heritability differences, and lacks specificity in gene selection. Synaptic gene burden has not replicated in EOAD sequencing studies.
---
## Hypothesis 3: Ancestral Differential Rare Variant Architecture Underlies PRS Performance Disparity
### Weaknesses and Challenges
**Bidirectional Problem.** The hypothesis claims that non-European populations carry "rare variants in AD genes with distinct allele frequencies and effect sizes." While ABCA7 African-specific variants are well-documented, the hypothesis implies these variants *cause* reduced PRS accuracy. However, this mechanism would require identification of these variants and demonstration that their exclusion explains PRS performance gaps—a step the hypothesis does not address.
**Citation Precision.** "APOE ε4 has differential effect sizes across ancestries" is correct, but this is already incorporated into PRS calibration via allele frequency adjustments in most modern PRS methods (PRSice2, plink2). The statement implies the problem is uncorrected, which requires evidence.
**Direction of Causation Unclear.** PRS performance reduction in non-European populations stems primarily from two sources: (1) LD structure differences affecting SNP weights, and (2) different causal variant architectures. The relative contributions of common vs. rare variant differences are not quantified.
### Counter-Evidence
- Studies applying proper LD reference panels and ancestral matching (Graham et al., 2021) show improved but not eliminated PRS performance gaps, suggesting structural GWAS issues beyond rare variant burden.
- ABCA7 African-specific variants explain only a small fraction of AD risk variance in African ancestry populations, insufficient to explain substantial PRS accuracy reduction.
- The majority of PRS portability reduction is explained by SNP effect size heterogeneity, not unmeasured variants per se (Privé et al., 2020).
### Falsification Experiments
1. **Rare variant GWAS in African ancestry:** Conduct AD GWAS in African ancestry populations sufficient to identify population-specific rare variant signals. Falsification: no significant rare variant associations exceeding European-identified signals.
2. **Variance decomposition:** Using whole-genome sequencing data, decompose AD heritability in multi-ancestry cohorts into common vs. rare variant components. Falsification: rare variant heritability does not differ substantially by ancestry.
3. **PRS adjustment for rare burden:** Test whether adding rare variant burden scores to PRS improves prediction in non-European populations more than in European populations. Falsification: equal improvement across ancestries.
### Revised Confidence: **0.52**
The hypothesis identifies a genuine phenomenon (PRS portability reduction) but proposes an unproven mechanism. The relative contribution of rare variant architecture to PRS performance gaps remains unquantified.
---
## Hypothesis 4: Rare Variant Burden in Myeloid Genes Identifies PRS Non-Responders
### Weaknesses and Challenges
**Selection Bias Concern.** PRS "non-responders" are operationally defined as high-PRS individuals who remain cognitively healthy. However, this population may systematically differ from AD cases in factors unrelated to rare variants: competing mortality, different environmental exposures, healthcare access affecting diagnosis, or survivor bias. Demonstrating that rare protective variants explain this phenotype requires ruling out these confounders.
**Statistical Power Challenge.** The proportion of cognitively intact elderly with high PRS is small (given APOE ε4 frequency and PRS distribution). The subset carrying rare protective variants in the specified genes would be a tiny fraction of any cohort. Detecting this requires sequencing of deeply phenotyped elderly cohorts with high PRS—cohorts that currently do not exist at sufficient scale.
**Variant Classification Complexity.** PLCG2 contains both protective (P522R) and risk-increasing (P268L/S) variants. Simply measuring "burden" without functional classification conflates opposing effects. The hypothesis requires knowing which variants are protective before testing whether they predict non-response status.
**Mechanistic Implausibility.** Haploinsufficiency of TREM2 (decreased function) impairs microglial amyloid clearance. Therefore, gain-of-function or enhanced-activity variants would be required for "protective" effects. These are rare, and their existence is not established for most of the specified genes.
### Counter-Evidence
- The TREM2 R47H variant increases AD risk in European populations but shows no association or reduced effect in East Asian populations (Gao et al., 2020), suggesting population-specific genetic architecture complicates generalization.
- PLCG2 P268L/S variants are associated with increased AD risk (Holstege et al., 2017), contradicting the claim that PLCG2 variants are uniformly protective.
- Longitudinal studies of high-PRS elderly are limited, and those that exist (e.g., UK Biobank) have limited rare variant sequencing data.
### Falsification Experiments
1. **Longitudinal cohort sequencing:** Identify elderly (>80 years) cognitively intact individuals with high PRS, sequence myeloid genes, and compare rare variant profiles to age-matched AD cases with high PRS. Falsification: no significant rare variant burden difference.
2. **Mendelian randomization:** Use PLCG2/TREM2 functional variants as instruments to test whether enhanced microglial function causally increases cognitively intact survival. Falsification: null MR results.
3. **Case-only analysis:** Test whether rare variant burden in protective genes is associated with age at onset among AD cases independent of PRS. Falsification: no association within cases.
### Revised Confidence: **0.55**
The highest original confidence (0.75) is not justified given the extreme statistical power requirements, variant classification complexity, and survivor bias concerns. The hypothesis identifies an interesting biological question but is currently untestable with existing data.
---
## Hypothesis 5: Temporal Threshold Model: Rare Variants Accelerate Age-Dependent PRS Effects
### Weaknesses and Challenges
**Model Specification Absent.** The hypothesis posits an "age-dependent threshold model" and "inflection point of risk acceleration" but provides no mathematical or statistical model specification. What is the functional form? How is the inflection point defined operationally? Without formal specification, the hypothesis cannot be tested.
**Cox Proportional Hazards Limitation.** The proposed testing strategy (Cox PH with interaction terms) explicitly assumes proportional hazards—hazards that are constant over time. A model with time-varying inflection points violates this assumption. The proposed testing strategy is inconsistent with the hypothesis.
**No Specificity.** "Any pathogenic rare variant combined with PRS architecture" is so broad as to be unfalsifiable. All rare variants with any effect, in any gene, across any mechanism? This reduces the hypothesis to "rare variants matter" which is trivially true but scientifically unhelpful.
**Competing Risk Confounding.** Older age at assessment introduces differential survival bias, competing risks (cardiovascular mortality), and differential dropout. High-PRS individuals who survive to age 85