Genetics & Risk Factors

1. The Genetic Architecture

Alzheimer’s disease has two genetically distinguishable forms, separated by age of onset, mode of inheritance, and the underlying mutational mechanism, yet converging on a shared molecular endpoint — pathological Aβ accumulation followed by tau spread (covered in Part III and Part IV).

Heritability estimates from twin studies (Gatz et al., Arch Gen Psychiatry 2006) place AD heritability at 58–79% — one of the more heritable common late-life diseases. Yet only a fraction of this is captured by current GWAS “chip heritability” (h²_SNP ~7–10%); the gap is the familiar missing heritability — partly rare variants, partly gene–environment interaction, partly assay limitations.

2. Familial AD: APP Mutations

The amyloid precursor protein gene (APP) sits on chromosome 21q21.3 — a position whose chromosomal context (Down syndrome) had already implicated it in AD before sequencing began. In 1991 Goate et al. (Nature) identified the first AD-causing mutation, the London mutation V717I, in a single English pedigree. Today >60 pathogenic APP mutations are catalogued (Alzforum mutation database). All cluster around the three sites where APP is cleaved — α, β, and γ-secretase — and act by tilting processing toward more or longer Aβ.

The Swedish mutation (KM670/671NL, a double mutation immediately N-terminal to the β-cleavage site) raises BACE1 cleavage efficiency several-fold, increasing total Aβ. The London and Indiana mutations sit at or just past the γ-cleavage site and shift cleavage toward the longer, more aggregation-prone Aβ42. The Arctic mutation (E693G) lies inside Aβ itself and accelerates protofibril formation; this mutation directly motivated development of anti-protofibril antibodies, the lineage that produced lecanemab (covered in Part VII).

The shift in Aβ42 / Aβ40 ratio caused by APP-V717I can be written as:

\( \Delta_{42:40} = \dfrac{[A\beta_{42}]_{mut}}{[A\beta_{40}]_{mut}} \Big/ \dfrac{[A\beta_{42}]_{wt}}{[A\beta_{40}]_{wt}} \approx 2.5\text{–}5 \)

APP duplication. A whole-locus tandem duplication of APP (Rovelet-Lecrux 2006 Nat Genet), increasing gene dosage from 2 to 3 copies, is sufficient to cause autosomal-dominant early-onset AD with cerebral amyloid angiopathy — a result that anticipates and parallels what happens in Down syndrome (next section).

3. Familial AD: Presenilin 1 (PSEN1)

Cloned in 1995 by Sherrington et al. (Nature) as the gene underlying chromosome-14-linked early-onset AD, PSEN1 on 14q24.2 encodes the catalytic subunit of γ-secretase — the intramembrane aspartyl protease that performs the third APP cleavage and generates Aβ. PSEN1 is by far the most common cause of FAD: more than 200 distinct pathogenic variants in >500 families worldwide are listed in the Alzforum database.

• Onset. Average age of symptom onset is ~45 years, with mutation-specific variation from late 20s (e.g. P264L, M146L) to early 60s. Within a family, age of onset is remarkably consistent — useful for prediction in DIAN cohorts.
• Mechanism. Almost all PSEN1 mutations shift γ-secretase processivity, raising Aβ42:Aβ40 ratio without necessarily changing total Aβ. They are gain-of-toxic-function for Aβ42, often combined with partial loss-of-function for endopeptidase trimming.
• Penetrance. Essentially complete — if you carry a pathogenic PSEN1 variant, you will develop AD at the family-typical age unless you die first of unrelated cause.
• Classical pedigrees. The Paisa kindred in Antioquia, Colombia (E280A, >6,000 individuals studied by Lopera, Arboleda, Quiroz) is the largest single-mutation FAD population on Earth and has driven much of our biomarker timeline. The Volga German PSEN2 N141I and the Japanese PSEN1 mutations are other notable founder populations.

The Aβ42 increase produced by representative PSEN1 mutations (e.g. E280A, M146V, L286V) is typically:

\( \dfrac{[A\beta_{42}]_{PSEN1\text{-}mut}}{[A\beta_{42}]_{wt}} \approx 1.5\text{–}3 \quad ; \quad \dfrac{[A\beta_{42}]}{[A\beta_{40}]} \approx 0.3\text{–}0.7 \)

4. Familial AD: Presenilin 2 (PSEN2)

PSEN2, cloned by Levy-Lahad et al. (Science 1995), sits on 1q42.13 and encodes the second presenilin paralogue. It is structurally and functionally near-identical to PSEN1 — either can serve as the catalytic core of γ-secretase — but is much less commonly mutated.

• Frequency. Only ~30 pathogenic variants in <100 families documented — well under 5% of FAD pedigrees.
• Onset. Generally later than PSEN1: average ~55 years, with broader range (40s–80s) and notably variable expressivity even within families.
• Founder mutation. N141I (the “Volga German” mutation) was first identified in ethnic Germans who emigrated to the Volga region of Russia under Catherine the Great in the 18th century, then to North/South America in the 19th — a striking example of allele tracking via human migration history.
• Penetrance. Less than complete (~95%) by age 80; some carriers reach late life unaffected, suggesting modulation by other genetic and environmental factors.
• Mechanism. Same as PSEN1 — altered γ-secretase processivity, ↑ Aβ42:Aβ40. Some PSEN2 variants (e.g. R71W, R62H) are now considered low-penetrance risk factors rather than fully Mendelian.

The relative rarity of PSEN2 compared to PSEN1 — despite their biochemical near-equivalence — is partly explained by tissue distribution: PSEN1 dominates γ-secretase activity in mature neurons, so its mutations show up more readily in disease.

5. Down Syndrome and AD

Trisomy 21 (Down syndrome, DS) carries an extra copy of chromosome 21 — and therefore an extra copy of APP. This 50% increase in APP gene dosage is sufficient on its own to drive an AD-equivalent process at population scale: by age 40, essentially every DS brain shows neuritic plaques and neurofibrillary tangles at autopsy (Wisniewski 1985, Mann 1986).

The DS connection is historically what put APP on the map: in the mid-1980s the AD pathology of DS adults pointed cytogeneticists to chromosome 21, and within five years APP had been cloned, mapped to 21q21, and shown to be the source of the Aβ peptide purified from plaques. Triplication of APP alone — without other chromosome-21 genes — is sufficient to cause AD (the rare partial trisomies and APP-locus duplications described in section 2).

Modern DS research provides a unique natural experiment in gene-dosage-driven AD. Because the genetic insult (and therefore biomarker trajectory) is similar to autosomal-dominant FAD, DS adults are now included as an “FAD-like” arm in trial design. The ABC-DS (Alzheimer’s Biomarker Consortium — Down Syndrome) cohort has produced biomarker trajectories that closely parallel DIAN autosomal-dominant carriers, with amyloid PET turning positive in the late 30s, tau PET ~5 years later, and dementia onset in the mid-50s.

6. APOE — The Major Late-Onset Risk Allele

In 1993 Strittmatter, Saunders, Roses et al. (PNAS 90:1977) reported that the APOE-ε4 allele was vastly over-represented in late-onset AD. Three decades on, APOE remains by an order of magnitude the largest single genetic determinant of sporadic AD risk. No common variant identified since has come close.

6.1 The three-allele system

APOE, on 19q13.32, encodes the 299-residue plasma lipoprotein apolipoprotein E. Three common alleles are defined by two non-synonymous SNPs at codons 112 and 158:

The single Cys112Arg substitution is enough to flip APOE from protective to risk — a remarkable example of a single human-readable amino acid change with population-wide neurological consequence.

6.2 Risk magnitude — dose effect

Pooled meta-analysis (e.g. ADGC, Reiman et al. Nat Commun 2020) gives odds ratios for AD relative to the ε3/ε3 reference genotype:

Risk is approximately log-linear in ε4 dose. Each ε4 allele also shifts age of onset earlier by ~7–9 years. In a recent re-analysis of ~3,200 ε4 homozygotes, Fortea et al. (Nat Med 2024) argued that ε4/ε4 is so consistently biomarker-positive by age 65 that it should be considered a genetic form of AD in its own right — a controversial but influential reframing.

Defining p as the population frequency of ε4, the equilibrium prevalence of the high-risk ε4/ε4 genotype under Hardy–Weinberg is:

\( P(\varepsilon4/\varepsilon4) = p^2 \approx 0.14^2 \approx 2\% \)

— matching the observed frequency. With ε4 OR ≈ 12, this 2% of the population accounts for a disproportionate share of the population-attributable fraction of AD: roughly

\( PAF_{\varepsilon4/\varepsilon4} \;=\; \dfrac{p^2(OR-1)}{1+p^2(OR-1)} \;\approx\; \dfrac{0.02 \cdot 11}{1+0.02 \cdot 11} \;\approx\; 18\% \)

6.3 APOE biochemistry — why ε4 raises risk

APOE is the principal lipid-transport protein of the central nervous system. In the periphery it shuttles cholesterol on chylomicrons and VLDL; in the brain it is secreted by astrocytes (and by activated microglia) and carries cholesterol and phospholipids to neurons via LDLR-family receptors. Multiple, partially overlapping mechanisms link APOE-ε4 to AD pathology:

1. Aβ clearance. APOE3 (and especially APOE2) bind Aβ and facilitate its clearance via LRP1, LDLR, and across the blood–brain barrier; APOE4 is less efficient at this. ε4 carriers have higher amyloid PET signal at any given age.
2. Aβ aggregation kinetics. APOE4 promotes Aβ fibrillisation and seeds plaque formation; APOE2 slows it.
3. Lipid & cholesterol transport. ε4 disturbs neuronal membrane lipid homeostasis, impairs synaptic plasticity, and disrupts myelination — with detectable phenotypes decades before pathology.
4. Tau / tauopathy modulation. APOE4 directly accelerates tau pathology and tau-driven neurodegeneration in tauopathy mouse models, independent of Aβ (Shi et al., Nature 2017).
5. Microglial activation. ε4 microglia adopt a pro-inflammatory state, with altered TREM2 and ApoE-dependent phagocytosis (covered in section 9).
6. Cerebral amyloid angiopathy (CAA). ε4 (and ε2 in a different way) increases risk of vascular Aβ deposition and CAA-related haemorrhage — clinically critical for anti-amyloid antibody dosing (ARIA risk).

6.4 ε4 in trial design and clinical care

• Trial enrichment. CLARITY-AD (lecanemab) and TRAILBLAZER-ALZ-2 (donanemab) genotyped all participants; the proportion of ε4 carriers in modern AD trials is ~65–70%.
• ARIA risk. Amyloid-related imaging abnormalities (ARIA-E oedema, ARIA-H micro-haemorrhage) on anti-amyloid antibodies are 2–4× more common in ε4 carriers, and ~5–8× more in ε4/ε4 homozygotes — driving the FDA label recommendation for genotyping before prescription.
• Clinical disclosure. Standard practice in 2026 is to genotype before lecanemab/donanemab and counsel patients accordingly. Disclosure outside this therapeutic context remains controversial; the REVEAL studies (Roberts et al.) showed ε4 disclosure does not produce sustained psychological harm in cognitively normal adults.
• Therapeutic targeting. ApoE-targeting strategies in trial: anti-ApoE antibodies (HAE-4, Holtzman lab), ASO-mediated APOE knockdown, gene-therapy delivery of APOE2, and small-molecule structure correctors that “ε3-ify” ε4 (Mahley/Huang).

7. The APOE-ε2 Protective Effect

The APOE-ε2 allele is the often-overlooked third arm of the system. Carrying one copy of ε2 reduces risk of AD to ~0.6× the ε3/ε3 reference, and homozygous ε2/ε2 to ~0.4× — arguably the largest known protective allele in human disease.

In a large recent ε2/ε2 ascertainment (Reiman et al. Nat Commun 2020) the lifetime AD risk by age 85 was estimated at ~5%, vs ~11% for ε3/ε3 and ~60% for ε4/ε4 — a ~12-fold spread between protective and risk homozygotes. ε2 also defers age of onset by ~5 years on average, and is associated with reduced amyloid PET signal at any age.

Mechanistically, APOE2 is the most efficient Aβ-clearance isoform and the slowest to seed Aβ aggregation; it also produces a less inflammatory microglial state. The trade-off: ε2 raises plasma triglycerides and predisposes to Type III hyperlipoproteinaemia, and there is some evidence ε2 increases primary age-related tauopathy (PART) and cerebral amyloid angiopathy — risk does not vanish, it redistributes.

8. The GWAS Landscape — >75 Loci

Genome-wide association studies of late-onset AD have grown from ~10 loci in 2009 (Harold, Lambert) to >75 independent risk loci in Bellenguez et al. (Nat Genet 2022, 111,326 cases / 677,663 controls) — the largest AD GWAS to date. After the towering APOE peak, individual non-APOE loci contribute small odds ratios (1.05–1.4), but their pathway architecture is enormously informative.

8.1 Pathway grouping of GWAS hits

The dominant signal is the microglial / innate immune axis. After APOE and the lipid pathway, more GWAS hits sit in microglia-expressed genes than anywhere else — an unbiased re-discovery of what was for decades a sideshow in AD biology and is now a central pillar (the immune hypothesis, section 9).

8.2 SNP heritability

From Bellenguez 2022 LD-score regression and similar analyses:

\( h^2_{SNP,\,AD} \approx 0.07\text{–}0.10 \quad\Rightarrow\quad \text{vs twin } h^2 \approx 0.6\text{–}0.8 \)

The gap is the “missing heritability”: rare variants (TREM2-R47H, SORL1 truncating mutations, ABCA7 loss-of-function), structural variants, gene–gene and gene–environment interactions, and population stratification not captured by chip-based GWAS. Whole-genome sequencing efforts (ADSP, NIAGADS) are progressively filling in the gap.

Cohorts driving modern AD genetics: ADGC (Alzheimer’s Disease Genetics Consortium, USA), EADI (European AD Initiative), CHARGE, GR@ACE, FinnGen, and UK Biobank (with proxy-AD via parental history). Bellenguez 2022 combined these into the IGAP3 meta-analysis.

9. TREM2 and Microglial Variants — The Immune Axis

In 2013 two parallel studies — Guerreiro et al. and Jonsson et al. (NEJM, back-to-back) — reported that the TREM2-R47H rare variant (minor allele frequency ~0.3% in Europeans) increased AD risk ~2.5–4-fold, an effect comparable in magnitude to one ε4 allele. This established beyond any reasonable doubt that microglia, not just neurons, sit at the causal heart of AD.

9.1 TREM2 biology

TREM2 (triggering receptor expressed on myeloid cells 2) on 6p21.1 encodes a single-pass transmembrane immunoglobulin-superfamily receptor expressed almost exclusively on myeloid cells — in brain, on microglia. Its ligands include a wide range of damage-associated molecular patterns: anionic lipids, apoptotic-cell phospholipids, myelin debris, oligomeric Aβ, and ApoE/Aβ complexes. Signalling proceeds via the adaptor DAP12 (TYROBP), and downstream: SYK, PI3K/AKT, mTOR, NF-κB, and metabolic switching to oxidative phosphorylation.

Functionally, TREM2 drives microglial transition into a disease-associated microglia (DAM) / activated response microglia (ARM) state — a phagocytically active, plaque-clustering, lipid-handling state characterised by Keren-Shaul, Amit and colleagues (Cell 2017). Without TREM2, microglia fail to compact plaques, fail to clear myelin debris, and fail to limit tau spread.

9.2 TREM2 risk variants

All AD-associated TREM2 variants are partial loss-of-function — they reduce TREM2’s ability to detect ligand or signal through DAP12. The homozygous complete-loss-of-function form gives Nasu-Hakola disease (polycystic lipomembranous osteodysplasia with sclerosing leukoencephalopathy) — presenile dementia plus bone cysts — the most extreme demonstration of how essential microglial TREM2 is to neural homeostasis.

9.3 Other microglial AD variants

• PLCG2-P522R — rare protective variant (OR ~0.7), gain-of-function in TREM2 downstream signalling. Validates the TREM2 → PLCG2 axis as a drug target.
• CR1 — complement receptor 1; AD risk allele alters microglial Aβ phagocytosis via complement opsonisation.
• CD33 / SIGLEC3 — risk allele increases CD33 surface expression, an inhibitory receptor that suppresses microglial Aβ clearance.
• MS4A4A / MS4A6A — influence soluble TREM2 levels and microglial activation state.
• SPI1 (PU.1) — master myeloid transcription factor; AD-protective allele lowers PU.1 expression and delays AD onset.
• INPP5D — microglial inositol phosphatase opposing TREM2/PI3K signalling; risk allele acts in microglia.

10. Polygenic Risk Scores and the Future of Prediction

A polygenic risk score (PRS) sums an individual’s allele dosages across GWAS-identified variants, weighted by their effect sizes:

\( PRS_i \;=\; \sum_{j=1}^{M} \hat{\beta}_j \, x_{ij} \)

where \(x_{ij}\in\{0,1,2\}\) is the dosage of risk allele \(j\) in individual \(i\) and \(\hat\beta_j\) the GWAS log-odds. For AD, two flavours are commonly reported:

• APOE-inclusive PRS. Dominated by the APOE region; AUC for distinguishing AD vs control ~0.72–0.78.
• APOE-stratified (non-APOE) PRS. Computed within ε4 strata; AUC ~0.65 on its own; provides clinically meaningful additional risk stratification beyond APOE.

In a typical clinical-risk model, the lifetime risk of AD can be modelled as a competing-risk Cox regression on age, sex, APOE dose, and PRS:

\( \lambda(t \mid X) \;=\; \lambda_0(t)\,\exp\!\left(\beta_{age} t + \beta_{\varepsilon 4} d_{\varepsilon 4} + \beta_{PRS}\,PRS_i \right) \)

In UK Biobank and similar cohorts, individuals in the top PRS decile have ~2–3× higher risk than the median, conditional on APOE. Combined with APOE-ε4/ε4, top-decile PRS pushes lifetime AD risk above 70%.

10.1 Clinical utility — what is PRS actually for?

• Trial enrichment. PRS-selected high-risk cognitively-normal cohorts can shorten primary-prevention trials by 30–50% by raising event rates.
• Risk stratification within ε4 carriers. A high-PRS ε4 carrier and a low-PRS ε4 carrier have meaningfully different absolute risks; PRS adds resolution beyond APOE alone.
• Combination with biomarkers. PRS + plasma p-tau217 + age now produces 5- and 10-year risk models with C-statistics >0.85 in cognitively normal adults — covered in Part VI.
• Population screening? Not yet justified at population scale — positive predictive value is modest in unselected populations and the actionable interventions (anti-amyloid antibodies) carry their own risks.

10.2 The disclosure question

Direct-to-consumer testing (23andMe and successors) reports APOE genotype to anyone who asks. Whether this is good practice in a disease that — until 2023 — had no disease-modifying therapy has been debated for two decades (Roberts and Green’s REVEAL studies; Goldman 2022 review). The arrival of lecanemab and donanemab, both with ε4-stratified ARIA risk and labelled to require genotyping, has effectively forced APOE testing into mainstream clinical care. Disclosure of non-APOE AD PRS in cognitively normal adults remains an active ethical and operational question.

10.3 Limitations

• European-ancestry bias. Most AD GWAS and PRS are derived from European-ancestry cohorts; performance in African, East Asian, Hispanic populations is materially worse. Closing this gap is a top-priority area of AD genetics 2024–2026.
• Common-variant ceiling. Even an “ideal” PRS captures only the SNP-heritability fraction (~10%); rare-variant scores will need to be added.
• Environmental gain. Biomarker-based prediction (plasma p-tau217, amyloid PET) now exceeds genetic prediction by a wide margin in symptomatic populations — PRS shines in pre-symptomatic, biomarker-negative populations and for trial enrichment.