Alzheimer’s disease is the most common form of dementia and one of the leading causes of death in aging populations. It is highly heritable, and genome-wide association studies have so far identified more than 70 risk loci, most of them common single-nucleotide polymorphisms, or SNPs. The strongest known risk factor, the APOE ε4 allele, carries an odds ratio of roughly 3.2 for heterozygous carriers, while the many other common variants exert individually weak but collectively substantial effects. Yet even taken together, these known variants do not fully explain the heritability observed in twin and family studies—a gap that has long hinted at unidentified genetic factors at work.

One obvious blind spot has been structural variants: deletions, insertions, duplications and repeat expansions that affect stretches of DNA longer than 50 base pairs. Structural variants touch far more nucleotide content across the genome than SNPs do, and rare examples are already known to be devastatingly influential—duplication of the APP gene, for instance, causes autosomal dominant Alzheimer’s in a handful of families. But the workhorse technologies of large-scale genetics, microarrays and short-read sequencing, cannot accurately discover or genotype more than half of structural variants, because their short reads cannot resolve the full sequence spanning a rearrangement. Long-read sequencing solves this problem by producing continuous reads of several kilobases, but its cost has kept sample sizes small.

The research team, led by Han Cao, Yifan Fu and Xiaopu Zhou under the senior authorship of Amy K. Y. Fu and Nancy Y. Ip at the Hong Kong University of Science and Technology, devised a hybrid strategy to get around that cost barrier. They first built a high-confidence catalog of structural variants from PacBio long-read data covering 11 Chinese genomes, generating 1.37 terabases of clean reads with an average length of 10.35 kilobases—roughly 40-fold coverage of the human genome. Alignments from two long-read aligners, NGMLR and minimap2, were fed into three structural variant callers, Sniffles2, cuteSV and SVIM, and the merged results were filtered through a random forest model trained to distinguish genuine variants from artifacts. The pipeline yielded 44,122 high-confidence variants—21,821 insertions and 22,301 deletions—with area-under-the-curve scores of 0.981 for insertions and 0.958 for deletions, and sharply reduced Mendelian error rates in parent-child trios.

Those variants were then genotyped in a Hong Kong Chinese cohort of 2,333 participants using high-coverage short-read data and the graph-based genotyper Paragraph. Because short-read genotyping is noisier, the team kept only variants whose genotypes showed perfect concordance between long-read and short-read calls in five samples sequenced both ways, leaving 17,294 variants with allele-frequency correlations of 0.93 to 0.94 between platforms. Benchmarking against PacBio HiFi sequencing of one sample confirmed the accuracy, with F1 scores of 0.938 for insertions and 0.934 for deletions. Functional annotation showed the variants were depleted in exons and other critical regulatory regions, consistent with purifying selection, and that the most deleterious ones—those scoring in the top 2.5 percent by the CADD-SV pathogenicity metric—were disproportionately found in known disease loci, with 61.3 percent of the very worst sitting in regions no previous association study had flagged.

The first headline result came from polygenic risk score analysis. Models built from common structural variants predicted Alzheimer’s status better than models built from common SNPs alone (pseudo r-squared of 0.056 versus 0.046), and the two score types were barely correlated with each other, indicating they capture largely independent genetic risk. Combining both variant classes explained nearly twice the disease variance of SNPs alone (r-squared of 0.098), and the improvement held up in conditional analyses designed to rule out confounding by linkage disequilibrium. In validation testing, individuals with mild cognitive impairment carried intermediate scores between controls and Alzheimer’s patients, and the structural-variant models distinguished mild cognitive impairment from normal cognition more accurately than the SNP-only model. Roughly 18 percent of common structural variants, particularly those inside tandem repeats, were not tagged by any nearby SNP—precisely the kind of signal conventional studies would miss.

Among the deleterious variants tested individually, one stood out: an insertion in the promoter and 5-prime untranslated region of FAM193B, a broadly expressed protein-coding gene on chromosome 5. PacBio HiFi sequencing of a carrier revealed the insertion’s true nature—an expanded CCG repeat, highly polymorphic in the population, with alleles ranging from four to more than 30 repeats. In the Hong Kong cohort, longer repeats were associated with increased Alzheimer’s risk in both additive and dominant models, and the association replicated in an independent mainland Chinese cohort. A fixed-effect meta-analysis of the two Chinese cohorts, spanning 1,516 Alzheimer’s patients and 1,674 controls, reached genome-wide significance, with an odds ratio of 2.20 per 10-repeat increase in the dominant model (P = 5.5 × 10⁻¹³). Alleles carrying at least 18 CCG repeats showed the strongest effect (odds ratio 2.11), and nearly a third of Chinese Alzheimer’s patients carried such an allele, compared with about 20 percent of controls.

Crucially, the signal extended beyond East Asia. In the UK Biobank’s European-ancestry participants, the repeat showed a nominally significant association with Alzheimer’s risk, with alleles of 19 or more repeats most strongly implicated (odds ratio 1.21). Conditional analyses showed the repeat’s effect was independent of rs335480, the genome-wide significant SNP previously reported in the FAM193B locus by European studies—and linkage patterns suggested that SNP was merely tagging the repeat itself. The effect size was smaller in Europeans, and a local ancestry analysis pointed to an interaction between the repeat and East Asian ancestry in an upstream regulatory region, which may explain the stronger penetrance observed in Chinese populations.

The team then traced the biological mechanism. Expression quantitative trait locus analysis showed the repeat is one of the strongest regulators of FAM193B in blood, and single-nucleus RNA sequencing of postmortem human brains from the UK Brain Bank Network revealed that longer repeats drive higher FAM193B expression across essentially every brain cell type, most significantly in neurons, astrocytes and oligodendrocytes. FAM193B expression itself was elevated in the brains of Alzheimer’s patients. Summary-data-based Mendelian randomization, including a multivariant test, indicated that the repeat is the shared causal variant linking neuronal FAM193B expression to disease risk. Mechanistically, the repeat region carries a promoter-like epigenetic signature, and the transcription factor MAZ binds longer repeats more avidly than shorter ones—an electrophoretic mobility shift assay showed stronger MAZ binding to a 25-repeat probe than an 11-repeat probe, and promoter constructs with 25 repeats drove significantly higher FAM193B expression in HEK293T cells than those with 11.

What does too much FAM193B do to neurons? The single-nucleus data showed that the repeat’s transcriptional footprint in neurons closely mirrors the transcriptional signature of Alzheimer’s itself (correlation r = 0.615), with affected genes enriched for neuron death, apoptosis and macroautophagy pathways. Longer repeats were also associated with reduced gray matter volume on brain MRI, a marker of neurodegeneration. In cultured IMR-32 neuroblastoma cells, overexpressing FAM193B increased γH2AX foci, a canonical marker of DNA damage, and RNA sequencing revealed upregulation of G1/S DNA damage checkpoints and apoptosis regulation alongside disruption of stress-response and mitochondrial pathways. Intriguingly, FAM193B protein is known to interact with GSK3α and GSK3β, the glycogen synthase kinases whose dysregulation promotes both amyloid-beta production and tau hyperphosphorylation—offering a plausible bridge to established disease mechanisms.

The implications reach well beyond a single gene. The study demonstrates that common structural variants, and common tandem repeats in particular, contribute measurable Alzheimer’s risk that conventional SNP-based analyses cannot capture, and that some of these variants carry effect sizes approaching those of APOE ε4—5.9 percent of Chinese individuals harbor at least 21 FAM193B-CCG repeats, a burden associated with an odds ratio of 2.89. Incorporating structural variants into polygenic risk models could sharpen lifetime risk prediction, especially in populations where long repeat alleles are frequent. And as long-read sequencing becomes cheaper, genome-wide surveys of common repeats may well uncover further regulatory variants of this kind, filling in the missing heritability of Alzheimer’s disease and pointing toward new therapeutic targets in the process.

Subject of Research: Genome-wide analysis of common structural variants and a risk-associated CCG repeat in FAM193B in Alzheimer's disease

Article Title: Genome-wide analysis of common structural variants in Alzheimer’s disease identifies a risk-associated regulatory CCG repeat in FAM193B

Article References: Genome-wide analysis of common structural variants in Alzheimer’s disease identifies a risk-associated regulatory CCG repeat in FAM193B. (n.d.). https://doi.org/10.1038/s41593-026-02481-z

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02481-z

Keywords: Alzheimer's disease, structural variants, FAM193B, CCG repeat, long-read sequencing, polygenic risk score, genetics, neurodegeneration, tandem repeats, gene regulation, DNA damage, MAZ