72. the continued value of proxy cases and improvements in snp-heritability and polygenic prediction of alzheimer's disease

Background Alzheimer's disease (AD) is the most common cause of dementia, with global case numbers projected to reach 153 million by 2050. AD is highly heritable, with twin-based heritability estimates of 60–80%, and approximately 1,200 causal loci are predicted to exist. In the latest GWAS, however, only around 80 loci have been associated with AD, suggesting that many remain undiscovered. Liability-scale SNP-heritability estimates have been low (∼5%), prompting suggestions that little additional genetic risk can be identified beyond APOE. Larger studies leveraging improved methodologies are needed to reconcile these observations. Methods We performed the largest GWAS of AD to date, meta-analyzing 109,479 cases, 74,141 proxy cases, 2,131,799 controls, and 499,708 proxy controls across 30 cohorts and five ancestries (total Neff = 416,493). Proxy cases were defined as genotyped individuals with a parent reported to have AD, with effect sizes transformed to the case-control scale. We conducted multi-ancestry, ancestry-specific, and phenotype-stratified (case-control vs. proxy) meta-analyses. SNP-heritability and polygenic prediction were estimated using SBayesRC, which models a mixture of effect-size distributions and incorporates functional annotations, and compared to LDSC. PGS performance was evaluated in five cohorts using a leave-one-cohort-out design. Results We identified 127 independent risk loci, 48 of which are new. The inclusion of proxy cases (40% of total cases) yielded 32 additional genome-wide significant loci compared to a case-control–only meta-analysis. The genetic correlation between case-control and proxy meta-analyses was 0.99 (s.e. = 0.14), and several sensitivity analyses suggest that including proxy cases did not bias our results. SBayesRC estimated liability-scale SNP-heritability at 16% in Europeans, substantially higher than the LDSC estimate of 6%. PGS achieved an average incremental liability-scale R² of 13% (maximum 17%), exceeding the LDSC heritability estimate. Discussion Our results highlight that the SNP-heritability of AD is considerably higher than previously estimated with LDSC. Methodologically, the LDSC–SBayesRC discrepancy highlights the importance of modeling large-effect variants in oligogenic disorders, and the strong agreement between proxy and case-control summary statistics supports the continued use of proxy cases to boost power in late-onset disorders. Summary statistics from all primary and stratified analyses will be made publicly available.