Multi-ancestry Transcriptome-Wide Association Study Reveals Shared and Population-Specific Genetic Effects in Alzheimer’s Disease

Abstract Alzheimer’s disease (AD) risk differs across ancestral populations, yet most genetic studies have focused on non-Hispanic White (NHW) cohorts. We conducted a multi-population transcriptome-wide association study (TWAS) using whole-blood RNA-seq and genotype data from NHW (n=235), African American (AA; n=224), and Hispanic (HISP; n=292) MAGENTA participants. Using SuShiE for multi-population cis-eQTL fine-mapping, we identified credible sets for 8,748 genes, improving fine-mapping precision relative to analyses using fewer populations. cis-eQTL effects were largely shared across populations, with a subset showing population-specific regulation. We performed population-stratified TWAS of AD and inverse variance-weighted meta-analysis, followed by gene-level TWAS fine-mapping (MA-FOCUS), prioritizing nine genes (FDR0.8), including established AD loci (BIN1, PTK2B, DMPK) with broadly consistent effects across populations. At BIN1, fine-mapped cis-eQTL variants used in the TWAS prediction model highlighted rs11682128, which is only modestly correlated with the GWAS index SNP rs6733839 (r2 ≈ 0.34), demonstrating how integrating eQTL fine-mapping with TWAS can refine signals beyond sentinel GWAS variants. We also identified an association between COG4 expression and AD in NHW, implicating Golgi-related pathways. Using independent SuShiE-derived models from TOPMed MESA (PBMC), several signals replicated directionally across ancestries, with the strongest statistical support in NHW. Overall, multi-population eQTL fine-mapping improves model interpretability and helps resolve shared and population-specific regulatory mechanisms relevant to AD. Competing Interest Statement The authors have declared no competing interest. Footnotes Compared with the old manuscript, the new manuscript is a more journal focused main manuscript with a reorganized structure and a more conservative presentation of results. The supplementary material block present in the old manuscript is removed from the main document. Main text tables drop from 12 to 3, while figure legends and tables are moved into dedicated end sections, and references increase from 57 to 67. The scientific scope remains largely the same, but the methods are expanded and clarified. The introduction adds more context on existing multi ancestry TWAS methods. The methods add fuller descriptions of harmonization, SuShiE modeling, FUSION modeling, TWAS fine mapping, and a new functional annotation component using FILER and chromatin interaction data. A major analytical revision is the switch from sample size weighted meta analysis in the old manuscript to inverse variance weighted meta analysis in the new manuscript, supported by newly added TWAS effect size estimation details. The results are tightened and in several places made more cautious. The sparse model section is retitled and no longer foregrounds GYPC as a key example. The abstract and discussion soften some novelty claims, including removal of the first multi ancestry TWAS claim, and COG4 is presented more cautiously as a candidate signal with added functional context. The shared meta analysis results also become more selective, dropping from 39 significant genes in the old manuscript to 23 in the new manuscript. The discussion is substantially expanded, especially on cross ancestry effect discordance, AD adjusted sensitivity analyses, functional support for COG4, limitations, and future directions. Back matter is also updated, with Data Availability changed to Data and code availability, Acknowledgement changed to Acknowledgments, and Declaration of Interests added.