Neuronal and microglial fine-tuned 3D DNA foundation models dissect genetic variants in Alzheimer’s disease

Abstract Although genetic variants are among the strongest risk factors for Alzheimer’s disease (AD), most associated variants reside in non-coding regions with unclear mechanisms. Here, we address that gap by fine-tuning the Akita DNA foundation model on high-resolution Hi-C data generated from neuronal progenitor cells (NPCs) and microglia to produce iNeuroAkita , a cell-context-aware three-dimensional (3D) genome model. iNeuroAkita shows improved correlation with experimental Hi-C contact maps and more accurate contact predictions at key AD loci (e.g. APOE and PSEN2) , and captures CTCF-dependent chromatin loops. Using in silico saturation mutagenesis (ISM) across 11,856 AD-associated single-nucleotide variants (SNVs) and 120 structural variants (SVs), we quantified impacts on chromatin spatial organization and identified two SNVs predicted to strongly disrupt 3D chromatin architecture: rs636317 at the MS4A locus that is predicted to disrupt local chromatic interactions and microglia eQTL rs7599054 that creates gained interactions connecting its surrounding region to CCNT2 and TMEM163 , consistent with enhancer–promoter activation in epigenomic and single-cell transcriptomic data. We also identified AD-associated deletions at the KANSL1 and ARMS2 that rewire local regulatory contacts and associate with cell-type-specific expression changes, demonstrating iNeuroAkita’s utility for systematic characterization of noncoding AD risk variants underlying 3D genome regulation.