Alzheimer’s risk loci concentrate in microglial immunity and proteopathic clearance
A new preprint turns 127 Alzheimer’s disease susceptibility loci into a pathway map with immediate translational intent. Instead of treating the loci as a static catalog of statistical associations, the analysis asks where the signal concentrates, which tissues carry the strongest regulatory evidence, and which targets already have human bioactivity data strong enough to support drug repositioning or genotype-stratified trial design.
The central message is hard to miss. Most non-APOE common-variant heritability is not scattered evenly across the genome; it clusters in immune-surveillance and proteostasis programs that are biologically plausible for sporadic Alzheimer’s disease.
Using summary statistics from large public genomic, transcriptomic, and pharmacologic resources, the work partitions functional heritability across five mechanistic programs: MICROGLIAL_IMMUNE, PROTEOPATHY_AUTOPHAGY, ENDOCYTIC_VESICULAR, SYNAPTIC_CALCIUM, and MITOCHONDRIAL_OXIDATIVE. Across 8,159,496 autosomal variants, the genome-wide mean test statistic was 1.5099, reflecting the broad polygenic burden of the trait. Within that framework, 89,589 cis-regulatory SNPs across 48 non-APOE effector genes showed a 12.88-fold enrichment over the autosomal background.
Heritability is not evenly distributed across biologic programs
The heaviest share of non-APOE functional heritability fell into two pathways that fit long-standing mechanistic hypotheses about Alzheimer’s disease. MICROGLIAL_IMMUNE accounted for 39.22% of the normalized non-APOE share, corresponding to h2_SNP = 6.28% and 13.02-fold enrichment. PROTEOPATHY_AUTOPHAGY followed closely at 36.34%, with h2_SNP = 5.81% and 14.98-fold enrichment. Together, those two programs represented more than three-quarters of the non-APOE common-variant liability signal.
The remaining burden was smaller but still informative. ENDOCYTIC_VESICULAR contributed 16.71% of the non-APOE share, SYNAPTIC_CALCIUM 5.40%, and MITOCHONDRIAL_OXIDATIVE 2.33%. The gradient matters because it suggests where the genetics are most densely wired to disease biology. Microglial immune response and proteopathic clearance are not side themes here; they are the dominant architecture outside the APOE/TOMM40 super-locus.
One locus still dwarfed the rest. The APOE/TOMM40 region reached a maximum chi-square of 3,267.15 and explained 74.82% of the top-locus additive variance. Even so, the non-APOE signal remained substantial enough to motivate pathway-specific therapeutic targeting rather than a one-size-fits-all strategy.
Cortical and myeloid signals diverge at the gene level
The most clinically interesting part of the analysis comes from the tissue-level colocalization work. Some loci behaved like cortical parenchymal signals, while others appeared to live primarily in myeloid compartments. That division is not semantic; it influences whether a target is likely to act through neurons and glia in the brain or through peripheral immune biology that may only secondarily affect neurodegeneration.
emerged as a clear cortical candidate. The rs2300880 signal showed Z_SMR = -4.62 with p_SMR = 3.93e-6, and colocalization was strong in Brain_Cortex with PP.H4_cis = 0.919 and PP.H4_local = 0.994, while Brain_Frontal_Cortex_BA9 also showed PP.H4_local = 0.979. AXL also colocalized directly in brain cortex, with rs12979764 yielding p_SMR = 2.13e-4, PP.H4_cis = 0.902, and PP.H4_local = 0.924. These are the kinds of signals that can plausibly translate into cortical target engagement in future interventional studies.
The myeloid genes told a different story. BIN1, CD33, TREM2, INPP5D, and SYK showed near-zero bulk cortical eQTL colocalization, with PP.H4_cis at or below 0.025 in Brain_Cortex, despite strong SuSiE Locus-to-Gene scores ranging from 0.875 to 0.968. In primary monocytes and spleen, however, the same biology became visible again, with INPP5D PP.H4_local = 0.992 and BIN1 PP.H4_local = 0.999. The implication is straightforward: some of the most genetically credible Alzheimer’s genes may be invisible in bulk cortex because their main regulatory action sits in immune cells rather than neuronal tissue.
Druggable targets already exist across the highest-priority pathways
The pharmacogenomic scan adds the most practical layer. ChEMBL and Open Targets data identified low-nanomolar human bioactivity for clinical-stage compounds across all five mechanistic programs. isradipine reached 2.0 nM for CACNA1S, nilvadipine 19.0 nM, omaveloxolone 2.2 nM for KEAP1, fostamatinib/R406 41.0 nM for SYK, and varoglutamstat/PQ912 61.0 nM for QPCT.
This is where the paper moves from genomic mapping toward trial design. If a patient carries a larger burden in a microglial or proteopathic pathway, the right drug may not be the one with the broadest prior neurologic branding; it may be the one that binds the relevant target with adequate human potency and aligns with the patient’s pathway-specific burden score. The logic is pharmacologic as much as genetic: a drug cannot repair a biologic program it does not reach, and a target that is not active in the relevant tissue will be a weak therapeutic lever even if the association signal is statistically strong.
What this means for genotype-stratified Alzheimer’s trials
The preprint argues for a more granular therapeutic framework in sporadic Alzheimer’s disease. Under that model, a Functional Pathway Burden Score would guide which patients are matched to pathway-selective drugs, rather than enrolling broad unstratified cohorts and hoping a monotherapy lands on the correct biologic substrate by chance. The proposal is attractive because the genetic architecture already points to discrete biology: microglial surveillance, proteopathic clearance, endocytic traffic, calcium signaling, and oxidative metabolism.
The study remains a preprint, and the findings are entirely summary-statistics based. No individual-level clinical phenotype was analyzed here, and no treatment response was measured. The tissue colocalization signals also depend on the quality and scope of the underlying reference datasets, which are strong but still incomplete representations of a highly heterogeneous brain disease.
Even so, the message from the data is unusually coherent. Outside APOE, Alzheimer’s genetics looks less like diffuse noise and more like a set of recurrent cellular liabilities, with immune and proteostatic systems carrying most of the common-variant burden. The next step is not another broad association scan. It is prospective validation in genetically stratified cohorts, where pathway burden can be tested against cognition, biomarker change, and drug response in the same patients.