--- title: "An Alzheimer's GWAS Finds 48 New Loci, but the Cell Biology Needs Careful Reading" description: "A Nature study combines multi-ancestry genetics, transcriptomics, and computational gene prioritization to map 127 Alzheimer's risk loci. Its microglial signal is stronger than its neuron-specific claims." date: 2026-10-29 topic: ai-for-science tags: ["genomics","bioinformatics","alzheimers-disease","genome-wide-association","transcriptomics","computational-biology","polygenic-risk"] production: ai-assisted source: https://notes.suffro.com/posts/alzheimers-genetics-127-loci-2026 --- # An Alzheimer's GWAS Finds 48 New Loci, but the Cell Biology Needs Careful Reading A genetic association is not a biological mechanism. Finding a region of DNA associated with Alzheimer's disease can be statistically convincing while leaving open which gene is responsible, which cell type it acts in, and whether changing that gene would help a patient. A [study published in *Nature* on October 28](https://www.nature.com/articles/s41586-026-11206-y) makes progress on several of those questions by combining a large genome-wide association study (GWAS) with computational gene prioritization and brain-cell expression data. The researchers report **127 Alzheimer's-associated loci, including 48 not previously identified**. They also find that genetic risk converges on immune activity in microglia and on patterns of altered neuronal gene expression. The strongest result is the expanded genetic map, not a new treatment or a diagnostic test. An important detail in the analysis is that the microglial signal remains significant after an additional adjustment, while the reported associations with particular neuronal subtypes become much less specific. That difference deserves as much attention as the headline locus count. ## A large study, with an important distinction between cases and proxies The researchers pooled data from 30 cohorts: 109,479 people classified as Alzheimer's cases, 74,141 proxy cases, 2,131,799 controls, and 499,708 proxy controls. A proxy case is not a clinically diagnosed patient in the same sense as a direct case. It is a genotyped participant reporting that a parent had Alzheimer's disease. Using family history increases statistical power for a late-onset disease, particularly when biobank participants have not yet reached the ages at which Alzheimer's commonly appears. It also introduces a different kind of phenotype: a person's genotype is being related to a parent's reported illness. The authors account for this distinction in their analyses and perform sensitivity checks, but proxy ascertainment is still a limitation when interpreting the combined results. The effective sample size, which reflects the statistical information available in the case-control comparisons, was 416,493. That is much smaller than the raw total of roughly 2.8 million participants. The distinction matters when comparing this work with other GWAS studies. The main analysis combined five broad ancestry groups. Despite that breadth, **91% of the effective sample was of European ancestry**, with approximately 5% African and 3% East Asian ancestry, and smaller contributions from admixed American and South Asian groups. This is a multi-ancestry study, but not one with balanced statistical power across populations. ## What the 127 loci actually mean The multi-ancestry GWAS identified 118 genome-wide significant loci, 41 of them new. Ancestry-specific analyses contributed nine additional loci, bringing the total to 127 and the novel count to 48. These are associated genomic regions, not 127 experimentally established causal genes. To move from regions to candidate genes, the authors combined several types of evidence. They used previous exome-wide association results, **FLAMES**, a machine-learning method for effector-gene prediction, and molecular quantitative trait loci that connect genetic variation with RNA expression, splicing, or protein abundance. Colocalization and Mendelian randomization helped prioritize plausible links. Existing biological literature supplied another check. This is where machine learning enters the scientific workflow. It helps rank possible effector genes among several candidates in a locus. The discovery itself is not the output of a single AI foundation model, and the statistical GWAS remains central. Nor does a model's preferred candidate establish causality: nearby genes can share association signals, regulatory effects can act at a distance, and different prioritization methods can disagree. The paper acknowledges that disagreement. For some loci, the authors retained multiple candidate genes when the evidence did not justify a unique choice. That is a scientifically useful outcome. A ranked hypothesis is better than pretending the mapping from variant to gene is settled. ## Microglia survive a stricter test than neuronal subtypes The researchers then integrated GWAS signals with cell-type expression data, including comparisons of Alzheimer's and control brains. They used MAGMA gene-property and gene-set analyses to ask whether genes showing particular expression patterns also carried stronger genetic association signals. Microglia, the brain's resident immune cells, stood out. Genes upregulated in microglia in Alzheimer's samples were enriched for genetic risk. The association remained significant after adjusting for the average differential-expression signal across cell types, with a reported P value of 1.58 × 10^-6. It was also robust to analyses excluding chromosome 19, which contains the influential APOE region, and to excluding proxy phenotypes. Several neuronal groups also showed an association: genes downregulated in certain inhibitory and excitatory neuronal subtypes were enriched for Alzheimer's risk signals. But after the authors adjusted for average differential expression across all cell types, those neuronal results were only nominally significant. In other words, the data did not strongly establish that the pattern was unique to those specific neuronal subtypes rather than part of broader neuronal downregulation. This does not make neuronal biology irrelevant. It means the evidence supports a stronger cell-type-specific claim for microglia than for the named neuronal populations. The study's own sensitivity analysis narrows the interpretation. There is another boundary. Genetic association combined with differential expression does not prove whether the expression changes initiate disease, follow disease progression, or reflect both. The authors explicitly leave the direction of the neuronal process unresolved. Longitudinal and perturbation experiments would be needed to test mechanisms more directly. ## Why the heritability estimate changed One of the paper's more consequential methodological results is an estimated **16% SNP heritability on the liability scale** for Alzheimer's disease in the European-ancestry analysis, using SBayesRC and an assumed population prevalence of 5%. That number is not the probability that a person will develop Alzheimer's. It is a population-level estimate of how much variation in an underlying disease-risk liability is captured by the common variants considered under a particular statistical model and prevalence assumption. The authors contrast SBayesRC with linkage disequilibrium score regression (LDSC), which gave an estimate near 6%. Their explanation is methodological: Alzheimer's has an unusually strong APOE association, and a method that models a mixture of genetic effect sizes can handle that architecture differently from one whose assumptions are less suitable for such a large-effect locus. The estimate also depends on prevalence assumptions. With an assumed prevalence of 20%, the reported European-ancestry liability-scale estimate rises to approximately 25%. The lesson is not that Alzheimer's suddenly became more genetic. It is that estimates of the genetic contribution depend on what variants are captured and how their effects are modeled. The researchers also evaluated polygenic scores in held-out clinical and biobank cohorts, using a leave-one-cohort-out procedure to avoid directly training the score on the cohort being tested. They report improved prediction, but the paper does not establish that these scores are ready for routine clinical screening. Performance and calibration across populations, ages, ascertainment settings, and clinical decision thresholds remain separate questions. ## Candidate drug targets are still hypotheses The analysis connected 17 genes at newly identified loci with drugs not approved for Alzheimer's disease. That is a prioritization result, not evidence that any of those drugs will treat the disease. The paper points to genes involved in immune responses, amyloid handling, and other pathways, including SYK and AXL. It also notes mixed existing evidence: some targets have encouraging animal-model findings, while other drug-target approaches have failed to improve Alzheimer's outcomes in previous human or animal studies. An association near a druggable gene is several steps away from a therapy. Researchers still need to identify the causal variant and effector gene, determine the direction of the relevant molecular effect, show that intervention changes disease-relevant biology, and establish safety and benefit in appropriate experiments and eventually clinical trials. The study is therefore best read as a map of experiments worth doing, rather than a shortlist of validated treatments. ## Reproducibility is stronger than clinical readiness The authors provide [analysis code on GitHub](https://github.com/euffelmann/paper-pgc_alzheimers), with an archived [Zenodo record](https://doi.org/10.5281/zenodo.22791372), and report the availability of [GWAS summary statistics](https://cncr.nl/research/summary_statistics/). The code repository includes analysis scripts and an R dependency lockfile. The paper states that the public summary statistics exclude 23andMe Research Institute data, while most individual-level data remain subject to cohort-specific access agreements. These are meaningful reproducibility resources, though they do not make the entire original multi-cohort analysis independently executable from public data alone. I have not rerun the authors' pipeline or reproduced its results. The study is peer-reviewed and considerably larger than many earlier genetic analyses. Its limitations are also concrete: uneven ancestry representation, reliance on proxy cases for part of the sample, heterogeneity between cohorts, computational rather than experimental assignment of many effector genes, and expression analyses that cannot by themselves establish the direction of causation. A [separate *Nature Genetics* meta-analysis published in June](https://www.nature.com/articles/s41588-026-02583-1) also expanded the Alzheimer's association map. The two studies are not interchangeable: their phenotype definitions, ancestry composition, and meta-analysis designs differ. Raw locus counts should not be treated as a head-to-head benchmark. ## The next scientific test is functional The contribution of this work is the connection between genetic association, candidate molecular mechanisms, and specific brain-cell expression patterns. It is particularly informative that a broad microglial signal withstands stronger statistical adjustment while the apparent specificity of several neuronal signals weakens. The next stage is experimental. Candidate variants and genes need perturbation in relevant human cell systems, with measurements of expression, cellular function, and disease-related phenotypes. Findings should be replicated in better-powered non-European cohorts. Polygenic prediction needs external calibration and evaluation for clearly defined clinical uses before anyone treats it as actionable. For computational biology, the study illustrates a useful division of labor: large-scale statistical genetics identifies where to look, machine learning and molecular data prioritize plausible genes, and biological experiments must determine what those genes actually do. The first two steps have advanced substantially. The third is still the hard part. ## Sources - [Uffelmann et al., "Alzheimer's disease genetics converges on microglial and neuronal dysregulation," Nature (October 28, 2026)](https://www.nature.com/articles/s41586-026-11206-y) - [Center for Neurogenomics and Cognitive Research, study announcement and research context (October 28, 2026)](https://cncr.nl/news-article/largest-genetic-study-on-alzheimers-identifies-48-new-risk-loci/) - [Author analysis code repository, euffelmann/paper-pgc_alzheimers](https://github.com/euffelmann/paper-pgc_alzheimers) - [Archived analysis code, Zenodo](https://doi.org/10.5281/zenodo.22791372) - [CNCR GWAS summary-statistics portal](https://cncr.nl/research/summary_statistics/) - [Consensus meta-analysis of genome-wide association studies for Alzheimer's disease and related dementias, Nature Genetics (June 2026)](https://www.nature.com/articles/s41588-026-02583-1)