NEWS
Alzheimer’s Genetics Was Never Finished, a 2.8 Million Scan Shows
A 2.8 million person scan puts Alzheimer’s common-variant risk at 16%, finds 48 new loci, and ties risk to microglia plus vulnerable neurons.
A 2.8 million person Alzheimer’s scan published on 28 October found 127 risk loci, 48 of them new, and put the common-variant share of risk at 16%. Standard tools had put that share near 5%, which had convinced some geneticists that genome-wide studies were largely finished after APOE.
Emil Uffelmann and Douglas Wightman at Vrije Universiteit Amsterdam led the work with Ole Andreassen at the University of Oslo and Danielle Posthuma, under the Psychiatric Genomics Consortium Alzheimer’s group. Co-first author Uffelmann posted the haul the same day.
1/n
Out today in @Nature: our Alzheimer’s disease GWAS of ~2.8M people (~184k cases). We find 48 new loci, ~3x higher SNP-heritability, enrichment in genes up-regulated in microglia and down-regulated in neurons, and we release public summary statistics. https://t.co/XKNxI8f0XF pic.twitter.com/sHIEO9liqU
— Emil Uffelmann (@euffelmann) October 28, 2026
A 2.8 Million Person Scan and 48 New Loci
The meta-analysis pooled 30 cohorts: 109,479 diagnosed cases, 74,141 proxy cases whose mother or father had Alzheimer’s or dementia, 2,131,799 controls, and 499,708 proxy controls. The effective sample size was 416,493. After quality control the team tested 8,152,095 common variants.
Wightman called it the largest genetic association study on Alzheimer’s to date. A multi-ancestry analysis produced 118 genome-wide significant loci, 41 of them new by a cutoff of October 2024. Nine further loci appeared only in the European-only scan, bringing the total to 127, of which 48 were new. No risk locus was found on the X chromosome.
Five of those newly nominated genes, EIF4G3, DOCK4, VMAC, LRRC25, and CEP89, had already been listed in a European-ancestry meta-analysis led by Céline Bellenguez at Institut Pasteur de Lille and published on 3 June 2026, which reported 91 loci in 128,681 cases or proxies and 849,833 controls. Uffelmann’s team did not fully reconcile that June list, so the unique new-gene haul is smaller than the raw count of 48 new Alzheimer’s risk loci first suggests.
THE SAMPLE AND THE LOCUS COUNT
| Ancestry or analysis | People | Significant loci |
|---|---|---|
| European | 2,569,498 | 114 |
| African | 132,125 | 3 |
| East Asian | 68,399 | 2 |
| Admixed American | 36,563 | 0 |
| South Asian | 8,542 | 0 |
| Multi-ancestry meta-analysis | 2,815,127 | 118 |
| All analyses combined | 2,815,127 | 127 |
African samples produced only TREM2, SORL1, and APOE. East Asian samples produced only SORL1 and APOE. The effective sample was 91% European, 5% African, and 3% East Asian, with admixed American and South Asian shares each under 1%.
HOW THE LOCUS LIST GREW
- September 2021: Wightman and colleagues report 38 loci in 1,126,563 people.
- April 2022: Bellenguez and colleagues report 75 loci in 111,326 cases or proxies and 677,663 controls.
- 3 June 2026: Bellenguez’s European consensus meta-analysis reports 91 loci, 16 of them new.
- 28 October 2026: Uffelmann, Wightman, Posthuma, and Andreassen report 127 loci, 48 of them new, in 2,815,127 people.
Proxy cases were 40% of all cases and were what let biobanks, whose volunteers are often younger than typical onset, add power. The case-control-only meta-analysis still found 95 loci. Genetic correlation between the case-control and proxy scans was 0.99.
Why LDSC Cut Alzheimer’s Heritability to 5%
Twin studies led by Margaret Gatz have long put Alzheimer’s heritability at 60 to 80%. Genome-wide scans kept returning a single-nucleotide share near 5%. That mismatch fed the claim, associated with work by Qiongshi Zhang in 2020, that little common-variant risk remained beyond APOE. Other models, including work by Dominic Holland, had expected on the order of 1,200 causal loci.
Linkage disequilibrium score regression, the workhorse estimator, assumes many small effects spread across the genome. APOE-e4 violates that assumption. In this dataset LDSC still returned 6% (standard error 1%) in European ancestry, in line with earlier papers. SBayesRC, which allows a mix of large and small effects and treats the APOE-e4 tag SNP rs429358 as a fixed effect, returned 16% (standard error 0.2%).
APOE-e4 alone accounted for 6% of liability-scale risk in European and East Asian ancestry, matching the entire LDSC estimate, and for 2% in African ancestry. SBayesRC point estimates were 13% in East Asian ancestry and 10% in African ancestry, with wide errors; the African figure was not clearly different from zero. If lifetime prevalence is set at 20%, as in people older than 84, the European estimate rises to about 25%.
SNP SHARE OF RISK, TWO TOOLS
| Estimator (5% prevalence) | European | East Asian | African |
|---|---|---|---|
| LDSC | 6% | not reported as a point estimate | not different from zero |
| SBayesRC | 16% | 13% | 10% |
| APOE-e4 tag SNP alone | 6% | 6% | 2% |
Brian Kunkle, a geneticist at the University of Miami who was not on the paper, said SBayesRC may be more appropriate for a trait such as Alzheimer’s, and called the revised estimate perhaps the most substantial result in the study. Rebecca Sims of Cardiff University, also independent of the team, said LDSC underestimates the SNP contribution because APOE breaks the small-effect assumption, and that a large gap still remains versus twin estimates.
This shows that large genetic studies still have more to find than some assumed. But it is a population-level figure. It does not mean that 16 percent of any one person’s Alzheimer’s is caused by their genes, and it does not mean a genetic test can determine whether someone will develop the disease.
Emil Uffelmann, co-first author, Vrije Universiteit Amsterdam
Simulations on UK Biobank data, built to mimic Alzheimer’s architecture, found LDSC biased downward and SBayesRC roughly unbiased once SNPs in linkage with APOE-e4 were held out. The common-variant map was not empty. The estimator had been.
Microglia Turn Up, and Vulnerable Neurons Do Too
Gene-set tests recovered four processes seen in earlier scans: lipid binding, immune activation, amyloid breakdown, and cell adhesion. They also flagged five processes that had not come through GWAS gene-set analysis before: EphrinB reverse signalling, glutathione metabolism, RAS signalling, sphingolipid metabolism, and synapse endocytosis. Eight of the 48 new loci could not be tied to a significant gene set.
Using expression from healthy brains, Alzheimer’s association clustered in genes that run high in microglia. Using a case-control single-cell map from Nakatsuka and colleagues, the GWAS signal also sat in genes that microglia overexpress in patients, a result that held after correction for average expression across cell types (P = 1.58 × 10^-6). Inflammatory microglial states were the strongest among 12 states tested, though none passed significance on their own.
The same analysis found GWAS signal in genes that patients underexpress in three neuronal types: Sncg-expressing neurons, somatostatin (Sst) inhibitory interneurons, and layer 6 intratelencephalic Car3 projection neurons. After a further correction for average neuronal change, those neuronal hits were only nominally significant. Sncg and Sst cells are known to be vulnerable early in the disease, which is why Uffelmann argued some risk variants may act through them rather than only through immune cells.
Among genes feeding that neuronal signal, APH1B encodes a subunit of gamma-secretase, which cuts amyloid precursor protein, and SPRED2 has been tied to neural repair in zebrafish. Colocalization and Mendelian randomization also pointed to higher astrocytic KIF21B and lower oligodendrocytic MAF as risk-raising patterns. That cell-type split sits beside other maps of prefrontal gene activity across aging, which track how cortical cells change their programs with age.
PATHWAYS THE NEW LOCI JOIN
- Amyloid and microglia: SNX2, DOK3, SYK, GGA2, and CEBPA physically or genetically interact with known Alzheimer’s genes in amyloid processing and the microglial response to amyloid-beta.
- Phagocytosis: INPP5D, SYK, BLNK, HAVCR2, PICALM, AXL, FCER1G, and PLCG2 were linked to microglial engulfment of debris.
- Shared neurodegeneration: UBE2V1-SPATA2 colocalized with amyotrophic lateral sclerosis (posterior probability 0.95); GBA1 is shared with Parkinson’s disease; TMEM163 had opposite effects in the two diseases, likely through different cell types.
- Pathology endophenotypes: HDAC9 has been tied to tangles and cerebral amyloid angiopathy; C16orf95 to phosphorylated tau in spinal fluid, though local genetic correlation with Alzheimer’s was not significant.
Posthuma said the value of the study is that new risk factors can now be tied to more specific processes, and that other brain cell types, not only immune cells, are in the picture. Uffelmann told colleagues the expression shifts are associations, but that where those genes rise or fall in these cell types, a genetic variant is a likelier driver than a late response to plaques or tangles.
QPCT, SYK, and Drugs Already on the Shelf
Seventeen genes at the new loci interact with drugs approved for other diseases. Uffelmann’s thread named six as priorities: KEAP1, SYK, AXL, RRM2B, and CACNA1S at new loci, and QPCT at a known locus, where a machine-learning model, FLAMES, preferred QPCT over two genes listed in earlier GWAS.
QPCT makes the pyroglutamate modification on position 3 of amyloid-beta, the same modified form bound by donanemab, an approved antibody. Small-molecule QPCT blockers have been in development for other neurodegenerative diseases as well. The furthest along, varoglutamstat (PQ912), did not beat placebo on cognition or function in early Alzheimer’s. The varoglutamstat Phase 2 trial was later terminated for strategic reasons, not safety.
SIX GENES THAT ALREADY HAVE DRUGS
- QPCT: Enzyme that creates the pyroglutamate amyloid species donanemab targets; the oral blocker varoglutamstat showed no Phase 2 benefit.
- SYK: Kinase downstream of TREM2 in microglia; inhibitors protected neurons in culture and showed promise in Alzheimer’s mouse models, while clinical SYK drugs have been developed for blood cancers and autoimmunity.
- AXL: TAM-family receptor in phagocytosis; an AXL blocker failed to improve symptoms in 5xFAD mice.
- CACNA1S: Calcium-channel gene; a related calcium-channel drug did not improve symptoms in people with mild to moderate Alzheimer’s.
- KEAP1: Oxidative-stress regulator; disrupting KEAP1-Nrf2 helped in animal models.
- RRM2B: Ribonucleotide reductase subunit; blocking the enzyme improved symptoms in a worm model of the disease.
Those six are starting points, not prescriptions. Posthuma’s group warned that a genetic association is not proof that a drug against that gene will work, a caution the QPCT result already illustrates. Global Burden of Disease forecasts still put the human load at 152.8 million dementia cases by 2050, up from 57.4 million in 2019, which is why a failed pill does not close the gene as a research target so much as show that pathway and product are not the same thing.
Top 10% Carries Seven Times the Case Rate
A multi-ancestry polygenic score, built with SBayesRC and tested in hold-out samples that had been left out of the discovery scan, reached an average incremental liability-scale R² of 12.8% in European cohorts, with a maximum of 16.9%. Two clinical collections (DemGene in Norway and the Gothenburg H70 and clinical Alzheimer’s samples in Sweden) reached 14.5% and 16.9%. Two biobanks, TwinGene and UK Biobank, reached 9.8% and 10.1%. A score built only from European proxy cases averaged 9.6%.
The SBayesRC score beat combined APOE-e2 and APOE-e4 allele counts in every European test cohort. In unrelated UK Biobank participants of European ancestry aged 65 and older, using a score that had not been trained on UK Biobank, case prevalence in the top 10% was sevenfold the rate in the bottom 10%, and threefold the rate in the whole sample. Discrimination was stronger at those extremes (area under the curve 0.74) than in the full sample (0.69).
WHAT THE SCORE DID IN UK BIOBANK
- Extremes: Sevenfold case prevalence, top 10% versus bottom 10%, in European-ancestry people aged 65 and older.
- Versus the sample: Threefold case prevalence, top 10% versus everyone in that age band.
- Discrimination: Area under the curve 0.74 at the extremes, 0.69 in the full sample.
- Ceiling in Europeans: Average incremental R² 12.8%, peak 16.9% in a clinical collection.
Uffelmann said the score is a research tool for now, and that whether it can be used in clinic is a separate question. Age, environment, and other biology still sit outside the 16% SNP figure.
African Samples Still Lack Power for New Loci
APOE-e4 odds ratios ran from 1.7 in African ancestry to 3.4 in East Asian ancestry. Female-only scans found 12 loci, male-only scans six, on much smaller effective samples (64,771 and 49,748). The APOE-e4 tag had a 13% larger odds ratio in women than in men (2.85 versus 2.52). Genetic correlation between European women and men was estimated at 0.77 by LDSC and at a lower bound of 0.82 by SBayesRC.
The multi-ancestry score performed like the European-only score in most hold-outs, except in the African subset of UK Biobank, which had only 550 people. Because APOE-e4 is a weaker marker in African ancestry, folding in scores from ancestries with larger APOE effects, especially East Asian, pulled African prediction down. Polygenic scores trained mainly on European cohorts will still mis-rank risk in African ancestry samples for that reason, even when the discovery table is labelled multi-ancestry.
Kunkle said the low locus count outside Europe is a function of sample size and that well-phenotyped non-European cohorts remain the gap. Uffelmann said including other ancestry groups makes the team slightly more confident that the signals apply more broadly, though it is early, and that they are seeking more non-European datasets. Proxy cases were 63% of African cases, 62% of East Asian cases, 68% of admixed American cases, and 82% of South Asian cases, versus 37% in Europeans, because non-European data came mostly from biobanks with family-history questionnaires rather than memory clinics.
The group posted public Alzheimer’s summary statistics for the multi-ancestry scan and for ancestry, sex, and case-versus-proxy splits, plus X chromosome results and versions that drop UK Biobank, with 23andMe Research Institute data withheld. Forest plots for each locus sit on Zenodo. The 16% figure reopened a map that LDSC had declared nearly closed. The people still missing from that map are the next limit.
Frequently Asked Questions
What is a proxy case in an Alzheimer’s GWAS?
A proxy case is a genotyped volunteer who is not diagnosed with Alzheimer’s but who reported that a mother or father had Alzheimer’s or dementia. Effect sizes and standard errors from those scans are multiplied by 2 so they sit on the same scale as case-control results, mothers and fathers are analysed separately then combined, and people who reported both parents affected are randomly split between the two. In this study, proxies added 32 genome-wide significant loci and made up 40% of all cases.
Why is the 16% SNP figure still far below twin estimates?
Twin studies put total heritability at 60 to 80%, which still includes rare variants, structural changes, and effects that common SNP arrays do not capture. SBayesRC’s 16% is the common-SNP share at a 5% lifetime prevalence; the same model rises to about 25% if prevalence is set at 20%, as in people older than 84. Sims noted that the remaining gap is a reason to keep looking beyond common SNPs.
Did the study find Alzheimer’s risk loci on the X chromosome?
No. A subset with an effective sample size of 196,822 was analysed for X, and no locus reached genome-wide significance. The public release still includes those X chromosome summary statistics, along with sex-stratified files.
How is the QPCT gene connected to donanemab?
QPCT catalyses formation of pyroglutamate at position 3 of amyloid-beta. Donanemab is an antibody aimed at that modified species, so the gene and the drug sit on the same chemical step, even though an oral QPCT blocker, varoglutamstat, did not improve clinical scores in Phase 2.
Which earlier 2026 Alzheimer’s GWAS already listed some of these genes?
Bellenguez and colleagues, in a European-ancestry consensus meta-analysis published on 3 June 2026, reported 91 loci. Five genes from Uffelmann’s new-locus list, EIF4G3, DOCK4, VMAC, LRRC25, and CEP89, were among them, which is why a count of 48 new loci overstates how many effector genes were unseen as of mid-2026.
Disclaimer: This article is news reporting and analysis of a published genetics paper. It is for information only and is not medical, genetic-counselling, or treatment advice, and it is not a recommendation to take or stop any drug, antibody, or polygenic score test. Readers who are weighing Alzheimer’s risk, symptoms, or medicines should speak with a qualified physician or genetic counsellor before acting on any figure or gene named here. Locus counts, heritability estimates, polygenic scores, and trial statuses reflect the papers and registries cited and can change as new cohorts, methods, or drug results appear.
-
BUSINESS3 months agoConsumer Sentiment Falls to 51.7 as Future Outlook Darkens
-
GAMING3 months agoTake-Two Subpoenas Microsoft While GTA 6 Preorders Soar
-
NEWS3 months agoPluto’s Heart Glacier Still Pushes Liquid Nitrogen Upward
-
NEWS3 months agoThe Extra Heat Month Hits Children Where Care Is Thinnest
-
ENTERTAINMENT3 months agoNetflix Weighs Hosting Peacock and Fox One in Its App
-
NEWS3 months agoFDA Clears Rasonque and Widens Who Can Take It
-
NEWS3 months agoUMMC Will Rebuild 30-Year-Old Cancer Labs With $2.4 Million
-
GAMING3 months agoGTA 6’s 80-Hour Run Counts Goals That Change the Story
