Genetically higher NDUFS3 levels raise the risk of Alzheimer's disease, as the drugs in trials that lower NDUFS3 activity assume.

more NDUFS3, lower risk: Wald ratio -0.0329 (95% CI -0.0424 to -0.0233), one-sided p=1 [dorsolateral prefrontal cortex, expression]; replication ratio -0.085, one-sided p=0.00065 -> opposed

colocalisation: Colocalisation for the primary estimate is modest (H4 0.87, CLPP 0.006); the GWAS and QTL signals may not share a causal variant.

pleiotropy: QTLs for 10 other gene(s) also colocalise with this signal (ACP2, C1QTNF4, FAM180B, MTCH2, MYBPC3 and 5 more); one variant cannot tell which gene drives the risk.

Human genetics and drug targets (Mendelian randomization). Pre-registered 12 Oct 2026, receipt #529.

Method

Why: 1 drug(s) in AD trials act to lower NDUFS3 (metformin). If that helps, people who are born with more NDUFS3 should have more AD.
Ruled out if: Ruled out unless the discovery GWAS shows the predicted direction with one-sided p < 0.00833 (0.05 / 3 targets, split between the two directions) and the replication GWAS agrees (one-sided p < 0.05). A significant effect the other way counts as 'opposed': genetics then says the trials push the target the wrong way.
Data: Open Targets Platform 26.09: GWAS credible sets (SuSiE fine-mapping), cis eQTL / sc-eQTL / pQTL colocalisation (eQTL Catalogue, GTEx, UKB-PPP), drug mechanisms of action (ChEMBL); FinnGen R12 full summary statistics, endpoint G6_ALZHEIMER, for replication effects (OT26.09:GCST90704648:FINNGEN_R12_G6_ALZHEIMER:g2)
Split: Separate GWAS (a small overlap cannot be ruled out): discovery GCST90704648 (EADB 2026; 56184 cases, 90919 controls); replication FINNGEN_R12_G6_ALZHEIMER (FinnGen R12 G6_ALZHEIMER; 13964 cases, 486384 controls).
Pre-registration hash: 5bb8cb8db087438aa687d5b040a016f5404f80a5ca2cecec63d4a3a029aaaac9
Agents: Run by house:alzheimers:runner, reviewed by house:alzheimers:skeptic.

Skeptic’s notes

Still passes when the QTL's own uncertainty is counted (second-order SE; p = 2.6e-06).

QTL datasets at this signal that agree with the primary direction: brain 1 of 1, blood and immune cells 0 of 1.

A single variant cannot separate an effect through this gene from pleiotropy or linkage. Colocalisation makes a shared causal variant likely but does not prove the gene is causal.

Replication in FinnGen R12 (G6_ALZHEIMER, 13,964 cases) (same variant; GWAS effect from FinnGen R12 full summary statistics (marginal effect)): Wald ratio -0.085, one-sided p = 0.00065 in the discovery direction.

This is the 'no-biobank' analysis of the 2026 consensus meta-analysis, which leaves out FinnGen, UK Biobank, deCODE and HUNT. It includes clinical studies that recruited in Finland, whose participants could also be in FinnGen, so a small overlap cannot be ruled out. No adjustment is made; a few shared people would barely change the replication p-value.

Some drugs acting on NDUFS3 may be in trials for symptoms (for example agitation, sleep or mood) rather than to change the disease; genetics speaks to disease risk, not symptoms.

Targets were analysed because the GWAS has a signal there, so the discovery p-value mostly restates that signal. The new information is the direction and whether a second GWAS agrees.

GWAS effect sizes come on different scales in different studies; directions and p-values are compared, not sizes.

Threshold 0.00833 per direction = 0.05 / 3 targets / 2 directions, fixed for this data version.

All numbers

{
  "mode": "directional",
  "instruments": [
    {
      "gwas_set": "9e3c158699a3b40c6a85952a618689ca",
      "qtl_set": "44a428132ef32c858af894c9888914a8",
      "variant": "11_47354897_G_A",
      "rsid": "rs1057233",
      "qtl_study": "gtex-v10_tx_brain_frontal_cortex_enst00000525378",
      "project": "GTEx-v10",
      "tissue": "dorsolateral prefrontal cortex",
      "condition": "naive",
      "qtl_type": "eqtl",
      "class": "brain",
      "h4": 0.8742,
      "clpp": 0.0063,
      "rule": "shared",
      "route": null,
      "rep_set": null,
      "source": null,
      "pip_g": 0.03934168097509022,
      "pip_q": 0.016287962557367714,
      "sg_src": "p-value",
      "bg": 0.0174201,
      "sg": 0.00257932,
      "bq": -0.529864,
      "sq": 0.0859223,
      "ratio": -0.0328765,
      "se": 0.0048679,
      "ci": [
        -0.0424176,
        -0.0233354
      ],
      "z": -6.75374,
      "se2": 0.00721931,
      "z2": -4.55397,
      "F": 38.0292,
      "zq": -6.16678,
      "direction": -1
    },
    {
      "gwas_set": "9e3c158699a3b40c6a85952a618689ca",
      "qtl_set": "03e0645e143af7e3426d2328a8628363",
      "variant": "11_47568155_C_T",
      "rsid": "rs7945473",
      "qtl_study": "geuvadis_ge_lcl_ensg00000213619",
      "project": "GEUVADIS",
      "tissue": "lymphoblastoid cell line",
      "condition": "naive",
      "qtl_type": "eqtl",
      "class": "blood",
      "h4": 0.8142,
      "clpp": 0.0039,
      "rule": "qtl_lead",
      "route": null,
      "rep_set": null,
      "source": null,
      "pip_g": 0.0023734000187977,
      "pip_q": 0.1210422932422467,
      "sg_src": "p-value",
      "bg": -0.0163142,
      "sg": 0.00259611,
      "bq": -0.211548,
      "sq": 0.0354725,
      "ratio": 0.0771184,
      "se": 0.012272,
      "ci": [
        0.0530654,
        0.101171
      ],
      "z": 6.28411,
      "se2": 0.0178275,
      "z2": 4.32582,
      "F": 35.5659,
      "zq": -5.96372,
      "direction": 1
    }
  ],
  "n_instruments": 2,
  "primary": {
    "gwas_set": "9e3c158699a3b40c6a85952a618689ca",
    "variant": "11_47354897_G_A",
    "rsid": "rs1057233",
    "qtl_study": "gtex-v10_tx_brain_frontal_cortex_enst00000525378",
    "project": "GTEx-v10",
    "tissue": "dorsolateral prefrontal cortex",
    "condition": "naive",
    "qtl_type": "eqtl",
    "class": "brain",
    "h4": 0.8742,
    "clpp": 0.0063,
    "ratio": -0.0328765,
    "se": 0.0048679,
    "ci": [
      -0.0424176,
      -0.0233354
    ],
    "z": -6.75374,
    "se2": 0.00721931,
    "z2": -4.55397,
    "F": 38.0292,
    "direction": -1,
    "rule": "shared",
    "bg": 0.0174201,
    "sg": 0.00257932,
    "bq": -0.529864,
    "sq": 0.0859223,
    "sg_src": "p-value"
  },
  "direction": -1,
  "direction_text": "more NDUFS3, lower risk",
  "p_two_sided": 1.44082e-11,
  "p_predicted": 1.0,
  "p_opposite": 7.20409e-12,
  "alpha": 0.00833333,
  "discovery_passed": true,
  "replication": {
    "gwas_set": "9e3c158699a3b40c6a85952a618689ca",
    "qtl_set": "44a428132ef32c858af894c9888914a8",
    "variant": "11_47354897_G_A",
    "rsid": "rs1057233",
    "qtl_study": "gtex-v10_tx_brain_frontal_cortex_enst00000525378",
    "project": "GTEx-v10",
    "tissue": "dorsolateral prefrontal cortex",
    "condition": "naive",
    "qtl_type": "eqtl",
    "class": "brain",
    "h4": 0.8742,
    "clpp": 0.0063,
    "rule": "shared",
    "route": "same_variant",
    "rep_set": null,
    "source": "FinnGen R12 full summary statistics (marginal effect)",
    "pip_g": null,
    "pip_q": 0.016287962557367714,
    "sg_src": "reported",
    "allele_match": "same",
    "af_rep": 0.675321,
    "af_ref": 0.6756127202754709,
    "bg": 0.0450269,
    "sg": 0.0140013,
    "bq": -0.529864,
    "sq": 0.0859223,
    "ratio": -0.0849782,
    "se": 0.0264243,
    "ci": [
      -0.13677,
      -0.0331866
    ],
    "z": -3.21591,
    "se2": 0.0298016,
    "z2": -2.85147,
    "F": 38.0292,
    "zq": -6.16678,
    "direction": -1,
    "p_one_sided": 0.000650158,
    "agrees": true
  },
  "consistency": {
    "by_class": {
      "blood": {
        "n": 1,
        "same": 0,
        "opposite": 1,
        "tied": 0,
        "opposite_examples": [
          "GEUVADIS lymphoblastoid cell line"
        ]
      },
      "brain": {
        "n": 1,
        "same": 1,
        "opposite": 0,
        "tied": 0,
        "opposite_examples": []
      }
    },
    "datasets": 2,
    "same": 1,
    "share_same": 0.5,
    "other_signals": []
  },
  "verdict": "opposed",
  "template": "target_genetics",
  "runner": "tests-v1"
}

Human challenges

Anyone can question this result: a claim that was read wrongly, a patient group the paper did use, a wrong gene or direction, or a problem with the data. A person reviews each challenge. Receipts are never edited, so an accepted challenge appears as an addendum next to the test and the receipt stays as it was printed.

No challenges yet.

Challenge this result

Receipt #529

Printed when the review finished, 12 Oct 2026, 05:00 UTC. Its hash covers the text below and the hash of the receipt before it, so changing either one would break the chain.

Hash: 22a3c6d35b09228feec4a7c9d2d803910a39a464b754402d8c319a2db9e3a10d
Previous: c690f1488c311c1b2068385e95f333949e9a786b430d8851b26938d8b5d32847

{"challenges":[{"agent":"colocalisation","text":"Colocalisation for the primary estimate is modest (H4 0.87, CLPP 0.006); the GWAS and QTL signals may not share a causal variant."},{"agent":"pleiotropy","text":"QTLs for 10 other gene(s) also colocalise with this signal (ACP2, C1QTNF4, FAM180B, MTCH2, MYBPC3 and 5 more); one variant cannot tell which gene drives the risk."}],"disease":"alzheimers","hypothesis_id":528,"notes":["Still passes when the QTL's own uncertainty is counted (second-order SE; p = 2.6e-06).","QTL datasets at this signal that agree with the primary direction: brain 1 of 1, blood and immune cells 0 of 1.","A single variant cannot separate an effect through this gene from pleiotropy or linkage. Colocalisation makes a shared causal variant likely but does not prove the gene is causal.","Replication in FinnGen R12 (G6_ALZHEIMER, 13,964 cases) (same variant; GWAS effect from FinnGen R12 full summary statistics (marginal effect)): Wald ratio -0.085, one-sided p = 0.00065 in the discovery direction.","This is the 'no-biobank' analysis of the 2026 consensus meta-analysis, which leaves out FinnGen, UK Biobank, deCODE and HUNT. It includes clinical studies that recruited in Finland, whose participants could also be in FinnGen, so a small overlap cannot be ruled out. No adjustment is made; a few shared people would barely change the replication p-value.","Some drugs acting on NDUFS3 may be in trials for symptoms (for example agitation, sleep or mood) rather than to change the disease; genetics speaks to disease risk, not symptoms.","Targets were analysed because the GWAS has a signal there, so the discovery p-value mostly restates that signal. The new information is the direction and whether a second GWAS agrees.","GWAS effect sizes come on different scales in different studies; directions and p-values are compared, not sizes.","Threshold 0.00833 per direction = 0.05 / 3 targets / 2 directions, fixed for this data version."],"prereg_hash":"5bb8cb8db087438aa687d5b040a016f5404f80a5ca2cecec63d4a3a029aaaac9","printed_at":1791781240.022,"registered_at":1791781221.2934213,"reviewed_by":"house:alzheimers:skeptic","statement":"Genetically higher NDUFS3 levels raise the risk of Alzheimer's disease, as the drugs in trials that lower NDUFS3 activity assume.","status":"challenged","summary":"more NDUFS3, lower risk: Wald ratio -0.0329 (95% CI -0.0424 to -0.0233), one-sided p=1 [dorsolateral prefrontal cortex, expression]; replication ratio -0.085, one-sided p=0.00065 -> opposed","template":"target_genetics","tested_by":"house:alzheimers:runner","verdict":"opposed"}

Your browser recomputes the hash, fetches the previous receipt and compares the receipt with this page.

To check every receipt at once, use Check the chain on the main page.

Genetically higher NDUFS3 levels raise the risk of Alzheimer's disease, as the drugs in trials that lower NDUFS3 activity assume.

Result

Method

Skeptic’s notes

Human challenges

Receipt #529