Every few weeks, a protein with a name like a flight number arrives in the news as a possible blood test for dementia. GDF15. CDCP1. ITGB2. CXCL9. Somewhere a daughter reads the write-up on her phone in a hospital corridor and wonders whether this is the one, whether she should print it out for the memory clinic appointment in six weeks. The honest answer has always been that it is too early to know. What has been missing is a simple count of how often any of these proteins ever come back.

Anne-Li TB Lind, of Olink Proteomics in Uppsala, Sweden, did that count and published it as a conference abstract in Alzheimer's & Dementia on 8 January 2026. She reviewed 39 dementia studies published between 2021 and 2025 that used the proximity extension assay, a method that measures hundreds or thousands of proteins in one go, in cerebrospinal fluid and in blood, including work drawing on UK Biobank. Between them, those 39 studies flagged more than 180 proteins as altered in human biofluids. Twenty-six of them, including IL6, MMP10, GDF15, NEFL, CDCP1, ITGB2, CXCL9, IGFBP2 and GFAP, turned up in more than one study. One hundred and four were unique to a single study. This is a conference abstract, not a peer-reviewed full paper: it counts how often a protein appears in the literature, it does not pool effect sizes or retest anybody's blood, and its published figures account for 130 of the more than 180 proteins without explaining the remainder. The author's affiliation is with the company that makes the assay under review, which is stated openly in the abstract and worth knowing when you read it.

Being named once is not a mark against a protein. Every biomarker that ever worked started as a single result in a single cohort. GFAP, a marker of astrocyte activation, and NFL, a marker of nerve cell damage, were confirmed repeatedly across these studies, and both began exactly this way. The trouble is that a list of 180 promising proteins reads, from the outside, like 180 discoveries, when what the arithmetic shows is a field in which most findings have been seen exactly once by exactly one group. Lind's abstract also notes that 13 of the 39 publications used UK Biobank data, which means some of the repetition may come from different teams analysing the same enormous resource rather than from independent confirmation in a new population. The abstract does not break that down, so nobody should assume either way.

The flood has an obvious cause. In October 2023, the UK Biobank Pharma Proteomics Project reported plasma proteomic profiles for 54,219 participants in Nature, and made the data available far beyond the consortium that generated it. When thousands of proteins can be measured across tens of thousands of people, and the data can be queried by anyone with a hypothesis, candidate markers appear faster than anyone can check them. The Alzheimer's Association's own conference is the same story at human scale: AAIC 2025, in Toronto in July, ran 148 scientific sessions for more than 8,000 people in the room and more than 3,000 watching remotely. Volume is not the problem. Volume without replication is.

For a family, the practical meaning is about distance rather than doubt. Lind's abstract sorts these candidates into at least four regulatory categories: risk, diagnostic, prognostic and pharmacodynamic. Those are four different jobs, and they are not interchangeable. A protein that shifts the odds of dementia across a population of 50,000 people tells you very little about the person sitting in the clinic; a marker that tracks a drug's effect in a trial is not a test for whether you have anything. The abstract's own conclusion is that these candidates warrant further validation, which is the field saying, in its own careful register, not yet. Between a protein name in a results table and a blood test a clinician can act on lie the unglamorous years: independent cohorts, standardised assays, agreed thresholds, and evidence that the number changes what anyone should do next. That work is the point at which twenty-six matters more than one hundred and eighty.