What should you require from an Alzheimer’s tau biomarker testing partner?

Plasma p-tau217 now separates amyloid-positive from amyloid-negative patients about as well as a lumbar puncture does. That result has changed how sponsors think about diagnosis, and it has quietly created a procurement problem. Teams assume that because the analyte question is settled, the partner question is settled too. It is not. A trial-grade tau measurement answers a different question than a diagnostic classification, and most of the real differentiation between testing partners sits in the gap between the two.

An Alzheimer’s tau biomarker testing partner should be evaluated on four things: evidence that the assay measures tau originating from the tissue you care about, precision data at your decision threshold rather than across the full dynamic range, a documented pre-analytical protocol your sites can follow, and a validation package structured for the regulatory question your program will face.

Key takeaways

• Diagnostic accuracy statistics describe classification at a single timepoint. Enrichment, target engagement and pharmacodynamic monitoring each require different performance evidence.
• The same tau species measured on different assays does not give the same answer, so “p-tau217 testing” is not a specification.
• Analyte provenance — where in the body the measured tau came from — is the binding constraint on interpretation, not assay sensitivity.
• Pre-analytical handling determines whether data from multiple sites can be pooled at all.
• Partner selection should be driven by the decision the biomarker has to support, not by the headline analyte.

Diagnostic accuracy and trial-grade performance are different specifications

The evidence base for blood-based tau measurement has moved fast. The underlying biology of tau and p-tau biomarkers in Alzheimer’s disease has been settled for years; what changed recently is the ability to measure it reliably in plasma. In a study of 1,767 patients with cognitive symptoms across four secondary care cohorts and one primary care cohort, plasma p-tau217 detected Alzheimer’s pathology with areas under the receiver operating characteristic curve of 0.93 to 0.96, reaching 85% accuracy in the primary care setting [1]. Earlier work in primary and secondary care reported that physicians identified Alzheimer’s correctly in 58% of patients after a standard workup, against 89% for a plasma-based score [2]. The 2024 revised criteria from the Alzheimer’s Association workgroup formalized what those numbers imply: Alzheimer’s disease is now defined biologically, with fluid biomarkers accepted alongside imaging as evidence of underlying pathology [3].

None of that tells a sponsor which testing partner to use, for two reasons.

The first is that the analyte is not the assay. A head-to-head comparison of ten plasma phospho-tau assays measuring p-tau181, p-tau217 and p-tau231 in the same prodromal Alzheimer’s samples found meaningful differences in how well each detected abnormal amyloid status and predicted progression to dementia [4]. Two partners can both offer “p-tau217” and return data that behave differently at your cutoff. Any request for proposal that specifies an analyte without specifying the assay, the antibody pair and the calibration approach has not specified anything.

The second reason is that classification and change detection are separate problems. A screening or enrichment application needs a threshold that reliably sorts patients into two groups, and area under the curve is a reasonable summary of that. A pharmacodynamic application needs something else entirely: the ability to resolve a within-subject change over weeks or months, in the presence of biological and analytical variability, at a magnitude a drug might plausibly produce. That is a question about within-subject coefficient of variation, longitudinal drift, and lot-to-lot consistency — and it is almost never answered by the diagnostic accuracy figures a partner puts in a capability deck.

Ask for the second set of numbers explicitly. A partner who has run longitudinal panels for sponsors will have them. A partner who has only ever supported cross-sectional diagnostic studies will not, and the difference will surface at the interim analysis rather than during vendor selection.

Provenance is the constraint that sensitivity data cannot address

Total plasma tau is a mixed pool. Some of it reflects central nervous system pathology; some of it reflects peripheral biology, clearance and comorbidity. Improving the detection limit of an assay does not change that composition; it measures the same mixture more precisely. Plasma p-tau specificity is therefore a question about what is being measured, not about how little of it the instrument can detect.

The clearest illustration is renal function. In a retrospective analysis of two observational cohorts, plasma p-tau217 concentrations rose as estimated glomerular filtration rate fell, with the effect most pronounced in individuals who were amyloid-PET negative; the authors concluded that measurements in patients with an eGFR below 45 carry a real risk of misclassifying pathology status [5]. In an older trial population, chronic kidney disease at that stage is common. A biomarker that shifts with kidney function is a biomarker that will enrich a study arm for renal impairment as well as for Alzheimer’s pathology, and no amount of analytical sensitivity corrects for it.

This is why provenance — the tissue of origin of the analyte being measured — belongs at the top of a partner evaluation rather than in an appendix. For a diagnostic classification supported by imaging and clinical assessment, a confounded signal may be tolerable. For a target engagement or pharmacodynamic readout, where the entire inferential claim is that a change in the blood reflects a change in the brain, it is not.

Neuron-derived extracellular vesicles are one approach to that constraint. Because the vesicles are captured on the basis of neuronal surface features before their cargo is measured, the tau being quantified carries a claim about where it came from. The supporting literature is now reasonably deep. Tau and phospho-tau measured in blood neuronal-derived vesicles show concordance with the corresponding cerebrospinal fluid measures in Alzheimer’s and amnestic mild cognitive impairment cohorts [6]. Vesicle-associated synaptic proteins have distinguished individuals who later converted to Alzheimer’s dementia at the asymptomatic stage [7], and vesicle biomarkers have tracked subclinical cognitive decline in late middle age [8]. Tau released within vesicles from Alzheimer’s synapses retains seeding activity, which argues that the vesicle-associated fraction is biologically consequential rather than incidental debris [9]. In a controlled clinical comparison, vesicle-associated total tau and phospho-tau species moved differentially between treatment groups — the behavior a pharmacodynamic marker has to show to be useful [10]. A recent meta-analysis of general and central-nervous-system-derived vesicle biomarkers for Alzheimer’s and related dementias summarizes the diagnostic performance evidence across this literature [11].

ExoSORT™ was built for that specific problem: isolating the neuron-derived vesicle fraction from plasma so that downstream tau measurement carries a provenance claim rather than an assumption. Whether that matters for a given program depends on the decision the biomarker has to support. It matters a great deal for target engagement and much less for a one-time eligibility screen.

The practical version of this criterion is a question, not a preference: what evidence can you show me that the fraction you measure originates where you say it does? A capable partner will answer with characterization data. A less capable one will answer with a diagram.

Pre-analytical control decides whether multi-site data can be pooled

The most common failure in multi-site biomarker programs is not assay performance. It is sample handling.

The Standardization of Alzheimer’s Blood Biomarkers working group tested variations in routine collection and handling across a panel of Alzheimer’s-related blood biomarkers and documented measurable effects from ordinary procedural choices [12]. A subsequent consensus effort from the Global Biomarker Standardization Consortium assessed collection tube type, hemolysis, centrifugation settings, delays before centrifugation and storage, tube transfers and freeze–thaw cycles, measuring phospho-tau isoforms across four analytical platforms, and issued a standardized handling protocol on the basis of the results [13].

Those documents are the reference standard a partner should be measured against. Three questions follow:

• Is your handling protocol aligned to the current consensus guidance, and where does it deviate? Deviation is not automatically a problem. An undocumented deviation is.
• Who trains the sites, and what does that training consist of? A protocol that lives in a binder at the central laboratory does not control what happens in a phlebotomy room in month fourteen of enrolment.
• What happens to a sample that arrives outside protocol? The answer should be a documented decision rule established before the study starts, not a case-by-case judgement made once the data are in hand.

Programmes that skip this conversation tend to discover the consequences as unexplained inter-site variance, which is expensive to diagnose retrospectively and sometimes impossible to remediate.

Nine questions to put to a prospective tau biomarker testing partner

The following list is deliberately ordered. The early questions determine whether a partner is a candidate at all; the later ones separate candidates from each other.

1. What decision is this measurement supporting, and what evidence do you have that the assay supports that specific decision? Eligibility screening, stratification, target engagement, pharmacodynamic monitoring and safety monitoring are five different evidentiary requirements.
2. What is the provenance of the analyte you measure, and what characterization data supports that claim? This is the question that most cleanly separates partners.
3. What is your within-subject and inter-assay precision at our decision threshold? Not across the dynamic range — at the concentration where the study’s decisions are made.
4. What is your longitudinal performance across reagent lots and over the duration of a multi-year study? Ask for bridging data between lots and the plan for maintaining comparability as the study runs.
5. How is your pre-analytical protocol aligned to current consensus handling guidance, and how do you handle deviations? As above.
6. How much plasma does the full panel require, and what does that do to our sampling schedule? Sample volume is a protocol constraint disguised as a technical detail, particularly in trials with frequent timepoints or elderly participants.
7. Can you measure the analytes we need simultaneously from a single aliquot?Sequential single-analyte measurement multiplies volume requirements, freeze–thaw exposure and cross-run variability. A multiplex readout removes several sources of variance at once.
8. What does the validation documentation look like, and has it supported a regulatory interaction before? Ask to see the structure of an analytical validation package, not a summary of it. If your program is heading toward a companion diagnostic, ask this question in the first meeting rather than the fifth.
9. What is your reporting standard for method detail? For vesicle-based measurements, the field’s consensus reporting framework sets out what should be documented about separation and characterization [14]. A partner who reports against a published standard is a partner whose data another group can reproduce.

A partner who answers all nine without hesitation is not necessarily the right choice. But a partner who cannot answer questions two, three and five is not a candidate for a program where the biomarker carries inferential weight.