Biomarker Selection in Alzheimer’s and Prodromal AD Cohorts

Prodromal AD is a recruitment label, not a biological population

The 2024 revised criteria from the Alzheimer’s Association workgroup define the disease biologically rather than syndromically, and set out an integrated biological and clinical staging scheme that explicitly accommodates common co-pathologies, cognitive reserve and resistance [1]. That framing is a considerable advance for diagnosis. It also sharpens a problem for trial design that was previously easy to overlook.

“Prodromal AD” in a protocol usually resolves to two operational criteria: a mild cognitive syndrome, and evidence of amyloid pathology. Both are necessary. Neither is sufficient to describe a biologically coherent group. Within any cohort assembled that way sit people at different points on the tau trajectory, people whose cognitive decline is driven substantially by a second proteinopathy, people with vascular contributions, and people whose symptoms will not progress on the study timeline at all. They are pooled because the entry criteria could not distinguish them.

The workgroup’s companion analysis of how the criteria apply to drug development makes the same point from the regulatory side: staging that separates biological change from clinical change creates an obligation to specify which stage a given trial is intervening in, and to measure accordingly [12]. A cohort definition built entirely on a pathology-presence marker satisfies eligibility review and leaves the mechanism question open.

Conversion behaviour makes the heterogeneity concrete. Some participants meeting prodromal criteria progress to dementia within two years; others remain stable for a decade or revert to normal cognition. A trial powered on an assumed decline rate is therefore powered on an assumption about cohort composition, and that assumption is rarely measured. When the observed decline in the placebo arm undershoots the projection, the trial has lost its ability to detect slowing regardless of whether the compound worked. This is the recurring structural failure in Alzheimer’s blood biomarker clinical development: the selection question and the power calculation are answered by different pieces of information, and only one of them was collected.

This is where enrollment strategies quietly break. The trial has selected for a shared diagnostic label, then tested a hypothesis about a shared mechanism. Those are not the same selection.

What plasma phosphorylated tau does well, and where its job ends

Plasma p-tau217 is the strongest pathology-detection tool the field has produced for blood. In a three-cohort study, a commercially available p-tau217 immunoassay reached accuracies comparable to CSF biomarkers for identifying abnormal amyloid-beta and tau pathology, a three-range reference held across all three cohorts, and over eight years of follow-up the largest longitudinal change appeared in participants positive for both amyloid and tau [2]. For screening, enrichment and reducing the number of confirmatory scans a program has to pay for, this is the right instrument. Programs that are not using it at screening are spending money they do not need to spend.

The limits are not limits of sensitivity. They are limits of what the measurement is about.

It reports pathology state, not target engagement. A p-tau217 value tells you that amyloid and tau pathology are present and roughly how advanced. For a program whose mechanism is autophagy-lysosomal, synaptic, inflammatory or metabolic, it says nothing about whether the compound reached its target or changed the pathway it was designed to change. Using it as a proxy for engagement substitutes a staging marker for a mechanism readout.

It is shifted by the comorbidities of the population being studied. Among cognitively unimpaired participants in a population-based sample, a higher Charlson Comorbidity Index and chronic kidney disease were associated with higher levels of every plasma marker examined, and several relationships differed between participants with mild cognitive impairment and those with dementia [3]. Prodromal cohorts skew older and carry more renal impairment, more cardiovascular disease and more polypharmacy than the research cohorts in which cut-points were derived. The consequence is not random noise around a true value. It is a systematic shift in the distribution, in the direction that produces false positives at screening and spurious movement on repeat measurement.

It is compartment-agnostic. Whatever the assay detects in plasma, the measurement carries no information about which tissue released it. That is tolerable when the question is binary and the reference standard is imaging. It stops being tolerable the moment the question becomes “did this change happen in neurons.”

None of this argues against plasma p-tau. It argues against asking one marker to carry four different decisions.

Co-pathology is the stratification variable that goes unpriced

Across 766 autopsied individuals spanning the neurodegenerative disease spectrum, pure single-proteinopathy disease was the exception. Co-pathology was frequent in every group, and even the minimal-pathology aging comparison group carried a co-pathology burden of 48 percent [4]. The authors were direct about the implication for trials built around single-target monotherapy.

The specific co-pathologies matter for how a prodromal cohort behaves. Limbic-predominant age-related TDP-43 encephalopathy is common enough that in community-based autopsy cohorts roughly a quarter of brains carried sufficient LATE neuropathological change to be associated with discernible cognitive impairment, and its amnestic presentation mimics Alzheimer’s-type dementia closely enough that antemortem separation on clinical grounds is unreliable [5]. Combinations extend further: tau, amyloid-beta, alpha-synuclein and TDP-43 co-occur in the same brain often enough to constitute a described phenotype with its own clinical profile [6].

Put those figures against an enrollment list. A prodromal cohort selected on amyloid positivity contains a meaningful fraction of people whose near-term cognitive trajectory is driven substantially by a proteinopathy the study drug does not address. They are amyloid-positive, so they qualify. They are also close to unresponsive to the mechanism under test. Their inclusion does not just add variance; it moves the mean toward the null and requires the sponsor to buy back the lost power with sample size.

The arithmetic is unforgiving. If a quarter of an enrolled cohort cannot respond to the mechanism under test, the observed effect size in the full population is roughly three quarters of the true effect in the responsive subgroup, and the sample size required to detect it rises by a factor of about 1.8. That is the difference between a 300-participant study and a 540-participant study, before accounting for the additional screening burden. A baseline measurement that identified the non-responsive fraction would have been cheaper than the participants it would have excluded.

Treating co-pathology as an analysis-stage covariate rather than a selection-stage measurement is a costly ordering. The information that would have improved the cohort is collected after the cohort has already been assembled.

Match the marker to the decision it has to support

The selection problem becomes tractable once each decision gate is written down with its own measurement requirement. Three gates cover most early Alzheimer’s programs, and the discipline that distinguishes durable Alzheimer’s blood biomarker clinical development from expensive data collection is refusing to let one marker straddle two of them.

Most programs specify the first gate carefully, treat the second as an afterthought, and discover at the third that nothing collected along the way supports the claim they want to make. The third gate is a separate problem with its own literature, and we have treated it at length in our discussion of blood-based panels and CNS trial endpoints.

A selection sequence that holds up

1. Write the decisions before the panel. List every go/no-go the biomarker data will inform, with the timepoint at which each is made. Panels assembled from available assays rather than from stated decisions are the most common source of unusable datasets.
2. Use a pathology-detection marker for eligibility and stop there. Plasma phosphorylated tau earns its place at screening. Do not extend it into mechanism claims it cannot support.
3. Measure co-pathology at baseline, not at analysis. If a second proteinopathy would predict non-response to your mechanism, it belongs in the stratification scheme.
4. Specify a brain-specific pharmacodynamic readout tied to your mechanism.Autophagy-lysosomal, synaptic and inflammatory pathways all have measurable correlates; the question is whether the measurement can be attributed to neurons.
5. Bank longitudinally from the first visit. Retrospective questions are answerable only from samples that already exist, and blood permits a sampling density that serial lumbar puncture does not.

Where provenance changes the answer

The constraint in early Alzheimer’s is not assay sensitivity. Detection limits have moved steadily for a decade. The constraint is that a plasma concentration is a sum across every tissue that contributed to it, and neurons are a small contributor to most of what circulates. Brain specificity is a property of how the sample is prepared before measurement, not of which analyte is measured afterward.

Enriching extracellular vesicles of neuronal origin from plasma by immunoaffinity capture against a neuron-specific surface epitope addresses that constraint directly: the measurement is made on vesicles that came from neurons, so a change in cargo is attributable to neurons. In the methodological work establishing the approach, vesicles enriched for neuronal origin were shown to provide a more sensitive and accurate basis for biomarkers than plasma, serum, or non-enriched total plasma vesicles [10]. The strategy generalizes beyond the brain; a recent review of tissue-specific vesicle enrichment from circulation, co-authored with investigators at Johns Hopkins, Columbia and the Harvard T.H. Chan School of Public Health, sets out the approach across organ systems and the characterization standards it requires [11].

For prodromal cohorts specifically, the relevant evidence is that this compartment carries signal early. In a case-control study using blood drawn at two timepoints — once when participants were cognitively normal and again one to ten years later when they had been diagnosed with Alzheimer’s — neuron-derived vesicle levels of total tau, phosphorylated tau species and amyloid-beta 1-42 were significantly elevated at the presymptomatic draw [7]. Autophagy-lysosomal cargo behaved similarly: cathepsin D, LAMP-1 and ubiquitinylated proteins were higher and heat-shock protein 70 lower in preclinical disease, which points at a mechanism rather than a stage [8]. A protein profile from the same compartment separated participants with mild cognitive impairment who later converted to dementia from those who remained stable [9].

Those three findings map onto the three gates above. Presymptomatic signal supports enrichment. Pathway-level cargo supports pharmacodynamic readouts for mechanisms that plasma p-tau cannot address. Conversion prediction supports stratification within a cohort that looks homogeneous on entry criteria.

The practical argument for blood as the sampling matrix is unchanged and worth restating, because it drives what a pharmacodynamic design can actually do. Serial lumbar puncture realistically yields two or three timepoints per participant across a trial, which is enough to describe a before-and-after and not enough to describe a trajectory. Venipuncture places no comparable ceiling on sampling density. A pharmacodynamic question that needs six timepoints is answerable in blood and effectively unanswerable in CSF, whatever the relative analyte concentrations.

NeuroDex built the ExoSORT™ platform around this argument, and the same logic extends to the co-pathology problem: work supported by the Alzheimer’s Drug Discovery Foundation’s Diagnostics Accelerator applies the platform to measuring TDP-43 and alpha-synuclein in the plasma of patients already diagnosed with Alzheimer’s, with the explicit aim of characterizing mixed pathology in that population. Teams building comparable strategies in other indications may find the structure of our Lewy body dementia biomarker discussion a useful parallel.

The direction of travel in early Alzheimer’s is toward cohorts defined by more than one axis. The revised criteria already anticipate this by building co-pathology into the staging scheme, and regulators reviewing enrichment strategies will reasonably ask what a sponsor knew about mixed pathology at randomization. Answering that question from stored samples after a readout is difficult and sometimes impossible; answering it from a baseline stratification plan is routine.

The next several years of Alzheimer’s blood biomarker clinical development will be shaped less by improvements in detection limits than by how sponsors allocate measurements across decisions. Pathology-detection assays in blood are already good enough to take most of the screening burden off imaging. What remains underbuilt is the layer beneath: measurements that can say which tissue produced the signal, which second proteinopathy is present, and whether a compound changed the pathway it was designed to change. The programs that answer those questions well will be the ones that decided, before the first participant was screened, which measurement was responsible for which decision.