TDP-43-High Enrichment: A Stratification Model for ALS Trials

Key Takeaways

• Plasma NfL stratifies ALS trials by general neuroaxonal injury, not by TDP-43 pathology — a mismatch for TDP-43-targeting ASOs, degraders, and small molecules.
• Adaptive enrichment designs require a biomarker with a defensible cutpoint, serial-sampling feasibility, and a demonstrated link to the drug’s mechanism, not just disease association.
• Neuron-derived extracellular vesicle (NDEV) assays can isolate TDP-43 signal that originates specifically from neurons, addressing the tissue-of-origin ambiguity that limits whole-plasma TDP-43 measurement.
• Diagnostic-accuracy data (AUC greater than 0.9) for EV-based TDP-43 detection already exists in peer-reviewed literature, offering a template for trial-ready cutoffs.

A sponsor moving a TDP-43-targeting ASO or small molecule into a Phase II ALS trial faces a specific version of an old problem: the biomarker most readily available wasn’t built for the mechanism being tested. Plasma neurofilament light chain (NfL) is the default enrichment tool in ALS trial design today, and for good reason — it is well-validated, widely available, and correlates with disease progression. But NfL measures axonal injury in general. It does not tell a sponsor whether a given patient’s disease is driven by TDP-43 pathology specifically, which matters enormously when the investigational drug’s entire mechanism of action depends on that pathology being present and active.

This is not a hypothetical concern. TDP-43 mislocalization and aggregation are present in the substantial majority of ALS cases, making it one of the field’s most consistent pathological findings — and also one of its most stubbornly hard to measure in a living patient. Every enrichment strategy built on an indirect proxy for that pathology inherits the proxy’s blind spots: patients with genuinely TDP-43-driven disease can be missed, while patients whose neurodegeneration stems from a different mechanism can be included simply because their NfL happens to be elevated. For a trial testing whether clearing or stabilizing TDP-43 changes outcomes, that mismatch between eligibility criterion and drug mechanism is not a minor statistical inconvenience. It can determine whether a genuinely effective drug looks like it failed.

Why NfL Under-Selects for TDP-43-Driven Disease

NfL’s diagnostic value in ALS is well established. In a prospective cohort of 231 patients with suspected ALS, plasma NfL distinguished ALS from ALS mimics with a sensitivity of 84.7% and specificity of 83.3%, and elevated baseline levels predicted shorter survival, with fast progressors showing significantly higher plasma and CSF NfL compared to intermediate and slow progressors [1]. That is a legitimate prognostic signal. But NfL is released whenever axons are damaged, regardless of which upstream protein is driving the degeneration. It cannot distinguish a TDP-43 proteinopathy from a SOD1-driven or FUS-driven case, because it sits downstream of the pathology rather than reflecting it directly.

This distinction became directly measurable in a 2024 study of plasma extracellular vesicles across ALS, FTD, and progressive supranuclear palsy (PSP). Researchers found that plasma EV TDP-43 distinguished ALS from healthy controls, PSP, and behavioral variant FTD with AUC values of 0.99, 0.99, and 0.91 respectively, and critically, plasma EV TDP-43-based AUC values exceeded plasma NfL-based AUC values for the same comparisons [2]. In other words, when the two markers were measured head-to-head in the same cohort, the EV-based TDP-43 signal outperformed NfL specifically on the axis that matters for enrichment: identifying who actually has TDP-43 pathology, not just who has neurodegeneration of some kind.

The same study found something a trial statistician should find useful: plasma TDP-43 levels measured directly, without EV enrichment, did not differ between the diagnostic groups [2]. The signal was only recoverable once the extracellular vesicle fraction was isolated. That single finding explains why whole-plasma TDP-43 assays have underperformed in the clinic for over a decade — the protein exists throughout the body, and peripheral background swamps the neuronal signal unless something separates the two.

What an Enrichment Design Actually Requires

Regulatory guidance on enrichment strategies is explicit that a biomarker used for patient selection needs more than an association with disease — it needs performance characteristics adequate for classification, a defined cutoff, and evidence tying it to the mechanism under study [7]. Statistical frameworks for adaptive enrichment, including two-stage designs formalized by Simon and Simon [3], allow a trial to begin with a broader population and narrow enrollment to a biomarker-defined subgroup at a planned interim analysis, provided the biomarker’s cutoff has been prespecified and its measurement is reproducible across sites.

That reproducibility requirement is where many candidate ALS biomarkers stall. A biomarker can show a statistically significant group difference in a single academic cohort and still fail as an enrichment tool if the cutoff doesn’t transfer across sites, instruments, or patient populations. The EV TDP-43 work addresses this directly: cutoff values derived by Gaussian mixture modeling in one cohort were tested in an independent validation cohort of 292 patients including 34 genetically confirmed cases, where cutoffs proved nearly interchangeable between the two cohorts [2]. For a sponsor building an enrichment design, that kind of cross-cohort cutoff stability is closer to what a statistical analysis plan actually needs than a single-cohort association.

There is also a sample-size tradeoff that enrichment designs force sponsors to confront directly. Restricting enrollment to a biomarker-positive subgroup can shrink the number of patients needed to detect a treatment effect, because the population is more homogeneous with respect to the mechanism the drug targets — but only if the biomarker’s classification is accurate enough that the “positive” group is not itself diluted with misclassified patients. A biomarker with modest specificity effectively re-introduces the heterogeneity the enrichment step was meant to remove, just at a smaller scale, and can leave a trial under-powered despite appearing more targeted on paper. This is precisely why adaptive two-stage designs are attractive for ALS: rather than committing to a single enrichment cutoff before any data exists, a sponsor can enroll a broader population initially, observe the biomarker-positive and biomarker-negative response patterns at a planned interim look, and then restrict further enrollment to whichever subgroup shows the larger effect. That structure only works, however, if the biomarker measurement itself is stable enough between the interim analysis and full enrollment that the “positive” definition does not shift under the trial’s feet — which returns to the same reproducibility requirement discussed above.

The HEALEY ALS Platform Trial illustrates why this matters operationally, not just statistically. As an adaptive multi-regimen platform, HEALEY evaluates several investigational products against a shared placebo group, with new regimens added over time — a structure that depends on consistent, transferable eligibility criteria across regimens added years apart [4]. Each new regimen added to the platform requires its own eligibility criteria, sample size, and endpoint specification, negotiated between the sponsor, the platform’s statistical team, and the shared infrastructure that already governs the Master Protocol. A regimen built around a TDP-43-targeting mechanism would need an enrichment criterion that plugs into that existing framework without requiring a bespoke validation exercise each time — which is a much easier ask if the biomarker’s cutoff has already demonstrated cross-cohort stability, as EV-based TDP-43 measurement has, than if the sponsor is proposing a criterion validated only in its own preclinical dataset. A biomarker whose cutoff drifts between cohorts is a poor fit for that kind of infrastructure; one that has already been shown to hold across independent cohorts is a better starting point for a regimen-specific enrichment criterion layered onto a shared platform.

Why Vesicle Enrichment Solves the Tissue-of-Origin Problem

TDP-43 pathology is central to ALS — cytoplasmic mislocalization and aggregation are present in the substantial majority of cases — but TDP-43 the protein is not brain-specific. It is expressed throughout the body, which is precisely why whole-plasma measurements have struggled to separate disease signal from peripheral background. The 2024 EV study addressed this by isolating L1CAM-positive vesicles from plasma before measuring TDP-43, on the reasoning that vesicles carrying this surface marker are enriched for a neuronal origin, though the neuronal specificity of L1CAM and its association with extracellular vesicles remains a point of active scientific discussion [2]. The enrichment step, not the antibody alone, is what recovered the group differences that whole-plasma TDP-43 could not detect.

This is also where NeuroDex’s own recent data adds a second, independent data point. Using ExoSORT™, an automated platform that isolates neuron-derived extracellular vesicles from blood using a scalable, high-specificity antibody workflow, NeuroDex has reported treatment-responsive changes in NDEV-associated TDP-43 in ALS patients, with the enrichment step increasing neuronal signal substantially over unenriched plasma analysis [6]. The result echoes the core finding from the independent academic literature: TDP-43 pathology becomes a usable stratification marker only after the neuronal vesicle fraction has been separated from the rest of circulating plasma. Two different enrichment approaches converging on the same conclusion is a stronger basis for trial design than either result alone.

For a sponsor designing a TDP-43-targeting Phase II program, the practical implication is straightforward: an enrichment biomarker built on total plasma TDP-43 is measuring mostly noise. One built on the neuron-derived vesicle fraction is measuring a signal that has already shown AUC values above 0.9 for distinguishing ALS from controls in peer-reviewed, cross-validated cohorts — and that can, in principle, also serve a second function post-enrollment as a pharmacodynamic readout of target engagement, since the same vesicle-based measurement responds to treatment.

Why This Extends Beyond ALS

TDP-43 pathology is not confined to ALS, and neither is the enrichment problem it creates. An estimated substantial proportion of Alzheimer’s-spectrum disease involves TDP-43 co-pathology, either as a contributing factor or as the primary driver in limbic-predominant age-related TDP-43 encephalopathy (LATE), a condition that can closely mimic typical Alzheimer’s disease clinically while following a distinct molecular course. A pilot study measuring TDP-43 in plasma neuronal-derived exosomes found levels were markedly elevated in Alzheimer’s disease patients compared with healthy controls, isolated using the same neuron-derived exosome approach applied in ALS [5]. Notably, that study also found no correlation between TDP-43 level and global cognitive scores [5], reinforcing the point that TDP-43 burden and clinical presentation do not move in lockstep — a patient can look clinically similar to their neighbor in the waiting room while having a very different underlying proteinopathy, which is exactly the scenario an enrichment biomarker is supposed to resolve.

For sponsors running combination programs or exploring TDP-43-directed mechanisms across more than one indication, this cross-disease relevance matters practically. A stratification approach validated in ALS cohorts and separately supported in Alzheimer’s-spectrum cohorts suggests the underlying biology, and the vesicle-based measurement strategy built around it, is not an ALS-specific artifact. That has implications for population screening and natural history studies as well as interventional trials: a marker capable of flagging TDP-43-high subpopulations could, in principle, support enrichment not just for a Phase II ALS trial but for AD-spectrum programs where TDP-43 co-pathology is suspected of blunting response to amyloid- or tau-targeted therapies, and where patient selection failures have already been documented for exactly this reason.

What This Means for Trial Design Going Forward

None of this displaces NfL. Disease progression and general neurodegeneration remain relevant secondary endpoints, and NfL’s prognostic value for survival and progression rate is well established. But for enrollment criteria in a trial testing a TDP-43-specific mechanism, the field is moving toward biomarkers that reflect the actual molecular target rather than its downstream consequences. As adaptive platform infrastructure like HEALEY continues to add regimens with increasingly specific mechanisms — RNA-targeted therapies, TDP-43 clearance enhancers, aggregation inhibitors — the demand for a stratification marker that maps directly onto the mechanism, rather than a proxy for injury in general, will only grow. Cross-cohort validation of EV-based TDP-43 measurement gives sponsors an evidence base to build enrichment criteria on now, rather than waiting for a single-source biomarker to accumulate that history on its own.

The practical next step for a sponsor evaluating this approach is not to replace an existing biomarker plan wholesale, but to ask a narrower question during protocol development: does the trial’s primary hypothesis depend on TDP-43 pathology being present, or on neurodegeneration in general? If the answer is the former, an enrichment criterion built on a proxy for the latter is a mismatch worth correcting before the protocol is finalized, not after an underpowered or diluted trial result raises the question retroactively. The data now exists, across independent cohorts and across at least two distinct vesicle-enrichment approaches, to build that correction into a trial design rather than discover the gap after the fact.