Three Gates Between an ALS Biomarker and a Trial Endpoint

Gate two: from fit for purpose to pharmacodynamic

This is the gate that fails quietly, and the reason is attribution.

A pharmacodynamic biomarker has to answer a narrow question: did the drug engage its target and produce the intended molecular consequence? Answering it requires that a change in the measurement be attributable to the drug’s mechanism rather than to disease progression, to a change in the patient’s physiology, or to variation in a compartment the drug never touched. That is a different property from precision, and no amount of analytical performance supplies it. Plasma NfL makes the distinction concrete. It is a well-characterized marker of axonal injury and it is measured with excellent analytical precision. It also varies with renal function in older adults, an association that persists after adjusting for age, sex, and body mass index [5], and separately with blood volume and body mass index [6]. None of that is measurement error. It is real biology entering a total-plasma measurement from outside the compartment of interest. In a mechanism-of-action readout, a change of that origin is indistinguishable from the change you are trying to detect.

Timing is the second half of the attribution problem and gets less attention than it deserves. A pharmacodynamic readout only means something if the sampling schedule is aligned to the expected time course of the mechanism. Sample too early and the molecular consequence has not propagated. Sample too late and it has been absorbed into progression, at which point the readout is measuring disease course rather than drug action. For a mechanism with an unknown time course, the honest position at phase 1 is to sample densely and accept that you are characterizing the curve rather than testing a hypothesis about it. Programs that skip this step arrive at phase 2 with a schedule chosen for site convenience and a signal they cannot interpret.

The reproducibility literature in fluid biomarkers reaches the same conclusion from a different direction, identifying cohort composition, sample handling, and analytical variation as interacting sources of non-replication that have to be designed against rather than assessed afterwards [9]. The effect-size problem compounds it: biomarker associations reported in highly cited papers are consistently larger than the same associations in subsequent meta-analyses [10], so a pharmacodynamic effect size taken from the literature is probably optimistic.

What clears this gate is knowing where the analyte came from. Measuring a neuronal analyte inside a neuron-derived vesicle population rather than in whole plasma changes the denominator of the measurement, which is what makes attribution possible. Published work has demonstrated the approach for TDP-43 specifically, detecting elevated levels in plasma neuronal-derived vesicles in an amyloid-confirmed cohort where cerebrospinal fluid concentrations are too low for practical measurement [11]. The underlying measurement problem for TDP-43 is worth understanding separately; the point here is narrower, that provenance is the specific property gate two demands.

Gate three: from pharmacodynamic to endpoint

ALS has the clearest precedent in neurology for this transition, and it is worth reading carefully rather than triumphantly.

In the phase 3 VALOR trial, tofersen was compared with placebo over 24 weeks of dosing in adults with SOD1 ALS. The primary endpoint, change in the ALS Functional Rating Scale-Revised at week 28 in participants predicted to progress faster, was not met. The trial did show reductions in cerebrospinal fluid SOD1 as an indirect measure of target engagement, and reductions in plasma NfL [1]. Accelerated approval followed in April 2023 on the basis of the plasma NfL reduction, with a confirmatory trial in presymptomatic carriers as the condition.

Two things in that sequence matter for anyone designing a program now. The clinical endpoint failed and the biomarker carried the approval, which is the outcome biomarker strategy exists to produce. And the biomarker in question was not tissue-specific. Gate three was cleared on a marker that would struggle at gate two for a mechanism-specific readout, because what a regulator needed here was a link to clinical benefit in a defined population, not molecular attribution.

It is also worth being clear about what made that argument available. SOD1 ALS is genetically defined, which removes diagnostic ambiguity from the population. Plasma NfL had years of accumulated natural history data behind it, establishing how it behaves in untreated disease. And the condition’s severity and rarity limited the feasibility of running another placebo-controlled trial, which shifted the benefit-risk calculation. Remove any one of those and the same biomarker data supports a weaker case. A program without a genetically defined population or a natural history dataset is not one step behind the precedent; it is missing a precondition of it.

That distinction is what the qualification framework formalizes. Qualification is a conclusion that within a stated context of use, a marker can be relied on to have a specific interpretation in drug development [3]. The context of use is the operative unit: population, disease stage, the decision the marker informs, and the measurement method. A marker is never qualified in the abstract. The standardized vocabulary for these distinctions, including the separation between pharmacodynamic, prognostic, and surrogate roles, is set out in the joint FDA-NIH glossary [2], and using it precisely in early interactions with a regulator is cheaper than discovering later that your program and your reviewer meant different things by the same word.

Where ALS programs actually stall

ALS makes all three gates harder than Alzheimer’s does, for structural reasons rather than scientific ones.

Trials are short and populations are small, so there is less room to accumulate evidence across phases. Progression is heterogeneous, which widens the variance a pharmacodynamic signal has to clear. And the functional rating scale that anchors clinical assessment is coarse relative to the changes a mechanism-specific marker can detect over a short window.

There is also a compounding effect specific to small indications. In a large disease area, a marker accumulates evidence across many programs and sponsors, and any single program inherits that base. In ALS the base is thinner, which means individual programs carry more of the evidentiary burden themselves and get less benefit from work done elsewhere. That argues for generating more evidence earlier than the phase would seem to require, and for treating natural history data as an asset worth building rather than a cost to minimize.

The stalls follow a recognizable pattern. A program validates a method to exploratory standard, generates an encouraging phase 1 signal, and then finds at phase 2 that the method cannot support the decision it now needs to support. Or it builds a pharmacodynamic case on a marker that moves for reasons unrelated to the mechanism, and cannot separate drug effect from progression. Or it reaches end of phase 2 with good biomarker data and no context-of-use statement, having never specified in advance what claim the data was meant to support. The broader case for biology-based rather than symptom-based assessment in ALS is well established, as is the current therapeutic landscape these programs sit inside. The gap is procedural: knowing that blood biomarkers help does not tell you what evidence to generate, in what order.