New biomarkers

The central nervous system has a feature that is awkward for medicine: it is enclosed. Whenever it has been necessary to know what was happening inside, the route has been an image or an operation. The promise of biomarkers is a third way: to read, in an accessible fluid, the molecules the brain releases when something happens to it.

In recent years that promise stopped being rhetorical on at least three fronts. They are worth walking through with their figures and, above all, with their limits, because they are an exemplary case of how a real result can be told misleadingly.

Head injury: a test for saying no

Two proteins released into the bloodstream when there is brain damage — GFAP and UCH-L1 — rise in the first hours after a head injury. The study that validated them, in almost two thousand patients with mild injury, showed very high sensitivity and a negative predictive value close to the maximum: if the test is negative, the probability of a lesion visible on CT is very low¹. There is now a version that works at the bedside, on whole blood, in minutes².

Now the small print, which is where the whole lesson lies. The test's specificity is around forty per cent². That is: it serves to rule out, never to confirm. In the emergency population where the laboratory version was validated, nine out of ten positive results corresponded to patients with no lesion on CT².

And there is a general principle this example illustrates better than any other: negative predictive value is not a property of the test. The same panel performs differently depending on who it is applied to: in a population with few severe cases a negative is nearly definitive; in a population with a heavier case mix, the same negative lets considerably more through². Transferring the published figure to a service with a different population is not a nuance: it is an arithmetic error.

Alzheimer's: the field that moved most

Here the advance is genuine and large. A phosphorylated form of the tau protein, p-tau217, discriminates amyloid pathology in blood with an accuracy that in a Swedish study of more than twelve hundred patients with cognitive symptoms reached the low nineties, in both primary and specialist care³. The figure that frames the discussion is the comparison: in that same series, primary care physicians were right about six times in ten and dementia specialists a little over seven. The biomarker was right nine.

In May 2025 the US agency authorised the first blood test of this kind⁴. It is worth reading exactly what was authorised, because the distance from the headline is large. It is indicated as an aid to identifying amyloid pathology, in adults aged fifty or over, who already have symptoms, in a specialist care setting. The document itself states it is not intended for screening or as a stand-alone diagnostic test⁴.

There is also a grey zone that is almost never mentioned. In the study supporting that authorisation, roughly one patient in five received an indeterminate result, whose informational value is essentially nil⁴. An article reporting sensitivity and specificity without saying that a fifth of patients get no answer is describing the test badly.

Nor is this a standardised field. A head-to-head comparison of five different p-tau217 assays in almost a thousand people showed that all detect the pathology well but do not perform equally, and that their cut-offs are not interchangeable⁵. That the warning is not theoretical was demonstrated seven months after authorisation, when the manufacturer recalled lots of the test itself for falsely elevated results⁴.

Neuro-oncology: the right fluid is the other one

In brain tumours, liquid biopsy in blood performs poorly, and there is an anatomical reason: the blood-brain barrier retains tumour DNA. The foundational work comparing fourteen cancer types found gliomas to be the worst case of all, with detectable tumour DNA in fewer than one in ten patients⁶.

In cerebrospinal fluid the performance is different. Around half of glioma patients have detectable tumour DNA, and when they do, the genomic profile faithfully reproduces that of the tissue⁷. For a specific alteration, the gap between one fluid and the other can be enormous⁸.

Even so, not even cerebrospinal fluid is standard practice yet. The international group that systematised the topic lists four applications — diagnosing when there is no tissue, measuring residual disease, distinguishing progression from pseudoprogression, estimating prognosis — and presents them explicitly as proposals requiring standardisation and trials before entering practice⁸. In particular, distinguishing progression from pseudoprogression with liquid biopsy is today a stated goal, not an available service. And since half of patients test negative, a negative result tells you nothing.

A marker useful for what it does not do

Neurofilament light chain deserves mention. It is a reliable marker of axonal damage and is at the threshold of routine practice, above all in multiple sclerosis⁹. But it rises in multiple sclerosis, Alzheimer's disease, frontotemporal dementia, amyotrophic lateral sclerosis, stroke, trauma and Parkinson's disease alike, and it rises with age. It answers "is there neuronal damage, and how much?". It does not answer "what disease does this patient have?".

What to take away

The dominant error in coverage of this field is a single one with many faces: confusing "an aid to diagnosis in people with symptoms" with "diagnoses", and jumping from there to "detects the disease twenty years early in anyone".

The two great consensus documents published in 2024 expressly forbid that, and from opposite sides of the debate. The Alzheimer's Association group, which redefined the disease in biological terms, nonetheless recommends against diagnostic testing in people without cognitive impairment outside a research context¹⁰. The international working group goes further: it holds that most cognitively normal people with positive biomarkers should not receive the disease label but be considered at risk of developing it¹¹. The 2025 clinical practice guideline completes the picture with hard performance thresholds for replacing current tests, and confines its use to people with objective cognitive impairment, in specialist care, within a full clinical assessment¹².

Detecting a molecule is not diagnosing a disease. Between the two there is a person, a history and an examination — and that distance, for now, no test travels alone.