Personalized medicine

For a century, classifying a brain tumour meant looking at it. A pathologist examined the specimen under the microscope, described cell shape, counted mitoses, looked for necrosis and abnormal vessels, and from that description came a name and a grade. It was a reasonable method with a serious problem: two expert observers could look at the same slide and disagree¹.

In 2021 the World Health Organization published a classification that formally incorporated molecular markers into diagnosis². It did not add them as supplementary information: it put them in the name of the disease. The consequences were immediate and counterintuitive. The word glioblastoma became reserved for tumours without IDH mutation³. An astrocytoma with IDH mutation and certain genetic alterations can be classified as grade 4 even when the microscope shows none of the features that historically defined that grade⁴.

Put differently: appearance lost the final word.

The classifier that redrew the map

The tool that consolidated that shift does not analyse genes one by one but the DNA methylation pattern — a kind of chemical fingerprint each tumour type retains from the cell it came from. The foundational paper, published in Nature in 2018, showed something hard to ignore: compared with the standard method, the methylation classifier changed the diagnosis in up to one in eight prospective cases¹.

There is also an intraoperative version. Combining nanopore sequencing with a neural network trained to work with sparse data, a molecular classification can be obtained while surgery is under way⁵. Across twenty-five real operations the system returned an answer in under ninety minutes and was correct in most. The number almost never quoted is the other one: in seven of those twenty-five it did not reach its confidence threshold and chose to abstain rather than err. That is a design virtue and, at the same time, an exact description of where the technology stands.

What did change in the operating room

Here personalized medicine already alters conduct, with formal backing. The joint guideline of the European neurosurgical and neuro-oncological societies, published in 2025, grades the surgical objective according to molecular subtype⁶. In glioblastoma without IDH mutation the recommendation is the highest level: resect all contrast-enhancing tumour whenever it can be done safely, and there are data suggesting that going beyond the visible margin is associated with better survival. In IDH-mutant gliomas the recommendation drops a step, and in oligodendroglioma the results are frankly mixed. In recurrence, the evidence is weaker still.

The practical reading is uncomfortable and worth stating: the same tumour volume on the same scan justifies different degrees of surgical aggressiveness depending on the molecular profile. And — against intuition — in the subtype with the better prognosis the evidence for resecting more is weaker, not stronger.

This gives new meaning to a gesture that looked purely technical. The specimen removed is no longer only tissue to be taken out. It is the source of the information that will decide everything that follows.

What has not changed

Here it is worth being direct, because this is where enthusiasm detaches from fact.

In 2023 a phase 3 trial showed that an IDH inhibitor substantially delays progression of low-grade gliomas carrying that mutation, and defers radiotherapy and chemotherapy⁷; the US agency approved it in August 2024⁸. That is genuinely good news. But the outcome measured was progression-free survival, not overall survival — and given the trial design, with crossover to the active arm, there may never be a clean overall survival figure. "Delays progression" and "increases survival" are not the same sentence.

In glioblastoma the balance is harsher. The 2025 international consensus describes therapeutic progress as modest and places median population survival at around a year³. The example that best illustrates it is a trial that did everything right: it selected more than six hundred patients by the correct gene amplification — that is, personalized medicine executed properly — and overall survival came out identical to placebo⁹. Choosing the right patient is not enough if the drug does not reach the tumour.

Where targeted therapy does work is in niches defined by a specific alteration rather than by organ: gliomas with BRAF V600E mutation¹⁰, tumours with NTRK fusions¹¹. These are real responses, in small, single-arm series without a comparator. Results that deserve enthusiasm and do not admit the language of "new standard".

The problem least discussed

All of the above assumes that sequencing exists and is available. In much of the world, it is not.

The American Society of Clinical Oncology puts it without hedging: not all populations have benefited equally from access to these tests¹². And the most revealing figure comes from the best-equipped country in the world: in a survey of US hospitals with neurosurgical residencies, only a minority could perform every relevant molecular test, and the number of tests available correlated with the median income of the county where the hospital sat¹³.

That access inequality is already stratified by local wealth inside US academic centres says a good deal about what to expect outside that world. It is worth noting that the WHO classification itself anticipated this: it incorporated specific descriptors for cases where molecular testing is unavailable or insufficient². A report that does not reach the molecular subtype is not necessarily a careless report. It is the classification working under resource constraints — which are the conditions most of the world works in.