A Brain Biopsy From a Blood Draw?

 Here is a thought experiment that, until recently, belonged firmly in science fiction. You go to the doctor, give a small vial of blood, and a computer hands back a detailed molecular portrait of what is happening inside your brain. Not a scan, not a guess from symptoms, but a simulated readout of which genes are active in which cells, deep in tissue no needle will ever safely reach.

This week a San Francisco company said it is building exactly that. Verge Labs, which until now went by Verge Genomics, relaunched as what it calls a frontier AI lab for human disease biology. The centerpiece of the pitch is a phrase worth sitting with: the virtual biopsy.

The problem they are aiming at

Start with a sobering statistic that drives the whole effort. Roughly nine out of ten drug trials fail. The most common reason is not that the drug is useless in some absolute sense, but that the wrong target was tested in the wrong patients. We intervene on a mechanism that turns out not to matter for the disease, or we give a promising compound to a group of patients too biologically varied for any signal to show through the noise.

Nowhere is this harder than in the brain. Neurological diseases are notoriously difficult to study because you cannot simply biopsy a living person's brain to see what is going wrong at the molecular level. So the field has leaned on proxies: animal models, cells in a dish, indirect markers. These proxies are useful, but they are also where a great deal of neuroscience drug development quietly goes to die.

What Verge is actually proposing

Verge's bet is that a decade spent gathering the right data now pays off. The company says its proprietary dataset includes more than twelve thousand brain transcriptomes across six thousand patients, fifteen million single-cell profiles, and matched genomic, proteomic, and clinical data. That is brain tissue, the thing the field has been missing, at meaningful scale.

The idea is to train a generative AI model on that tissue, then connect it to the kinds of measurements you can actually take from a living person: blood, cerebrospinal fluid, imaging, clinical records. If the model has learned the relationship between what shows up in blood and what is happening in brain tissue, then in principle you can feed it a blood draw and have it generate a plausible molecular picture of the brain it could never directly see. From there the stated goals are the ones that would matter most: match a patient to the therapy most likely to work for them, predict how their disease will progress, and forecast how their biomarkers will move before a trial is run rather than after.

The framing the company uses is that neuroscience drug development should start to look like modern precision oncology, where treatment is increasingly matched to the specific molecular profile of a tumor rather than to a one-size-fits-all diagnosis. That comparison is ambitious, and also a useful way to understand the goal.

Where I want to be careful

This is the part where I have to separate what is genuinely interesting from what is, for now, a claim.

A company launch is not a peer-reviewed result. The striking numbers in the announcement, including a reported eighty-three percent validation rate for targets surfaced by the platform across a decade of programs, are the company's own figures, framed the way companies frame things at a launch. That does not make them wrong. It does mean they have not yet been independently kicked around by the rest of the field, and that is the process that turns a claim into a finding. The company also notes that its first AI-discovered drug completed an early-stage trial in late 2025, and that the trial's data is now feeding back into the model. A single early trial is encouraging and also a very long way from proof.

There is also a deeper scientific tension worth naming. The whole approach depends on the assumption that a blood draw carries enough information to reconstruct what is happening in the brain. Sometimes it does. Sometimes the most important changes are local, confined to a region or a cell type, and leave little trace in the blood. Knowing when the virtual biopsy is trustworthy and when it is confidently inventing detail is exactly the hard part, and it is the kind of thing that only careful, published validation can settle.

Why it still caught my attention

Even with all those caveats, the shape of this is telling, and it connects to a pattern I keep noticing. We are watching the same architectural idea, the foundation model trained to capture the deep regularities of biological data, get pointed at one domain after another. Proteins one week, cells the next, and now the molecular state of a whole organ inferred from the periphery. The bet underneath all of them is the same: that biology has learnable structure dense enough for a model to reconstruct the parts we cannot directly observe.

What makes the Verge version interesting is the insistence on human tissue as the ground truth, rather than the cheaper proxies the field usually settles for. If that bet pays off, the payoff is real, because the brain is precisely where our proxies have failed us most. If it does not, it will likely be because the gap between blood and brain was wider than the data could bridge.

I do not know which way it goes. But the question it raises is the right one, and it is the question this whole era keeps circling back to: not whether we can build a model that produces a confident answer about a patient, but whether we can know when that answer is true. For something as consequential as a brain you cannot see, that distinction is everything.