The drugs did not come from a hypothesis about amyloid. They came out of a sorting exercise. Researchers at UC San Francisco and the Gladstone Institutes started with gene expression signatures of Alzheimer's disease read cell type by cell type from three published single-cell studies of post-mortem human brain, then asked the Broad Institute's Connectivity Map, a library of gene expression responses to more than 1,300 compounds tested in human cell lines, a blunt question: which of these pushes the diseased pattern back the other way? Out of 1,300, 86 reversed the signature in at least one cell type, 25 in several, and only 10 of those were already approved for use in people. "Thanks to all these existing data sources, we went from 1,300 drugs, to 86, to 10, to just 5," lead author Yaqiao Li said in the institute's announcement of the work. Two of the five went into mice: letrozole, an aromatase inhibitor used in breast cancer, predicted to act on the neuronal signature, and irinotecan, a topoisomerase inhibitor used in colon and lung cancer, predicted to act on glia.
Before the animals, there was a detour through the clinic, or rather through its paperwork. The team ran the shortlist against anonymised electronic records from 1.4 million adults aged 65 and over across the University of California health systems, matching for demographics, other conditions and cancer indication, and found that people who had been exposed to letrozole or to irinotecan were diagnosed with Alzheimer's less often than their matched comparisons. Li called it "kind of like a mock clinical trial", and the phrase is worth holding onto in both directions: it is a signal that a drug is worth a look, and it is not a trial. People treated for cancer differ from people who are not in a hundred ways records cannot capture, and the authors note that cancer patients carry a lower baseline Alzheimer's risk to begin with.
The experiment that anchors the paper, published in Cell on 21 July 2025, was small and specific. Eighty double transgenic mice, crosses of the 5xFAD amyloid line and the PS19 tau line so that they carry both hallmark pathologies, were split sex-balanced into four groups of 20: vehicle, letrozole at 1 mg/kg, irinotecan at 10 mg/kg, or both, injected into the abdomen every other day for three months from the age of four to five months. Over six days of hidden platform training in the Morris water maze the groups learned alike. In the probe trials, at 24 hours and again at 72 hours, only the combination group showed a statistically significant preference for the right quadrant. In the brains, the combination group had the strongest overall rescue: less amyloid deposition and less phosphorylated tau, less microglial activation (the irinotecan effect), and better preserved neuronal layers in the hippocampal CA1 region (the letrozole effect). Single-nucleus RNA sequencing of the treated animals showed the disease-associated gene networks shifting back, cell type by cell type, which is what the screen had predicted and is the part of the result the authors treat as proof of concept for the whole approach.
The cracks are in the paper too, and they matter more than the headline. The benefit was sex-dependent: male mice improved, female mice barely did, an awkward finding for a drug that works on the estrogen pathway. Single-drug treatment helped males somewhat in the maze but not females, which sits oddly beside the human records, where letrozole alone looked protective. And these are cytotoxic and hormone-blocking cancer medicines with real toxicity in older bodies; milligram-per-kilogram doses injected into a mouse are not a prescription, not a schedule, and not a dose anyone has established for a human brain. The research team has said publicly that they want a clinical trial and are working to fund one. As of now there is no published trial of this pair in people with Alzheimer's.
What lands this in a news column rather than a journal club is the 2026 NIH Alzheimer's Disease and Related Dementias Research Progress Report, released 25 September 2026, which lists the work under expanding the therapeutic pipeline through drug repurposing: two FDA-approved cancer drugs, found computationally, improving memory and reducing pathology in mouse models. That framing is the point. The antibody era has taught families to ask about a target, usually amyloid. This route ignores the target question and asks instead which existing pill perturbs a diseased brain's gene networks in the useful direction, then looks for corroboration in millions of ordinary medical records. If the method holds, the payoff is speed: molecules whose human safety profiles already exist, and a screening loop that can be rerun as better human brain data arrives.
For a family reading this at a kitchen table, the practical content is thin and should be. Nothing here changes a treatment plan, and an oncology drug obtained for a dementia indication outside a trial carries risks that a mouse maze says nothing about. What is genuinely new is a shortlist and a method, arrived at without a single preconception about what Alzheimer's is supposed to be. Anyone curious about repurposing trials, when they open, should raise it with the clinician who knows the whole chart.