Every pathology lab knows the trick. Drop Thioflavin T onto a brain section, shine the right light, and the amyloid deposits answer back in green. The dye has been the field's cheap, reliable flashlight for decades: it finds plaques, it says nothing about them. A team led by Yu Liu and Lihua Zhang at the Dalian Institute of Chemical Physics, with Xin Zhang's group at Westlake University, asked a blunt question. What if the dye stopped wasting its energy on light and did chemistry instead?
That is, more or less, what they built. Reported in Nature Communications on 9 January 2026 (received April 2025, accepted December 2025), the method they call Amyloid-ID rests on a redesigned ThT scaffold. The authors walked through a series of nine probes, adding heavy atoms, shifting charge, extending conjugation, until the excited state energy that normally leaves the molecule as fluorescence was diverted instead into a triplet state that makes reactive oxygen species. The winning probe, P9, keeps ThT's habit of sticking to amyloid but now catalyses labelling of whatever protein sits within reach, through a type-I radical pathway, using a small nucleophilic tag (propargylamine) that can later be clicked to a reporter and pulled out for mass spectrometry. Two details matter for anyone judging the chemistry: P9 preferred fibrils over the loose monomer form of the same proteins, and tau fibrils imaged before and after illumination kept their fibrillar shape, so the light was not simply destroying the thing being measured. The probe was only moderately good at labelling oligomers, the smaller assemblies many researchers consider the more toxic form.
The point of all this is a problem that has dogged plaque proteomics. To find out what is inside a deposit, you usually have to destroy the tissue that holds it, grinding brain into a slurry and sorting the pieces afterwards. Amyloid-ID works on the section itself, with the architecture intact. In female mice at 12 months from three widely used models (3xTg, rTg4510 and APP/PS1), with age and sex matched wild-type C57BL/6 animals as controls and three independent biological replicates, the method profiled deposits in place. It recovered tau, the biomarker that these workflows are notorious for losing. Compared against four methods, about 80 percent of identified proteins were shared, while MAPT (tau) and Macf1 turned up only with Amyloid-ID and were then checked by immunofluorescence imaging. Of the proteins identified, 59, or 17 percent, were classified as neurodegeneration related by KEGG analysis, and 8 of the top 10 were linked to neurodegenerative disease.
Then came the part that reads like a finding rather than a demonstration. Across the three models, which differ in what they are engineered to accumulate, the team reported 410 proteins in common inside the deposits, and the cellular component analysis kept pointing at the same compartment: mitochondria, including electron transport chain components. They did not leave it on the spreadsheet. The mitochondrial protein mt-Nd2 was imaged alongside amyloid stained with Thioflavin S in all three models, and it sat where the deposits were. The authors connect this to a long-standing observation that Alzheimer's symptom severity tracks with cerebral hypometabolism, and note the mitochondrial tricarboxylic acid cycle as the brain's bioenergetic hub. Connect is the right verb. Proximity labelling reports what was near enough to be tagged; it does not say what arrived first, what dragged what in, or whether the entanglement is cause, consequence or bystander.
For a family reading this after a diagnosis, the honest translation is short: nothing here changes a clinic visit, a scan, a prescription or a prognosis. There is no drug in this paper, no test, no person. What there is instead is a better pair of tweezers, and in a field where the argument over amyloid has run for years on incomplete inventories, knowing what else is trapped in the deposit is not a small thing. The plaque has mostly been discussed as a lump of one protein. This work treats it as a crowded neighbourhood and starts reading the names on the doors.
The authors are unusually candid about the boundary. They write that all their models sit on a single genetic background, that the choice of control strain strongly shapes any proteomics result normalised against it, and that the robustness of their findings is limited by the lack of multiple backgrounds. They warn explicitly that anyone using the method for biomarker discovery should weigh the effects of the genetic engineering in the animals themselves. That is a tool paper doing the thing tool papers rarely do: telling you where it will break before someone else does.