Mold on bread: a spot today, half a loaf tomorrow. That's roughly how amyloid-beta spreads through the brain — a sticky protein that clumps up in Alzheimer's disease and disrupts neuron function. Doctors can see it on PET scans (like an X-ray, but instead of bones, it shows clusters of molecules), but a scan is just a snapshot. What happened a year ago? What will happen in three? Previously, you could only guess.

The authors of the new study built a model that learns not from words but from thousands of real scans. They used data from 1461 people in the ADNI project and 1070 from the A4 study. The model, named LNODE, works like a disease-spread map: it sees a few points (where and how much protein on different scans) and fills in the entire trajectory — both past and future. It doesn't memorize each patient individually but looks for common patterns. From ten raindrops on a window, you can figure out where the cloud came from — same principle here.

The secret is that the model identifies hidden states — something like the brain's internal "climate," which either promotes or inhibits amyloid accumulation. These states change over time according to universal rules, but each person has a unique starting point. So LNODE doesn't just predict the average picture across the clinic; it sees that one patient follows a slow trajectory while another follows a fast one. On scans taken four years apart, the model predicts protein levels almost without error.

This feels less like a medical handbook and more like a weather forecast: you look at today's clouds and know whether to expect a storm next week. Same with amyloid: its movement follows strict, though invisible, laws. Once you capture them in an equation, you can peer into the disease's future without extra scanners.