New AI Reconstructs a Neuron's Hidden Signals From Its Branches

A neuron has two main parts: branching arms called dendrites that collect incoming signals, and a main body called the soma that decides whether the cell fires. In many lab experiments, tiny electrodes can only reach the dendrites, so the soma's electrical activity — the part that really matters — stays invisible. Researchers trained an AI that is forced to follow the known physics of neurons (a "physics-informed neural network") to work backwards and rebuild that hidden signal.

The results were striking. With no soma data at all, the AI produced a smooth, wrong curve: it missed every firing spike, and its error was roughly 10 to 14 millivolts, which is large at this scale. But feeding it just 1 percent of the soma's real measurements recovered every spike. At 5 percent, the error dropped to about 2 millivolts, and the AI could also estimate the cell's internal ion channels — the tiny gates that control firing — to within a fraction of a percent.

The catch: all of this happened on synthetic data. The team generated the recordings from a mathematical model of a neuron rather than measuring a real one. Living brain tissue is messier — electrical noise, damage from electrodes, and cells that don't match textbook equations. The method also still needs at least a small amount of actual soma data to anchor it, so it isn't a way to skip measurements entirely.

Still, the direction matters. Neuroscience is limited by what it can measure: you can watch a neuron's inputs or its output, rarely both at once. Tools that infer the missing half could make experiments cheaper and more informative, and they feed into longer-term efforts like brain-computer interfaces and better models of how neurons compute. For now, treat it as a promising lab technique, not a brain-reading device.