Neuroscientist proposes $20 million project to collect recordings from neurons of known types and train AI to identify cell types from spikes
Neuroscientist proposes $20 million project to collect recordings from neurons of known types and train AI to identify cell types from spikes
On November 10, Josh Siegle of the Allen Institute published a five-year plan through the Institute for Progress think tank.
In 2023, the Allen Institute and its partners compiled a catalog of mouse brain cells containing more than 5300 types. A cell's type determines whether its signal excites or inhibits neighboring neurons and how far that signal travels. The catalog was built from individual cells and brain slices, whereas electrodes in the living brain record neuronal spikes. Such recordings, Siegle writes, “show what neurons are saying, but do not reveal who is speaking.”
Currently, only a few cell types can be identified from recordings, using simple rules: a narrow spike in the cerebral cortex identifies one type of inhibitory cell. Other clues to cell type may lie in the timing of spikes and their relationships to spikes from neighboring neurons. Neural networks can detect these patterns, but they need recordings from neurons whose types are already known. Siegle proposes starting with a database of these recordings, an “ImageNet for cell types.” ImageNet, a database of labeled images, enabled neural networks to learn to distinguish thousands of categories without hand-written rules.
In mice, cells of a particular type are made to produce a light-sensitive protein, so a flash of light triggers spikes only in those cells. This labels their type in the recording. A neural network classifier learns to recognize the cell type from its activity outside the light flashes, then identifies that type in any recordings made without flashes. These experiments have become easier with the Neuropixels Opto probe, which Siegle helped develop. It combines electrodes and light sources on a single shank.
This kind of labeling is not yet possible in the living human brain, so Siegle proposes building the database primarily from mouse recordings. Cell types in mice, macaques, and humans have similar gene expression, shapes, and connections. If their functions are also conserved, their spikes may be similar, allowing a classifier trained on mouse recordings to be applied to human recordings. Macaque recordings would help establish which features of cell types are conserved in primates. Before use on patient data, the classifier would need to be tested on human brain tissue removed during surgery, with measurements of both its electrical properties and gene expression.
A classifier of this kind already exists for one brain region, the cerebellum. In a 2025 study, researchers trained it on 202 mouse recordings. The authors estimate that it distinguishes four cell types with accuracy above 95%; when its predictions are confident, they almost always agree with expert judgments on mouse and monkey recordings. Siegle considers this insufficient for a general classifier. He proposes recording tens of thousands of neurons representing dozens of cell types across brain regions and species, then making the data openly available.
Siegle estimates that the project would cost $20 million over five years. About $15 million would support a team of 13 working on mouse experiments and the classifier, with the remainder going to two macaque laboratories. Projects like this, he writes, rarely fit conventional grant programs and more often rely on consortia or private philanthropists.
Realistic brain models require both a map of connections and activity recordings from cells of known types. The MICrONS project recorded the activity of tens of thousands of neurons in a cubic millimeter of mouse visual cortex and reconstructed their connections, but mapping an entire brain this way remains too slow and expensive. Siegle argues that a classifier offers a shortcut: a neuron's type suggests which cells it is likely to connect to, potentially linking conventional recordings to existing connection maps. Simulations based on these connections could predict the effects of activating or suppressing particular cells.
- https://ifp.org/brain-cell-types/
- https://alleninstitute.org/news/scientists-unveil-first-complete-cellular-map-of-adult-mouse-brain
- https://alleninstitute.org/news/neuropixels-opto-sheds-new-light-into-deepest-regions-of-the-brain
- https://pmc.ncbi.nlm.nih.gov/articles/PMC12191827/
- https://www.nature.com/articles/s41586-025-08790-w