CircuitATLAS preprint: An AI agent nominated ATP1A3, and a mouse experiment showed a weaker cortical response to a chemical challenge

CircuitATLAS preprint: An AI agent nominated ATP1A3, and a mouse experiment showed a weaker cortical response to a chemical challenge

On 7 October, the authors posted a preprint on arXiv describing CircuitATLAS, a knowledge graph that an AI agent uses to identify molecular targets for neural circuit dysfunction. The system nominated ATP1A3, and an experiment in awake mice showed how increased expression of this protein changes the cortical response to chemically induced hyperexcitability.

The same pattern of neural circuit dysfunction can have different molecular causes. CircuitATLAS searches for a target by starting with measurable circuit activity: the agent traces a path from an altered rhythm through a brain region and cell type to a protein. The graph contains 3,83 million nodes and 7,66 million links connecting diseases, their manifestations, electrical activity, brain regions, cells, and molecules. The authors removed predefined disease-to-gene and disease-to-protein links from the searchable portion of the graph so that the search would proceed through circuit function.

For a model of cortical hyperexcitability, the system selected ATP1A3, a form of the sodium-potassium pump found in neurons. During rapid firing, it helps inhibitory neurons restore their baseline electrical charge and restrain cortical activity. The authors hypothesized that increasing ATP1A3 expression in these cells would reduce the response to an acute chemical challenge.

To test this hypothesis, the researchers increased ATP1A3 expression in cortical inhibitory neurons. They then locally administered 4-aminopyridine, a substance that sharply increases excitability, to awake mice and measured the increase in the power of cortical electrical oscillations. The experiment included five control animals and six animals with increased ATP1A3 expression. The median increase was lower in the second group across all three frequency bands: 1,33 versus 11,16 in controls for beta, 1,19 versus 9,73 for low gamma, and 1,02 versus 7,81 for high gamma.

After the experiment, the authors conducted a computational search for small molecules and identified 250 compounds predicted to bind to ATP1A3. A subsequent functional test should establish which compounds enhance pump activity and which inhibit it.