Graph neural network approach for inferring neuronal connectivity

CircuitGNN infers neural effective connectivity from single-neuron population activity with high precision using recordings as short as 4 min and scaling to ~750 neurons.

• Why it matters: Anatomical mapping is resource-intensive and does not directly capture effective connectivity, so there is a need for activity-based inference that can reflect learning-related changes.
• What they did: The authors built a graph neural network, CircuitGNN, and tested it on simulated population activity from an invertebrate central pattern-generating circuit and a cortical microcircuit, plus networks varying in size, duration, and hidden-neuron proportion.
• The result: CircuitGNN accurately inferred connectivity in small circuits, detected both strengthening and weakening during learning, outperformed other methods on the simulated datasets, and stayed robust with 20% or 40% hidden neurons.

AI summary of the abstract · may contain errors · not medical advice · read the paper.