Neuroscience has spent two decades mapping how the brain is wired. Diffusion MRI traces the large fibre bundles between regions in living people.¹ Electron microscopy has mapped about 150 million synapses in a single cubic millimetre of human cortex.² Cell atlases list the types of neurons in each region.³ These maps are extraordinary, and on their own they still do not tell you what a circuit will do.

A wiring diagram is static. The same connections can produce very different activity depending on neuromodulators, cell state, medication and disease.⁴ Human connectomes from imaging are coarse and are usually averaged across many people,⁵ and synapse-level maps come from small pieces of tissue that will never be stimulated.² When a stimulating lead sits in a patient's subthalamic nucleus, the map can say which pathways run nearby. It cannot say how this patient's circuit will respond to 2 milliamps at 130 hertz.

That is the problem we work on. We use the connectome as a scaffold and build, patient by patient, a circuit model of how activity moves through it when it is perturbed. A circuit model is only useful if it predicts what happens next, so prediction is how we build it and how we judge it.

The data to do this has only recently started to exist. Sensing-enabled stimulators now record local field potentials from the same leads they stimulate through.⁶ Every change of setting is a perturbation with a recorded response, in a living human brain, repeated in every implanted patient. Very little of it is used today.

Each prediction can be checked. The model proposes a setting, the device records how the circuit responds, and the next prediction starts from a better model of that patient.

This is a long project with many unknowns. If you work on machine learning for neural data, run a DBS program, build stimulation hardware, or want to spend years on this problem, get in touch.