With target locations given, one-synapse source means cut model feature error by 0.059% in one MICrONS mouse

We tested whether links with one recorded synapse add visual feature information beyond links with two or more synapses. We kept all 142,345 directed links among 8,219 selected MICrONS v1507 neurons. Each target had a unit vector with 512 digital-twin readout components. A fixed linear estimator used input feature means from training cells only. Controls included target soma and readout locations, cell class, area, scan, model performance, and source counts. Adding the one-synapse mean reduced test error by 0.0586%. The conditional 95% range from resampling soma cubes was 0.0461–0.0714%. The gain was 0.0523% with cube groups held out, 0.0804% among targets with both source channels, and 0.0309% beyond an added local source-group reference. An extended penalty grid left the primary result unchanged. All 303,164,967 source connection rows were scanned. Matched sources supplied only 0.5682% of recorded input counts to these targets. The result concerns a restricted feature estimator in one mouse, with target locations and source features given. It does not measure physiological strength or establish causal function. The package contains the selected graph, feature array, errors, approximate predictions, figures, code, source hashes, plan changes, and repeat instructions.

connectome · MICrONS · structure and function · visual cortex · digital twin · synapse count · feature prediction · source coverage