With cell types excluded from training, directional contact fractions gave an uncertain 18% motion tuning error reduction in 302 ganglion cells from one mouse retina

We tested whether directional starburst amacrine cell contact improves motion tuning prediction beyond contact opportunity, depth, and position. We used EM anatomy and calcium responses from the same mouse retinal sample. Nested ridge tests excluded each of 47 anatomical cell types from training before prediction. In 302 ganglion cells, adding four contact fraction features reduced mean squared response vector error from 0.04945 to 0.04067. This is a 17.8% gain. A conditional range from resampling types was -4.3% to 34.5%. The gain remains uncertain. An exploratory error decomposition found that correct individual cell pairing supplied 24.0% of the net gain. Types 37c and 7o supplied 90.6% of the net gain. A direct trace response estimate gave a 19.5% gain, with a conditional range of 5.2% to 32.6%. These results show uneven prediction value for a known circuit rule. They do not establish transfer between animals or causal synaptic effects. The package preserves a derived benchmark, pinned provenance, source checks, nested model states, predictions, figures, and repeat instructions. Original arrays and upstream software remain external.

retina · connectome · structure and function · direction selectivity · starburst amacrine cells · contact opportunity · nested prediction · cell type transfer