With independent repeat errors, starburst contact features captured 1.25% of residual motion-vector energy in 302 cells from one mouse retina

We tested whether starburst contact features carry stable motion response information beyond cell type and imaging recording. The data came from one mouse retina. We used 302 matched ganglion cells in 47 types and four repeats. With independent errors between repeats, cross-products estimate stable response energy without within-repeat noise bias. Four contact features captured 1.25 percent of residual stable energy. The mean after within-type cell-label shuffles was 2.09 percent. The one-sided shuffle tail was 0.6596. A descriptive recording-weight range was -0.74 to 15.95 percent. This range does not give an animal-population confidence interval. A planned check of 42 known direction-selective cells gave 43.73 percent alignment and an unadjusted tail of 0.0132. No planned sensitivity passed a 0.05 threshold after an eight-test correction. Added spatial controls reduced the subgroup estimate to 26.19 percent, with tail 0.3390. Shuffles within both type and recording gave tail 0.2462, with only 14 movable subgroup cells. The main test gives no clear evidence of added stable information in this global contact feature space. It does not test new-cell prediction or prove no contact effect. Fixed direction order, possible errors shared between repeats, one retina, and a bounded calcium statistic limit the result. We provide the trial-level feature data, a cross-repeat projection tool, checks of its assumptions, all planned and exploratory results, source provenance, code, and figures.

retina · structure and function · connectomics · starburst amacrine cells · motion coding · repeat noise · cross-repeat energy · reproducible dataset