Initial neighbor activity gave an uncertain 0.4% gain in additive five-day spine retention models across 13 mice

We tested whether initial neighbor activity improves prediction of spine retention after structure and own activity controls. We used 1,404 five-day records from 22 neurons in 13 mice in released binocular visual cortex data. Some spines occur in both intervals. Each test excluded its mouse from model fitting and penalty choice. Additive logistic models used initial spine area, location, calcium activity, trial quality, selectivity, and soma alignment. Adding nearest-neighbor distance, response status, and correlations reduced mean test log loss from 0.63483 to 0.63232. The gain was 0.00251, or 0.40%. Its 95% conditional mouse-resample range was -0.00425 to +0.01153. Five of 13 mice improved. Adding soma alignment after own activity gave a gain of 0.00010, with range -0.00309 to +0.00302. Subsets by age interval, neighbor distance, and somatic response retained wide ranges. Initial event rate was higher in later retained spines in 11 of 13 mice. Thus, the tested neighbor terms gave little clear extra prediction after the stated controls. This result does not reject local plasticity. The ranges keep fitted predictions fixed. Small animal numbers, optical measurements, structural tracking labels, and additive models limit the result. The package has compact input data, source hashes, code, all predictions, full refit checks, figures, and repeat instructions.

dendritic spine · structural plasticity · calcium imaging · spine retention · visual cortex · neighbor activity · animal exclusion · prediction benchmark