Count weights gave no clear movie response prediction gain in 169 selected MICrONS targets

We tested whether contact counts improve prediction of measured movie responses from fixed anatomical input sets in one MICrONS mouse. We used 169 targets with at least five measured inputs, 1036 connections and 135 input cells. In five held-out response-entry blocks, mean scaled error was 2.11541 with count weights and 2.11852 with equal weights. The improvement was 0.00312, with a nominal target bootstrap interval of -0.03309 to 0.03410. Observed count weights were near the center of 499 within-target shuffle scores, with lower-tail fraction 0.45. Total cleft volume reduced block error but gave no clear gain with interleaved folds. Count weights gave worse block error than a training-mean baseline. Sparse inputs, one selected animal, shared sources and stimuli, trial-mean data and a limited linear model restrict inference. The test does not establish causal transmission or transfer across animals. We provide original code, aggregate evidence, source hashes and a retrieval recipe. Source and derived record-level data remain outside the signed package.

MICrONS · structure and function · synapse count · response prediction · null result · Allen Institute