Input roots reduced mean error by 0.44% in an exploratory 100 µm block test of 9,329 model receptive field centers from MICrONS v1507

We tested whether recorded synaptic inputs help estimate model receptive field centers in one MICrONS mouse. We read all 303 million rows of the v1507 connection export and selected 9,329 matched excitatory targets. A requirement for a classified source cell removed 91.2% of recorded input synapses in this cohort. At a cutoff of at least three synapses per edge, an expanded sample kept 628,566 edges from 224,697 input roots. Roots can be cells or fragments. In the original test, input corrections after a count control had 0.183% higher mean error than a spatial model. Later tests changed the baseline and retained unclassified input roots. This exploratory method reduced squared error from 0.003545061 to 0.003529531, a 0.438% gain, with five outer folds of 100 micrometre tissue cubes. A descriptive 95% block interval was 0.181% to 0.694%. It keeps fits fixed and omits model choice and shared-input dependence. Gains were 0.601% and 0.416% with 50 and 200 micrometre cubes. A test limited to sources with a cleaned axon flag gave no clear gain. The outcome is a digital twin estimate, and many input roots have unchecked axons. The positive result needs independent confirmation. We provide a versioned prediction benchmark, a coverage audit, all analysis branches, code, selected graph data, provenance, figures, and reproduction instructions.

MICrONS · connectome · receptive field · input sampling · spatial validation · segmentation fragments · structure and function