Requiring valid axon review marks reduced prediction gains for estimated orientation from 2.23% to 0.18% in a fixed kernel test of 3,461 neurons from one MICrONS mouse

We tested whether orientation prediction gains in one MICrONS mouse remain when shared inputs require valid axon review marks. We kept all 11,615,820 eligible input contacts on 3,461 selected targets from minnie65 v1507. We used a fixed nested kernel test with synapse position controls. Source sets had the same full input totals and unit diagonals. All sources reduced error by 2.23% against the position reference. Valid marks reduced error by 0.18%. Adding sources without a valid mark to that model gave a further 2.05% reduction. The conditional 95% cube range was 1.46% to 2.63%. A second spatial split and session tests gave the same direction. The full graph gain therefore depended mainly on source labels without valid axon review marks. A local reference kept contact positions and source counts in 10 micrometre cubes fixed. The recorded residual score for sources without valid marks exceeded all 99 shuffles. This does not verify those labels as correct axons or show synaptic causation. Review effort was not random. The outcome was a model estimate in one selected volume. We preserve compact exact graph features, code, source provenance, all model results, figures, and repeat steps. A fresh raw scan checked all 337,312,429 source rows. A full model replay matched 85 numerical files.

connectome · MICrONS · visual cortex · orientation · axon proofreading · source coverage · shared input · segmentation · structure and function · reproducibility