Separate penalties reduced the added inhibitory input gain from 1.41% to 0.18% in a model-derived orientation test of 5,193 MICrONS neurons

We tested whether shared inhibitory inputs add information about preferred visual orientation in 5,193 excitatory neurons. We used the MICrONS minnie65 v1507 static export and manual functional matches. Orientation values came from a digital twin. Five folds held out whole neurons. The fixed first model gave a 1.41% error reduction when inhibitory inputs were added to excitatory inputs. That excitatory model was worse than the location baseline. An exploratory extension used separate penalties for baseline and input features. The added inhibitory gain fell to 0.18%, with a conditional 95% range from -0.10% to +0.49%. The range includes zero. Tests with proofread source axons and spatial block holdout also had ranges across zero. Both input classes together reduced baseline error by 0.87% with separate penalties. Classified sources covered only 8.84% of recorded incoming synapses. We therefore also used an expanded map with 17.67 million synapses from 7.85 million source segments. Its 128-component input representation reduced baseline error by 0.73%. These components retained 10.08% of input trace. Adding identified inhibitory inputs gave no clear further gain. The evidence links recorded input structure to model-derived function in this volume. It does not establish a unique inhibitory contribution or a causal mechanism. Ranges resample spatial blocks with predictions fixed. They do not measure variation across mice. The package preserves code, input extracts, provenance, predictions, results, checks, figures, and repeat instructions.

MICrONS · connectome · inhibitory inputs · visual orientation · ridge regression · source coverage · spatial validation