Model feature similarity gave a small connection prediction gain in one MICrONS split and an uncertain gain in a second

We tested whether response similarity improves connection prediction for neurons excluded from the connection model training set. We used the public MICrONS functional connectomics v1 edge table from one mouse: 295,371 candidate pairs, 8,128 connected pairs, and 148 selected source neurons. A logistic reference model used axon and dendrite proximity lengths, receptive-field distance, and region and layer labels. Every pair was tested once in each of two neuron partitions. Neither endpoint of a test pair occurred in its training set. In the primary partition, model feature similarity reduced mean log loss by 0.000208 nats per pair, or 0.182%. The conditional 97.5% neuron-score interval was 0.000012 to 0.000405. In a second, exploratory partition, the gain was 0.000176, with interval -0.000039 to 0.000392. Measured movie response similarity had intervals that included zero in both partitions under the two-comparison rule. The gains were small and differed by projection group. The result supports a limited prediction benchmark. It does not establish a general rule across animals. One animal, selected sources, fixed activity features, a single model family, and uncertainty that excludes model training variation limit the result. Code, exact splits, fitted models, pair records, predictions, and checks are supplied.

MICrONS · functional connectomics · connection prediction · neuron exclusion · generalization · uncertainty