A contact position model gave 4.3% more coherence test error across six trials from 35 inputs to one leech motor neuron

We tested whether synapse count and position predict rhythmic activity in inputs to one leech motor neuron. The panel had 35 mapped input neurons, 271 synapses, two swim trials, and four local bend trials. The target was complex coherence referred to motor neuron DE-3R. Each test excluded all activity from one neuron. Inner tests selected ridge penalties and distance kernels. The count model selected the training mean in all 210 tests. The count plus path position model gave 4.32% more mean squared error than the training mean: 0.44487 against 0.42644. Whole-neuron label permutations with selection repeated gave an upper-tail fraction of 0.4955. Euclidean distance, higher quality matches, removal of the largest input, and magnitude-only tests also gave no gain. A separate check transferred measured coherence between the two swim trials for each neuron. Error fell by 81.62% relative to the pooled swim training-mean baseline. This check uses activity from each test neuron and shows repeatability in these recordings. We provide pinned input hashes, independent code, all test predictions, and reproduction instructions. Results apply to these released trials and this smooth model family. The test has one animal and excludes crawling.

functional connectomics · synapse position · leech · coherence · kernel regression · neuron exclusion · reproducibility