Anatomical features gave no clear added shape prediction gain in a fixed ridge test of 686 C. elegans transfer kernels

We tested whether anatomical routes or neuron class predicts normalized signal transfer shape. We used a pinned C. elegans calcium response export and all four available anatomical maps. All 302 canonical neurons could act as relays. The main sample had 686 WT transfer pairs with q < 0.05, at least three finite trial values, and a nonzero kernel. We applied one common input, aligned response sign, and removed gain. Five outer folds kept reverse directions and bilateral copies of each class pair together. Inner folds chose ridge penalties. Common controls used trial count, soma distance, and pharynx labels. Adding anatomical counts, routes, and walks reduced mean error from 0.409596 to 0.406678, or 0.71%. The descriptive 95% node-weight range was -2.86% to +3.86%. It includes zero and keeps predictions fixed. Sender, receiver, and combined class models also had main ranges across zero. In a prespecified q < 0.01 subset of 398 pairs, both classes reduced error by 6.29%, with a range of +1.74% to +12.15%. Anatomical gain ranges included zero in every specified check. An added exact convolution check gave a 0.7145% gain. Shortest route length had rank correlation 0.037 with temporal center. The package provides a temporal benchmark, source hashes, code, selection records, predictions, model states, figures, and a fresh-directory repeat. The results concern averaged calcium kernels and fixed additive models. They do not establish prediction in new animals, absent anatomical timing information, or a mechanism of signal transmission.

connectome · Caenorhabditis elegans · signal propagation · temporal kernels · calcium imaging · structure and function · ridge regression