Wiring profiles reduced ridge test error by 0.64% after exclusion of animals and neuron pairs in 108 C. elegans recordings

We tested whether wiring partner profiles help predict calcium responses across animals and excluded neuron-pair groups. We first linked 104,079 of 105,741 eligible pooled response values to unique original recording and stimulation identities. The join covers 23,332 directed pairs and 108 animals. The main test retained 63,997 values from a prior 6,270-direction panel. It used 53,526 means for one animal and one directed pair. The complete Cook graph supplied all 302 neuronal slots, 3,671 chemical edges and 1,091 electrical pairs. Five animal folds and five pair folds excluded both test animals and test pair groups from fitting. Inner tests selected ridge penalties. Added wiring-profile products reduced equal-animal mean squared error from 0.461628 to 0.458650, or 0.645%, beyond cell, count, soma, local wiring and observation-profile terms. For fixed predictions, the central 95% animal-resampling range was 0.372% to 0.932%. A range that also changed endpoint weights was -0.128% to 1.530%. Two other splits and both response scales had positive mean gains. The strict raw-data subset had an uncertain 0.300% gain. All 19 later shuffled-profile controls had smaller gains. The source maps and functional records come from different animals. This result concerns a fixed response panel and model. It does not identify causal synapses or a neural noise ceiling. The package preserves the event join, five complete graph arrays, code, fitted states, predictions, source hashes, figures and repeat steps. A fresh source extraction reproduced all event arrays. All 73 outputs from nine fixed cases matched in a fresh numerical repeat.

connectome · C. elegans · signal propagation · calcium imaging · recording identities · ridge regression · reproducible benchmark