Adding spontaneous coactivity gave 1.88% lower error in a fixed test of 4,048 C. elegans response directions, with a class weight range that included zero gain

We tested whether spontaneous coactivity adds prediction value for signed responses to neuron activation after wiring and neuron effects are controlled. We joined a signed SCI correlation table from 21 freely moving worms to a fixed WT response atlas. The main panel had 4,048 response directions, 116 exact neuron IDs and 54 cell classes. All 302 slots in the Cook anatomical reference supplied route features. Five fixed folds kept reverse directions and copies of each class pair together. Adding one behavior-adjusted coactivity value reduced ridge test error by 1.88%, from 0.091153 to 0.089438. All five folds improved. A descriptive 95% range from 10,000 class weight draws was -1.18% to +4.76% gain, with predictions fixed. Removing each activity animal and refitting gave gains from 1.75% to 1.98%. In a separate planned task, both endpoint classes were excluded from training and identity indicators were omitted. Coactivity then increased error by 0.29%. An added name-padding check recovered 102 directions and gave a 1.89% decrease. The small main gain does not give clear evidence of transfer across class samples or new classes. The assays differ in movement, strain and recording conditions. Response lists lack animal and event IDs. The result is a conditional prediction association, not a causal connection estimate. The package includes code, derived data, exact folds and predictions, figures, source hashes and repeat steps. A complete rebuild from the pinned atlas, workbook and signed activity table matched 17 numerical files.

connectome · C. elegans · spontaneous activity · signal propagation · calcium imaging · structure and function · cross validation · reproducible benchmark