Terms for wiring direction raised Brier error in fixed WT and unc-31 models of 3,261 C. elegans atlas pairs

We tested whether wiring direction adds more response prediction information in unc-31 mutants than in WT. We used the same 3,261 ordered pairs from a fixed C. elegans signal propagation atlas. Four anatomy exports supplied a graph with 301 native neuron names. The baseline logistic model used symmetric wiring, electrical counts, two-step paths, measurement count, distance and endpoint cell effects. The full model added direct and two-step direction terms. Opposite directions and mirror copies shared five fixed test folds. Adding direction raised WT Brier error from 0.057062 to 0.057309 and unc-31 error from 0.043006 to 0.043160. The gain difference, normalized by detection variance, was -0.000096. Its conditional central 95% network reweighting range was -0.00936 to +0.01016. All 14 planned mutant cases gave small losses. One of ten later fold assignments gave a very small mutant gain. The software recovered a strong artificial direction signal in all 20 control runs. These are tests on withheld atlas entries, not separate animals. Sparse mutant measurements, detection labels and cross-animal anatomy limit the result. The package preserves code, numerical input tables, complete native edge facts, source hashes, predictions, model states, checks and figures. Fresh-source and offline builds reproduced the numerical results.

connectome · Caenorhabditis elegans · signal propagation · unc-31 · directionality · structural connectivity · calcium imaging · prediction benchmark