Heat group labels increased prediction error by 1.79% in a fixed test of calcium correlation changes in 40 C. elegans recordings

We tested whether heat exposure helps predict persistent changes in shared neural activity in another worm. We used 19 heat and 21 baseline recordings from the Atanas and Kim dataset, with all 302 slots of the Cook reference graph available for structural features. The main panel had 118,121 entries for 5,773 named neuron pairs. We removed behavior and additive endpoint changes. Each test excluded one complete worm from training. Adding heat group labels to a fixed, pooled pair forecast increased mean squared error by 1.793%. Error increased in every worm. The approximate 95% animal jackknife interval was 1.473% to 2.113%. Random group labels also increased error; the recorded-label upper permutation tail was 0.1025. Fixed anatomy models gave gains below 0.1% over a no-change forecast, with intervals that included zero. An exploratory correction on 4,941 pairs removed estimated sampling variance from the squared cohort contrast. The corrected mean was 0.000037, with an approximate interval from -0.000280 to 0.000354. The result limits these fixed forecasts. It does not show that heat has no shared neural effect. Cohorts differed in date, anatomy came from other animals, and the target excluded broad neuron changes and the acute heat response. The package preserves attributed inputs, five reference graph arrays with 302 slots, source hashes, exact forecasts, code, figures, and an offline replay.

connectome · Caenorhabditis elegans · functional connectivity · state change · heat · cross-validation · sampling variance · calcium imaging