Neural network features distinguish chemosensory stimuli in Caenorhabditis elegans
Fig 7
The Simplified Network made by removing trivial interactions between putative sensory neurons improves robustness of the logistic regression classifier, and shows that Graph features alone are capable of discriminating stimulus identity.
Graph features from the Whole Network allowed above-chance accuracy when training on Data Set 2 (DS2) and testing on Data Set 1 (DS1; B, * p-value = 0.04), but not vice-versa (A); the same was true of Graph features from the Simplified Net- work, which produced the same accuracies as the Whole Network (C,D, * p-value = 0.04). Moreover, Graph features from the Simplified Network led to above-chance classification accuracy on the combined data set of 48 worms (E,F, * p-value = 0.01), while Activity features could not be used to exceed chance-accuracy. The process of network simplification is depicted in panel G via a cartoon, where we removed the (trivially) strong edges between neurons correlated to stimulus onset (i.e., “putative sensory neurons”, yellow circles). No features were standardized. The dashed red line refers to chance accuracy of 25%. Statistical significance was assessed using a permutation test with n = 100 permutations. N = 24 worms for Data Set 1 (all worms except those exposed to diacetyl), 24 worms for Data Set 2.