A recurrent neural network model of C. elegans responses to aversive stimuli

Although the synaptic connections of the C. elegans connectome have been precisely mapped with detailed electron microscopy studies of nematode slices, their excitatory or inhibitory character is mostly unknown. This makes the C. elegans neural dynamics unpredictable, and limits our understanding of how specific sub-circuits work. The present study proposes a recurrent neural network model that reproduces known escape behaviours of C. elegans. To do this, our model uses the known information about the connectome, and makes guesses on the excitatory or inhibitory character of the synapses, which are iteratively updated until the model is able to reproduce the known behaviours. Specifically, we apply this model-based approach to study the neuronal sub-circuits involved in the association of aversive stimuli with escape responses. To form an escape response dataset that allows us to adjust our model parameters, we performed a meta-study on the C. elegans stimulus–response behaviours reported in literature, focusing on robust escape reactions triggered by aversive stimuli. Given a wide sample of all possible parameter sets that satisfy the behavioural constraints of the dataset, we find that more than 75% of the synaptic connections reach a univocal optimal assignment of the inhibitory or excitatory character. To validate our model, we show that in the few cases where the excitatory/inhibitory character is already known, our retrieved optimum in synaptic characters matches the results reported in literature. Finally, we assess the accuracy of this approach by applying it on recurrent neural networks that have the same connectivity structure of C. elegans but random inhibitory or excitatory character. These findings confirm the predictive power of the proposed method.