Study on C. elegans behaviors using recurrent neural network model

Abstract

With the complete knowledge on the anatomical nerve connections of the nematode Caenorhabditis elegans (C. elegans), the chemotaxis behaviors including food attraction and toxin avoidance, are modeled using dynamic neural networks (DNN). This paper first uses artificial DNN, with 7 neurons, to model chemotaxis behaviors with single sensor neurons. Real time recurrent learning (RTRL) is carried out to train the DNN weights. Next, this paper split the single sensor neuron into the left and right pair (dual-sensor neuron), with the assumption that C. elegans can distinguish the input difference between left and right, and then the model is applied to learn to reproduce the chemotaxis behaviors. The simulation results conclude that DNN can well model the behaviors of C. elegans from sensory inputs to motor outputs both in single sensor and dual-sensor neuron networks.