State-dependent navigation strategies in C. elegans vary with olfactory learning

Animals flexibly adjust behavior depending on the environmental context and learned experience. The nematode C. elegans modulates its preference towards an odor source depending on its pairing with food or starvation. However, precisely how the worm alters its navigation strategy in response to olfactory learning has not been characterized, in part due to challenges with measuring the exact odor cues in the environment. Here, we used a novel apparatus and optogenetics to measure and then model learned odor-navigation strategies in worms. Specifically, we modelled navigation using a Mixture of Generalized Linear Models (MoGLMs), where each GLM corresponds to a particular strategy known to exist in worms: (1) biased random walk, in which the probability of a turn is modulated by odor concentration and (2) weathervaning, in which the head angle depends on the local odor gradient. We fit this model to detailed locomotion data under a precisely measured butanone odor concentration, after the worm had olfactory learning experiences. The inferred parameters show that learning bidirectionally modulates the sensing kernels for both strategies. We corroborated this finding using optogenetics, confirming that the behavioral response to optogenetic activation of an olfactory neuron is bidirectionally modulated by learning. We then extended our model to a hidden-Markov model (HMM), with each state corresponding to a distinct MoGLM. This model provided a substantially better fit to data, in contrast to prior work assuming worms employ a fixed navigation strategy. Finally, we showed that the inferred HMM outperforms a fixed strategy, indicating that a worm with multi-state switching achieves a normatively better solution to the olfactory navigation problem. By combining a new modeling approach with new measurements of odor-guided behavior, we provide a new paradigm for understanding state-dependent olfactory navigation and learning in C. elegans.