Quantifying the behavioral dynamics of C. elegans with autoregressive hidden Markov models

In order to fully understand the neural activity of Caenorhabditis elegans, we need a rich, quantitative description of the behavioral outputs it gives rise to. To this end, we quantify the behavioral dynamics of the worm with autoregressive hidden Markov models (AR-HMMs), a class of models that has recently yielded some insight into mouse behavior [1]. These models explicitly encode three hypotheses: (i) while the instantaneous posture of the worm is represented as a high-dimensional vector of points along the body, the first four principal components, or eigenworms, capture a significant fraction of the postural variance; (ii) within this four dimensional space, the postural dynamics are well-approximated with linear autoregressive models; and (iii) the linear autoregressive model switches over time as the worm transitions between different discrete behaviors, like forward crawling, reverse crawling, pausing, and turning. We show how AR-HMMs segment recordings of freely crawling C. elegans into meaningful discrete behaviors, providing a quantitative description of postural dynamics and a rigorous framework for assessing, comparing, and simulating worm behavior.