Decomposed Linear Dynamical Systems (dLDS) models reveal context-dependent dynamic connectivity in C. elegans

Despite innumerable efforts to characterize C. elegans neural activity and behavior, the roles of each individual neuron and the compositional mechanisms of all their interactions are still not completely known. In particular, there is evidence that C. elegans neurons perform changing roles over time, in different behavioral contexts, etc.; however, existing stationary anatomical and functional connectivity can average across time and obfuscate the nonstationary nature of the neural dynamics. We contribute to these efforts by leveraging recent advances in decomposed linear dynamical systems (dLDS) models. dLDS models neural dynamics in a latent space with a set of linear operators that can be recombined and reused over time, enabling the discovery of multiple parallel neural processes on different timescales. We leverage the ability to identify reused patterns of dynamical neural interactions, which we call “dynamical connectivity,” to 1) identify contextually dependent roles of neurons; 2) discover the underlying variability of neural representations even under discrete behaviors; 3) quantify differences between anatomical, functional, and dynamical connectivity; and 4) learn a single aligned latent space underlying multiple individual worms’ activity. These results highlight the importance of defining a neuron’s functions not solely by its internal activity or place in a time-averaged network, but by its evolving interactions across context-dependent circuits.