Data-driven discovery of long timescale behavioral strategies during sensory evoked locomotion

Survival requires the brain to generate flexible and adaptive behavioral mechanisms across multiple timescales. To understand how short timescale behaviors are chained to generate long sequences, we leverage high throughput behavioral recordings of the larval zebrafish subjected to diverse sensory stimuli. Fish swim in short and discrete burst-like motions bouts. We construct a maximally predictive state space by stacking consecutive bouts, and then investigate the time evolution of state-space densities through transfer operators. Their spectral decomposition reveals two stereotyped long timescale behaviors that last tens to hundreds of seconds - a “Cruising” strategy dominated by forward locomotion and a “Roaming” strategy, where the animal repeats fast changes in orientation. These strategies are reliably found across arenas of different geometries. Surprisingly, visual or chemical stimuli do not yield novel strategies but drive the animal into either strategy. While appetitive stimuli during prey capture shift the behavior to cruising, aversive stimuli promote roaming. Our analysis reveals a simple dichotomy in the structure of motor strategies adopted by larval zebrafish to adjust navigation to external stimuli.