Latent factors capturing evolution of behavior in larval zebrafish

Animal behavior exhibits remarkable diversity across spatial and temporal scales, showing both significant regularity and high variability. Understanding such multiscale dynamics is essential for revealing the mechanisms and organizing principles that generate behavior and shape motor control. However, conventional analyses largely rely on static or statistical descriptors, which fail to capture the intrinsic dynamical rules governing behavioral evolution over time. Larval zebrafish, whose body posture can be quantified as spatial coordinates, provide an ideal vertebrate model for studying behavioral dynamics. Here, we develop a dynamical reconstruction framework that combines delay embedding with an echo state network to recover the latent state space of zebrafish behavior from high-dimensional postural time series. The framework preserves temporal dependencies and maps postural variations onto latent factors whose combinations define distinct behavioral states-turning, routine swimming, and slow swimming-matching feature-based clustering. These states evolve as continuous trajectories on a low-dimensional attractor, with transitions marked by nonlinear tipping points that reflect reorganization of underlying factors. Leveraging this reconstructed state space, the framework also enables short-term behavioral prediction. Overall, our framework provides a unified quantitative tool for revealing the latent dynamical principles of zebrafish behavior, with potential extensions to neural and bio-inspired control.