Drift-diffusion dynamics of hippocampal replay

Replay in the hippocampus during sharp-wave ripples is thought to play major roles in learning and memory. However, existing analysis methods often lead to inconsistent and inaccurate metrics for characterizing replay dynamics. We develop a novel computational framework that models the replay dynamics using a drift-diffusion process. Further, to capture the potentially rich population-level dynamics during sharp-wave ripples, our model allows switching between multiple well-motivated types of dynamics. Applications of our method to rat hippocampal recordings lead to a number of insights. First, our results reveal that a small fraction of SWRs are stationary, and those with drift-diffusion dynamics generally have a much higher speed than the animal’s movement. Second, we find that replay dynamics are heterogeneous and inconsistent with a random walk model proposed previously. The mean-squared displacement of replay trajectories scales quadratically with time, supporting the presence of substantial drifts. Third, we find that only a tiny fraction of SWR events (less than 1%) exhibit sequential structure at the timescale of 100 ms before the animal had spatial experience in an environment. This suggests that, while neural activity in the hippocampus may be coordinated before the animal’s spatial experience (i.e., preplay), the level of coordination is much weaker than that after spatial experience. Overall, our approach enables precise characterizations and unambiguous interpretations of population dynamics during sharp-wave ripples, which more broadly can provide a better understanding of the functions and underlying mechanisms of replay.Author summary: During sharp-wave ripples (SWRs), neural populations in the mammalian hippocampus sometimes replay the neural activity patterns reminiscent of previous experience. Such replay events have been implicated in various cognitive functions, such as learning, memory and planning. Despite extensive prior research, the field still lacks appropriate inferential methods to accurately characterize the potentially rich neural dynamics during SWRs. This issue has led to substantial limitations in interpreting the experimental data collected in previous studies. To address this gap, we have developed a computational modeling framework to infer and quantify the neural dynamics during SWRs. This approach is flexible in that it captures different types of neural dynamics as well as the switching of neural dynamics in a unified model. Meanwhile, the parameters in the model are highly interpretable. Applications of this method to neural population recordings from a one-dimensional spatial task in rats reveal new insights into a number of questions that are currently under debate regarding the structure and property of hippocampal replay.