Prioritized nonlinear modeling of shared dynamics across neural populations

Advances in neural recording technology enable simultaneous measurements across multiple brain regions, providing new opportunities to study inter-regional interactions. However, several challenges arise when developing nonlinear dynamical models of cross-region interactions. In particular, we identify three key desired properties for a nonlinear framework. First, it should prioritize extraction of cross-region dynamics to avoid confounds from within-region dynamics. Second, it should enable localization of nonlinear structure within the model to better interpret cross-regional interaction models. Third, it should support both causal and non-causal inference of shared dynamics and do so using source region activity alone. Current cross-regional models do not satisfy all these properties.

Here, we address these challenges by developing cross-population prioritized dynamical nonlinear interaction model (CroP-DYNO), a nonlinear dynamical framework that prioritizes the learning of shared cross-population dynamics to avoid confounds from within-population activity. Further, CroP-DYNO allows individual components of the dynamical model to be independently specified as nonlinear or linear. Finally, it supports both causal and non-causal inference of latent states, using only source region activity.

We validate our method on datasets across species and distinct brain regions. We find that both the prioritized learning and the nonlinear modeling in CroP-DYNO are important for accurately extracting cross-population dynamics. As such, CroP-DYNO outperforms baseline nonprioritized and linear prioritized models in predicting target neural population activity from source activity. Further, CroP-DYNO enables systematic localization of nonlinear structure and quantifies dominant interaction pathways between brain regions, with interaction strengths that align with known circuit anatomy.

Overall, these results establish CroP-DYNO as a flexible nonlinear framework for studying interactions between neural populations and brain regions, enabling both accurate modeling and interpretable dissection of nonlinear structure in cross-regional communication models.