Building data-driven multi-regional models for decision-making at a single-trial scale

Recent technological advances allow us to measure large populations of neurons across multiple brain regions with millisecond precision and over extended periods while monitoring behavior. These data reveal that both behavior and neural dynamics change over minutes and hours, leading to different behavioral states at the level of single trials. However, many of the current methods and models for linking neural activity and behavior rely on trial-averaged data and often focus on one brain region, limiting our ability to causally link multiregional interactions to flexible behaviors that span short and long timescales. We developed a data-driven modeling framework to gain a mechanistic understanding of how multiregional neural activity drives moment-to-moment decisions and is conditioned by behavioral states. Our framework fits multiregional recordings at the level of single trials and at single-cell resolution by factorizing overall neural activity into distinct yet recurrently connected selective populations that encode task variables, such as sensory stimuli or choice, as well as additional populations that are unselective for the behavior. We applied this framework to multi-regional neural recordings of ~66000 neurons during a memory-guided decision-making task. Analysis of the trained models revealed that the choice-selectivity of the network is resistant to perturbations along the choice-selective factor, while, paradoxically, perturbations to a few unselective populations in the motor cortex significantly impact the choice-selectivity by collapsing the dynamics of the choice-selective populations. This finding mirrors recent results from photostimulation experiments. We systematically dissect the sources of trial-by-trial variability in terms of low-dimensional population dynamics and its possible effect on behavior. Thalamocortical interactions route unselective latent activity, driving choice-selective dynamics across hemispheres. Our modeling framework extracts testable mechanisms from multimodal data, serving as a new tool in systems neuroscience.