Overcoming non-identifiability issues in brain-wide communication models

Advancements in large-scale neural recordings are now enabling opportunities to probe cognitive computations across simultaneously recorded brain regions, necessitating modeling tools for deciphering multi-region interactions. In an emerging class of dynamical models, brain-wide neural activity is reconstructed by region-specific modules that interact through communication channels and may receive additional information as inputs. These models aim to recover the effective connectome, which includes the inter-region connection diagram and the messages communicated between regions. However, without proper constraints, model fitting may arrive at one of many solutions that fit the data well, despite most of those solutions being inconsistent with the ground truth connectome. Therefore, identifiability is crucial for drawing reliable scientific conclusions from a learned model. In this work, we highlight non-identifiabilities issues stemming from insufficient constraints on model architecture and intrinsic properties of neural data. Specifically, since inputs and communication are typically unobservable, models must apply strong assumptions about how to set these quantities or infer them from data, resulting in somewhat arbitrary model constraints. Additionally, the presence of globally distributed signals and correlated activity across regions make it difficult to pinpoint where those signals originate and how they propagate across regions. In this work, we propose a novel multi-regional communication model (termed MR-LFADS) specifically designed to overcome these challenges. We show that, in situations emphasizing each issue, our model accurately recovers both the communication diagrams and the messages communicated.