Rethinking brain-wide interactions through multi-region ‘network of networks’ models

Abstract

The neural control of behavior is distributed across many functionally and anatomically distinct brain regions even in small nervous systems. While classical neuroscience models treated these regions as a set of hierarchically isolated nodes, the brain comprises a recurrently interconnected network in which each region is intimately modulated by many others. Uncovering these interactions is now possible through experimental techniques that access large neural populations from many brain regions simultaneously. Harnessing these large-scale datasets, however, requires new theoretical approaches. Here, we review recent work to understand brain-wide interactions using multi-region "network of networks" models and discuss how they can guide future experiments. We also emphasize the importance of multi-region recordings, and posit that studying individual components in isolation will be insufficient to understand the neural basis of behavior. Highlights• Brains can be viewed as a composition of anatomically-distinct modules called regions • Brain regions interact in complex ways to form a "network of networks"• Untangling this network of networks could lead to new views of functional modularity

Inferring brain-wide interactions using data-constrained recurrent neural network models

“…Mathematically, these weights correspond to interaction strengths, summarized by a vector with each source unit represented as a single entry in the vector. However, neural circuits in biological brains are typically intricately, recurrently connected 2 . This feature prompted common use of RNNs to model their computational functions 8,9 .…”

“…This feature prompted common use of RNNs to model their computational functions 8,9 . RNNs trained to produce desired behaviors [10][11][12] and tasks [13][14][15][16][17][18] or match neural data 2,17,19 (or both 20,86 ) can be reverse-engineered to generate hypotheses for how biological neural circuits could implement similar functions 21,22 . As in the single-layer network, the activity of any unit in an RNN can be computed as a weighted sum of the activity of all other units in the network, which are the sources of its input (Figure 1b).…”

“…Given the high degree of recurrent connectivity between regions, interactions between active neurons in different brain regions can be conceived as an RNN 2 . To implement CURBD, we model brain-wide circuitry as multiple interconnected RNNs forming a "network of networks" 2 .…”

“…Third, correlative analyses typically describe interactions between two regions and are difficult to extend to data from multiple interacting regions, though recent work on switching dynamical systems shows promise 53,54 . CURBD addresses these limitations by building and analyzing RNNs that are trained to match the entire timeseries from experimentally collected data 2,86 . Thus, CURBD explicitly models the recurrence between all recorded neurons, capturing all possible multi-region interactions in the dataset.…”

“…Brain regions, in particular, have striking specialization and unique functional characteristics. However, individual brain regions also frequently interact with numerous other regions throughout the brain 2 . These macroscopic circuits are recurrently connected via direct projections, multi-synapse loops, and more widespread, indirect effects such as neuromodulator release 3 .…”