Multisensory integration in the mouse cortical connectome using a network diffusion model

Abstract

Having a structural network representation of connectivity in the brain is instrumental in analyzing communication dynamics and neural information processing. In this work, we make steps towards understanding multisensory information flow and integration using a network diffusion approach. In particular, we model the flow of evoked activity, initiated by stimuli at primary sensory regions, using the asynchronous linear threshold (ALT) diffusion model. The ALT model captures how evoked activity that originates at a given region of the cortex "ripples through" other brain regions (referred to as an activation cascade). We find that a small number of brain regions-the claustrum and the parietal temporal cortex being at the top of the list-are involved in almost all cortical sensory streams. This suggests that the cortex relies on an hourglass architecture to first integrate and compress multisensory information from multiple sensory regions, before utilizing that lower dimensionality representation in higher level association regions and more complex cognitive tasks.

Keywords: Asynchronous linear threshold model; Claustrum; Hourglass effect; Mouse connectome; Network diffusion cascade; Parietal temporal cortex.

Figure 1.

Illustration of ALT model and τ-core analysis. (A) A toy example of a five-node network on which we run the ALT model. Each edge is marked with a communication delay, followed by a weight. The activation threshold is θ = 1. The black edges represent the underlying structural network, while the red unidirectional edges represent the activation cascade as it unfolds over time. (B) The activation cascade (a directed acyclic graph) for the previous toy example. The source of the cascade is n₁. (C) A different toy example with three activation cascades (the sources are nodes u, v, and y). The total number of source-target paths is 12 (5 at the left, 3 at the middle, and 4 at the right). Node w has the highest path centrality (P(w) = 10/12). If τ ≤ 10/12, the τ-core consists of only that node.

Figure 2.

VSD data-processing pipeline. (A) Lower: The Allen Reference Atlas (ARA). Upper left: A sample VSD image covering most of the left cortical surface five frames after visual stimulation. Upper right: The ROIs at the left ARA cortical surface mapped to the native cortical surface of an animal. (B) The activation time of a pixel is defined as the frame of maximum poststimulus VSD signal at that pixel. (C) The activation time of an ROI is defined as the activation time of most pixels in that ROI. (D) The output of this pipeline is an activation time for each ROI, depicted here with a gray scale map (black for the first ROI activation and white for the last).

Figure 3.

Effect of parameter θ on cascade size, and similarity between the 10 cascades. (A) Each row of the heat map shows the fraction of activated nodes after the stimulation of a single source, for different values of the threshold θ. The selected threshold is marked with the dashed vertical line. (B) Similarity between the 10 sensory cascades using the average-linkage hierarchical clustering method.

Figure 4.

The visual activation cascade, according to ALT (θ = 0.98). The source for this cascade is the primary visual cortex (VISp). The red edges form the activation cascade, while the underlying blue edges show anatomical connections that do not participate in this cascade, those connections may be present in other sensory cascades or they may play a role in feedback (or second-order) interactions that are not captured by the “first ripple” scope of the ALT model. To help with the visualization, we place the nodes in eight layers, so that cascade edges only point from a layer to a higher layer (never to the same or lower layer). The vertical position of each node is slightly “jittered” to avoid cluttering due to anatomical connections between nodes of the same layer.

Figure 5.

Comparison between model-based and experimental temporal ordering of ROI activations. (A) The y-axis shows the percentage of (X,Y ) ROI pairs that show temporal agreement (green), temporal disagreement (red), and insufficient temporal resolution (blue) between the activation order of X and Y in the modeling results and the mouse experiments. The plot shows results for five animals and for five sensory stimulations (a touch at the whiskers, forelimb, and hindlimb, as well as an auditory and a visual stimulation). (B) The same comparison, but here we have randomized the ROIs that are active during each frame, preserving the number of ROI activations in each frame.

Figure 6.

Path centrality and τ-core analysis. (A) Path Centrality (PC) histogram for the 67 regions in N_c, considering all source-target paths across the 10 activation cascades. (B) Cumulative path coverage by the top X core nodes for X = 1, ⋯ ,67. Nine regions are sufficient to cover τ = 90% of all paths. (C) Core regions for τ = 90% also showing the path coverage contributed by each of them and its path centrality rank. (D) The location of the top five core regions.

Figure 7.

Location-related metrics. In both matrices, a column represents one of the 10 activation cascades, originating at the node shown at the top of the column. (A) Each row represents the source-distance of the corresponding node from the source of that column’s cascade. White denotes a distance of one hop, while black denotes the maximum distance for that cascade. Rows are ordered in terms of the average distance (in number of hops) of the corresponding node from the sources of all activation cascades (excluding the MOB cascade, which is very different). (B) Each row represents the influence of the corresponding node, that is, the number of nodes that are reachable from that node in the activation cascade that the column represents. White denotes an influence of one (only that node), while black denotes an influence that covers all network nodes. Rows are ordered in terms of the average influence of the corresponding node across all activation cascades (excluding the MOB cascade).

Figure 8.

Robustness results. The effect of different connectome randomization methods on the core size. Light blue shade marks the 5% to 95% values among 100 randomization runs, while the solid blue line is the median of these runs. The red line represents the τ-core size for the original connectome. The dotted green line marks the τ-core size for τ = 90%. The table at the bottom shows the fraction of random networks that include each of the eight τ-core nodes.