Brain architecture-based vulnerability to traumatic injury

Abstract

The white matter tracts forming the intricate wiring of the brain are subject-specific; this heterogeneity can complicate studies of brain function and disease. Here we collapse tractography data from the Human Connectome Project (HCP) into structural connectivity (SC) matrices and identify groups of similarly wired brains from both sexes. To characterize the significance of these architectural groupings, we examined how similarly wired brains led to distinct groupings of neural activity dynamics estimated with Kuramoto oscillator models (KMs). We then lesioned our networks to simulate traumatic brain injury (TBI) and finally we tested whether these distinct architecture groups' dynamics exhibited differing responses to simulated TBI. At each of these levels we found that brain structure, simulated dynamics, and injury susceptibility were all related to brain grouping. We found four primary brain architecture groupings (two male and two female), with similar architectures appearing across both sexes. Among these groupings of brain structure, two architecture types were significantly more vulnerable than the remaining two architecture types to lesions. These groups suggest that mesoscale brain architecture types exist, and these architectural differences may contribute to differential risks to TBI and clinical outcomes across the population.

Keywords: Kuramoto model; brain networks; lesions; structural connectivity; traumatic brain injury.

FIGURE 1

Modeling pipeline overview. Our methodology used two modeling approaches: (1) an approach to group similar structural connectivity (SC) matrices into distinct types, and (2) a method to convert any SC into an estimate of the neural dynamics using a Kuramoto oscillator model (KM). Our first approach examined the different architectural groupings among the SC matrices extracted from the Human Connectome Project dataset (top). To reduce the bias from sparse connectivity matrices on the subsequent groupings, we reparcellated Schaefer 100 into Yeo 7 representations, and then computed Pearson correlation for every pair of Yeo 7 networks in our sample. These pairwise similarity scores were collated in a similarity matrix. Modularity calculations on this similarity matrix identified which networks exhibited the greatest similarity. Our second modeling approach started with the Schaefer 100 SC network for a subject (bottom), used a KM to predict brain activity.

FIGURE 2

Modules of brain networks with similar architectures exhibited distinct structural graph metrics. (A) As visualized in the similarity matrices for both male and female structural connectivity (SC) networks, four modules (outlined in dashed lines) appeared. Within a module similarity was greater than outside (0.942 and 0.906 respective mean correlations). (B) The mean Yeo 7 edge weights for all networks within a module produced representative matrices. (C) SC metrics (MSPL, GE, CC, BC) produced module-dependent distributions (for a given row, cells containing the same color did not significantly differ). A key is also provided. For GE, CC, and BC, M1-F1 and M2-F2 modules produced significant pairings. Abbreviations used: M1, Male 1; M2, Male 2; F1, Female 1; F2, Female 2; G1, Group 1; G2, Group 2; V, Visual; SM, Somatomotor; DA, Dorsal Attention; VA, Ventral Attention; L, Limbic; C, Control; D, Default; MSPL, Mean shortest path length; GE, global efficiency; CC, Clustering coefficient; BC, Betweenness centrality.

FIGURE 3

Tuning the gamma parameter in the modularity algorithm produced differing numbers of modules. (A) The number of modules remained relatively stable at gamma values below 1. Increasing gamma allowed us to detect more—but smaller—modules. (B) We measured the distributions of various graph metrics (MSPL, GE, CC, BC) for both two and three male and female modules. The third module was comparable to the second module for both sexes, while M1 and F1 remained unchanged. (C) Multiple comparisons significance testing showed that for GE, CC, and BC, M1 and F1 did not significantly differ as a pair, nor did M2, F2, M3, and F3. For a given row in a table, cells containing the same color did not significantly differ. A key for grouping is provided.

FIGURE 4

Modularity exhibited nonuniform accuracy when considering different subsets of edges from the original Schaefer 100 parcellation. (A) Accuracy, or the fraction of subjects placed in the correct module, was dependent on the subset of nodes included in the similarity matrix formulation. Asterisks indicate subsets where males and females had significantly different accuracies. (B) Group identification according to each modularity analysis. Visual/somatomotor, visual/ventral attention, and somatomotor/limbic edges failed to produce comparable modules. Dashed lines delineate between results from edges from one Yeo 7 system (left) and results from edges between two Yeo 7 systems (right). System pairs for each subset listed as System 1/System 2. Abbreviations used: V, Visual; SM, Somatomotor; DA, Dorsal Attention; VA, Ventral Attention; L, Limbic; C, Control; D, Default.

FIGURE 5

Modules produced distinct patterns of Kuramoto oscillator outputs. (A) Synchrony reached its peak at a lower critical coupling strength for M1 and F1 networks than M2 and F2. (B) Metastability achieved a greater peak in M2 and F2 modules, but at greater critical coupling values. Critical coupling for synchrony for a given structural connectivity (SC) network was computed as the first global coupling input that yielded synchronization within 10% of the maximum synchrony. Critical coupling for metastability was computed as the global coupling input that corresponded to the Kuramoto oscillator model with the greatest metastability. Asterisks between Group 1 and Group 2 modules indicate that the distributions significantly differed between pairings but not within pairings.

FIGURE 6

Lesioning altered both the structure and neural dynamics of sampled networks. (A) As lesioning increased, synchrony decreased for all coupling strengths in all modules. M1 and F1 maintained greater levels of peak synchronization compared to M2 and F2. (B) Max metastability initially increased as result of lesion, but decreased after further lesioning. (C) Global efficiency (GE) decreased as a function of lesion severity. Between 10% and 25% lesioning, GE did not exhibit the paired M1-F1 and M2-F2 significance previously seen. Beyond 25% lesioning, M1 and F1 networks exhibited greater GE than M2 and F2. Asterisks denote significant pairings consistent with previous results. Group 1 and Group 2 modules distributions significantly differed between pairings but not within pairings.