Cliques and cavities in the human connectome

Diffusion spectrum imaging (DSI) data and T1-weighted anatomical scans were acquired from eight healthy adult volunteers on 3 separate days (27 ± 5years old, two female, and two left-handed) (Gu et al. 2015a). All participants provided informed consent in writing according to the Institutional Review Board at the University of California, Santa Barbara. Whole-brain images were parcellated into 83 regions (network nodes) using the Lausanne atlas (Hagmann et al. 2008), and connections between regions (network edges) were weighted by the number of streamlines identified using a determistic fiber tracking algorithm. We represent this network as a graph G(V,E)on V nodes and E edges, corresponding to a weighted symmetric adjacency matrix A. For clique calculations in the main text, the original network (ρ = 0.9552) was thresholded at ρ = 0.25(corresponding to a weight = 261) to remove spurious connections (Zalesky et al. 2010; Zalesky et al. 2016; van den Heuvel et al. 2012) and for consistency with previous work (Sizemore et al. 2016). See Supporting Information and Refs (Cieslak and Grafton 2014; Gu et al. 2015a) for detailed descriptions of acquisition parameters, data preprocessing, and fiber tracking. In the supplement, we provide additional results for the case in which we correct edge weight definitions for the effect of region size Fig. 23.

In a graph G(V,E)a k-clique is a set of k all-to-all connected nodes. It follows that any subset of a k-clique is a clique of smaller degree, called a face. Any clique that is not a face we call maximal. To assess how individual nodes contribute to these structures, we define node participation in maximal k-cliques as P _k (v), and we record the total participation of a node as \(P(v) = {\sum }_{k = 1}^{n}P_{k}(v)\).

To detect cycles which enclose topological cavities, we computed the persistent homology using (Henselman and Ghrist 2016). We restrict our attention to dimensions 1–2 after finding no persistent features in dimension 3 (Sizemore et al. 2016).

Computing persistent homology involves first decomposing the weighted network into a sequence of binary graphs beginning with the empty graph and adding one edge at a time in order of decreasing edge weight (also called a Weight Rank Clique Filtration (Petri et al. 2013a, b). Formally, we translate edge weight information into a sequence of binary graphs called a filtration, beginning with the empty graph G ₀ and adding back one edge at a time following the decreasing edge weight ordering. To ensure all edge weights are unique we added random noise uniformly sampled from [0,0.0001]. However, this has essentially no effect on the persistence diagrams, as stability theorems ensure that small perturbation of the filtration leads to small perturbation of the persistent homology (Chowdhury and Mémoli 2016; Cohen-Steiner et al. 2007). Noise can have a small effect on cycle representatives but in this study a great majority of edges within the thresholded networks are unique so the noise is not expected to largely alter cycle representatives – only to order those edges with tied edge weights.

Within each binary graph of this filtration, we extract the collection of all k-cycles, families of (k + 1)-cliques which, when considered as a geometric object, form a closed shell with no boundary. Formally, as we are working with coefficients in \(\mathbb {Z}_{2}\), these are collections of (k + 1)-cliques {σ ₁,…σ _n } such that every k-subclique of some σ _i (called a boundary) appears as a subclique in the collection an even number of times. Two k-cycles are equivalent if they differ by a boundary of k + 1-cliques. This relation forms equivalence classes of cycles with each non-trivial equivalence class representing a unique topological cavity. (In the mathematical literature, these are called non-trivial homology classes. However, due to the extensive and potentially confusing collision with the use of the word “homology” in the study of brain function, here we elect to use this new terminology outside of references and necessary mathematical discussion in the Methods and Supplementary Information. Throughout, the word “homology” refers to the mathematical, rather than the biological, notion.)

Constructing the sequence of binary graphs allows us to follow equivalence classes of cycles as a function of the edge density ρ. Important points of interest along this sequence are the edge density associated with the first G _i in which the equivalence class is found (called the birth density, ρ _{b i r t h} ) and the edge density associated with the first G _i in which the enclosed void is triangulated into higher dimensional cliques (called the death density, ρ _{d e a t h} ). One potential marker of the relative importance of a persistent cavity to the weighted network architecture is its lifetime (ρ _{d e a t h} − ρ _{b i r t h} ). A large lifetime indicates topological cavities that persist over many edge additions, suggesting a greater importance of that cavity to the intrinsic structure of the complex. An alternative measure is the death to birth ratio π = ρ _{d e a t h} /ρ _{b i r t h} which highlights topological cavities that survive exceptionally long in spite of being born early, a feature that is interesting in geometric random graphs (see Bobrowski et al. 2015 and Supporting Information).

To study the role of each topological cavity in cognitive function, we extract the minimal representatives of each non-trivial equivalence class at the birth density. For unfiltered complexes, the problem of finding a minimal generator for a given homology class is well known to be intractable (Chen and Freedman 2011; Dey et al. 2011). However, leveraging the filtration, we are able to answer the corresponding question in this context with relative ease. We used the persistent homology software Eirene (Henselman and Ghrist 2016) which returns the birth density and consequentially the starting edge of each persistent homology class. To recover the minimal cycle, we threshold the network at the density immediately preceding ρ _{b i r t h} , then perform a breadth-first search (Rubinov and Sporns 2010) for a path from one vertex to the other, taking all minimum length paths as solutions. If for one persistent cavity we find multiple possible minimum-length paths arising from different equivalence classes, we still record and analyze each of the possible minimal generators, since any could be the homology class. For higher dimensional cycles we perform a similar process by hand, but we note that they could be algorithmically identified using appropriate generalizations of the graph search method and other approaches (Dey et al. 2011).

In addition to the notions of cliques and cavities from algebraic topology, we also examined corresponding notions from traditional graph theory including communicability and rich-club architecture, computed using the Brain Connectivity Toolbox (Rubinov and Sporns 2010).

We first considered nodes that participated in many maximal cliques, and we assessed their putative role in brain communication using the notion of network communicability. The weighted communicability between nodes i and j is with D := diag(s _i ) for s _i the strength of node i in the adjacency matrix A, providing a normalization step where each a _{i j} is divided by \(\sqrt {d_{i}d_{j}}\) (Crofts and Higham 2009; Estrada and Hatano 2008). This statistic accounts for all walks between node pairs and scales the walk contribution according to the product of the component edge weights. The statistic also normalizes node strength to prevent high strength nodes from skewing the walk contributions. We refer to the sum of a node’s communicability with all other nodes as node communicability, C _i .

Intuitively, nodes that participate in many maximal cliques may also play a critical role in the well-known rich club organization of the brain, in which highly connected nodes in the network are more connected to each other than expected in a random graph. For each degree k we compute the weighted rich club coefficient where W _>k is the summed weight of edges in the subgraph composed of nodes with degree greater than k, E _>k is the number of edges in this subgraph, and \(w_{l}^{ranked}\) is the l-th greatest edge weight in A. Rich club nodes are those that exist in this subgraph when ϕ ^w (k) is significantly greater (one sided t-test) than \(\phi ^{w}_{random}(k)\), the rich club coefficient calculated from 1000 networks constructed by randomly rewiring the graph A while preserving node strength (Rubinov and Sporns 2010).

Furthermore, highly participating nodes may also contribute to a hierarchical organization of the network. To evaluate this contribution, we compute the k-core and s-core decompositions of the graph (Hagmann et al. 2008; Chatterjee and Sinha 2007). The k-core is the maximally connected component of the subgraph with only nodes having degree greater than k. The s-core is similarly defined with summed edge weights in the subgraph required to be at least s.

We sought to compare the empirically observed network architecture to that expected in an appropriate null model. Due to the well-known spatial constraints on structural brain connectivity (Klimm et al. 2014; Lohse et al. 2014; Bullmore and Sporns 2012; Betzel et al. 2016) as well as the similarity in mesoscale homological features to the Random Geometric network (Sizemore et al. 2016) we considered a minimally wired network in which nodes are placed at the center of mass of anatomical brain regions. Each pair of nodes are then linked by an edge with weight w _i,j = 1/d(i,j), where d(i,j)is the Euclidean distance between nodes i and j. For consistency with the empirical data, we threshold this complete weighted network at an edge density of 0.25 for analyzes in which the DSI network is also thresholded. In each scan, the locations of region centers were collected. Thus, we considered a population of 24 model networks where differences between model networks arise from differences between scans. This null model allows us to assess what topological properties are driven by the precise spatial locations of brain regions combined with a stringent penalty on wiring length. Note that defining edge weights to be the inverse pairwise distance between points creates a filtered complex similar to that of either the Vietoris-Rips (Vietoris 1927; Hausmann et al. 1995) or Čech complex with an axis adjusted for edge rank instead of weight. We use the edge rank filtration for the null model here for consistency with the empirical data. Many ways of constructing simplicial complexes from graphs exist (Bergomi et al. 2017) but we have chosen the above methods because they are reletaively well understood and do not require further assumptions about the data.

Though we detected persistent cavities in the group-averaged DSI network using persistent homology, we also ask whether these patterns of connectivity and the corresponding cavities exist in multiple individuals and in multiple scans acquired from the same individual. To address this question, we asked whether a similar geometric loop is seen and whether a similar topological cavity is present in each scan. However, identifying similar topological cavities is not trivial and we next thoroughly discuss our method including our definition of “similar topological cavities”.

Here, we use the group-averaged network thresholded at an edge density (ρ) of 0.25 to remove spurious edges (Zalesky et al. 2010, 2016; van den Heuvel et al. 2012) and for consistency with previous studies (Sizemore et al. 2016). Results at other densities are similar, and details can be found in the Appendix. As a null-model, we use minimally wired networks (Fig. 1d) created from assigning edge weights to the inverse Euclidean distance between brain region centers (see Methods) observed in each of 24 scans. This model mimics the tendency of the brain to conserve wiring cost by giving edges that connect physically close nodes higher weight than edges between distant nodes.

The first step in a topological analysis is an enumeration of all maximal k-cliques in the average structural network. Recall that a k-clique is a set of k nodes having all pairwise connections (see Fig. 1b for 2-, 3-, and 4-cliques representing edges, triangles, and tetrahedra, respectively.) By definition, a subgraph of a clique will itself be a clique of lower dimension, called a face. A maximal clique is one that is not a face of any other (see Fig. 1c for a maximal 4-clique, which contains 3-, 2-, and 1-cliques as faces).

To understand the anatomical distribution of maximal cliques in both real and null model networks, we count the number of maximal k-cliques in which a node is a member, and refer to this value as the node participation, P _k (v) (see Methods). Summing over all k gives the total participation, P(v). We observe that the distribution of maximal clique degrees is unimodal in the minimally wired null model and qualitatively bimodal in the empirical data (see Fig. 2a), though we report statistically that we cannot reject that it is unimodal (p = 0.210, dip test (Hartigan and Hartigan 1985)). Anatomically, we observe a general progression of maximal clique participation from anterior to posterior regions of cortex as we detect higher degrees (Fig. 2a, bottom and Fig. 8). Indeed, maximal cliques of 12–16 nodes contain nearly all of the visual cortex. This spatial distribution suggests that large interacting groups of brain regions are required for early information processing, while areas of frontal cortex driving higher-order cognition utilize smaller working clusters. We also observe that the human brain displays preferences for small (4–6 node), and large (12–16 node) processing units instead of medium-sized (approximately 8 node) units as in the minimally wired null model.

The anterior-posterior gradient of maximal clique size can be complemented by additionally analyzing regional variation in the cognitive computations being performed. Specifically, we ask whether node participation in maximal cliques differs in specific cognitive systems (Power et al. 2011) (Fig. 2b). We observe that the largest maximal cliques are formed by nodes located almost exclusively in the subcortical, dorsal attention, visual, and default mode systems, suggesting that these systems are tightly interconnected and might utilize robust topologically-local communication. This spatial distribution of the participation in maximal cliques differs significantly from the minimally wired null model, particularly in the cingulo-opercular and subcortical systems. We hypothesized that these differences may be driven by the excess of maximal 8-cliques in the minimally wired network (Fig. 2a). Expanding on the difference in node participation (\(P_{k}^{DSI}(v) - P_{k}^{MW}(v)\)), we see that the large discrepancies between empirical and null model networks in cingulo-opercular and subcortical systems are caused by a difference in maximal cliques of approximately eight nodes (Fig. 2b, bottom). Finally, we observe that the systems involved in the two peaks of the maximal clique distribution shown in Fig. 2a differ greatly from one another. The first peak composed of smaller cliques involves regions from nearly all systems, while the second peak is almost exclusively composed of regions in the default mode, subcortical, and visual systems. We observe the largest cliques in the subcortical, default mode, dorsal attention, and visual systems, though only the visual and dorsal attention systems have maximal clique distributions with significantly higher means than the rest of the brain regions (p << 0.001, p < 0.05, respectively). These data suggest that small, local processors may be a common feature across systems, while larger cliques may allow for rapid multi-system cross-talk.

We next check that the building blocks, here k-cliques, behave consistently with more common graph theoretic metrics. A node with high participation in maximal cliques must in turn be well connected locally (though the converse is not necessarily true – consider a node that only participates in one maximal 16-clique). Therefore we expect the participation of a node to act similarly to other measures of connectivity. To test this expectation, we examine the correlation of node participation with node strength, the summed edge weight of connections emanating from a node, as well as with node communicability, a measure of the strength of long distance walks emanating from a node (Fig. 3a). We find that both strength and communicability exhibit a strong linear correlation with the participation of a node in maximal cliques (Pearson correlation coefficient r = 0.957 and r = 0.858, respectively).

These results indicate that regions that are strongly connected to the rest of the brain by both direct paths and indirect walks also participate in many maximal cliques. Such an observation suggests the possibility that brain hubs – which are known to be strongly connected with one another in a so-called rich-club – play a key role in maximal cliques. To test this, we measure the association of brain regions to the rich-club using notions of coreness. A k-core of a graph G is a maximal connected subgraph of G in which all vertices have degree at least k, and an s-core is the equivalent notion for weighted graphs (see Methods). Using these notions, we consider how the k-core and s-core decompositions align with high participation (Fig. 3b). In both cases, nodes with higher participation often achieve higher levels in the k- and s-core decomposition. Moreover, we also observe the frequent existence of rich club connections between nodes with high participation (Fig. 3b, bottom). Together, these results suggest that rich-club regions of the human brain tend to participate in local computational units in the form of cliques.

Whereas cliques in the DSI network act as neighborhood-scale building blocks for the computational structure of the brain, the relationships between these blocks can be investigated by studying the unexpected absence of strong connections, which can be detected as topological cavities in the structure of the brain network. Because connections are treated as communication channels along which brain regions can signal one another and participate in shared neural function, the absence of such connections implies a decreased capacity for communication which serves to enhance the segregation of different functions.

To identify topological cavities in a weighted network, we construct a sequence of binary graphs, each included in the next (Fig. 4a), known as a filtration. Beginning with the empty graph, we replace unweighted edges one at a time according to order of decreasing edge weight, and we index each graph by its edge density ρ, given by the number of edges in the graph divided by the number of possible edges. After each edge addition, we extract motifs of k-cliques called (non-trivial) (k − 1)-cycles, each of which encloses a k-dimensional topological cavity in the structure. This shift in index is due to geometry: a 2-clique is a 1-dimensional line segment, a 3-clique is a 2-dimensional triangle, etc. When k is clear or not pertinent, we will suppress it from the notation, and refer simply to “cycles” and “cavities”. While any cavity is surrounded by at least one cycle, often multiple cycles surround the same cavity. However, any two (k-1)-cycles that detect the same cavity will necessarily differ from one another by the boundaries of some collection of (k + 1)-cliques (see Supporting Information and Fig. 15). Any two such cycles are called topologically equivalent, so each topological cavity is detected by a non-trivial equivalence class of cycles. The equivalence class containing the cycle consisting of a single vertex is called trivial and bounds the “empty” cavity. We can represent a topological cavity using any of the cycles within the corresponding equivalence class, but for purposes of studying computational architectures it is reasonable to assume information will primarily travel along paths of minimal length; thus, in this analysis we will consider the collection of cycles in an equivalence class with the minimal number of nodes and call these the minimal cycles representing the cavity. Note in the absence of a filtration, there are serious computational issues involved in locating minimal-size representatives of equivalence classes. However, in this setting the computation is easily performed using standard algorithms (see Methods).

As we move through the filtration by adding edges, the structure of the cycles, and thus of the cavities they represent, will evolve. We consider an example in Fig. 4a, showing a green minimal cycle surrounding a 2D cavity which first appears (is born) in the graph sequence at ρ _{b i r t h} (cyan). As an edge completing a 3-clique is added, the minimal cycle representative shrinks to four nodes in size, then finally is tessellated by 3-cliques (dies) at ρ _{d e a t h} (orange). We record ρ _{b i r t h} and ρ _{d e a t h} for all topological cavities (e.g., non-trivial equivalence classes of cycles) found within the filtration, and display them on a persistence diagram (Fig. 4b). Cavities that survive many edge additions have a long lifetime, defined as ρ _{d e a t h} − ρ _{b i r t h} , or a large death-to-birth ratio, ρ _{d e a t h} /ρ _{b i r t h} . Such cycles are commonly referred to as persistent cavities and in many applications are considered the “topological features” of the system.

We investigate the persistence of 2D and 3D cavities (respectively represented by equivalence classes of 1- and 2-cycles) in the group-average DSI network and minimally wired null networks (see Fig. 4c). There are substantially fewer persistent cavities in the group-average DSI network than in the null models. To illustrate the structure of these cavities, we select four representative cavities with exceedingly long lifetimes or a high ρ _{d e a t h} to ρ _{b i r t h} ratio (Fig. 4c, d) in the empirical data, and for each we find the minimal-length representative cycles at ρ _{b i r t h} (Fig. 4e). Such cycles for all of the persistent cavities found in the empircal data are illustrated in Figs. 20 and 21. The first persistent cavity appears as early as ρ = 0.003and is minimally enclosed by the unique blue cycle composed of the thalamus and caudate nucleus of both hemispheres. The green cycle connecting the medial and lateral orbitofrontal, rostaral anterior cingulate, putamen, and superior frontal cortex is the only minimal cycle surrounding a long-lived cavity in the left hemisphere. The final persistent 2D cavity in the average DSI data is found in the right hemisphere between the medial orbitofrontal, accumbens nucleus, any of the subcortical regions hippocampus, caudate nucleus, putamen, thalamus, and amygdala, and any of the rostral middle frontal, lateral orbitofrontal, medial orbitofrontal of the left hemisphere, and rostral anterior cingulate from both hemispheres (see Fig. 4e for all 12 minimal representatives). Finally, the purple octahedral cycle made from 3-cliques contains the inferior and middle temporal, lateral occipital, inferior parietal, supramarginal, superior parietal, and either of the superior temporal and insula of the left hemisphere, and encloses the longest-lived 3D cavity in the structural brain network. Though each minimal generator may have distinct biological implications, we observe a global pattern of subcortical–cortical connections within cycles. Indeed, 18 of the 20 recovered 1-cycles and both 2-cycles contain this motif. Additionally, the two persistent cycles that do not follow this motif comprise a third of persistent cycles robustly seen in the minimally wired network, suggesting that within-subcortical loops are more probable in this maximally efficient scheme.

It is important to ask whether the architectural features that we observe in the group-averaged DSI network can also be consistently observed across multiple individuals, and across multiple scans of the same individual to ensure these cavities are not artifacts driven by a few outliers. Comparison of persistent cavities arising from two different networks is complicated by our notion of equivalence of cavities, and our desire to work with particular representative cycles. To capture the extent to which the cavities and their minimal representatives in the average DSI data are present in the individual scans, we record the collection of cliques that compose each minimal cycle representing the equivalance class (as seen in Fig. 4e), and check both for the existence of one of those collections of cliques, corresponding to the existence of the same strong fiber tracts, and, more stringently, for the presence of a topological cavity represented by that cycle in each individual’s DSI network (see Supporting Information for more details). We observed that the subcortical cycle (Fig. 4e, blue) exists and these nodes (thalamus and caudate nucleus of both hemispheres) surround an equivalent 2D cavity in at least one scan of all individuals and the late-developing subcortical-frontal cycle (Fig. 4e, red) surrounds a cavity found in seven of the eight individuals in at least one of three scans (Fig. 5b, f). The earlier arriving subcortical-frontal cycle (Fig. 4e, green) is present in all individuals and a similar cavity is seen at least once in all individuals (Fig. 5d). Finally, we observe that the octahedral connection pattern in posterior parietal and occipital cortex (Fig. 4e, purple) is present at least once in seven of eight individuals and these regions enclose a similar cavity at least once in six of these individuals (Fig. 5h). In the opposite hemisphere, the cyclic connection patterns and similar cavities appear though not as regularly (Fig. 5). Finally we check the existence of similar cavities within the minimally wired null models, and see cavities denoted by the green and purple cycles are never seen (Fig. 5). However, similar cavities to those represented by the red and blue minimal cycles appear frequently in the null model, though with different birth/death densities and lifetimes. In summary we find topological cavities observed in the group-averaged DSI network appear consistently across individuals, suggesting their potential role as conserved wiring motifs in the human brain.

In addition to consistency across subjects and scans, it is important to determine whether the known high connectivity from subcortical nodes to the rest of the brain may be artificially obscuring non-trivial cortico-cortical cavities important for brain function. To address this question, we examined the 66-node group-average DSI network composed only of cortical regions, after removing subcortical regions, insula, and brainstem. We recovered a long-lived topological cavity surrounded by four cycles of minimal length composed of nine nodes connecting temporal, parietal, and frontal regions (Fig. 6). Note in the schematic of Fig. 6a we see clearly two 2D cavities. The birth edge here was between the lateral orbitofrontal and superior temporal regions, which prevents us from determining whether the exact minimal cycle surrounding this cavity follows the superior frontal (LH)/posterior cingulate or the superior frontal (RH)/caudal middle frontal branch of the top loop. Following either of these two branches (then either of the banks of the superior temporal sulcus or middle temporal route) gives four cycles in which two are equivalent to each other but not to either cycle in the other pair. We will accept all of these four as minimal maroon cycles since any of the four could be minimal representatives. Moreover, at least one of these minimal cycles and corresponding cavity was observed in each scan of every individual (Fig. 26c), and often in the opposite hemisphere as well (Fig. 26d). These results reveal that cortico-cortical cycles are indeed present and suggest their potential utility in segregating function across the brain.

Algebraic topology is a relatively young field of pure mathematics that has only recently been applied to the study of real-world data. However, the power of these techniques to measure structures that are inaccessible to common graph metrics has gained immediate traction in the neuroscience community. Here, we highlight a few notable examples from the growing literature; a more comprehensive recent account can be found in Giusti et al. (2016). At the neuron level, persistent has been used to detect intrinsic structure in correlations between neural spike trains (Giusti et al. 2015), expanding our understanding of the formation of spatial maps in the hippocampus (Dabaghian et al. 2012). Moreover, at the level of large-scale brain regions, these tools have been exercised to characterize the global architecture of fMRI data (Stolz 2014). Based on their unique sensitivity, we expect these algebric-topological methods to provide novel contributions to our understanding of the structure and function of neural circuitry across all scales at which combinatorial components act together for a common goal: from firing patterns coding for memory (Rajan et al. 2016; Leen and Shea-Brown 2015) to brain regions interacting to enable cognition.

Our study uses algebraic topology in the classical form to obtain a global understanding of the structure, and in conjunction, it investigates particular topological features themselves and relates these features to cognitive function. Cycle representatives have previously been considered in biology (Chan et al. 2013; Petri et al. 2014; Lord et al. 2016; Kim et al. 2014; Emmett et al. 2016; Mamuye et al. 2016), but to our knowledge this is a first attempt to compare topological features in multiple brains.

Cliques and minimal cycles representing cavities are structurally positioned to play distinct roles in neural computations. Cliques represent sets of brain regions that may possess a similar function, operate in unison, or share information rapidly (Sizemore et al. 2016). Furthermore, the hierarchical organization of small cliques located more anteriorly and larger cliques connecting multiple systems allows for swift global sharing of information produced by local processing. Conversely, minimal cycles correspond to extended paths of potential information transmission along which computations can be performed serially to affect cognition in either a divergent or convergent manner. Indeed, the capsule-like or chain-like nature of cycles is a structural motif that has previously been – at least qualitatively – described in neuroanatomical studies of cellular circuitry. In this context, such motifs are known to play a key role in learning (Hermundstad et al. 2011), memory (Rajan et al. 2016), and behavioral control (Levy et al. 2001; Fiete et al. 2010). The presence of cycles suggests a possible role for polysynaptic connections and their importance to neural computations, consistent with evidence from the field of computational neuroscience highlighting the role of highly structured circuits in sequence generation and memory (Rajan et al. 2016; Hermundstad et al. 2011). Indeed, in computational models at the neuron level, architectures reminiscent of chains (Levy et al. 2001; Fiete et al. 2010) and rings are particularly conducive to the generation of sequential behavioral responses. It is interesting to speculate that the presence of these structures at the much larger scale of white matter tracts could support diverse neural dynamics and a broader repertoire of cognitive computations than possible in simpler and more integrated network architectures (Tang et al. 2016).

Another consideration concerns the apparent asymmetry of our results with respect to left and right cerebral hemispheres. While unanticipated, we note that in some cases they have intuitive mathematical underpinnings. For example, in Fig. 3, we explicitly count maximal cliques, so one edge difference between a region in the left and right hemisphere could result in a large difference in the number of observed maximal cliques. Interestingly, despite this fact we still observe a strong correlation between node strength and P(v), instilling confidence in these results. From a neuroscience point of view, brain asymmetries are not wholly unexpected. There is a storied and ever-growing literature describing the lateralization (i.e., asymmetries) of brain function (Galaburda et al. 1978). While speech generation (Rasmussen and Milner 1977) and language processing (Desmond et al. 1995; Thulborn et al. 1999) are among the most commonly-cited functions to exhibit lateralization (Doron et al. 2012; Chai et al. 2016), such effects have also been linked to a diverse group of other cognitive domains. These include emotion (Wager et al. 2003), processing of visual input (Sandi et al. 1993), and even working memory (Carpenter et al. 2000). In addition, a number of studies have also reported the emergence of pathological lateralization or the disruption of asymmetries with neurocognitive disorders including ADHD (Oades 1998). Our study does not offer a conclusive demonstration that the observed asymmetries arise from the lateralization of any specific brain function; we merely wish note that there is a precedent for such observations.

Network filtration revealed several persistent cavities in the macroscale human connectome. While each minimal cycle surrounding these cavities involved brain regions interacting in a distinct configuration, we also observed commonalities across these structures. One such commonality was these minimal cycles tended to link evolutionarily old structures with more recently-developed neo-cortical regions (Rakic 2009). For example, the green cycle depicted in Fig. 4e linked the putamen, an area involved in motor behavior (Middleton and Strick 2000), with the rostral anterior cingulate cortex, associated with higher-order cognitive functions such as error-monitoring (Braver et al. 2001) and reward processing (Kringelbach and Rolls 2004). This observation led us to speculate that the emergence of these cavities may reflect the disparate timescales over which brain regions and their circuitry have evolved (Gu et al. 2015b), through the relative paucity of direct connections between regions that evolved to perform different functions. This hypothesis can be investigated in future work comparing the clique and cavity structure of the human connectome with that of non-human connectomes from organisms with less developed neocortices.

Though we highlighted minimal cycles in the brain, by nature persistence describes the global organization of the network. Often regions in the brain wire minimally to conserve wiring cost (Bassett et al. 2010; Bullmore and Sporns 2012; Klimm et al. 2014; Lohse et al. 2014), though there are exceptions that give the brain its topological properties such as its small-world architecture (Bassett and Bullmore 2006; Pessoa 2014; Hilgetag and Goulas 2016; Muldoon et al. 2016a; Bassett and Bullmore 2016). Following this idea, we could interpret the difference in the number of persistent cavities between the minimally wired and DSI networks as a consequence of the non-minimally wired edges, which tessellate cavities in the brain itself. Yet when the subcortical regions are removed, the persistent cavities of the minimally wired and DSI networks are much more similar (Fig. 6b). This suggests that the wiring of cortical regions may be more heavily influenced by energy conservation than the wiring of subcortical regions. Additionally the drop in the number and lifetime of persistent cavities when subcortical regions are included indicates that these subcortical regions may prematurely collapse topological cavities. The often high participation of subcortical regions in maximal cliques suggests these well-connected nodes may have hub-like projections to regions involved in cortical cycles, thus tessellating the cortical cavity with higher dimensional cliques (topologically these subcortical nodes are cone points). Previous studies have found that networks with “star-like” configurations are optimally efficient in terms of shortest-path efficiency, but also efficient in terms of a random walk-based measure of efficiency (Goni et al. 2013). That is, networks optimized to have one or the other type of efficiency tend to have stars. Thus, stars appear to be useful configurations for fast communication, both along shortest paths and also in an unguided sense along random walks. The fact that we see star-like projections to cycles from subcortical regions may suggest that they are useful for efficient communication.

An important consideration relates to the data from which we construct the human structural connectome. DSI and tractography, non-invasive tools for mapping the brain’s white-matter connectivity, have some limitations. Tractography algorithms trade off specificity and sensitivity, making it challenging to simultaneously detect true connections while avoiding false connections (Thomas et al. 2014), fail to detect superficial connections (i.e. those that do not pass through deep white matter) (Reveley et al. 2015), and have challenges tracking “crossing fibers”, connections with different orientations that pass through the same voxel (Wedeen et al. 2008). Nonetheless, DSI and tractography represent the only techniques for non-invasive imaging and reconstruction of the human connectome. While such shortcomings limit the applicability of DSI and tractography, they may prove addressable with the development of improved tractography algorithms and imaging techniques (Pestilli et al. 2014).

Though comparing persistent homology of weighted networks at the global level has been successful (for example Benzekry et al. 2015; Horak et al. 2009), scrutinizing individual persistent features may have more clinical relevance due to their size and understandability. Yet, multiple questions remain to be answered before this goal can be achieved.

The first question pertains to the choice of representative cycle. As the current study presents an initial consideration of the persistent features of the structural connectome, we record all minimal generators, which reduces the number of choices made, and we define minimality using topological (hop) distance, which simplifies our analysis. However, a case could be made for using the representative with the minimal summed edge weight (Dey et al. 2011). Such a definition would further simplify the analysis by potentially giving a unique ‘minimal’ generator for each equivalence class. Additionally one might ask if a ‘minimal’ generator is even the appropriate representative cycle in the first place. Perhaps cycles of longer length have cognitive or clinical relevance beyond information distribution.

Second, it will be necessary to further develop the concept of similar persistent cavities. Here we used a region-matching process in order to incorporate perspectives from neuroscience and topology. An important open question is whether a more algorithmic matching could be devised that is better suited to the perspectives from both fields. Along the same lines, it is important to consider the birth, death time, and lifetime of a given persistent cycle (Stolz et al. 2017). We interpret longer-lived and earlier-born persistent cycles as more essential to the global architecture, and we hypothesize that this translates to healthy cognitive control and function as well. Then if two cavities are similar in terms of their regional composition, but are not similar in terms of birth or death times (for example, the blue cycle in the DSI versus MW networks in Fig. 5), it remains an open question whether the two cavities should be considered truly similar in a biological context.

Thirdly, with the development of algebraic-topological tools as described above, we speculate that comparing late-arriving persistent features could be important for clinical applications. Weaker connections have been shown to distinguish between health individuals and those with schizophrenia (Bassett et al. 2012), and have also been shown to predict individual differences in intelligence (Cole et al. 2012). Since late-born persistent cycles are a very particular arrangement of weak edges, we hypothesize that such cavities may be powerful biomakers of individual brains, capable of distinguishing between diseased and normal connectomes.