A connectomicsbased taxonomy of mammals
Abstract
Mammalian taxonomies are conventionally defined by morphological traits and genetics. How species differ in terms of neural circuits and whether interspecies differences in neural circuit organization conform to these taxonomies is unknown. The main obstacle to the comparison of neural architectures has been differences in network reconstruction techniques, yielding speciesspecific connectomes that are not directly comparable to one another. Here, we comprehensively chart connectome organization across the mammalian phylogenetic spectrum using a common reconstruction protocol. We analyse the mammalian MRI (MaMI) data set, a database that encompasses highresolution ex vivo structural and diffusion MRI scans of 124 species across 12 taxonomic orders and 5 superorders, collected using a unified MRI protocol. We assess similarity between species connectomes using two methods: similarity of Laplacian eigenspectra and similarity of multiscale topological features. We find greater interspecies similarities among species within the same taxonomic order, suggesting that connectome organization reflects established taxonomic relationships defined by morphology and genetics. While all connectomes retain hallmark global features and relative proportions of connection classes, interspecies variation is driven by local regional connectivity profiles. By encoding connectomes into a common frame of reference, these findings establish a foundation for investigating how neural circuits change over phylogeny, forging a link from genes to circuits to behaviour.
Editor's evaluation
This important article uses an impressively rich data set (obtained and curated by the authors) to compare the structural brain connectomes of many animals spanning six taxonomic orders. The approach is innovative and relies on graph theoretical measures to describe the connectivity, which means it can be done without the need to spatially/functionally match the brains. The authors find compelling evidence that there is more variability between than within order. They attribute this effect to changes in local connectivity features, whereas global patterns are preserved. The approach can potentially be a useful way to study phylogeny and brain evolution.
https://doi.org/10.7554/eLife.78635.sa0Introduction
Anatomical projections between brain regions form a complex network of polyfunctional neural circuits (Sporns et al., 2005). Signalling on the brain’s connectome is thought to support cognition and the emergence of adaptive behaviour. Advances in imaging technologies have made it increasingly feasible to reconstruct the wiring diagram of biological neural networks. Thanks to extensive international datasharing efforts, these detailed reconstructions of the nervous system’s connection patterns have been made available in humans and in multiple model organisms (van den Heuvel et al., 2016), including invertebrate (White et al., 1986; Chiang et al., 2011; Towlson et al., 2013; Worrell et al., 2017), avian (Shanahan et al., 2013), rodent (Oh et al., 2014; Bota et al., 2015; Rubinov et al., 2015), feline (Scannell et al., 1995; de Reus and de Reus and van den Heuvel, 2013; Beul et al., 2015), and primate species (Markov et al., 2012; Majka et al., 2016; Liu et al., 2020).
The rising availability of connectomics data facilitates crossspecies comparative studies that identify commonalities in brain network topology and universal principles of connectome evolution (van den Heuvel et al., 2016; Barker, 2021; Barsotti et al., 2021). A common thread throughout these studies is the existence of nonrandom topological attributes that theoretically enhance the capacity for information processing (Sporns, 2013). These include a highly clustered architecture with segregated modules that promote specialized information processing (Watts and Strogatz, 1998; Hilgetag and Kaiser, 2004), as well as a densely interconnected core of highdegree hubs that shortens communication pathways (van den Heuvel et al., 2012), promoting the integration of information from distributed specialized domains (ZamoraLópez et al., 2010; AvenaKoenigsberger et al., 2017). These universal organizational features suggest that connectome evolution has been shaped by two opposing and competitive pressures: maintaining efficient communication while minimizing neural resources used for connectivity (Bullmore and Sporns, 2012).
While comparative analysis can focus on commonalities among mammalian connectomes and identify universal wiring principles, it can also be used to systematically explore differences among connectomes that confer specific adaptive advantages. Indeed, despite commonalities, architectural variations are also observed even among closely related species (Barker, 2021). Factors such as the external environment, genetics. and distinct gene expression programs also account for diversity in neural connectivity patterns (Martinez and Sprecher, 2020). Subtle variations in connectome organization may potentially account for speciesspecific adaptations in behaviour and cognitive function.
But how does the connectome vary over phylogeny? Traditionally, mammalian taxonomies were built on morphological differences among species (Darwin, 1959). Besides physical commonalities, species within the same taxonomic group also tend to share similar behavioural repertoires (York, 2018; Bendesky and Bargmann, 2011; Yokoyama et al., 2021). Modern highthroughput wholegenome sequencing has further delineated phylogenetic links and relationships among mammalian species (Murphy et al., 2021; Zoonomia Consortium, 2020; ÁlvarezCarretero et al., 2021; Seehausen et al., 2014). In addition to refining the overall classification of mammals, wholegenome comparative analyses have established the genetic basis of phenotypic variation across phylogeny (Murphy et al., 2021). Whether interspecies differences in the organization of connectome wiring conform to this taxonomy remains unknown. How do genes sculpt behaviour across evolution? Could speciation events in the genome leading to variations in connectome architecture be the missing link between genomics and behaviour? Rigorously addressing these questions is challenging due to the lack of methodological consistency in the acquisition and reconstruction of neural circuits, or the limited number of available species.
Here, we comprehensively chart connectome organization across the mammalian phylogenetic spectrum. We analyse the mammalian MRI (MaMI) data set, a comprehensive database that encompasses highresolution ex vivo diffusion and structural (T1 and T2weighted) MRI scans of 124 species (a total of 225 scans including replicas) (Assaf et al., 2020). All images were acquired using the same scanner and protocol. All connectomes were reconstructed using a uniform parcellation scheme consisting of 200 brain areas, including cortical and subcortical regions. Notably, the MaMI data set spans a wide range of categories across different taxa levels of morphological and phylogenetic mammalian taxonomies (Assaf et al., 2020). Specifically, it includes animal species across 5 different superorders (Afrotheria, Euarchontoglires, Laurasiatheria, Xenarthra, and Marsupialia) and 12 different orders (Cetartiodactyla, Carnivora, Chiroptera, Eulipotyphla, Hyracoidea, Lagomorpha, Marsupialia, Perissodactyla, Primates, Rodentia, Scandentia, and Xenarthra).
Taking advantage of the harmonized imaging and reconstruction protocols, we quantitatively assess the similarity of species’ connectomes to construct datadriven phylogenetic relationships based on brain wiring. We compare these interspecies wiring similarities with conventional morphologically and genetically defined mammalian taxonomies. We determine the extent to which connectome topology conforms to established taxonomic classes, and identify network features that are associated with speciation.
Results
The MaMI data set consists of highresolution ex vivo diffusion and structural (T1 and T2weighted) MRI scans of 124 species. Since there is no speciesspecific template, all connectomes were reconstructed using a uniformly applied 200node parcellation. Having equally sized networks facilitates graph comparison but also implies a lack of direct correspondence between nodes across species. However, because our focus is on the statistics of connectomes’ topology, this does not impact our analyses. As the size of the network is kept constant across all species, voxel size is normalized to brain volume. Figure 1 shows the distribution of connectomes across 10 mammalian orders (out of the 12 present in the data set). We focus on the 6 orders that contain 5 or more distinct species (within the Laurasiatheria and Euarchontoglires superorders); these include Chiroptera, Rodentia, Cetartiodactyla, Carnivora, Perissodactyla, and Primates, resulting in a total of 111 different animal species and 203 brain scans. A complete list of the animal species included in the data set is provided in Figure 1—figure supplement 1.
Connectomebased interspecies distances
Similarity between species’ network architectures is estimated using two networkbased distance metrics: spectral distance, based on the eigenspectrum of the normalized Laplacian of the connectivity matrix (see Figure 2—figure supplement 1; de Lange et al., 2014), and topological distance, based on a combination of multiscale graph features of the binary and weighted connectivity matrices (Figure 2—figure supplements 2 and 3 show the distribution of individual local and global graph features, respectively; Rubinov and Sporns, 2010). For completeness, Figure 2—figure supplements 4 and 5 show the cumulative distribution of binary and weighted local features, respectively, for individual species. Both methods measure how similar the architectures of two connectomes are. To identify brain connectivity differences across species, we need to be able to analyse data in a shared frame of reference. The normalized Laplacian eigenspectrum and the graph features of the connectivity matrix allow us to translate connectomes into a common feature space in which they are comparable, despite the fact that they come from different species, and that the nodes do not correspond to one another (Mars et al., 2021). To account for the fact that some of the species have more than one scan, we randomly select one sample per species and estimate (spectral and topological) interspecies distances. We repeat this procedure 10,000 times and report the average across iterations.
Figure 2a shows the spectral distances between species’ connectomes. In general, we observe smaller distances among members of the same order (outlined in yellow). Figure 2b confirms this intuition by showing that spectral distances within orders (i.e. values along the diagonal) tend to be smaller than distances between orders (i.e. values off the diagonal). Figure 2c shows the distributions of intra and interorder distances. The mean/median intraorder distance is significantly smaller than the mean/median interorder distance (twosample Welch’s ttest: mean intra and interorder distances are 0.43 and 0.55, respectively, $\mathrm{p}<{10}^{4}$ twotailed, and Cohen’s $d$ effect size = 0.67; twosample Mann–Whitney Utest: median intra and interorder distances are 0.44 and 0.55, respectively, $\mathrm{p}<{10}^{4}$ twotailed, and commonlanguage effect size = 68%; Figure 2c). We find comparable results when estimating species similarity using topological distance (twosample Welch’s ttest: mean intra and interorder distances are 0.41 and 0.53, respectively, $\mathrm{p}<{10}^{4}$ twotailed, and Cohen’s $d$ effect size = 0.59; twosample Mann–Whitney Utest: median intra and interorder distances are 0.41 and 0.53, respectively, $\mathrm{p}<{10}^{4}$ twotailed, and commonlanguage effect size = 66%; Figure 2d–f). Figure 2—figure supplement 6 shows the same results as in Figure 2, but using all samples including replicas (i.e. without random resampling). Altogether, results suggest that species with similar genetics, morphology, and behaviour tend to have similar connectome architecture. In other words, variations in connectome architecture reflect phylogeny.
Architectural features differentiate species
Next we consider which network features contribute to the differentiation (Figure 2—figure supplements 2 and 3 show the distributions of local and global graph features, respectively). To address this question, we recompute interspecies topological distances using different sets of graph features (Figure 3). We find that the difference between intra and interorder topological distances tends to be larger when only local (nodelevel) features are included in the estimation of the topological distance (i.e. degree, clustering coefficient, betweenness, and closeness; Figure 3b and e) compared to when only global features are considered (i.e. characteristic path length, transitivity, and assortativity; Figure 3c and f). This is the case for both the binary and weighted versions of these features (top and bottom rows in Figure 3, respectively). Figure 3—figure supplement 1 shows the same results as in Figure 3, but using all samples including replicas (i.e. without random resampling). These results suggest that differentiation of orders is better explained by differences in local network topology; conversely, global network topology appears to be conserved across species. An illustration of this principle is depicted in Figure 3—figure supplement 2 showing that the relative local connectivity of the anterior and the posterior ends of the cortex changes across taxonomic orders (Barrett et al., 2020; Krubitzer and Kaas, 2005; Krubitzer and Kahn, 2003).
A similar conclusion can be drawn when the eigenvalue distributions of the (normalized) Laplacian of the connectivity matrices are compared across species (Figure 2—figure supplement 1). In spectral graph theory, the presence of eigenvalues with high multiplicities or eigenvalues symmetric around ${\lambda}_{i}=1$ provides information about the network’s local organization that results from the recursive manipulation of connectivity motifs (Banerjee and Jost, 2008; Banerjee and Jost, 2009; de Lange et al., 2014). For instance, node duplication (i.e. the presence of nodes with the same connectivity profile) results in an increase of ${\lambda}_{i}=1$. The duplication of edge motifs (i.e. the multiple presence of pairs of connected nodes with the same connectivity profile), on the other hand, produces eigenvalues at equal distances to ${\lambda}_{i}=1$. Visually inspecting their Laplacian eigenspectra, one can notice that, across taxonomic orders, species tend to differ mostly around the interval $0.5\le {\lambda}_{i}\le 1.5$, both in terms of the multiplicity of ${\lambda}_{i}=1$, as well as in the width of the bellshaped curve around ${\lambda}_{i}=1$. While differences in the multiplicity of ${\lambda}_{i}=1$ indicate differential amounts of duplicated node motifs present in the network, differences in the value and multiplicity of eigenvalues around ${\lambda}_{i}=1$ indicate the presence of distinct edge motifs with disparate numbers of duplications in the network. Therefore, differences across taxonomic orders are most likely due to the presence of different local connectivity fingerprints in the connectivity matrix (Figure 2—figure supplement 1; de Lange et al., 2014; Mars et al., 2018a; Mars et al., 2018b). Determining which are specifically these node and edge motifs cannot be done by simply examining the Laplacian eigenspectra, and is out of the scope of this study. Additional evidence supporting the idea that spectral distance captures mostly differences in local network topology is the fact that the correlation between spectral and topological distance is maximum when only local features are included in the estimation of the topological distance (Figure 3—figure supplement 3).
We also observe that that the difference between intra and interorder topological distances is greater for binary than for weighted features (Figure 3a–c and d–f, respectively), independently of being local or global. This suggests that the strength of the connections is less important than the binary architecture of the connectivity matrix.
Some of the features used for the estimation of the topological distance depend on network density, which varies across taxonomic orders (Figure 3—figure supplement 4). To determine whether the observed differences between intra and interorder distances are above and beyond differences due to network density, we perform the same analysis shown in Figure 3, after controlling for density (Figure 3—figure supplement 5). Results, shown in Figure 3—figure supplement 6, suggest that differences between intra and interorder topological distances are not driven by differences in network density, but variations in wiring patterns, as captured by topological features, play a role in the observed phylogenetic variations in connectome organization.
Altogether, our results show that the subset of features that best differentiate species across taxonomic orders are the binary local topological features. We perform a set of complementary analyses to assess which subset of features produces the best partition of animal species relative to traditional taxonomies. To do so, we (1) project the data on a 2D plane using multidimensional scaling (Figure 3—figure supplement 7) and (2) apply hierarchical clustering to interspecies distance matrices (Figure 3—figure supplement 8). Visual inspection of these results suggests that, consistent with our previous results (Figure 3), local features compared to global features (ignoring panel a, centre vs. right column, respectively, in Figure 3—figure supplements 7 and 8), as well as binary features compared to weighted features (ignoring panel a, centre vs. bottom row, respectively, in Figure 3—figure supplements 7 and 8), yield species partitions that more closely reflect established phylogenetic relationships, further supporting the idea that connectome organization recapitulates traditional taxonomic relationships that are based on morphology and genetics.
Conservation of smallworld architecture
Anatomical brain networks are thought to simultaneously reconcile the opposing demands of functional integration and segregation by combining the presence of functionally specialized clusters with short polysynaptic communication pathways (Tononi et al., 1994; Sporns, 2013; Sporns et al., 2005; Bassett and Bullmore, 2006). Such architecture is often referred to as smallworld and is observed in a wide variety of naturally occurring and engineered networks (Watts and Strogatz, 1998). Here, we explore whether these principles of segregation and integration in global connectome organization are consistent across phylogeny. To do so, we estimate for each species the ratio of clustering coefficient to characteristic path length, normalized relative to a set of randomly rewired graphs that preserve the degree sequence of the nodes (Humphries and Gurney, 2008; Maslov and Sneppen, 2002; Rubinov and Sporns, 2010; Figure 4). Consistent with previous reports in individual species’ connectomes (Hilgetag and Kaiser, 2004; Sporns and Zwi, 2004; Bassett and Bullmore, 2006), we find that all connectomes display high and diverse levels of smallworldness, suggesting that simultaneously highly segregated and integrated networks is a global trait conserved across mammalian brains.
Conservation of edge classes across species
The topological and spatial arrangement of connections in connectomes is thought to shape the segregation and integration of information and, ultimately, their computational capacity (Faskowitz et al., 2021). To investigate interspecies differences in the topological and spatial distribution of connections, we stratify edges into different classes in four commonly studied partitions. Partitions include inter and intramodular connections (Figure 5a), inter and intrahemispheric connections (Figure 5b), connection length distribution (short, medium, and longrange connections; Figure 5c and Figure 5—figure supplement 1), and richclub (richclub, feeder and peripheral connections; Figure 5d). Overall, we find that, along the four partitions, the relative proportions of each connection class are conserved across taxonomic orders, despite differences in connection density. Collectively, this is consistent with the results from the previous sections showing that global architectural features of connectomes are consistent across phylogeny.
Discussion
In this study, we chart the organization of wholebrain neural circuits across 111 mammalian species and 5 superorders. We find that connectome organization recapitulates to a large extent traditional taxonomies. While all connectomes retain hallmark global features and relative proportions of edge classes, interspecies variation is driven by local regional connection profiles.
Conventional mammalian taxonomies are delineated based on the concept that a species is a group of organisms that can reproduce naturally with one another and create fertile offspring (Mallet, 1995). As a result, classical taxonomies based on animal morphology have largely been reconciled with emerging evidence from wholegenome sequencing Baker and Bradley, 2006; namely, organisms with similar genomes display similar physical characteristics and behaviour. Our work shows that interspecies similarity – as defined by morphology, behaviour, and genetics – is concomitant with the organization of neural circuits. Specifically, species that are part of the same taxonomic order tend to display similar connectome architecture, suggesting that brain network organization is under selection pressure (see also Butler and King, 2004; Lande, 1976; Wright, 1931 for an alternative mechanism of trait evolution characterized by pure drift models based on Brownian motion), analogous to size, weight, or colour.
Which network features drive differences across taxonomic orders? Interestingly, all connectomes display consistent global hallmarks that were previously documented in tracttracing studies, including high clustering and nearminimal path length characteristic of smallworld organization, as well as segregated network communities and densely interconnected hub nodes (van den Heuvel et al., 2016). The conservation in global wiring and organizational principles is further supported by a reduced difference between intra and interorder topological distances estimated exclusively from global features compared to the case in which only local features are considered. Thus, relative differences between connectomes across taxonomic orders are mainly driven by local regional features. These results are in line with the idea that a brain region’s functional fingerprint – the specific computation or function that it performs by virtue of its unique firing patterns and dynamics – is determined by its underlying corticocortical connectional fingerprint (Mars et al., 2021; Mars et al., 2018b; Mars et al., 2016; Passingham et al., 2002). Accordingly, interspecies differences in functional and behavioural repertoire are likely supported by changes in local connectivity patterns. Along the same lines, our results are also consistent with the notion that neural circuit evolution involves random local circuit modifications that may have provided species with behavioural adaptations, allowing them to face specific challenges (Barker, 2021), such as extreme environmental pressures Park et al., 2008; Smith et al., 2011; Eigenbrod et al., 2019, or to support specific behaviours, such as courtship (O’Grady and DeSalle, 2018; Markow and O’Grady, 2005; Ding et al., 2019; Seeholzer et al., 2018; Khallaf et al., 2020; Barkan et al., 2018; Ding et al., 2016; York et al., 2019), social bonding (Insel and Shapiro, 1992; Winslow et al., 1993; Jaggard et al., 2020; Loomis et al., 2019), or foraging (Vanwalleghem et al., 2018; Pantoja et al., 2020). How computations and cognitive functions emerge from these speciesspecific circuit modifications remains a key question in the field (Buckner and Krienen, 2013; Suárez et al., 2021).
These results highlight the importance of developing speciesspecific, anatomicalbased parcellations, as well as new ways to align connectomes from different species. Understanding how homologous regions correspond to one another will allow further investigation of regional interspecies differences in connectome topology, which is a fundamental step for advancing comparative connectomics. A variety of emerging methods are contributing to further resolve correspondence between brain regions across species, facilitating finegrained comparisons at the level of individual regions. These methods implement regional comparisons based on different data modalities including measures of cytoarchitecture (J Garey, 1999; Bianchi et al., 2013), receptor distribution (Levant, 1998), functional and structural connectivity fingerprints (Passingham et al., 2002; Mars et al., 2016), patterns of gene expression (Warrington et al., 2022; Beauchamp et al., 2022), and macroscale gradients of functional activation (Buckner and Margulies, 2019).
It is noteworthy that the relative proportion of edge classes (inter vs. intramodular, inter vs. intrahemispheric, short vs. medium vs. longrange and richclub vs. feeder vs. peripheral) are preserved across species. This result is reminiscent of recent work on allometric scaling that investigates how white matter connectivity scales with brain size (Bullmore and Sporns, 2012). For example, species with fewer commissural interhemispheric connections exhibit lower hemispheric mean shortest path (i.e. stronger intrahemispheric connectivity), suggesting a similar connectivity conservation principle (Assaf et al., 2020). Likewise, using diffusionweighted MRI data across 14 different primate species, another study reported negative allometric scaling of cortical surface area with white matter volume and corpus callosum crosssectional area (Ardesch et al., 2021). This scaling results in less space for white matter connectivity with increasing brain size, translating into larger brains with a relatively higher proportion of shortrange connections than longrange connections when compared with smaller brains (Ardesch et al., 2021). These results, however, do not contradict studies showing a positive allometric scaling between white matter and grey matter volume (Zhang and Sejnowski, 2000; Theunissen, 1988; Schlenska, 1974; Frahm et al., 1982); while the proportion of total volume devoted to cerebral white matter is higher in larger brains, it does not keep pace with the rapid cortical expansion that occurs with larger brain size. Collectively these studies highlight that the architecture of neural circuits and their physical embedding are intertwined, and the distribution of connections is such that it retains consistent global architectural features across phylogeny.
The present results contribute to the emerging field of comparative connectomics (van den Heuvel et al., 2016; Mars et al., 2021; Barker, 2021; Tendler et al., 2021). Adopting a harmonized imaging protocol in a large number of mammalian species facilitates a rigorous quantitative comparison of neural circuits. Central to this are network analytic methods that map connectomes to a common space and quantify similarities across local and global levels of organization (de Lange et al., 2014; de Lange et al., 2016; Bassett and Sporns, 2017; Mars et al., 2018b, Mars et al., 2021; Warrington et al., 2022). By comprehensively charting taxonomies of connectome architectures, we may uncover the principles that govern the wiring of neural circuits (AvenaKoenigsberger et al., 2015). In particular, quantitative analysis of connectome architecture across phylogeny may help to link genomics and behaviour (Mišić and Sporns, 2016). Traditionally, taxonomic groups were defined in terms of physical morphology and behavioural repertoire (Burke, 1968), but these are now understood to be driven by speciation events in the genome (Murphy et al., 2021; Zoonomia Consortium, 2020; ÁlvarezCarretero et al., 2021). However, we do not yet understand how genes influence neural circuit architecture, which in turn shapes the behavioural repertoire of an organism. By understanding how neural circuits change over phylogeny, we can fill this gap and forge a link from genes to circuits to behaviour. Ultimately, the confluence of genomics, connectomics, and behaviour may help to triangulate towards a more wellrounded view of speciation (HernándezHernández et al., 2021), and their simultaneous investigation can further illuminate the link between structure and function in brain networks (Bassett et al., 2010; Stiso and Bassett, 2018; Suárez et al., 2020).
This work must be considered with respect to multiple limitations. First, uniformly parcellating brains of different size into the same number of nodes facilitates comparison of network architecture, but potentially obscures biologically important regional differences. Second, many species are represented by a single individual. Although we focus on orders rather than individual species and there is high withinspecies reliability (Assaf et al., 2020), the analyses do not capture individual variability within species. Third, all connectomes are reconstructed using diffusionweighted imaging, which is subject to both systematic false positives and false negatives (MaierHein et al., 2017; Schilling et al., 2019). While a uniform, highresolution ex vivo scanning protocol allows for systematic comparisons among species, and our results recapitulate findings from tracttracing studies, future work in comparative connectomics will benefit from technological and analytical advances in neural circuit mapping. Fourth, evolutionary circuit modifications may not occur at the level of largescale white matter, but at finer scales involving smaller nuclei or physiological events not accessible by diffusion imaging, such as up or downregulation of neurotransmitter receptors (Barker, 2021). Nevertheless, the strikingly consistent taxonomic and phylogenetic relationships revealed by connectome analysis remain and suggest that macroscale connectivity, as measured by diffusion MRI, is informative of species similarities and differences across taxonomic orders.
By encoding connectomes into a common frame of reference, we quantitatively assess neural circuit architecture across the mammalian phylogeny. We find that connectome organization recapitulates previously established taxonomic relationships. Collectively, these findings set the stage for future mechanistic studies to trace the link between genes to neural circuits and ultimately to behaviour, and offer new opportunities to explore how changes in brain network structure across phylogeny translate into changes in function and behaviour.
Materials and methods
Brain samples
Request a detailed protocolThe MaMI database includes a total of 225 ex vivo diffusion and T2 and T1weighted brain scans of 125 different animal species (Figure 1—figure supplement 1). No animals were deliberately euthanized for this study. All brains were collected based on incidental death of animals in zoos in Israel or natural death collected abroad, and with the permission of the national park authority (approval no. 2012/38645) or its equivalent in the relevant countries. All scans were performed on excised and fixated tissue. Animals’ brains were extracted within 24 hr of death and placed in formaldehyde (10%) for a fixation period of a few days to a few weeks (depending on the brain size). Approximately 24 hr before the MRI scanning session, the brains were placed in phosphatebuffered saline for rehydration. Given the limited size of the bore, small brains were scanned using a 7T 30/70 BioSpec Avance Bruker system, whereas larger brains were scanned using a 3T Siemens Prisma system. To minimize image artefacts caused by magnet susceptibility effects, the brains were immersed in fluorinated oil (Flourinert, 3M) inside a plastic bag during the MRI scanning session.
MRI acquisition
Request a detailed protocolA unified MRI protocol was implemented for all species. The protocol included highresolution anatomical scans (T2 or T1weighted MRI), which were used as an anatomical reference, and diffusion MR scans. Diffusion MRI data were acquired using high angular resolution diffusion imaging (HARDI), which consists of a series of diffusionweighted, spinecho, echoplanarimaging images covering the whole brain, scanned in either 60 (in the 7T scanner) or 64 gradient directions (in the 3T scanner) with an additional three nondiffusionweighted images (B0). The b value was 1000 smm^{2} in all scans. In the 7T scans, TR was longer than 12,000 ms (depending on the number of slices), TE was 20 ms, and Δ/δ = 10/4.5 ms. In the 3T scans, TR was 3500 ms, with a TE of 47 ms and Δ/δ = 17/23 ms.
To linearly scale according to brain size the twodimensional image pixel resolution (per slice), the size of the matrix remained constant across all species (128 × 96). Due to differences in brain shape, the number of slices varied between 46 and 68. Likewise, the number of scan repetitions and the acquisition time were different for each species, depending on brain size and desired signaltonoise ratio (SNR) levels. To keep SNR levels above 20, an acquisition time of 48 hr was used for small brains (∼0.15 ml) and 25 min for large brains (>1000 ml). SNR was defined as the ratio of mean signal strength to the standard deviation of the noise (an area in the nonbrain part of the image). Full details are provided in Assaf et al., 2020.
Connectome reconstruction
Request a detailed protocolThe ExploreDTI software was used for diffusion analysis and tractography (Leemans et al., 2009). The following steps were used to reconstruct fibre tracts:
To reduce noise and smooth the data, anisotropic smoothing with a 3pixel Gaussian kernel was applied.
Motion, susceptibility, and eddy current distortions were corrected in the native space of the HARDI acquisition.
A spherical harmonic deconvolution approach was used to generate fibreorientation density functions per pixel (Tournier et al., 2004), yielding multiple (n ≥ 1) fibre orientations per voxel. Spherical harmonics of up to fourth order were used.
Wholebrain tractography was performed using a constrained spherical deconvolution (CSD) seed point threshold similar for all samples (0.2) and a step length half the pixel size.
The end result of the tractography analysis is a list of streamlines starting and ending between pairs of voxels. Recent studies have shown that fibre tracking tends to present a bias where the vast majority of end points reside in the white matter (Tournier et al., 2004). To avoid this, the CSD tracking implemented here ensures that approximately 90% of the end points reside in the cortical and subcortical grey matter.
Network generation and analysis
Request a detailed protocolBefore the reconstruction of the networks, certain fibre tracts were removed from the final list of tracts. These include external projection fibres that pass through the cerebral peduncle, as well as cerebellar connections. Innerhemispheric projections, such as the thalamic radiation, were included in the analysis. Brains were parcellated into 200 nodes using a kmeans clustering algorithm. All the fibre endpoint positions were used as input, and cluster assignment was done based on the similarity in connectivity profile between pairs of end points. Therefore, vertices with similar connectivity profile have a higher chance of grouping together. The clustering was performed twice, once for each hemisphere. Nodes were defined as the mass centre of the resulting 200 clusters. Connectivity matrices were generated by indexing the number of streamlines between any two nodes (Assaf et al., 2020). The resulting connectivity matrices are hence sparse and weighted adjacency matrices. For the analysis of the Laplacian eigenspectrum, connectivity matrices were binarized by setting connectivity values to 1 if the connection exists and 0 otherwise.
Even though the sizes of the regions differ across species, we opted for a uniform parcellation scheme (i.e. 200 nodes) for several reasons. First, to our knowledge, there is no MRI parcellation for the brains of the majority of the species studied here. Second, how brain regions correspond to one another across species (i.e. homologues) is still not completely understood for many regions and for many species. Third, comparing networks of different sizes introduces numerous analytical biases because most network measures trivially depend on size, making the comparison challenging. We therefore opted to implement a uniform parcellation scheme across species, allowing us to translate connectomes into a common reference feature space in which they can be compared (see ‘Spectral distance’ and ‘Topological distance’ sections). Note that this approach does not take into account speciesspecific regional delineations, nor does it capture homologies between nodes across species, which are still not completely understood. To ensure that the results are not idiosyncratic to the choice of parcellation and parcellation scale, we replicated all results using a lower resolution (100 node; Figure 2—figure supplement 7 and Figure 3—figure supplement 9) and higher resolution (300 node; Figure 2—figure supplement 8 and Figure 3—figure supplement 10) parcellation.
Controlling for the scanning resolution and acquisition parameters
Request a detailed protocolAs the size of the matrix was kept constant across species (i.e. 200 nodes), voxel dimensions were linearly scaled with brain volume, thus resulting in different scanning resolutions across the samples. To verify that this was a reasonable assumption, several tests were performed: (1) the diffusionbased connectome of the mouse was previously compared against one derived from tracttracing (see Assaf et al., 2020 for details), obtaining a strong correlation between both networks. (2) Results on the connectivity conservation principle presented in Assaf et al., 2020 were invariant to different scanning and parcellation parameters across nine different species. (3) The diffusionweighted imaging method was able to reconstruct specific ground truth fibre systems across brains, and these fibre bundles scaled in size with brain volume.
Spectral distance
Request a detailed protocolTo estimate similarities between species’ connectome organization, we computed the Laplacian eigenspectrum of each graph. The Laplacian eigenspectrum acts like a spectroscopy of the graph and summarizes distinct aspects of the underlying topology (Banerjee and Jost, 2009; Banerjee and Jost, 2008; Newman, 2001; Grone et al., 1990; Grone and Merris, 1994; Das, 2004). We considered the normalized Laplacian matrix $L$ for undirected graphs with binary adjacency matrix $A$ defined as $L=I{D}^{1}A$, where $D$ is a diagonal matrix with $D(i,i)=$ deg $i$, and deg $i$ is the binary degree of vertex $i$.
with $i$ and $j$ representing two vertices of the graph. The Laplacian spectrum is then given by the set of all the eigenvalues of $L$. Importantly, the eigenspectrum of the normalized Laplacian has the advantage that all eigenvalues are in the range $[0,2]$ (Chung, 1996), facilitating comparison across species. Furthermore, the normalized Laplacian is unitarily equivalent to the symmetric normalized Laplacian (Chung, 1996), that is, ${L}_{symm}=I{D}^{\frac{1}{2}}A{D}^{\frac{1}{2}}$, thus the eigenvalues of both Laplacians are real. The spectral distance between every pair of animal species was then estimated as 1 minus the cosine similarity of their Laplacian eigenvalue distributions, where eigenvalue distributions were assumed to be vectors in a highdimensional space.
To allow comparison of our results with previous reports (de Lange et al., 2014), in addition to comparing species using their connectome’s Laplacian eigenspectra straightaway, we smoothed the eigenvalue distribution (i.e. ${\lambda}_{1}$, ${\lambda}_{2}$, …, ${\lambda}_{n}$) by convolving eigenvalue frequencies with a Gaussian kernel. The new estimated density is given by
with $n$ being the number of eigenvalues in the approximated distribution, and $\sigma $ being a smoothing factor of 0.015. We used a step of 0.001, which resulted in a total of $n=2000$ points. The approximated distribution was normalized such that area under the curve is 1. For the smoothing, we used the KernelDensity function in the neighbors module of the Scikitlearn Python package (Pedregosa et al., 2011). Details of the implementation can be found in the publicly available code repository. As with the Laplacian eigenspectrum, the spectral distance between every pair of animal species was estimated as 1 minus the cosine similarity of their smoothed (normalized) Laplacian eigenvalue distributions. Results of this supplementary analysis can be found in Figure 2—figure supplement 9.
Topological distance
An alternative way to estimate interspecies distances in connectome organization is to compute the correlation between their network features. We estimated a set of local and global graph theory measures of the connectivity matrix. Local measures include node degree, clustering coefficient, node betweenness, and closeness. Global measures include characteristic path length, transitivity, and assortativity. We included both the binary and the weighted versions of these measures. We constructed a vector of local and global topological features for every animal species. Because there are as many local features as nodes in a network, we only used the average and the standard deviation of these measures. Similar to the spectral distance, the topological distance between every pair of animal species was estimated as 1 minus the cosine similarity of their topological feature vectors. All local and global features were estimated using the Python version of the Brain Connectivity Toolbox (https://github.com/aestrivex/bctpy; Sporns et al., 2022; Rubinov and Sporns, 2010). Definitions of these topological metrics can be found below.
Local features
Request a detailed protocolDegree (bin). Number of connections that a node participates in:
${k}_{i}={\displaystyle \sum _{j\ne i}{a}_{ij}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}$where ${a}_{ij}=1$ if nodes $i$ and $j$ are connected, otherwise ${a}_{ij}=0$. $N$ corresponds to the set of all nodes in the graph (Rubinov and Sporns, 2010).
Degree (wei). Sum of connection weights that a node participates in:
${s}_{i}={\displaystyle \sum _{j\ne i}{w}_{ij}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}$where ${w}_{ij}$ corresponds to the connection weight between nodes $i$ and $j$. $N$ corresponds to the set of all nodes in the graph (Rubinov and Sporns, 2010).
Clustering (bin). Proportion of transitive closures (closed triangles) around a node, that is, the fraction of neighbour nodes that are neighbours of each other:
$\begin{array}{rl}{c}_{i}(A)& ={\displaystyle \frac{\frac{1}{2}\sum _{j\ne i}\sum _{h\ne (i,j)}{a}_{ij}{a}_{ih}{a}_{jh}}{\frac{1}{2}{k}_{i}({k}_{i}1)}}\\ & \\ & =\frac{({\displaystyle {A}^{3}{)}_{ii}}}{{k}_{i}({k}_{i}1)}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N\end{array}$where $A$ corresponds to the binary adjacency matrix of the graph, ${({A}^{3})}_{ii}$ is the ith element of the main diagonal of ${A}^{3}=A\cdot A\cdot A$, and k_{i} is the degree of node $i$ (Watts and Strogatz, 1998; Rubinov and Sporns, 2010).
Clustering (wei). Mean ‘intensity’ of triangles around a node:
$\begin{array}{rl}{c}_{i}(W)& ={\displaystyle \frac{\frac{1}{2}\sum _{j\ne i}\sum _{h\ne (i,j)}{w}_{ij}^{\frac{1}{3}}{w}_{ih}^{\frac{1}{3}}{w}_{jh}^{\frac{1}{3}}}{\frac{1}{2}{k}_{i}({k}_{i}1)}}\\ & \\ & ={\displaystyle \frac{({W}^{[\frac{1}{3}]}{)}_{ii}^{3}}{{k}_{i}({k}_{i}1)}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}\end{array}$where $W$ corresponds to the weighted connectivity matrix, ${({W}^{[\frac{1}{3}]})}_{ii}^{3}$ is the ith element of the main diagonal of ${({W}^{[\frac{1}{3}]})}^{3}={W}^{\frac{1}{3}}\cdot {W}^{\frac{1}{3}}\cdot {W}^{\frac{1}{3}}$, and k_{i} is the degree of node $i$ (Onnela et al., 2005; Rubinov and Sporns, 2010).
Shortest path length (bin). Minimum geodesic distance between pairs of nodes:
${d}_{ij}={\displaystyle \sum _{{a}_{uv}\in {g}_{i\leftrightarrow j}}{a}_{uv}}$where ${g}_{i\leftrightarrow j}$ is the shortest geodesic path between nodes $i$ and $j$. Note that ${d}_{ij}=\mathrm{\infty}$ for all disconnected pairs $i,j$(Rubinov and Sporns, 2010).
Shortest path length (wei). Minimum weighted distance between pairs of nodes:
${d}_{ij}^{\text{w}}={\displaystyle \sum _{{a}_{uv}\in {g}_{i\leftrightarrow j}}f({w}_{uv})}$where $f$ is a map (e.g. the inverse) from weight to length and ${g}_{i\leftrightarrow j}$ is the shortest weighted path between nodes $i$ and $j$ (Rubinov and Sporns, 2010).
Betweenness (bin). Proportion of shortest (geodesic) paths in the graph that traverse a node:
${b}_{i}={\displaystyle \frac{1}{(n1)(n2)}\sum _{j\ne i}\sum _{h\ne (i,j)}\frac{{\rho}_{hj}(i)}{{\rho}_{hj}}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}$where ${\rho}_{hj}$ is the number of shortest (geodesic) paths between nodes $h$ and $j$, ${\rho}_{hj}(i)$ is the number of shortest (geodesic) paths between nodes $h$ and $j$ that pass through $i$, and $n$ is the number of nodes in the graph (Freeman, 1978; Brandes, 2001; Kintali, 2008; Rubinov and Sporns, 2010).
Betweenness (wei). Same as betweenness (bin), but shortest paths are estimated on the respective weighted graph (Freeman, 1978; Brandes, 2001; Kintali, 2008; Rubinov and Sporns, 2010).
Closeness (bin). Mean shortest (geodesic) path length from a node to all other nodes in the network:
${e}_{i}(A)={\displaystyle \frac{\sum _{j\ne i}\sum _{h\ne (i,j)}{a}_{ij}{a}_{ih}[{d}_{jh}({N}_{i}){]}^{1}}{{k}_{i}({k}_{i}1)}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}$where ${d}_{jh}({N}_{i})$ is the shortest (geodesic) path length between nodes $j$ and $h$, which contains only neighbours of $i$, and is estimated on the corresponding binary adjacency matrix $A$ (Latora and Marchiori, 2001; Rubinov and Sporns, 2010).
Closeness (wei). Mean shortest (weighted) path from a node to all other nodes in the network:
${e}_{i}(W)={\displaystyle \frac{\sum _{j\ne i}\sum _{h\ne (i,j)}({w}_{ij}{w}_{ih}[{d}_{jh}^{\text{w}}({N}_{i}){]}^{1}{)}^{\frac{1}{3}}}{{k}_{i}({k}_{i}1)}\phantom{\rule{1em}{0ex}}\phantom{\rule{1em}{0ex}}\mathrm{\forall}i\in N}$where ${d}_{jh}^{\text{w}}({N}_{i})$ is the length of the shortest (weighted) path between nodes $j$ and $h$, which contains only neighbours of $i$, and is estimated on the corresponding weighted adjacency matrix $W$ (Latora and Marchiori, 2001; Rubinov and Sporns, 2010).
Global features
Request a detailed protocolCharacteristic path length (bin). Average shortest (geodesic) path length between all pairs of nodes in the graph:
$L={\displaystyle \frac{1}{n}\sum _{i\in N}\frac{\sum _{j\ne i}{d}_{ij}}{n1}}$where ${d}_{ij}$ is the shortest (geodesic) path length between nodes $i$ and $j$, and $n$ is the number of nodes in the graph (Watts and Strogatz, 1998; Rubinov and Sporns, 2010).
Characteristic path length (wei). Average shortest (weighted) path length between all pairs of nodes in the graph:
${L}^{\text{w}}={\displaystyle \frac{1}{n}\sum _{i\in N}\frac{\sum _{j\ne i}{d}_{ij}^{\text{w}}}{n1}}$where ${d}_{ij}^{\text{w}}$ is the shortest (weighted) path length between nodes $i$ and $j$, and $n$ is the number of nodes in the graph (Watts and Strogatz, 1998; Rubinov and Sporns, 2010).
Transitivity (bin). Ratio between the observed number of closed triangles and the maximum possible number of closed triangles:
$\begin{array}{rl}T(A)& ={\displaystyle \frac{\sum _{i\in N}\sum _{j,h\in N}{a}_{ij}{a}_{ih}{a}_{jh}}{\sum _{i\in N}{k}_{i}({k}_{i}1)}}\\ & \\ & ={\displaystyle \frac{\sum _{i\in N}({A}^{3}{)}_{ii}}{\sum _{i\in N}{k}_{i}({k}_{i}1)}}\end{array}$where $A$ is the binary adjacency matrix, and k_{i} is the degree of node $i$ (Newman, 2003; Rubinov and Sporns, 2010).
Transitivity (wei). Ratio between the ‘intensity’ of observed closed triangles and the maximum possible ‘intensity’ of closed triangles:
$\begin{array}{rl}T(W)& ={\displaystyle \frac{\sum _{i\in N}\sum _{j,h\in N}{w}_{ij}^{\frac{1}{3}}{w}_{ih}^{\frac{1}{3}}{w}_{jh}^{\frac{1}{3}}}{{k}_{i}({k}_{i}1)}}\\ & \\ & ={\displaystyle \frac{\sum _{i\in N}({W}^{[\frac{1}{3}]}{)}_{ii}^{3}}{\sum _{i\in N}{k}_{i}({k}_{i}1)}}\end{array}$where $W$ is the weighted connectivity matrix, and k_{i} is the degree of node $i$ (Newman, 2003; Rubinov and Sporns, 2010).
Assortativity (bin). Correlation coefficient between the degree of a node and the mean degree of its neighbours:
$r=\frac{{l}^{1}\sum _{(i,j)\in L}{k}_{i}{k}_{j}[{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}({k}_{i}+{k}_{j}){]}^{2}}{{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}({k}_{i}^{2}+{k}_{j}^{2})[{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}({k}_{i}+{k}_{j}){]}^{2}}$where ${l}^{1}$ is the inverse of the number of links in the graph, k_{i} is the degree of node $i$, and $L$ is the set of links in the graph (Newman, 2002; Rubinov and Sporns, 2010).
Assortativity (wei). Correlation coefficient between the weighted degree of a node and the mean weighted degree of its neighbour:
$r}^{\text{w}}=\frac{{l}^{1}\sum _{(i,j)\in L}{w}_{ij}{s}_{i}{s}_{j}[{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}{w}_{ij}({s}_{i}+{s}_{j}){]}^{2}}{{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}{w}_{ij}({s}_{i}^{2}+{s}_{j}^{2})[{l}^{1}\sum _{(i,j)\in L}\frac{1}{2}{w}_{ij}({s}_{i}+{s}_{j}){]}^{2}$where ${l}^{1}$ is the inverse of the number of links in the graph, s_{i} is the weighted degree of node $i$, and $L$ is the set of links in the graph (Leung and Chau, 2007; Rubinov and Sporns, 2010).
Smallworld organization
Request a detailed protocolWe use the index proposed in Humphries and Gurney, 2008 to measure connectomes’ smallworldness level. The index is given by
where $L$ and $C$ are the corresponding characteristic path length and clustering coefficient of each connectome, respectively, and ${L}_{rand}$ and ${C}_{rand}$ are the corresponding average quantities for a set of 1000 randomly rewired graphs that preserve the degree sequence and distribution of the nodes (Maslov and Sneppen, 2002). A network is said to possess a smallworld architecture if $S\ge 1$, that is, if it situates above the identity line in a $\gamma $ vs. $\lambda $ plot.
Multiresolution community detection
Request a detailed protocolWe used the Louvain algorithm to determine the optimal community structure of connectomes (Blondel et al., 2008). Briefly, this algorithm extracts communities from large networks by optimizing a modularity score. Here, we use the Qmetric as the objective function (Blondel et al., 2008; Fortunato and Barthélemy, 2007):
where $A$ corresponds to the adjacency matrix of the network, k_{i} is the degree of node $i$, $m$ is the sum of all connections in the graph, c_{i} is the community affiliation of node $i$, and $\delta $ is the Kronecker delta function (i.e. $\delta (x,y)=1\mathit{\hspace{1em}}\text{if}\mathit{\hspace{1em}}x=y,0\mathit{\hspace{1em}}\text{otherwise}$). The size of the partition is controlled by a resolution parameter $\gamma $ (higher $\gamma $ values result in a larger number of modules). Because the Louvain method is a greedy algorithm, we first found multiple (250) optimal partitions at $\gamma =1$, and then we determined a single partition using a consensus clustering approach (Bassett et al., 2013).
Classification of edges
Connectomes’s connections were classified into different categories based on four criteria. The first criterion is based on the modular structure of the network and classifies connections depending on whether they link brain regions within the same module (i.e. intramodular) or regions across different modules (i.e. intermodular); the second criterion is whether connections link brain regions within the same hemisphere (i.e. intrahemispheric connections) or across hemispheres (interhemispheric connections); the third criterion is based on the physical length of the connections (i.e. short, medium, and longrange connections); and the fourth criterion is based on the richclub structure of the network (i.e. richclub, feeder, and peripheral connections).
Inter vs. intramodular connections
Request a detailed protocolTo classify connections as being either inter or intramodular, a consensus clustering algorithm was applied on each connectome to determine a partition of the network into different modules (see ‘Community detection’). Once modules are identified, intermodular connections correspond to those linking brain regions across different modules, whereas intramodular connections correspond to those linking brain regions belonging to the same module.
Connection length
Request a detailed protocolEuclidean distance between regions’ centres was used as a proxy for connection length. To subdivide connections into short, medium, and longrange, connection lengths were estimated as a percentage with respect to the maximum distance between regions. Shortrange connections correspond to those that are less than or equal to 25% of the maximum distance; mediumrange connections are above 25% but less than or equal to 75% of the maximum distance; and longrange connections are those above 75% of the maximum distance.
Richclub vs. feeder vs. peripheral connections
Request a detailed protocolTo classify edges as being either richclub, feeder, or peripheral connections, it is necessary to identify first the richclub of hubs in the network, that is, the densely interconnected core of nodes that have a disproportionately high number of connections (van den Heuvel and Sporns, 2011; van den Heuvel et al., 2012). To do so, we compute the richclub coefficients $\mathrm{\Phi}(k)$ across a range of degree $k$ of the unweighted (binary) connectomes. For binary networks, all nodes that show a degree $\le k$ are removed from the network, and for the remaining set of nodes (i.e. a subgraph), the richclub coefficient is estimated as the ratio of connections present in the subgraph, to the total number of possible connections that would be present if the resulting subgraph was fully connected. Formally, the richclub coefficient is given by Zhou and Mondragon, 2004; Colizza et al., 2006; McAuley et al., 2007
In random networks, such as the Erdős–Rényi model, nodes with a higher degree have a higher probability of being interconnected by chance alone, thus showing an increasing function of $\mathrm{\Phi}(k)$. For this reason, the richclub coefficient is typically normalized relative to a set of $m$ comparable random networks of equal size and node degree sequence and distribution (Maslov and Sneppen, 2002; Colizza et al., 2006; McAuley et al., 2007). The normalized richclub coefficient is then given by
where ${\varphi}_{random}$ corresponds to the average richclub coefficient over the $m$ random networks. In our particular case, $m=1000$. An increasing normalized coefficient ${\mathrm{\Phi}}_{norm}>1$ over a range of $k$ reflects the existence of richclub organization.
To assess the statistical significance of richclub organization, we used permutation testing (Bassett and Bullmore, 2009; van den Heuvel et al., 2010). Briefly, the population of the $m$ random networks yields a null distribution of richclub coefficients. For the range of $k$ expressing richclub organization (i.e. ${\mathrm{\Phi}}_{norm}>1$), we tested whether $\varphi (k)$ significantly exceeds ${\mathrm{\Phi}}_{random}(k)$. Next, we identify the kth level at which the maximum significant ${\mathrm{\Phi}}_{norm}$ occurs. Nodes with a degree $\ge k$ are said to belong to the kthcore of the network. Next, we identify the hubs, that is, nodes whose degree is above the average degree of the network plus 1 standard deviation. Therefore, richclub nodes were identified as those nodes that are both hubs and belong to the kthcore of the network.
Once richclub nodes are identified, richclub connections are defined as edges between richclub nodes; feeder connections are edges connecting richclub to nonrichclub nodes, and peripheral connections are edges between nonrichclub nodes.
Controlling for replicas
Request a detailed protocolBecause some of the species have multiple scans, this could bias the distribution of intra and interorder distances, which could be dominated by those species with a large number of replicas. To account for that, we randomly sample a single connectome per species, and we calculated interspecies distances. We repeated this procedure iteratively 10,000 times. The reported intra and interorder distance distributions correspond to the average distances across iterations.
Controlling for density
Request a detailed protocolSome of the graph features used for the estimation of the topological distance are highly dependent on the density of the network. To regress out the effects of network density on a graph feature, a univariate linear and an exponential model are fitted using density as the explanatory variable, and each feature as response variable. That is,
where $y$ represents network features and $x$ corresponds to density. Network features are then replaced by the residuals of the model. The decision to fit either a linear, an exponential, or no model at all was based on the variance explained by the model or ${R}^{2}$. The model with the largest ${R}^{2}$ is selected. Only those features with ${R}^{2}>0.1$ were controlled to account for density (features in Figure 3—figure supplement 5 with a regression line).
Code availability
Request a detailed protocolAll codes used for data analysis and figure generation are publicly available on GitHub (https://github.com/netneurolab/suarez_connectometaxonomy; Suarez, 2022 copy archived at swh:1:rev:0d8e98f65a51a77784b31ec3ca59176d9119d927) and are built on top of the following opensource Python packages: rnns (https://github.com/estefanysuarez/rnns.git; Suarez, 2021), Netneurotools (https://github.com/netneurolab/netneurotools; Markello et al., 2022), Numpy (Harris et al., 2020; van der Walt et al., 2011; Oliphant, 2006), Scipy (Virtanen et al., 2020), Pandas (McKinney, 2010), Scikitlearn (Pedregosa et al., 2011), bctpy (https://github.com/aestrivex/bctpy; Sporns et al., 2022; Rubinov and Sporns, 2010), Matplotlib (Hunter, 2007), and Seaborn (Waskom et al., 2016).
Data availability
The MaMI data set was originally collected and analyzed by Assaf and colleagues in Assaf et al., 2020. We have included the connectivity matrices used in this study in a public repository available at https://doi.org/10.5281/zenodo.7143143.
References

Scaling principles of white matter brain connectivityNeuroscience 32:bhab384.https://doi.org/10.1101/2021.05.31.445808

Conservation of brain connectivity and wiring across the mammalian classNature Neuroscience 23:805–808.https://doi.org/10.1038/s4159302006417

Network morphospaceJournal of the Royal Society, Interface 12:20140881.https://doi.org/10.1098/rsif.2014.0881

Communication dynamics in complex brain networksNature Reviews. Neuroscience 19:17–33.https://doi.org/10.1038/nrn.2017.149

Speciation in mammals and the genetic species conceptJournal of Mammalogy 87:643–662.https://doi.org/10.1644/06MAMMF038R2.1

On the spectrum of the normalized graph LaplacianLinear Algebra and Its Applications 428:3015–3022.https://doi.org/10.1016/j.laa.2008.01.029

Graph spectra as a systematic tool in computational biologyDiscrete Applied Mathematics 157:2425–2431.https://doi.org/10.1016/j.dam.2008.06.033

Premotor neuron divergence reflects vocal evolutionThe Journal of Neuroscience 38:5325–5337.https://doi.org/10.1523/JNEUROSCI.008918.2018

Brains and speciation: control of behaviorCurrent Opinion in Neurobiology 71:158–163.https://doi.org/10.1016/j.conb.2021.11.003

Differences in frontal network anatomy across primate speciesThe Journal of Neuroscience 40:2094–2107.https://doi.org/10.1523/JNEUROSCI.165018.2019

Neural architectures in the light of comparative connectomicsCurrent Opinion in Neurobiology 71:139–149.https://doi.org/10.1016/j.conb.2021.10.006

Human brain networks in health and diseaseCurrent Opinion in Neurology 22:340–347.https://doi.org/10.1097/WCO.0b013e32832d93dd

Genetic contributions to behavioural diversity at the geneenvironment interfaceNature Reviews. Genetics 12:809–820.https://doi.org/10.1038/nrg3065

A predictive model of the cat cortical connectome based on cytoarchitecture and distanceBrain Structure & Function 220:3167–3184.https://doi.org/10.1007/s004290140849y

Fast unfolding of communities in large networksJournal of Statistical Mechanics 2008:10008.https://doi.org/10.1088/17425468/2008/10/P10008

A faster algorithm for betweenness centrality*The Journal of Mathematical Sociology 25:163–177.https://doi.org/10.1080/0022250X.2001.9990249

The evolution of distributed association networks in the human brainTrends in Cognitive Sciences 17:648–665.https://doi.org/10.1016/j.tics.2013.09.017

The economy of brain network organizationNature Reviews. Neuroscience 13:336–349.https://doi.org/10.1038/nrn3214

A brief history of the taxonomy of mammalsMCV/Q, Medical College of Virginia Quarterly 4:77–80.

Phylogenetic comparative analysis: a modeling approach for adaptive evolutionThe American Naturalist 164:683–695.https://doi.org/10.1086/426002

ConferenceSpectral graph theory. FresnoProceedings of the American Mathematical Society.https://doi.org/10.1090/cbms/092

Detecting richclub ordering in complex networksNature Physics 2:110–115.https://doi.org/10.1038/nphys209

The Laplacian spectrum of a graphComputers & Mathematics with Applications 48:715–724.https://doi.org/10.1016/j.camwa.2004.05.005

The laplacian spectrum of neural networksFrontiers in Computational Neuroscience 7:189.https://doi.org/10.3389/fncom.2013.00189

Rich Club organization and Intermodule communication in the cat connectomeThe Journal of Neuroscience 33:12929–12939.https://doi.org/10.1523/JNEUROSCI.144813.2013

Neural evolution of contextdependent fly songCurrent Biology 29:1089–1099.https://doi.org/10.1016/j.cub.2019.02.019

Comparison of brain structure volumes in insectivora and primates. I. neocortexJournal Fur Hirnforschung 23:375–389.

Centrality in social networks conceptual clarificationSocial Networks 1:215–239.https://doi.org/10.1016/03788733(78)900217

The laplacian spectrum of a graphSIAM Journal on Matrix Analysis and Applications 11:218–238.https://doi.org/10.1137/0611016

The laplacian spectrum of a graph IISIAM Journal on Discrete Mathematics 7:221–229.https://doi.org/10.1137/S0895480191222653

Speciation across the tree of lifeBiological Reviews of the Cambridge Philosophical Society 96:1205–1242.https://doi.org/10.1111/brv.12698

Clustered organization of cortical connectivityNeuroinformatics 2:353–360.https://doi.org/10.1385/NI:2:3:353

Matplotlib: a 2D graphics environmentComputing in Science & Engineering 9:90–95.https://doi.org/10.1109/MCSE.2007.55

Nature versus nurture revisited: an old idea with a new twistProgress in Neurobiology 70:33–52.https://doi.org/10.1016/s03010082(03)000881

The evolution of the neocortex in mammals: how is phenotypic diversity generated?Current Opinion in Neurobiology 15:444–453.https://doi.org/10.1016/j.conb.2005.07.003

Natural selection and random genetic drift in phenotypic evolutionEvolution; International Journal of Organic Evolution 30:314–334.https://doi.org/10.1111/j.15585646.1976.tb00911.x

Efficient behavior of smallworld networksPhysical Review Letters 87:198701.https://doi.org/10.1103/PhysRevLett.87.198701

ConferenceExploredti: a graphical toolbox for processing, analyzing, and visualizing diffusion mr dataIn Proc Intl Soc Mag Reson Med.

Network topology of the marmoset connectomeNetwork Neuroscience 4:1181–1196.https://doi.org/10.1162/netn_a_00159

A species definition for the modern synthesisTrends in Ecology & Evolution 10:294–299.https://doi.org/10.1016/01695347(95)900314

Evolutionary genetics of reproductive behavior in Drosophila: connecting the dotsAnnual Review of Genetics 39:263–291.https://doi.org/10.1146/annurev.genet.39.073003.112454

Comparing brains by matching connectivity profilesNeuroscience and Biobehavioral Reviews 60:90–97.https://doi.org/10.1016/j.neubiorev.2015.10.008

Connectivity fingerprints: from areal descriptions to abstract spacesTrends in Cognitive Sciences 22:1026–1037.https://doi.org/10.1016/j.tics.2018.08.009

A common space approach to comparative neuroscienceAnnual Review of Neuroscience 44:69–86.https://doi.org/10.1146/annurevneuro100220025942

Of circuits and brains: the origin and diversification of neural architecturesFrontiers in Ecology and Evolution 8:82.https://doi.org/10.3389/fevo.2020.00082

Richclub phenomenon across complex network hierarchiesApplied Physics Letters 91:084103.https://doi.org/10.1063/1.2773951

ConferenceData Structures for Statistical Computing in PythonPython in Science Conference.https://doi.org/10.25080/Majora92bf192200a

From regions to connections and networks: new bridges between brain and behaviorCurrent Opinion in Neurobiology 40:1–7.https://doi.org/10.1016/j.conb.2016.05.003

Phylogenomics and the genetic architecture of the placental mammal radiationAnnual Review of Animal Biosciences 9:29–53.https://doi.org/10.1146/annurevanimal061220023149

Assortative mixing in networksPhysical Review Letters 89:208701.https://doi.org/10.1103/PhysRevLett.89.208701

The structure and function of complex networksSIAM Review 45:167–256.https://doi.org/10.1137/S003614450342480

Intensity and coherence of motifs in weighted complex networksPhysical Review. E, Statistical, Nonlinear, and Soft Matter Physics 71:065103.https://doi.org/10.1103/PhysRevE.71.065103

Rapid effects of selection on brainwide activity and behaviorCurrent Biology 30:3647–3656.https://doi.org/10.1016/j.cub.2020.06.086

The anatomical basis of functional localization in the cortexNature Reviews. Neuroscience 3:606–616.https://doi.org/10.1038/nrn893

Scikitlearn: machine learning in pythonJournal of Machine Learning Research 12:2825–2830.

Analysis of connectivity in the cat cerebral cortexThe Journal of Neuroscience 15:1463–1483.

Volumenund oberflachenmessungen an gehirnen verschiedener saugetiere im vergleich zu einem errechneten modellJournal Fur Hirnforschung 15:401–408.

Genomics and the origin of speciesNature Reviews. Genetics 15:176–192.https://doi.org/10.1038/nrg3644

LargeScale network organization in the avian forebrain: a connectivity matrix and theoretical analysisFrontiers in Computational Neuroscience 7:89.https://doi.org/10.3389/fncom.2013.00089

The small world of the cerebral cortexNeuroinformatics 2:145–162.https://doi.org/10.1385/NI:2:2:145

The human connectome: a structural description of the human brainPLOS Computational Biology 1:e42.https://doi.org/10.1371/journal.pcbi.0010042

Network attributes for segregation and integration in the human brainCurrent Opinion in Neurobiology 23:162–171.https://doi.org/10.1016/j.conb.2012.11.015

Spatial embedding imposes constraints on neuronal network architecturesTrends in Cognitive Sciences 22:1127–1142.https://doi.org/10.1016/j.tics.2018.09.007

Linking structure and function in macroscale brain networksTrends in Cognitive Sciences 24:302–315.https://doi.org/10.1016/j.tics.2020.01.008

Learning function from structure in neuromorphic networksNature Machine Intelligence 3:771–786.https://doi.org/10.1038/s42256021003761

SoftwareA connectomicsbased taxonomy of mammals, version swh:1:rev:0d8e98f65a51a77784b31ec3ca59176d9119d927Software Heritage.

BookEugène Dubois and the ApeMan from Java: The History of the First ‘Missing Link’and Its DiscovererSpringer Science & Business Media.https://doi.org/10.1007/9789400922099

The rich club of the C. elegans neuronal connectomeThe Journal of Neuroscience 33:6380–6387.https://doi.org/10.1523/JNEUROSCI.378412.2013

Aberrant frontal and temporal complex network structure in schizophrenia: a graph theoretical analysisThe Journal of Neuroscience 30:15915–15926.https://doi.org/10.1523/JNEUROSCI.287410.2010

Richclub organization of the human connectomeThe Journal of Neuroscience 31:15775–15786.https://doi.org/10.1523/JNEUROSCI.353911.2011

Comparative connectomicsTrends in Cognitive Sciences 20:345–361.https://doi.org/10.1016/j.tics.2016.03.001

The numpy array: a structure for efficient numerical computationComputing in Science & Engineering 13:22–30.https://doi.org/10.1109/MCSE.2011.37

Integrative wholebrain neuroscience in larval zebrafishCurrent Opinion in Neurobiology 50:136–145.https://doi.org/10.1016/j.conb.2018.02.004

The structure of the nervous system of the nematode Caenorhabditis elegansPhilosophical Transactions of the Royal Society of London. Series B, Biological Sciences 314:1–340.https://doi.org/10.1098/rstb.1986.0056

Optimized connectome architecture for sensorymotor integrationNetwork Neuroscience 1:415–430.https://doi.org/10.1162/NETN_a_00022

The richclub phenomenon in the Internet topologyIEEE Communications Letters 8:180–182.https://doi.org/10.1109/LCOMM.2004.823426
Article and author information
Author details
Funding
Natural Sciences and Engineering Research Council of Canada
 Bratislav Misic
 Guillaume Lajoie
Canadian Institutes of Health Research
 Bratislav Misic
Fondation Brain Canada (Future Leaders Fund)
 Bratislav Misic
Canada Research Chairs
 Bratislav Misic
 Guillaume Lajoie
Michael J. Fox Foundation for Parkinson's Research
 Bratislav Misic
Healthy Brains for Healthy Lives
 Bratislav Misic
Canadian Institute for Advanced Research
 Guillaume Lajoie
National Science Foundation  BSF
 Yaniv Assaf
National Institute of Mental Health (R01MH122957)
 Olaf Sporns
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
BM acknowledges support from the Natural Sciences and Engineering Research Council of Canada (NSERC), Canadian Institutes of Health Research (CIHR), Brain Canada Foundation Future Leaders Fund, the Canada Research Chairs (CRC) Program and the Healthy Brains for Healthy Lives (HBHL) initiative. GL acknowledges support from NSERC, the CRC Program, and the Canada CIFAR AI Cahir Program.
Version history
 Preprint posted: March 12, 2022 (view preprint)
 Received: March 14, 2022
 Accepted: November 4, 2022
 Accepted Manuscript published: November 7, 2022 (version 1)
 Version of Record published: November 22, 2022 (version 2)
Copyright
© 2022, Suarez et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 1,605
 views

 257
 downloads

 14
 citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Immunology and Inflammation
 Neuroscience
Autoimmune diseases of the central nervous system (CNS) such as multiple sclerosis (MS) are only partially represented in current experimental models and the development of humanized immune mice is crucial for better understanding of immunopathogenesis and testing of therapeutics. We describe a humanized mouse model with several key features of MS. Severely immunodeficient B2mNOG mice were transplanted with peripheral blood mononuclear cells (PBMCs) from HLADRB1typed MS and healthy (HI) donors and showed rapid engraftment by human T and B lymphocytes. Mice receiving cells from MS patients with recent/ongoing Epstein–Barr virus reactivation showed high B cell engraftment capacity. Both HLADRB1*15 (DR15) MS and DR15 HI mice, not HLADRB1*13 MS mice, developed human T cell infiltration of CNS borders and parenchyma. DR15 MS mice uniquely developed inflammatory lesions in brain and spinal cord gray matter, with spontaneous, hCD8 T cell lesions, and mixed hCD8/hCD4 T cell lesions in EAE immunized mice, with variation in localization and severity between different patient donors. Main limitations of this model for further development are poor monocyte engraftment and lack of demyelination, lymph node organization, and IgG responses. These results show that PBMC humanized mice represent promising research tools for investigating MS immunopathology in a patientspecific approach.

 Neuroscience
Animals can use a repertoire of strategies to navigate in an environment, and it remains an intriguing question how these strategies are selected based on the nature and familiarity of environments. To investigate this question, we developed a fully automated variant of the Barnes maze, characterized by 24 vestibules distributed along the periphery of a circular arena, and monitored the trajectories of mice over 15 days as they learned to navigate towards a goal vestibule from a random start vestibule. We show that the patterns of vestibule visits can be reproduced by the combination of three stochastic processes reminiscent of random, serial, and spatial strategies. The processes randomly selected vestibules based on either uniform (random) or biased (serial and spatial) probability distributions. They closely matched experimental data across a range of statistical distributions characterizing the length, distribution, step size, direction, and stereotypy of vestibule sequences, revealing a shift from random to spatial and serial strategies over time, with a strategy switch occurring approximately every six vestibule visits. Our study provides a novel apparatus and analysis toolset for tracking the repertoire of navigation strategies and demonstrates that a set of stochastic processes can largely account for exploration patterns in the Barnes maze.