Hubs disruption in mesial temporal lobe epilepsy. A resting‐state fMRI study on a language‐and‐memory network

Abstract Mesial temporal lobe epilepsy (mTLE) affects the brain networks at several levels and patients suffering from mTLE experience cognitive impairment for language and memory. Considering the importance of language and memory reorganization in this condition, the present study explores changes of the embedded language‐and‐memory network (LMN) in terms of functional connectivity (FC) at rest, as measured with functional MRI. We also evaluate the cognitive efficiency of the reorganization, that is, whether or not the reorganizations support or allow the maintenance of optimal cognitive functioning despite the seizure‐related damage. Data from 37 patients presenting unifocal mTLE were analyzed and compared to 48 healthy volunteers in terms of LMN‐FC using two methods: pairwise correlations (region of interest [ROI]‐to‐ROI) and graph theory. The cognitive efficiency of the LMN‐FC reorganization was measured using correlations between FC parameters and language and memory scores. Our findings revealed a large perturbation of the LMN hubs in patients. We observed a hyperconnectivity of limbic areas near the dysfunctional hippocampus and mainly a hypoconnectivity for several cortical regions remote from the dysfunctional hippocampus. The loss of FC was more important in left mTLE (L‐mTLE) than in right (R‐mTLE) patients. The LMN‐FC reorganization may not be always compensatory and not always useful for patients as it may be associated with lower cognitive performance. We discuss the different connectivity patterns obtained and conclude that interpretation of FC changes in relation to neuropsychological scores is important to determine cognitive efficiency, suggesting the concept of “connectome” would gain to be associated with a “cognitome” concept.


| INTRODUCTION
Temporal lobe epilepsy (TLE) is characterized by seizures arising from a dysfunctional region known as epileptogenic zone (or epileptic focus) situated in temporal lobe and particularly, in temporal medial structures (Burianová et al., 2017). Given that language and memory networks (LMNs) include temporal regions, recurrent seizures can modify the function and the structure of these networks. These changes are based on the neural plasticity phenomenon that can take place in TLE patients over the years (Berg & Scheffer, 2011). The reorganization patterns can be more or less cognitively efficient and various degrees of language and memory deficits have been described in patients with mTLE (Alessio et al., 2013;Jaimes-Bautista, Rodríguez-Camacho, Martínez-Juárez, & Rodríguez-Agudelo, 2015;McAndrews & Cohn, 2012;Metternich, Buschmann, Wagner, Schulze-Bonhage, & Kriston, 2014). For instance, Hoppe et al. determined that language and memory were the most affected functions in a large cohort of epileptic patients mainly composed of mTLE (Hoppe, Elger, & Helmstaedter, 2007). Nearly half of the patients showed significant deficits of episodic memory (56%) and language (43%; including naming, speech comprehension, verbal fluency) and around 70% showed minor disorders of these functions.
Previous studies support the idea of close interconnections between left fronto-temporal language areas and hippocampal verbal memory networks in healthy subjects (Weber, Fliessbach, Lange, Kügler, & Elger, 2007) and in adults with epilepsy (Wagner et al., 2008). In the same line, a previous review (Baciu & Perrone-Bertolotti, 2015) pointed out the possible models of TLE reorganization wherein the left hippocampus (mainly involved in long-term memory functions) interacts with ipsilateral and contralateral language areas to modulate language networks (i.e., interhemispheric shifting). The proposed models (Baciu & Perrone-Bertolotti, 2015) correspond to the language-memory interface described by Duff and Brown-Schmidt (2012).
Functional connectivity (FC) is a powerful indicator of the intrinsic functional changes occurring in patients' brain, especially in epilepsy which is a pathology of networks (Besson et al., 2017;van Diessen, Diederen, Braun, Jansen, & Stam, 2013). Among the different FC measures that are available, FC at rest estimated from BOLD signals in fMRI is particularly robust for the description of the brain networks (van den Heuvel & Hulshoff Pol, 2010). Recent studies showed a very strong spatial similarity between intrinsic resting-state networks and networks recruited by a variety of fMRI activation paradigms (Rasero et al., 2018). For instance, Cole et al. found that cognitive task activations can be predicted in certain regions via estimated activity flow over resting-state FC networks, for basic motor tasks but also for higher level tasks such as reasoning (Cole, Ito, Bassett, & Schultz, 2016). Evidences are in favor of "distributed set of core regions active across multiple task and integrates more specialized regions, altering baseline communication dynamics in service of task specific computations" (Shine et al., 2018). In this framework, although flexible components associated with on-task reconfiguration have been suggested (Mill, Ito, & Cole, 2017), there is still large network components that remains "stable" across tasks. These stable components could be the core regions of synchronous networks at rest. Importantly, next to the well-known "default mode network" (Raichle, 2015), independent resting-state networks have been identified in healthy subjects (Abela et al., 2014;Doucet et al., 2011;Power et al., 2011), involving regions normally dedicated to low-level processes (sensorimotor, visual, and auditory) or higher level processes such as language functions (van den Heuvel & Hulshoff Pol, 2010). There is therefore a wide variety of resting networks that are not always studied. This leaves the field open to a broader and more varied study of patterns of brain connectivity at rest, especially in the pathological condition that is accompanied by neurocognitive reorganization.
Without focusing on a specific rest networks, patients with TLE show global reduction of BOLD FC at rest (Fahoum, Lopes, Pittau, Dubeau, & Gotman, 2012;Liao et al., 2010;Tracy et al., 2014) as well as significant alterations of spontaneous activity for specific nodes (i.e., specific brain regions; Zhang et al., 2010). In the same vein, Besson et al. in diffusion MRI tractographic studies found global and large alteration of structural connectivity in networks even far from the dysfunctional hippocampus (Besson et al., 2017(Besson et al., , 2014, reinforcing the idea that anatomical cabling generally directly supports FC (Hervé, Zago, Petit, Mazoyer, & Tzourio-Mazoyer, 2013). Depending on the spatiotemporal dynamics and the methodology used, networks modifications in TLE patients may be reflected by both loss (Luo et al., 2012;Pittau, Grova, Moeller, Dubeau, & Gotman, 2012;Vlooswijk et al., 2011) and gain of FC in comparison to healthy individuals (Bettus et al., 2008;Bonilha et al., 2012). Some modulating factors such as the hemispherical side of the epilepsy (left or right) are also important to consider. Namely, TLE with left seizure foci (dysfunctional hippocampus in the left hemisphere) showed more extensive and widespread changes in connectivity than TLE with right seizure foci both in language networks and in general (i.e., whole brain studies; de Campos, Coan, Lin Yasuda, Casseb, & Cendes, 2016;Dinkelacker, Dupont, & Samson, 2016;Ridley et al., 2015). However, these modulating factors are not always methodologically controlled for in the studies that can have relatively large but heterogeneous samples. In a machine learning study, Su, An, Ma, Qiu, and Hu (2015) investigated FC at rest in right TLE patients and matched healthy subjects to identify connections that distinguish the patients from the controls. Interestingly, their results showed reduced FC within the right hemisphere along with FC strengthening within the preserved left hemisphere, which was interpreted as a compensatory mechanism (Su et al., 2015). Current methods allow for the identification and description of networks in a remarkable complexity, there remains scope for a clearer explanatory understanding of how and importantly what these networks compute (Mill et al., 2017). Little is indeed currently known about the network mechanics responsible of systemwide brain states subserving the large spectrum of cognitive behaviors (Shine et al., 2018). Moving beyond the simple description of networks changes is essential, in particular in patients as considering the association between patterns of FC reorganization and behavioral performance may allow comprehension of the compensatory or deleterious functional roles on cognition.
Considering all findings mentioned above, this study set out to evaluate the reorganization of LMNs in terms of FC (LMN-FC) as assessed with rs-fMRI data in patients with mTLE, compared to healthy participants. We were also interested to determine the effect of the dysfunctional hippocampus lateralization on the LMN-FC in mTLE patients. For that purpose, we explored FC changes in two separate groups of matched TLE patients with seizures starting from the hippocampal complex either to the left (left mTLE; L-mTLE) or to the right (R-mTLE). We have generated our embedded LMN, based on the results of tasks-fMRI studies (a cross-sectional study proposed by Labache et al. (2019); and a meta-analysis published by Spaniol et al. (2009)) in order to obtain the core regions that can compose a stable components for language and memory. Two complementary analyses were applied in order to assess LMN-FC in mTLE patients: (a) region of interest (ROI)-to-ROI analysis to obtain precise information in terms of modifications of individual connections; and (b) graph theory (GT) analyses to estimate possible topological changes occurring on the two main network-specific properties (Sporns, 2013), namely, the segregation (i.e., communities of highly interconnected regions that permit performing tasks in parallel) and the integration (i.e., hubs, areas, or subnetworks able to maintain connections with different groups in order to quickly integrate information). The GT analyses were performed on efficiency parameters at both network and nodal (nodes are LMN regions) level. Spearman correlations were then calculated between selected FC parameters and cognitive scores to assess the effectiveness of FC reorganization. All FC analyses have been carried out using CONN toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012)

| Neuropsychological and clinical data in patients
All patients underwent complete cognitive evaluation including neuropsychological and language assessment carried out by a neuropsychologist and a speech therapist. The general cognitive evaluation (IQ, WAIS-IV: Wechsler, D, 2008) as well as the global executive functioning (Trail Making Test: Godefroy et al., 2008;Stroop test: Stroop, 1935) were used as the inclusion criteria and according to them all patients had normal IQ and executive scores. The efficiency of cerebral reorganization was estimated using correlations between cognitive scores for language and memory and FC parameters. Specifically, the following cognitive features were used to perform correlations: These test scores were then standardized by gender, age and sociocultural level. Information about neuropsychological tests is provided in Appendix S1 and Table 1 details the clinical information and cognitive performance obtained by patients.
On average, the two patient groups did not differ significantly in their clinical data: age (Mann-Whitney U = 153, p = .6); educational level (U = 152.5, p = .6); epilepsy duration (U = 155.5, p = .6); and number of AEDs (U = 160, p = .8). We observed significant differences between the two groups of patients for the left hippocampal volume (U = 97, p = .02). Regarding the volume of the right hippocampus the difference was not significant at a threshold of p < .05 (U = 231, p = .07). Nevertheless, there is a significant intragroup difference between the left and right hippocampi for the both groups of patients (L-mTLE: t(17) = −3.89, p < .001); R-mTLE: t(18) = 4.48, p < .001). For the L-mTLE group, the left hippocampus was significantly smaller (m = 3.4) than the right (m = 3.89). Conversely, for the R-mTLE group, the right hippocampus (m = 3.45) was significantly smaller than the left (m = 4.02). However, both groups were matched regarding the mean sizes of their respective dysfunctional hippocampi (i.e., left hippocampus for L-mTLE vs. right hippocampus for R-mTLE; U = 162.5, p = .8); as well as of their respective "healthy" hippocampi (right hippocampus for L-mTLE vs. left hippocampus for R-mTLE; U = 159, p = .7). In addition, none of the patients had a total IQ or executive performances below or equal to the pathological scores and there were no statistical differences between the two groups of patients (IQ: U = 163, p = .8; EF total: U = 164, p = .8).    Note: Z scores: mean, 0, SD = 1. A pathological z score is equal or below −1.65 SD (percentile 5); Index (standardized composite scores): mean = 100, SD = 15. A pathological index score is here equal or below 70 (−2 SD). Red stars highlight significant differences between the two groups of patients (p < .05); NS indicates clearly nonsignificant differences. Abbreviations: F, female; M, male; age, age at the examination time; EL, education Level (1, undergraduate, 2, graduate; 3, bachelor degree and more); handedness: R, right, L, left, Edinburgh quotient (Oldfield, 1970); HS, hippocampal sclerosis (No, MRI-negative HS); Vol. hippo R, volume in cm 3 of the right hippocampus; Vol. hippo L, volume in cm 3 of the left hippocampus; age onset, age of onset of seizures (age and duration in years); Seizure frequency: seizures per month; Nb. AEDs: number of antiepileptic drugs (by days); IQ, total IQ (Wechsler, D, 2008); EF total, average scores for executive function tests (TMT A, TMT B-A, Stroop interference); mTLE, mesial temporal lobe epilepsy; TMT A, performance (z score) for trail making test Part A (speed processing); TMT B-A, performance (z score) for the difference between trail making test Part B and Part A (mental flexibility); Stroop, performance (z score) for Stroop interference (automatic inhibition); Naming DO80, performance (z score) for French version of picture naming; Semantic fluency, performance (z score) for categorical word generation; Phonological fluency, performance (z score) for alphabetical word generation; VCI, verbal comprehension index (standardized composite score) for verbal semantic memory (WAIS-IV, Wechsler, D, 2008); AMI, auditory memory index (standardized composite score) for verbal memory (immediate and delayed; WMS-IV, Wechsler, D, 2009); VMI, visual memory index (standardized composite score) for visual memory (immediate and delayed; WMS-IV, Wechsler, D, 2009

LMN: Parcellation and node definition
Before performing the FC analyses (ROI-to-ROI and GT analyses), we first defined the LMN network. The LMN was composed of multiple brain regions provided by task-fMRI: one cross-sectional study for language (Labache et al., 2019) and one meta-analysis for memory (Spaniol et al., 2009). We selected MNI coordinates of the activation peaks identified by these studies and converted them into the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA) functional atlas (Joliot et al., 2015). Altogether, the LMN network is composed of 36 homologous brain regions (72 ROIs in both hemispheres), some of them being more specific for language (n = 10), some for memory (n = 20), or involved in the both language and memory (n = 6). The LMN network was therefore composed of 72 AICHA ROIs. We provide a detailed description of the ROIs in the Supplementary Material (Table S1) and Figure 1 shows the LMN in a brain rendering. In addition, to demonstrate the robustness of the chosen network, we have conducted an in-depth analysis of the correspondence and overlap between the LMN and maps derived from the Neurosynth Initiative (http://neurosynth.org/analyses/ [Yarkoni et al., (2011)]) for language and memory (Appendix S3).

Connectivity analyses
In order to evaluate LMN-FC, we have used the CONN toolbox for both ROI-to-ROI and GT analyses. The FC analyses included following steps: noise source reduction, first level individual analysis including correlation analyses, and second level random-effect group analysis.
Noise reduction analysis. The denoising step was applied on the previously preprocessed fMRI for the patients and the control group in order to reduce the noise and to increase sensitivity. Noise reduction analysis used the anatomical component-based noise correction (aCo-mpCor) implemented in CONN (Behzadi, Restom, Liau, & Liu, 2007).
For that purpose, a principal component analysis approach was applied to extract the BOLD signal from the white matter and the CSF, and use them as confounds. In addition, the output matrices generated by ART, as well as movement parameters generated by SPM were entered into CONN as covariates. After the CompCor regressing out, the resulting BOLD time series were band-pass filtered The resultant 72 × 72 matrices have then been used for statistical analyses described below.
GT analysis: For GT analyses, unweighted graphs were constructed by computing binary adjacency matrices for each participant at different connection cost (or sparsity) ranging between 5 and 20%. These thresholds were selected to account for representing the known sparsity of functional connections (economical brain functional networks; "small-world organization," (Achard & Bullmore, 2007), by controlling for the small-world parameter. Graph properties were calculated to derive estimates of global efficiency (E glob ) and local efficiency (E loc ), parameters that quantify networks integration and networks segregation, respectively (Rubinov & Sporns, 2010). These two parameters are thought to represent two core properties of a network and could be computed at the level of the whole network or at the node level.
E glob illustrates how efficiently is the information transmitted within the whole network (i.e., functional integration) and allows rapid integration of information within subnetworks. Global efficiency is computed as: where N is the total number of nodes in the network G, and d ij is the minimum average number of links (shortest path) that connect the node i and the node j (Latora & Marchiori, 2007). At a nodal level, the global efficiency is also known under the term of nodal efficiency (E nod ; Liu et al., 2017) and characterizes the extent to which a node is integrated within the entire network (hub integration; Fornito, 2016).
Nodal efficiency is computed as: As much for E glob or E nod , the higher the value, the faster the transfer of information.
E loc represents the efficiency of local communications that allows a specialization of processing within a densely interconnected group of regions. This parameter estimates to what extent the nodes tend to group of "cluster" together (i.e., functional segregation) and constitute connected local structures. Local efficiency is computed as: F I G U R E 1 Panel a: Language-and-memory network (LMN) to assess functional connectivity (FC). The LMN is composed of 72 homotopic areas (36 in each hemisphere) reported by two task-fMRI studies, one cross-sectional study for language (Labache et al., 2019) and one metaanalysis for memory (Spaniol et al., 2009) and adapted to Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA; Joliot et al., 2015) coordinates. Regions are projected as spheres onto 3D anatomical render templates. Sphere size reflects the AICHA region volume. Color code: dark blue, regions involved in language; light blue, regions involved in episodic memory (encoding and retrieval); green, regions involved in both language and memory. Panel b: Connectogram of mean FC correlation values in controls between regions of interest (ROIs) of the LMN network. Positive correlations are represented in orange-red. Negative correlations are represented in blue. The line width indicates the strength of the correlation. Strongest positive correlations are mostly intrahemispherical. Negative correlations are mostly interhemispherical. The first circle starting from the inside of the connectogram shows mean correlation coefficients for a given region (correlation between regions with all others with which it could be functionally connected). Dark red indicates high average of the correlation coefficient of the corresponding region. The second circle to outside classifies homotopic ROIs of the LMN into different lobes to which they may belong. Color code: Green, lobes and ROIs in the left hemisphere; purple, lobes and ROIs in the right hemisphere where G i is the induced graph obtained by the neighbors of node i, E glob (G i ) is the global efficiency of G i (Latora & Marchiori, 2007;Liu et al., 2017). The higher the value, the more locally efficient the network will be.

Second level statistical analyses
Statistical analyses for both FC and GT parameters were performed using R statistic packages through R studio software v1.     Figure 2).

| GT: Global network and nodes
We did not observe any significant differences between groups of patients and controls at the network level, neither for E glob nor E loc parameters. However, despite some variability observed in patients, the HDI (κ) based on E nod was significantly more negative in patients (L-mTLE and R-mTLE) than in controls at a 5% cost (κ 6 ¼ 0, p < .05).
This result indicates an improvement and/or a decrease of nodal parameter values in inverse proportion to the estimates of the controls, suggesting that there is still a network-wide pattern of LMN disruption in mTLE patients. Although there were some differences in the organization or arrangement of individual hubs between the two groups of mTLE (each of them compared to controls) there was no significant difference between the two groups of patients on the HDI mean. Figure 3 shows the boxplots of E glob , E loc , and HDI distributions for each group, computed at the network scale.
At the nodal level, the results for E loc were very sensitive to the sparsity threshold used and essentially not significant at an adjusted p-value. However, we obtained robust, stable, and significant results for the E nod parameter. Thus, considering the E nod results, we found several clusters of differences between patients and controls including decreases and increases in terms of LMN-FC ( Figure 4, Panel a).
For the L-mTLE group, significant E nod decreases implied bilateral  For the R-mTLE, we found E nod increases for temporal medial structures (bilateral amygdalo-hippocampal complex) and fusiform gyri ( Figure 5, Panels b and c). We also found a negative correlation between the sizes of the right hippocampus and the E nod values estimated for the same region (r = −.8, p < .05; Figure 5, Panel d and Figure S4). The decreases identified in R-mTLE concern only posterior networks (MTG, ITG, and angular gyri). In addition, we have observed an improvement of the E nod capacity for some bilateral frontal regions (mainly IFG and insula). Figure 5 shows and details the main GT results obtained at the nodal level for the R-mTLE group.

| Cognitive scores and correlations with FC parameters
Regarding language and memory scores of interest, we did not find statistical differences between L-mTLE and R-mTLE on naming The distribution of the cognitive standardized scores for the two groups of patients is presented in Figure 6.
At an adjusted-threshold (p FWE-corrected), no significant correlations were found between ROI-to-ROI FC results and language and memory scores. However, significant correlations (p FWE-corrected) were obtained between GT FC parameters and cognitive scores. Specifically, for the L-mTLE group negative correlations were found between E nod values for the left hippocampus and AMI (r = −.95, p < .01), as well as for the left fusiform gyrus and VMI (r = −.9, F I G U R E 2 Connectogram of significant pairwise functional connectivity (FC) differences obtained in left mesial temporal lobe epilepsy (L-mTLE) patients (n = 19) and right mesial temporal lobe epilepsy (R-mTLE) patients (n = 18) compared to controls (n = 48). Specifically, it shows a chord diagram of results obtained with region of interest (ROI)-to-ROI analyses at p false discovery rate (FDR)-corrected. Note: Red links = "hyperconnectivity" (significant gain of FC); blue links = "hypoconnectivity" (significant reduction of FC) between two ROIs in L-mTLE versus healthy. We found increased FC from or to limbic regions (including the dysfunctional hippocampus). Results were reported at p FDRcorrected p < .01). We did not observe significant positive correlations for the L-mTLE patients. Conversely, we found significant and positive correlations between the E nod values for: the left IFG and the phonological fluency (r = .89, p < .01), the left supplementary motor area (SMA) and the semantic fluency (r = .84, p < .01), the left hippocampus and both naming (r = .81, p < .01) and VMI (r = .78, p < .01), and for the right fusiform gyrus and VMI (r = .86, p < .01) but no negative correlation for the R-mTLE group. Figure 6 shows the heat maps of the correlations observed between the E nod values and the cognitive scores.

| DISCUSSION
The first main objective of the study was to estimate the reorganiza- In mTLE, the hippocampus is considered as a central core of abnormalities and is often structurally damaged (e.g., de Campos et al., 2016). Even in the case of the so-called cryptogenic epilepsy or MRInegative epilepsy, subtle lesions at the histological examination can be found (Bernasconi, Bernasconi, Bernhardt, & Schrader, 2011) and may sometimes be observed by using an ultrahigh-field 7-T (7 T) MRI (Obusez et al., 2018). In this study, we specifically found negative correlations across groups of patients between the size of the hippocampus involved in the epilepsy and the integration capacity of this region ( Figure S4). This hyperconnected pattern tended to mainly concern patients presenting with clear hippocampal sclerosis on the MRI (HS;  S2).

F I G U R E 3
Boxplots of the GT results obtained at the network scale. Top left: Representation of the global efficiency (E glob ) distribution according to the subjects groups. There were no differences between groups at p < .05 (sparsity 10%). Top right: Representation of the local efficiency (E loc ) distribution according to the subjects groups. There were no differences between groups at p < .05 (sparsity 10%). Bottom center: Boxplot of the hub disruption index (HDI; Achard et al., 2012) for healthy and patients. We obtained significant hubness imbalance between patients and controls at p < .05 (sparsity 10%). The HDI is different from 0 in patients, meaning a global language-and-memory network (LMN) hubs reorganization in patients compared to controls In line with our results, previous studies had found an increased hippocampal FC and of the core areas of the limbic network in TLE patients (e.g., . In addition, Englot et al. (2015) described a case of a patient with HS that showed specific increased FC for hippocampus, while the FC for lateral temporal network was reduced. Another study conducted by Ellmore, Pieters, and Tandon (2011) assessing the structural connectivity in mTLE patients found enhanced strength of the structural connections between the hippocampus and the rest of the brain, despite a reduced number of fibers.
This finding suggests that the hippocampal atrophy is accompanied by sparse but strong connections in these patients (Ellmore et al., 2011).
The study conducted by Bonilha et al. (2012) provides evidence supporting this phenomenon. MTLE was associated with a regional reduction in fiber density and absolute connectivity, especially in the ipsilateral limbic structures. Paradoxically, patients compared to controls exhibited a significant increase in structural connectivity of the hippocampus for the nodal degree or the betweenness centrality, GT parameters thought to reflect hubs in the network. The results of a more recent study (Besson et al., 2017) integrating intracranial EEG data to determine the location of epileptogenic foci and structural connectivity data are also fully consistent with the prior findings. They found hyperconnected epileptogenic regions at the expense of connectivity with the rest of the brain. Despite the damage, the hippocampus remains thus a structural and functional important hub in the patients' brain networks, which could be called the "hippocampal paradox." Interestingly, the hippocampal paradox (i.e., hyperfunctioning and/or hyperconnectivity despite damage) does not seem to be specific to mesiotemporal epilepsy since similar results were found in other pathologies affecting the hippocampal complex such as the MCI

F I G U R E 4 Illustrations of the main GT results obtained at the nodal level in left mesial temporal lobe epilepsy (L-mTLE) patients. Panel a:
Hierarchical clustermap based on the E nod values (raw data) obtained for each node of the language-and-memory network (LMN) and each subjects of the L-mTLE group. The hierarchical clustering was made using the Euclidean distance. There is a relative consistency between the subjects and two main clusters could be distinguished at the first level of the dendrogram. Panel b: Evolution of the E nod z scores observed in L-mTLE compared to controls depending on the evolution of the sparsity threshold (5, 10, 15, and 20%). Results are projected on a 3D brain render. The global pattern remains consistent and stable across the thresholds. We have observed a hyperconnectivity for the temporo-mesial structures (in red) of the LMN and a hypoconnectivity (in blue) for a large fronto-temporo-parietal network. Panel c: E nod results obtained for a sparsity threshold of 10%. The blue regions correspond to an E nod z score tending toward −1.65 SD. The red one, to an E nod z score that tends toward +1.65 SD. Regions with significant differences between L-mTLE and controls are surrounded in white (G_Frontal_Inf_Tri_1_2, G_Insula_Anterior_2_L; G_Angular_1_2, G, Parietal_Inf_1; G_Temporal_Mid_3, G_Temporal_Inf_4; G_Fusiform_1, G_ParaHippocampal_2, N_Amyglala_1, G_Hippocampus_2). Panel d: E nod values of the left hippocampus, projected on a 3D reconstruction of the specific left hippocampus of each of the L-mTLE patients. The 3D reconstruction of the hippocampi was made using the subject specific-ROIs segmentation provided by volbrain (http://volbrain.upv.es/). Hippocampi are classified according to their size in cm 3 , from the smallest to the largest. The darker the red color, the higher the E nod value. Thus, the smaller the hippocampus, the higher the E nod value tends to be. See Figure S4 for the scatterplot of the correlations between the hippocampus sizes and the E nod values or Alzheimer's disease (Celone et al., 2006;Kasper et al., 2016;Pasquini et al., 2015). In the case of epilepsy, Englot, Konrad, and Morgan (2016) proposed that the role of increased FC in the (peri-)dysfunctional regions may be related to the generation and the spreading of epileptic seizures rather than serving as a compensatory mechanism (Englot et al., 2016). Previous histological studies have shown that epileptic seizures may induce neuronal loss, but that are also followed by a development of new excitatory synapses and axonal sprouting, a phenomenon called "reactive plasticity" (Ben-Ari, Crepel, & Represa, 2008). However, the majority of these newly constituted synapses are anatomically and functionally aberrant (Esclapez, Hirsch, Ben-Ari, & Bernard, 1999;Represa, Tremblay, & Ben-Ari, 1987). This well- Regarding the spatially distant regions from the dysfunctional hippocampus, limbic seizures usually induced dysfunctions of neocortical regions (Englot, Mishra, Mansuripur, Herman, & Hyder, 2008). Beyond the dysfunctional hippocampus, the resting-state FC is generally decreased in TLE patients (Luo et al., 2012), suggesting disconnection of distal areas from the hippocampus. Our study results are also in favor of general FC decreased in the neocortical and remote regions of the dysfunctional hippocampus (Figures 2, 4, and 5). In line with our assumptions, we observed different patterns of FC changes according to the epilepsy lateralization. L-mTLE exhibited more F I G U R E 5 Illustrations of the main GT results obtained at the nodal level in R-mTLE patients. Panel a: Hierarchical clustermap based on the E nod values (raw data) obtained for each node of the language-and-memory network (LMN) and each subjects of the R-mTLE group. The hierarchical clustering was made using the Euclidean distance. There is a relative consistency between the subjects and two main clusters could be distinguished at the first level of the dendrogram. Panel b: Evolution of the E nod z scores observed in left mesial temporal lobe epilepsy (L-mTLE) compared to controls depending on the evolution of the sparsity threshold (5, 10, 15, and 20%). Results are projected on a 3D brain render. The global pattern remains consistent and stable across the thresholds. We have observed a hyperconnectivity for the temporo-mesial structures (in red) as well as for some frontal regions of the LMN and a hypoconnectivity (in blue) for a posterior network, limited to lateral temporal and parietal regions. Panel c: E nod results obtained for a sparsity threshold of 10%. The blue regions correspond to an E nod z score tending toward −1.65 SD. The red one, to an E nod z score that tends toward +1.65 SD. Regions with significant differences between R-mTLE and controls are surrounded in white (G_Frontal_Inf_Tri_1_2, G_Frontal_Mid_Orb-2_L, G_Insula_Anterior_3; G_Angular_1_2; G_Temporal_Mid_3_R; G_Fusiform_1_R, G_ParaHippocampal_2_R, N_Amyglala_1_R, G_Hippocampus_2). Panel d: E nod values of the right hippocampus, projected on a 3D reconstruction of the specific right hippocampus of each of the R-mTLE patients. The 3D reconstruction of the hippocampi was made using the subject specific-ROIs segmentation provided by volbrain (http://volbrain.upv.es/). Hippocampi are classified according to their size in cm 3 , from the smallest to the largest. The darker the red color, the higher the E nod value. Thus, the smaller the hippocampus, the higher the E nod value tends to be. See Figure S4 for the scatterplot of the correlations between the hippocampus sizes and the E nod values pronounced LMN-FC reorganization in comparison with R-mTLE patients. The major differences between groups of patients in the spatial dynamics of FC changes mainly concerned the regions beyond the dysfunctional hippocampus. More precisely, we observed a decreased FC at rest in a large fronto-temporo-parietal network for L-mTLE and a less extensive posterior (temporo-parietal) network for the R-mTLE group. The ENIGMA consortium study aiming to estimate the cortical modifications in a large sample of m-TLE patients show bilateral and significant reduction of thickness in neocortical regions distant from the hippocampus (Whelan et al., 2018). As we found in this study, the cortical thickness reductions were larger in L-mTLE (n = 415 patients) than in R-mTLE (n = 339 patients). Similar differences between L-mTLE and R-mTLE patients have also been reported in terms of structural connectivity at a whole brain level (e.g., Besson et al., 2014).
Two main hypotheses can explain this differential effect regarding areas remote to the dysfunctional hippocampus. First, the structural asymmetry is generally in favor of the left hemisphere. The left asymmetry (possibly due to a longer network maturation period; Keller, Schoene-Bake, Gerdes, Weber, & Deppe, 2012 cited by Besson et al. (2014)) could, indeed, be at the origin of the facilitation of the epileptic activity propagation through the brain explaining the wider modifications in the left hemisphere (Ridley et al., 2015). According to the second hypothesis, the right hemisphere would rather have a protective role, by being able to prevent the spread of m-TLE seizures to other cortices and compensate for brain dysfunctions induced by seizures (Besson et al., 2014 Table S1 in the Supplementary Material section. The boxplots show z scores for each group of mTLE patients. We found significant differences between groups (p < .05) for several language and memory tests. The significant differences between patients are framed, namely: phonological and semantic fluency, AMI (verbal memory) and VMI ( connections outside the dysfunctional hippocampus has also been reported (Cataldi, Avoli, & de Villers-Sidani, 2013). These stronger remote functional connections may have a compensatory role for the loss of FC in other regions of the network, a phenomenon known as "dynamic diaschisis" (Campo et al., 2012). The understanding of the functional role of the reorganization patterns in terms of cognitive efficiency (compensatory role; "positive" or "negative" plasticity) is one of the most important and current challenges.
The second main objective of this study was to consider the cognitive efficiency of LMN-FC reorganization patterns in patients. Associations between brain connectivity and cognition can be highly convoluted and reliable biomarkers of the cognitive phenotype in the pathological condition in particular must be sought. The concept of "cognitome" we propose, close to the one of connectome, seeks to further highlight the search for the nature and typology of the links that may exists between the level of brain networks (hardwaresoftware) and the level of cognition (output  Regarding the temporo-mesial hyperconnected cluster, we found a significant negative correlation between the increased E nod value of the left hippocampus and the AMI score ( Figure 6, Panel b) for the L-mTLE group, even though the left hippocampus traditionally plays an important role in verbal memory (Richardson et al., 2004;Travis et al., 2014). More specifically for the L-mTLE, the higher integrative parameter values for the left hippocampus were associated with lower scores for verbal memory. Voets et al. (2014) have also shown FC increase between the ipsilateral hippocampus and the parahippocampal and enthorinal complex in TLE (left and right combined).
This abnormal connectivity of the hippocampus with parahippocampal and enthorinal regions was associated with poor performance on a memory-encoding task, in line with our findings. Thus, at rest, the hyperconnectivity of the hippocampus with other cortical areas and in particular with language and memory regions does not always seem to be functionally useful, which suggests a "negative" or inefficient plasticity in this case. Some GT parameters seems to be good biomarkers to explain the cognitive phenotype presented by patients and importantly, similar FC patterns can be observed even though the cognitive consequences are considerably discordant. This highlights once again, both the richness and complexity of the brain patterns that can underlie cognitive behavior.
The heterogeneity of the epileptic pathologies is one of the main sources of inconsistent or conflicting results in the literature. Several authors even proposed that refractory mesial temporal epilepsy is a particular entity (e.g., No et al., 2017). We included in this study only patients with a clear diagnosis of m-TLE in order to maximize the homogeneity of the patient samples and minimize the variability that may be related to the location of the epileptogenic zone. In addition, given the differences reported by the previous studies between epilepsies involving the left or the right hemisphere (Besson et al., 2014;Dinkelacker et al., 2016), we constituted two distinct matched groups of patients by systematically excluding patients who may had bilateral seizure foci. However, even if on average our two groups were equivalent in terms of hippocampal size and left-right asymmetry, it is likely that different subtypes of hippocampal damage may have an influence on brain connectivity (Bernhardt, Hong, Bernasconi, & Bernasconi, 2015). Based on the location (i.e., hippocampal subfields) and on histological patterns of neuronal loss and gliosis, the ILAE proposed an HS classification system (Blümcke et al., 2013;Thom, 2014). It appears that there is a common, but also distinct FC between the different parts of the hippocampus and the rest of the brain in healthy individuals (Vos de Wael et al., 2018). Thus, different macroscale network modifications may appear in m-TLE patients depending on the hippocampal subregions affected by sclerosis. Furthermore, other factors could have an impact on the connectivity in patients: gender and age (Ridley et al., 2015), handedness (Bettus et al., 2010), age of seizure onset, or pathology duration (e.g., van Dellen et al., 2009), antiepileptic drugs (Haneef, Levin, & Chiang, 2014;Vlooswijk et al., 2011) or interictal epileptic discharges (Ibrahim et al., 2014). Although we controlled for these factors, their significance and especially the effect of their interactions on durable modulation of the FC should be assessed in future studies.
Beyond the physiological noises that could contaminate the rs-fMRI signal used for FC analyses (Birn, 2012, for a review), the choice of network can also influence the metrics. A majority of studies use an a priori anatomical template (e.g., automated anatomical labeling MarsAtlas: Auzias, Coulon, & Brovelli, 2016). However, anatomically defined areas may involve different subregions with distinct functional roles, which make it difficult to interpret the FC results obtained at the regional nodal level (see  for a detailed description of the limitations of using an anatomical template). Therefore, in defining our LMN ROIs, we used the functional atlas AICHA; Joliot et al., 2015) that is directly based on rs-fMRI data from hundreds of individuals with brain regions delineated according to the homogeneity of the intrinsic activity. With regard to GT analyses carried out on structural connectivity data,  showed that the complexity of the network in terms of number of regions anatomically defined (AAL 82 areas vs. random-seed generated templates comprising between 100 and 4,000 regions) did not have an impact on the global GT parameters. However, the comparisons between findings in terms of local metrics such as path length and clustering coefficient were affected by the parcellation scale (Zalesky, Fornito, Harding, et al., 2010). Since our network is composed of 72 ROIs, this could partly explain the lack of results based on local efficiency compute at the nodal level. In addition, the ROIs size variations can affect the connectivity estimates (Salvador et al., 2008) and we have indeed found a low, but significant, positive correlation between the size of the LMN regions and the estimated FC coefficients ( Figure S3).

| CONCLUSION
In conclusion, this study could help our understanding of topological changes of the brain connectivity in temporo-mesial epileptic patients.
The study of resting-state FC through an embedded LMN reveals large and differential connectivity changes in patients in regions and hubs traditionally involved in language and in memory. Interestingly, the hippocampus (and more generally the regions near the problematic epileptogenic zone) at the heart of dysfunctions in temporomesial epilepsy is atrophied for most patients but seems to be overconnected to the rest of the network. This paradox is even more interesting given that we observed different patterns of correlations, some suggesting a compensatory and others a deleterious role of the LMN-FC plasticity, according to patient groups. Our findings provide additional insights into the several forms of neuroplasticity emerging in the context of repeated epileptic seizures. In the last few years, interest in network sciences in neuroimaging and cognitive neuroscience, as well as knowledge related to brain connectivity, have increased exponentially. We hope that future connectomic studies will not only focus on the brain connectivity patterns, but also more systematically on the consequences and implications of connectivity on behavior and cognition, expending thus the concept of "connectome" to "cognitome."

ACKNOWLEDGMENTS
The authors would like to thank Sophie Achard for significant methodological support and for editing the final version of this manuscript.

CONFLICT OF INTEREST
The authors have no conflict of interest to declare.

ETHICS STATEMENT
Participants (controls and patients) provided written informed consent to participate in the study, which was approved by the local ethics committee (CPP: 09-CHUG-14/ANSM [ID RCB] 2009-A00632-55).

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the Grenoble Hospital Center (CHUGA). Restrictions apply to the availability of these data, which were used under license for this study.
Data are available from the authors with the permission of the CHUGA.