Population

In our cohort (n = 134), the studied non-synonymous coding single-nucleotide polymorphism (rs6265) was in Hardy-Weinberg equilibrium ( = 3.25, p = 0.07) with genotypic frequency of 0.6 (G/G), 0.31 (G/A) and 0.09 (A/A). The final study population comprised 36 healthy subjects aged 18–25. Fifteen (9 male) were identified as carrying the Met allele. The remaining 21 (9 male) were homozygotes for the Val allele and were referred to as the Val/Val group. The groups did not vary significantly in IQ or age, nor did their scores differ for a battery of psychological tests (Table S1 in File S1).

Network-based statistics

In our networks, with 1015 nodes and an average of 66,456 edges, we identified 387 connections in which the number of connecting tracks was significantly greater in carriers of the Met allele than in the Val homozygotes (p = 0.0122, permutation testing). The relative connection strengths for these edges are shown in Fig. 1a.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Significantly increased regional connectivity and topological changes in Met carriers.

(a) Track count increase for each connected edge in Met carriers (n = 15) versus Val/Val subjects (n = 21). (b) Region-to-region track pathways that are present only in the Met carriers.

https://doi.org/10.1371/journal.pone.0069290.g001

For Met carriers the strength at these edges was found to range between 1.75 and 48 times their strength in Val/Val. Of these edges, 41 (11) were found to have between 75 and 200 more tracks in Met carriers than in Val homozygotes. Met carriers had 200 to 400 more tracks in 123 (32) of the edges, 400 to 900 more tracks in 104 edges (27), and even greater factors in the remaining 23 edges (6). The affected edges were largely central connections and were not short or uncommon fiber pathways.

Roughly one quarter (96) of the edges that were identified were not present in any of the Val/Val subjects (i.e. the mean value in Met carriers was significantly greater than the value of zero, found in Val homozygotes). Fig. 1b shows the mean number of tracks for the 96 edges that were only present in the Met carriers. The connections unique to Met carriers appeared consistently across the group. We did not find any edges with significantly lower strength in Met carriers. The identified connectivity changes are unlikely to represent false positives because of the stringent non-parametric statistical method [11]. Moreover, the reported differences were specific to the BDNF polymorphism; subjects were also divided by gender (18 F, 18M), and by their adenosine deaminase (ADA) genotype (17 GA, 19 GG), and no significant results were obtained.

Global network metrics (graph density, number of connected components, transitivity) showed no variation between groups. Local nodal metrics (degree, clustering coefficient, number of triangles, [closeness, betweenness, degree] centrality, highest k-core number) were averaged for each participant and also did not vary. Wiring cost and network efficiency, compared both for the whole network as well as for only corticocortical connections, were unaffected by BDNF genotype. The total number of tracks per connectome, out of the generated 300,000 per subject, did not differ. The lack of significant variation in any of the network metrics is understandable because the total number of altered edges (387) is less than 1 of the mean number of edges (66,456) per network.

Classifier performance

The classifier was able to discriminate between Val homozygotes and Met carriers with 86.1 global accuracy. The predictive value for the Val/Val and Met carrier groups were 94.4 (p = 0.001) and 77.8 (p = 0.003), respectively. In Figure 2 the weights obtained by the classifier are visualized as edges in the brain network. For the classifier trained to identify gender, the global accuracy reached 63.9 (n.s.). Identifying the subjects' adenosine deaminase (ADA) genotype was only possible with an accuracy of 58.3 (n.s.).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Classifier weight distribution.

The weights obtained by the classifier have been plotted as network edges in order to show their spatial distribution. The thresholding procedure removed 99.75 of the edges for clarity. The remaining connections represent 21.69 of the absolute weight.

https://doi.org/10.1371/journal.pone.0069290.g002

Tractographic basis

Structural connectivity in Met allele carriers was found to be higher throughout the forebrain (Figure 3). Large increases were found in connections corresponding to the anterior and superior corona radiata, including the corticothalamic and corticospinal projections from the sensorimotor, premotor, and prefrontal portions of the internal capsule. General interhemispheric connectivity was increased via the corpus callosum and anterior commissure. Extremely high connectivity values were found between inferior medial frontal polar regions via the anterior forceps. The Met carriers also presented novel connections within the cingulum, corpus callosum, and anterior forceps, which were not found in the Val homozygotes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Connecting tracks in Met carriers.

The average number of tracks connecting two regions (i.e. the edge weight) in Met carriers was found to range from 1.75 to 48x the value found in the Val/Val group. The range was so broad that it had to be analyzed in separate stages. Fibers shown are filtered from a single Met carrier.

https://doi.org/10.1371/journal.pone.0069290.g003

Ethics Statement

Volunteers were recruited through advertisement on the University intranet. They gave their written informed consent to participate in the study, which was approved by the Ethics Committee of the Faculty of Medicine at the University of Liège.

Population

Participants were young (18–25 years old), healthy, and lean (Body Mass Index 26). They were all right-handed, as determined by the Edinburgh Inventory [26]. None complained about sleep disturbances, and this was reflected by the Pittsburgh Sleep Quality Index (PSQI score 6) [27]. Extreme chronotypes, according to the Horne and Ostberg morningness-eveningness questionnaire, were excluded (scores 31 or 69) [28]. Their sleep midpoint on free days was required to be between 3 and 5.99 as indicated by the Munich Chronotype Questionnaire [29]. They all scored in the normal range (0–9) on the Epworth Sleepiness Scale [30]. The absence of medical, traumatic, psychiatric, and sleep disorders was established through a semi-structured interview.

All participants had normal scores on the 21-item Beck Anxiety Inventory (score 11) and the 21-item Beck Depression Inventory-II (score 14) [31], [32]. They were non-smokers and moderate caffeine and alcohol consumers. None were on medication other than oral contraceptives. No caffeine was allowed during the experiment.

Volunteers complying with these criteria were invited to perform Raven's Progressive Matrices and a blood sample was obtained for BDNF genotyping [33]. Participants were eventually selected based on their BDNF genotype. Allelic groups were formed with participants that were matched according to gender, age, education level, chronotype, PSQI score and IQ (Table S1 in File S1). Subjects received financial compensation for their blood test and participation in the study.

Genotyping

Genomic DNA was extracted from blood samples using a MagNA Pure LC Instrument. The DNA sequence of interest was amplified by Polymerase Chain Reaction in a final volume of 50 l containing 0.6 M of each primer (Thermo Scientific, Germany), 0.5 l Faststart Taq DNA Polymerase (Roche Diagnostics, Germany), 0.8 mM of each deoxynucleotide triphosphate (Roche Diagnostics Germany) and 20 ng of genomic DNA. After 10 min of denaturation at 95°C, samples underwent 35 cycles consisting of denaturation (95°C, 30 sec), annealing (60°C, 40 sec), and extension (72°C, 30 sec), followed by a final extension of 7 min at 72°C. The amplified DNA product was then subjected to pyrosequencing (Pyromark Q96 Vacuum Workstation, PSQ 96MA, Pyromark Gold Q96 Reagents, Qiagen, Germany). The sequences of the primers are available upon request.

Data Acquisition

Data was acquired on a 3 T head-only scanner (Magnetom Allegra, Siemens Medical Solutions, Erlangen, Germany) operated with the standard transmit-receive quadrature head coil. A high-resolution T1-weighted image was acquired for each subject (3D modified driven equilibrium Fourier transform, repetition time  = 7.92 ms, echo time  = 2.4 ms, inversion time  = 910 ms, flip angle  = 15°, field of view  = 256×224×176 mm, 1 mm isotropic spatial resolution). Seven unweighted (b = 0) volumes were acquired followed by a set of diffusion-weighted (b = 1000) images using 61 non-collinear directional gradients.

Processing & Analysis

The processing workflow was developed in Python and imports modules from the Nipype project [34]. The pipelines used for both single subjects and groups have been detailed as part of the online Nipype documentation in order to improve transparency and promote reproducibility. Every piece of software (CMTK, ConnectomeViewer, Dipy, Freesurfer, FSL, Nipype, Nibabel, MRtrix) used to process data in this paper is currently operating under an open-source license. The process began by segmenting the structural MR images using the automated labeling of Freesurfer [15]. Segmented structural images were then further parcellated using the Lausanne2008 atlas for a total of 1015 regions of interest (ROIs) [14]. Diffusion-weighted images were corrected for image distortions arising from eddy currents using linear coregistration functions from the FMRIB Software Library (FSL) [35]. Fractional anisotropy maps were generated, and a small number of single-fiber (high FA) voxels were used to estimate the spherical harmonic coefficients of the response function from the diffusion-weighted images [36], [37]. Using non-negativity constrained spherical deconvolution, fiber orientation distribution (FOD) functions were obtained at each voxel. For our dataset with 61 directions, we used the maximum allowable harmonic order of 8 for both the response estimation and spherical deconvolution steps. Probabilistic tractography was performed throughout the whole brain using seeds from subject-specific white-matter masks and a predefined number of tracts. Fiber tracking settings were as follows: number of tracks  = 300,000, FA/FOD amplitude cutoff for terminating tracks  = 0.1, minimum track length  = 10 mm, maximum track length  = 200 mm, minimum radius of curvature  = 1 mm, tracking algorithm step size  = 0.2 mm. Using tools from Dipy (Diffusion in Python, http://nipy.sourceforge.net/dipy/), the tracks were affine-transformed into the subject's structural space. This procedure circumvents the common problem of having to downsample ROI image files – defined in structural space – so that they can be used in diffusion space for connectivity mapping, and therefore leads to more accurate connectomes. Connectome mapping was performed by considering every contact point between each tract and the outlined regions of interest. Unlike in some past papers (e.g. [14], [16]) which considered only fiber start and endpoints, we incremented our connectivity matrix every time a single fiber traversed between any two regions. This leads to a far denser network than we have seen before, presumably with more accurate network properties. The number of tracked fibers which remained in each subject's connectome was also recorded.

This method of connection mapping may need further optimisation, however, as it can potentially be linking gray matter cortical regions through unreliably tracked fibers. This is something that may be avoidable by placing limits on the propagation parameters, or with anatomically or otherwise constrained tractography approaches [38], [39]. The benefits and drawbacks of different mapping techniques should be explored by future studies.

Numerous network metrics were obtained for each connectome and compared at the group level. At the nodal level we calculated the degree, clustering coefficient, and number of triangles, as well as three measures of centrality (closeness, betweenness, degree), and the highest k (i.e. degree) value for each k-core the node is encompassed by. For the network as a whole we computed the average shortest path length (i.e. the inverse of efficiency), the wiring cost (using Euclidean distance between nodes), the graph density, the number of connected components, and the graph's transitivity [40]. For a complete description of all of these metrics, the reader is referred to the Python package NetworkX [41]. Tract and network visualization were performed in TrackVis (Ruopeng Wang, Van J. Wedeen, TrackVis.org, Martinos Center for Biomedical Imaging, Massachusetts General Hospital), MRtrix, and ConnectomeViewer [42]. Figure S2 provides, for visualization, the orientation distribution functions and generated fiber tracts for a midbrain coronal slice of a single subject. In Figure S3, the structural connectome and T1-weighted image are shown for the same subject, and thresholded across two distinct fiber-count ranges, so that both the core and the density of the network can be seen.

Statistical Analysis

The Network-based Statistic (NBS) was used to identify differences between BDNF allelic groups (Fig. S1a) [11]. For each permutation the t-values at each edge were thresholded above a value of 3. The supra-threshold components were then compared against the generated null distribution. The null distribution for each test was produced by permuting members of each population 5000 times and estimating the maximal component size.

A table describing a representative subject's connection matrix and edge weights is given in Table S2 in File S1. Since the networks in this study have a high number of regions, and we have performed whole-brain connectome mapping with a relatively low number of fibers, a large proportion of our network's edges have low fiber counts. This may be problematic for statistical testing with the NBS because these small-integer populations do not provide wide ranges for edge weights and can result in inaccurate t-values. In future, it may be prudent to generate a larger number of streamlines, reduce the number of nodes in the network, or restrict analysis to specific parts of the brain. Practically, it can be computationally intensive to deal with large streamline datasets and networks with high numbers of nodes. The trade-off between resolution and resources is something that must be decided by the researcher with the focus of the study in mind.

In Fig. S1b we projected the observed NBS component onto the tractography of a single subject. This projection is, in effect, a type of reverse connectome-mapping. Given the connectivity network, we filtered the set of tracts to show only those that traverse between regions with edges in the network. Global graph metrics, psychological metrics, and the total number of fibers per connectome, were compared directly between allelic groups via Student's t-test. Nodal measures were averaged for each subject and analyzed in the same manner. All distributions were plotted as combined histogram/kernel-density maps to evaluate gaussianity prior to statistical analysis. Apart from the results given by the network-based statistic, no significant differences were identified between the two genotypic groups for any of the graph-level measures. No significant differences were observed between allelic groups in any of the psychological metrics.

Classification

The multivariate statistical properties of our data were studied with a linear Gaussian Process Classification method [43] as interfaced by PRoNTo (Pattern Recognition for Neuroimaging Toolbox, http://www.mlnl.cs.ucl.ac.uk/pronto) [44]. The classifier was given the fiber-count connection matrices for each subject and their true classes (e.g. Met carrier, Female). No network metrics, topology, or spatial information was provided to the classifier. The accuracy and generalisability of the classification were assessed with a leave-one-out cross-validation procedure: one subject is left out at a time, the classifier is trained on the remaining data, and the true and predicted (by the trained classifier) classes of the left-out subject are compared. With this linear kernel method weights were also obtained indicating the contribution to the classification output (in favor of either genotypic group) of each edge in the network. The same method was employed to discriminate features related to the subjects' gender and genotype for the ADA gene. For the purposes of visualization, we thresholded the edges in Figure 2. The removed portion of the classification weights can be found in Fig. S4. Example calculations for the percent classification weight represented by the remaining edges can be found in the Supplementary Information in File S1.