Comparison of brain connectomes by MRI and genomics and its implication in Alzheimer’s disease

Data used in the preparation of this study were obtained from the ADNI database (http://​adni.​loni.​usc.​edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). There were 232 ADNI2 subjects and 621 ADNI3 subjects with both T1 and diffusion-weighted MRI images at baseline when we downloaded the data on October 2018 from the Laboratory of Neuro-Imaging (http://​adni.​loni.​usc.​edu) [11]. All images were converted from DICOM to NIFTI using DCM2NII software (University of South Carolina, SC, USA) [19]. All anatomical regions of interests (ROIs) were segmented using FreeSurfer 6.0, 64-bit version (Massachusetts General Hospital, MA, USA) [20]. The FreeSurfer pipeline included motion correction of volumetric T1-weighted imaged, stereotaxic space transformation, intensity non-uniformity correction, removal of non-brain tissue, tessellation of gray/white matter boundaries via surface modeling, automatic topology correction, and surface deformation followed by intensity gradient that optimally defined tissue borders where the greatest shift in intensity defined the transition into the other tissue. Image outputs were visually checked for each subject. Segmentation of ROIs was conducted based on “Desikan-Killiany” cortical atlas [21]. Diffusion-weighted images were preprocessed using FSL 5.0.10 (Wellcome Center, Oxford, UK) [22]. Diffusion imaging pipeline included brain extraction, susceptibility-induced distortion correction, eddy current and motion correction, individuals’ axial diffusivity (AxD), radial diffusivity (RD), mean diffusivity (MD), and fractional anisotropy (FA) estimation, and diffusion uncertainty map calculation using BEDPOSTX [23‐25]. Probabilistic tractography was performed using TRActs Constrained by UnderLying Anatomy (TRACULA), and 18 tracts were derived (Massachusetts General Hospital, MA, USA) [26]. The 18 tracts are forceps major, forceps minor, left anterior thalamic radiations (L-ATR), left cingulum—angular bundle (L-CAB), left cingulum—cingulate gyrus (L-CCG), left corticospinal tract (L-CST), left inferior longitudinal fasciculus (L-ILF), left superior longitudinal fasciculus parietal (L-SLFP), left superior longitudinal fasciculus temporal (L-SLFT), left uncinate fasciculus (L-UNC), right anterior thalamic radiations (R-ATR), right cingulum—angular bundle (R-CAB), right cingulum—cingulate gyrus (R-CCG), right corticospinal tract (R-CST), right inferior longitudinal fasciculus (R-ILF), right superior longitudinal fasciculus parietal (R-SLFP), right superior longitudinal fasciculus temporal (R-SLFT), and right uncinate fasciculus (R-UNC). For each tract, volume, average length, mean AxD, mean RD, mean MD, and mean FA were calculated.

Imaging data for ADNI2 and ADNI3 cohorts were processed under identical procedure. After all quality checks which include both systematic error checks and visual inspections, 593 out of 621 ADNI3 subjects and 220 out of 232 ADNI2 subjects successfully finished all imaging processes. Among 593 ADNI3 subjects, 550 subjects had qualifying clinical measures where age and gender were available and diagnosis record was within 60 days of scan date. Among 220 imaging processed subjects in ADNI2, 210 subjects had matching relevant phenotypes also within 60 days of scan date. Fifty-one subjects overlapped between ADNI3 and ADNI2 cohorts that had finished all the image processing and phenotype matched with scan date. Therefore, we omitted overlapping subjects from ADNI3 and analyzed 499 subjects for ADNI3 as discovery cohort and 210 ADNI2 subjects as the replication cohort. Association between AD diagnosis and diffusion measures of each tract (volume, average length, AxD, RD, MD, FA) was examined via linear model adjusted for age, gender, and total brain volume. Outliers that were 3.5 SD away from the mean were removed from the model, and all statistics were false discovery rate (FDR) adjusted for multiple comparisons. Effect sizes (β) of associations with AD (diffusion measure y = β*AD after adjusting covariates) was also calculated from the models.

The probability of tract reaching a Desikan-Killiany atlas defined grey matter ROIs at its white matter border was estimated. One voxel deep grey matter mask that is neighboring white matter was derived for all 82 ROIs and defined as ROI target masks. Path distribution for each tract was trimmed to include white matter and ROI target masks only. The probability of tract reaching ROI at its white matter boundary was calculated by dividing the number of paths passing through each voxel by the total path number in trimmed tract. Finally, the normalized probabilities within ROI target masks were averaged for all ROIs and this was repeated for each tract (Additional file 1: Figure S1A). These were defined as “reach probability” of tract connecting to grey matter ROIs (18 tracts towards 82 ROIs). The non-zero reach probability followed an extreme value distribution (Additional file 1: Figure S2A). We defined connections based on the empirical cumulative distribution function reflection point (reach probability = 0.002). At the cutoff, 203 of ROI-tract pairs were connected by a white matter tract. This effectively isolated some ROIs to specific tracts such as L-hippocampus was connected to L-CAB but not with L-SLFT or L-SLFP (Additional file 2: Table S1).

Post-mortem brain tissues curated by Mount Sinai Hospital were analyzed where gene expression for 17 brain regions limited to the left hemisphere were available [12, 13]. The transcriptome data was made up of maximum 63 subjects, and any two brain regions were shared by 30–51 subjects (Additional file 2: Table S2). The methods and cohort characteristics for this dataset have been described in detail [12, 13]. The 17 brain regions were frontal pole (FP), occipital visual cortex (OVC), inferior temporal gyrus (ITG), middle temporal gyrus (MTG), superior temporal gyrus (STG), posterior cingulate cortex (PCC), anterior cingulate cortex (ACC), parahippocampal gyrus (PHG), temporal pole (TP), precentral gyrus (PCG), inferior frontal cortex (IFC), dorsolateral prefrontal cortex (DLPFC), superior parietal lobule (SPL), prefrontal cortex (PFC), caudate nucleus (CN), hippocampus (HIP), and putamen (PUT) (Additional file 2: Table S2). The gene expression values were adjusted for age, sex, post-mortem interval, pH, ethnicity, and Braak staging scores. The adjustment removed potential batch-driven gene-gene correlations, such as both genes were up in an ethnic group or disease diagnosis group, but enhanced gene-gene correlations that were consistent among ethnic groups or disease diagnosis groups, and etc. Spearman correlation was used in identifying TTC gene pairs between 136 pairs of brain regions (17 × 16/2). Significant TTC gene pairs were identified at genome-wide threshold P-value < 1 × 10⁻⁸ as defined in previous TTC study [17]. Significant TTC gene pairs were counted for 136 brain region pairs (Additional file 2: Table S3).

Significant TTC signals were discretized as binary values, and Barber’s modularity was maximized which identifies two-mode networks of disjoint gene sets such that interaction only occurs with genes of another brain region [27]. LPAb+ algorithm outperforms other methods for bipartite networks [28, 29], and we utilized its two-stage procedure where first “bottom-up” step propagates labels iteratively to maximize node-by-node modularity and second “top-down” step joins modules together to increase network modularity [30]. Different random initialization of node selection was performed five times for all 136 ROI pairs and confirmed that the maximized modularity converged to same optimal solution. For each ROI pair, bipartite modules with more than 1000 interactions (TTC gene pairs) were selected and genes within modules were pooled for each tissue before conducting pathway enrichment analysis.

Curated pathways from Protein Analysis Through Evolutionary Relationships (PANTHER) database v.14.1 were analyzed [31]. Among 177 curated pathways available, eight pathways made up of drosophila-specific pathways were omitted (P06209, P06211, P06212, P06213, P06214, P06215, P06216, P06217). The Fisher exact test was performed to assess overrepresentation of our gene lists in each pathway, and all human genes (n = 20,996) were used as background. All pathway enrichment was corrected for FDR.

In order to infer broad biological insight from pathways overrepresented in the genes involved in TTC gene pairs, we created eight pathway categories that are biosynthesis, signaling, disease, physiology, development, gene regulation, metabolism, and catabolism for which detailed group identity for each pathway is listed in Additional file 2: Table S4. As post hoc analysis, after observing that a large number of associations were part of signaling pathway subgroup, we further divided signaling pathways into synaptic signaling, immune signaling, synaptic immune signaling, endocrine signaling, and unclassified (Additional file 2: Table S4). If the synthesized end product or degraded starting material served as a ligand in any synaptic, immune, or endocrine signaling, they were assigned as such (adrenaline and noradrenaline biosynthesis, aminobutyrate degradation, androgen/estrogen/progesterone biosynthesis, cobalamin biosynthesis, phenylethylamine degradation, vitamin B₆ metabolism, gamma-aminobutyric acid synthesis, histamine synthesis, vasopressin synthesis, vitamin D metabolism and pathway, bupropion degradation, nicotine degradation). Oxidative stress response (P00046) was categorized as immune signaling.

For a ROI pair, molecular pathways significantly enriched in genes involved in TTC gene pairs were binarized for each ROIs (FDR < 0.05) (Fig. 4). Binarized pathway associations were matched between ROI1 and ROI2 that are paired in bipartite clustering step and were transformed into adjacency matrix (Additional file 1: Figure S3A). This defines pathway interaction between ROI pairs. There were three types of ROI pairs: (1) ROI pairs not connected by tracts (not-bound), (2) ROI pairs bound by tracts (tract-bound), and (3) ROI pairs bound by AD-associated tracts (AD-tract-bound). Among 136 ROI pairs, there were 72 not-bound, 64 tract-bound, and 43 AD-tract-bound. Proportion of pathway interactions in each group was calculated by normalizing the summed adjacency matrices by the number of ROI pairs (Additional file 1: Figure S3B). The chi-square test was performed comparing the proportion of pathway interactions for the tract-bound and AD-tract-bound groups. Both comparisons were compared against the not-bound group (Additional file 1: Table S5–6). The P-values were −log₁₀ transformed and were hierarchically clustered using Ward’s method for further analysis.

The ADNI study collected whole blood samples for 811 subjects at baseline, which were processed using Qiagen PAXgene Blood RNA Kit (Germantown, MD, USA) [32]. Gene expression was profiled using Affymetrix Human Genome U219 Array (Affymetrix, Santa Clara, CA, USA) and was preprocessed using the Robust Multi-chip Average normalization method [33]. All quality check (QC) procedures were performed by ADNI Genetics Core which include RNA QC using Nanodrop and Agilent Bioanalyzer, overall array assay QC using Affymetrix Expression Console software and Partek Genomic Suite 6.6, sex verification, and sample identity prediction using Omni2.5 M genotype [32]. Quality-controlled transcriptome data was available for 744 subjects at 49,385 probe level and was downloaded from http://​adni.​loni.​usc.​edu. Blood transcriptome data was available for 102 ADNI2 subjects with successfully processed diffusion procedure and year at which PaxGene sample was collected matching with scan year. Toll receptor signaling pathway was represented by 49 genes spanning 129 probes in the transcriptome data [31], and probe-level expressions were collapsed to gene-level using mean-max method [34]. The association between diffusion measures for each tract and gene expression was examined according to the following linear model: Diffusion measure~α + Gene Expression + Sex + Age + RIN + (1| AffyPlate) + ε where α is intercept, ε is random error, and RIN is RNA integrity number. The aggregate effect of 49 Toll receptor signaling genes unto each diffusion measures was determined using sum of χ² method [35, 36], and its significance was evaluated by 100,000 permutations (P_permutation < 0.05).

All statistical analyses were performed using Julia 1.0.3 (MIT, Cambridge, MA) [37]. The networks were visualized using spring-affinity algorithm. All heatmaps were drawn in R using Ward’s method for hierarchical clustering (R Core Team, Vienna, Austria) [38].

The ADNI3 cohort (n = 449, the “Methods” section) [11], consisting of 347 healthy controls, 118 mild cognitive impaired (MCI), and 34 AD patients, was interrogated for characterizing diffusion measures in 18 tracts derived using TRACULA [26] (Table 1). The brain volumes were positively correlated with MD and RD in all tracts (Additional file 1: Figure S4). RD was more significantly correlated with volume than MD. The average length of tracts was negatively correlated with MD and RD where MD was more correlated with tract average lengths than RD. In all 18 tracts of interest, AxD was positively correlated with FA and RD was positively correlated with MD.

The diffusion measures were compared against age, sex, years of education, marriage status, APOE4 genotype, and total brain volume (TBV) and disease diagnosis (Fig. 1a). Age was significantly associated with AxD, RD, and MD in all tracts (P-values = 2.2 × 10⁻⁵ ~ 2.4 × 10⁻¹⁷) and with FA in a subset of tracts (16 out of 18 tracts below P-value < 0.05, P_min = 4.2 × 10⁻¹²), consistent with reports in literature [39]. TBV was associated with RD, MD, and FA in a number of tracts. Disease diagnosis status was associated with AxD, RD, and MD in a large number of tracts similar to findings in other studies [40]. After adjusting age, sex, and TBV effects, only disease diagnosis status remained significantly associated with the diffusion measures (Fig. 1b). Among all diffusion measures in 18 tracts, we identified 34 significant disease associations in a data-driven manner with AxD, RD, MD, and FA in a number of tracts at FDR < 0.05 (Table 2). There was no AD diagnosis status association with bilateral CST and FMajor which are responsible for motor and visual functions.

The ADNI2 cohort (n = 210, the “Methods” section) [41], consisting of 75 healthy controls, 91 MCI, and 44 AD patients, is a cohort independent from the ADNI3 cohort (the “Methods” section). The same 18 tracts were derived using TRACULA [26]. The similar inter-relationships among diffusion measures, covariates, and disease diagnosis were observed (Additional file 2: Table S7). After adjusting age, sex, and TBV effects, 11 diffusion measures were significantly associated with disease diagnosis at FDR < 0.05 (Table 2, right) involving 4 of 18 tracts, bilateral CABs, bilateral SLFTs, L-SLFP, and L-ILF. Note that not only all the 11 associations overlapped with the 34 associations identified in ADNI3 cohort, but the direction of measure changes in response to diagnosis status was also replicated (Fig. 2). AxD, RD, and MD increased while FA decreased with disease diagnosis status (Fig. 2). The effect sizes and directions in ADNI2 and ADNI3 cohorts for the 34 associations identified in the ADNI3 cohort and the 11 replicated associations exhibited higher effect sizes than non-replicated associations (Fig. 3, Additional file 2: Table S8), suggesting a larger sample size is needed for replicating these associations of small effect sizes.

We extracted path distribution information from each tract and calculated the probability of a tract reaching any Desikan-Killiany defined grey matter ROI [21]. Throughout this study, these measures were referred to as “reach probability.” Reach probability was developed to allow focusing on only major tracts and integrative analysis between neuroimaging and transcriptome datasets. The reach probability was limited to white matter boundary neighboring each respective ROI in order to avoid amplifying the connection to ROI based on within-ROI streamline propagations in the tractography processes. Reach probability was derived for 18 tracts towards 82 ROIs, and 607 out of 1476 (18 × 82) probabilities had zero reach probability (41%) (Additional file 2: Table S1). The non-zero reach probability followed an extreme value distribution (Additional file 1: Figure S2), and 203 ROI-tract pairs (33.4%) were identified (the “Methods” section).

In order to investigate molecular connections between brain regions, we analyzed tissue-tissue co-regulation [17, 18] of transcriptomic data covering 17 post-mortem brain regions (Fig. 4) [12, 13]. Subjects shared for each pair of brain regions were in the range of 30–51 subjects depending on post-mortem tissue availability. There were 136 (17 × 16/2) possible brain region pairs among 17 brain regions. Brain connectome is defined by TTC of all gene pairs after adjusting Braak score in order to examine consistent gene synchronization between brain regions with regard to different disease diagnosis groups. TTCs were adjusted for covariates such as age, sex, post-mortem interval, pH, and race [17] (detailed in the “Methods” section). The strength of brain region-region connections was measured by the number of significant TTC gene pairs. The distribution of significant TTC gene pair counts is shown in Fig. 5a for 136 region pairs, suggesting that only a fraction of brain regions were synchronized at the transcriptional level.

The post-mortem brain regions were originally labeled according to the Brodmann area map [12] and were matched to brain regions in Desikan-Killiany cortical atlas [21] that was used in the above tractography analysis (Additional file 2: Table S2). Among 136 possible brain region pairs, 64 region pairs were connected by white matter tracts defined by tractography. Among 64 tract-bound region pairs, 28 pairs were connected by AD-associated tracts: L-CAB, L-ILF, L-SLFT, and L-SLFP. Among the top 10 percentile of region pairs (n = 14) containing the highest number of significant tissue-tissue correlated gene pairs, 10 were tract-bound (Fisher’s exact test, P = 0.057; Fig. 5b), and 7 out of top 10 percentile of region pairs (n = 14) were bound by AD-associated tracts (Fisher’s exact test, P = 0.03; Fig. 5b). This suggests that the brain connectomes defined by two different approaches overlap, especially for connections related to AD.

Different brain regions were connected by white matter tracts and synchronized at the transcriptional level as shown above. To investigate whether any biological pathways were transcriptionally synchronized between brain region pairs, we constructed bipartite clusters of TTC gene pairs for all 136 ROI pairs and identified gene modules for each brain region in ROI pairs (Fig. 4). Genes in the modules were annotated using PANTHER database [31], and pathways enriched among these genes at FDR < 0.05 are listed in Additional file 2: Table S9. Among 169 pathways × 136 ROI pairs (22,984), 736 (3.2%) pathways to ROI pair associations were significant, covering 83 pathways and 69 ROI pairs (Fig. 6a). A large fraction (51/83 = 61.4%) of enriched pathways belonged to signaling pathways (Fisher exact test, P = 6.08 × 10⁻⁹, Fig. 6a). The 69 ROI pairs were clustered to 3 clusters according to enriched pathways (Fig. 6a). The ROI pairs in cluster I (Fig. 6a), which were connected by multiple pathways, were enriched for tract-bound ROI pairs and AD-associated tract-bound pairs (P = 0.04 and 0.01, respectively). This suggests that white matter tracts may serve as a mechanism of gene synchronization for signaling pathways, at least in brain regions present in ROI cluster I (PHG-TP, PHG-STG, ITG-MTG, ITG-PHG, IFC-ITG, PFC-STG, IFC-ITG).

Next, we examined whether any pathways were preferentially involved in TTCs of tract-bound or AD-associated tract-bound ROI pairs (Fig. 6b, detailed in the “Methods” section). Apoptosis signaling pathway (P = 0.006), EGF receptor signaling (P = 0.046), and metabotropic glutamate receptor (mGluR) I pathways (P = 0.049) were overrepresented in tract-bound region pairs (Additional file 2: Table S9). More pathways were preferentially involved in TTCs of AD-associated tract-bound ROI pairs, including apoptosis signaling (P = 0.008), muscarinic acetylcholine receptor (mAChR) 2 and 4 signaling pathway (P = 0.01), valine synthesis (P = 0.012), PI3 kinase pathway (P = 0.013), endothelin signaling pathway (P = 0.016), histamine synthesis (P = 0.016), p38 MAPK pathway (P = 0.02), mGluR I pathway (P = 0.024), inflammation mediated by chemokine and cytokine signaling pathway (P = 0.028), mGluR II pathway (P = 0.035), toll receptor signaling pathway (P = 0.045), adenine and hypoxanthine salvage pathway (P = 0.046) (Additional file 2: Table S9). The results suggest that signaling pathways, especially synaptic signaling and immune signaling pathways, involve in transcriptional synchronization between brain regions connected by white matter tracts.

Gene modules derived from bipartite clustering TTC gene pairs were enriched in specific molecular pathways, mostly related to signaling. However, biological pathways in one region may not reciprocally synchronize the same pathway in another brain region because each region is accountable for their own distinct roles. For instance, our analyses show that CN and ACC were structurally connected (Additional file 2: Table S1) and literature supports that they are functionally connected [42]. However, CN and ACC are enriched in different neuron types (dopaminergic [43] and spindle neurons [44], respectively) and are responsible for different biological processes that may be mediated by differing molecular functions. Using 169 curated pathways as generalizable domains of molecular functions [31], we investigate how pathways are differentially interacting between brain regions. We defined pathway interactions specific to tract-bound ROI pairs based on the chi-square test (P < 0.05, Additional file 2: Table S5) and the same analysis was performed for AD-associated tract-bound ROI pairs (Additional file 2: Table S6). Both pathway networks of tract-bound ROI pairs (G1) and AD-associated tract-bound ROI pairs (G2) were made up of nodes that are signaling related (Fig. 7a, b). G2 had a larger number of pathway interactions than G1 (Fig. 7c), and the node with the most number of edges in G2 was toll receptor signaling pathway (Additional file 2: Table S10). The top two nodes with the highest number of edges in the G1 were ionotropic glutamate receptor (iGluR) pathway and toll receptor signaling pathway (Fig. 7c).

Alternatively, TTC genes in ROI pairs may be involved in shared molecular functions (symmetric synchronization). We examined pathways associated in both brain regions in a pair (Additional file 2: Table S11). For genes in TTC gene pairs of tract-bound ROI pairs, iGluR pathway (P = 0.021), toll receptor signaling pathway (P = 0.021), inflammation mediated by chemokine and cytokine signaling pathway (P = 0.047) were significantly overrepresented (Fig. 7d). For genes in TTC gene pairs of AD-associated tract-bound region pairs, toll receptor signaling pathway (P = 0.003), iGluR pathway (P = 0.013), inflammation mediated by chemokine and cytokine signaling pathway (P = 0.031), PI3 kinase pathway (P = 0.031), mGluR group III pathway (P = 0.033), endothelin signaling pathway (P = 0.036), mGluR group II pathway (P = 0.036), mGluR group I pathway (P = 0.036), T cell activation (P = 0.036), 훽3 adrenergic receptor signaling pathway (P = 0.036), and mAChR 1 and 3 signaling pathway (P = 0.044) were preferentially involved (Fig. 7d).

Immune activities in the blood may reflect molecular states in the brain [45]. Because toll receptor signaling pathway was the most enriched pathway involved in symmetric synchronization between AD-associated tract-bound ROI pairs (Fig. 7d), we interrogated how toll receptor signaling-related genes’ expression in the blood associated with diffusion measures in the brain. There were 102 subjects with both blood expression data and dMRI scans in ADNI2 (the “Methods” section). We examined the pooled effect of 49 genes representing toll receptor signaling pathway [31] on diffusion measures of 18 tracts using sum of the chi-square method and compared them with the inferences based on 100,000 permutations [35, 36]. Multiple diffusion measures including AxD of R-ATR (P = 1.0 × 10⁻⁵), R-CCG (P = 1.0 × 10⁻⁵), L-UNC (P = 5.6 × 10⁻⁴), L-CCG (P = 3.5 × 10⁻³), L-ILF (P = 7.5 × 10⁻³), and R-SLFT (P = 7.6 × 10⁻³) were significantly associated with expression of genes in the toll receptor signaling pathway in the blood (Fig. 8a, Additional file 2: Table S12). RD (which measures diffusivity orthogonal to AxD) of forceps minor (P = 5.6 × 10⁻³) and L-CAB (P = 2.0 × 10⁻²) was significantly associated with the expression of toll receptor signaling-related genes in the blood (Fig. 8c). MD which captures diffusivity in all directions was associated in forceps minor (P = 7.6 × 10⁻³), bilateral CCGs (P = 9.7 × 10⁻³ and 4.7 × 10⁻³ left and right respectively), L-CAB (P = 3.4 × 10⁻²), L-SLFP (P = 1.5 × 10⁻⁵), and L-SLFT (P = 2.7 × 10⁻²) (Fig. 8b). FA which describes white matter integrity was associated in bilateral CABs (P = 1.0 × 10⁻² and 4.9 × 10⁻⁴ left and right respectively) and R-CCG (P = 4.7 × 10⁻²) (Fig. 8d). Similar to diagnosis associations (Table 2), we did not observe any toll receptor signaling gene expression association with bilateral CST and forceps major which are responsible for motor and visual functions.