Background
Neural connections comprising a nervous system are often described in complicated anatomical networks. Much of the human brain connectome has been assessed using magnetic resonance imaging (MRI) where functional MRI and diffusion MRI (dMRI) can measure correlated neural activity and structural connectivity of the brain in vivo, respectively [
1,
2]. Various neurological diseases such as Alzheimer’s disease (AD) are associated with disruption of the brain connectome and studies show that the course of AD continuum is associated with the changes in brain network architecture [
3‐
5]. Although our knowledge regarding the connectome changes in AD is abundant, understanding the molecular consequences or causes of brain connectome changes is lacking.
Gene expression signatures carry important information for understanding structural and functional brain connectivity. It has been shown that the connectivity in rodent brains can be predicted from mouse brain expression data [
6,
7]. Brain connectivity based on blood-oxygen-level-dependent signals at a resting state is significantly associated with correlations between gene expression of human brain segments [
8]. However, the transcriptomic studies of AD are often limited to isolated brain regions such as the hippocampus or dorsolateral prefrontal cortex alone and are difficult to interpret its findings in respect to the brain connectome when its relation is not examined together [
9,
10]. The mechanisms of how one brain region impacts molecular pathways in other regions, especially how the brain regions susceptible to AD pathology interact with each other at the transcriptome level, have not been systematically studied.
Here, we performed imaging-transcriptomic study analyses of brain connectomes based on dMRI imaging data from Alzheimer’s Disease Neuroimaging Initiative (ADNI) and a brain transcriptome dataset covering 17 brain regions [
11‐
13]. Unlike traditional imaging genetic association analyses, where the goal is to identify the relationship between genetic variation and the changes in neurological traits [
14,
15], the analyses here focused in spatial correlations between gene expression and structural brain connectivity. We hypothesize that different brain regions are synchronized at the molecular level (genomic connectome), partially facilitated by white matter tracts (structural connectome). Dysfunction of genomic connectome may associate with neurological diseases and reflect genetic propensity underlying AD etiology. To test our hypothesis, we (1) identified white matter tracts associated with AD based on dMRI and replicated them in an independent cohort [
16], (2) identified brain regions connected by white matter tracts, (3) compared structural brain connections and genomic brain connections defined as tissue-to-tissue correlations (TTCs) at the transcription level, and (4) identified biological pathways involved in TTCs in structurally connected brain regions [
17,
18].
Discussion
Tissues, organs and cell groups within organs, communicate with one another to perform biological functions in concert, and gene transcriptions are synchronized between tissues reflecting cross-tissue and cross-cell-group communications [
17,
18]. In the brain, white matter tracts serve as an important medium of brain regional cross-talk [
46‐
48], and we observed that large numbers of genes were synchronized at the transcriptional level in tract-bound brain regions (Fig.
5b). Gene modules derived from bipartite clustering of TTC gene pairs between tract-bound brain regions were significantly over-represented in signaling pathways (Fig.
6). Since axon bundles with synaptic connections constitute white matter tracts, identifying associations between mAChR, mGluR, and iGluR signaling pathways and TTC gene pairs were within our expectations (Fig.
6b). Toll receptor signaling pathway was the most enriched pathway in the symmetric gene synchronization between AD-associated tract-bound brain regions (Fig.
7). There are at least two potential mechanisms: (1) Toll-like receptor (TLR) signaling plays a role in brain region-to-region communication via white matter tract and (2) TLR signaling pathways in brain regions and in the blood are synchronized [
49]. The association between diffusion measures in major tracts and toll receptor signaling pathway activity in blood convolutes the two potential mechanisms. Although the mechanism is not clear, our results suggest the immune system’s involvement in AD-associated brain region-to-region cross-talk.
TLRs play important roles in innate immunity in humans, and TLR activation in microglia due to neuropeptide aggregation is well established [
50,
51]. However, the expression of TLRs is not limited to microglia [
52,
53], but is also present in astrocytes [
54], oligodendrocytes [
55], neural progenitor cells [
56,
57], and neurons [
58]. The biology of TLRs is complex and goes beyond just recognizing pathogen-associated molecular patterns [
59]. TLR3 can recognize double-stranded RNA for its activation [
60], and the signaling cascade of TLRs varies for different neuronal cell types [
61]. TLR2 and TLR4 are known to regulate hippocampal adult neurogenesis and neural progenitor cell differentiation [
62]. TLR3 is associated with increased mature neurons in the hippocampus and enlarged dentate gyrus and the CA1 region [
56]. TLR3 and TLR8 are present in the axonal tracts during the brain development and regulate neurite outgrowth and apoptosis [
63‐
65]. In addition, differential expression of TLRs in human post-mortem brains are associated with alcohol addiction [
66], depression [
67,
68], and schizophrenia [
69], and these neurological disorders are also associated with white matter abnormalities [
70‐
72]. However, it is not known how TLRs may act on axonal degeneration and cross-communication between brain regions via axon fibers
.
Diffusion-weighted imaging is a powerful tool in assessing microstructural changes of white matter in vivo, and diffusion parameters can capture white matter integrity [
1]. In our work, TLR signaling expressions were associated with FA in bilateral CABs (Fig.
8). Because CABs have a strong connection to the hippocampus, white matter integrity measured by FA may be regulated by TLR signaling in the hippocampus and TLR-dependent adult neurogenesis [
62]. AxD estimates parallel diffusivity along the direction of the highest diffusion and was significantly associated with expression of TLR signaling for bilateral-CCG, L-UNC, R-ATR, L-ILF, and R-SLFT. This suggests that TLR signaling may be involved in the loss of barriers restricting water diffusion in the associated tracts such as myelination level reduction or axon losses [
73‐
75]. Although the association between diagnosis and diffusion measures in L-ILF and R-SLFT was replicated in the ADNI2 cohort, L-UNC, R-ATR, and R-CCG findings failed to replicate in the ADNI2 cohort (Table
2). L-CCG was only nominally significant (FDR < 0.1) in both ADNI3 and ADNI2 cohorts (Table
2). This suggests that expression variation of genes in the TLR signaling pathway might be more powerful in detecting microscopic white matter abnormalities in comparison to diagnosis status, and further study may allow developing blood biomarkers relevant to disease-associated white matter changes in vivo.
The sample size of ADNI3 was larger than the size of ADNI2 so that the ADNI3 study had a higher power to identify AD associations in diffusion imaging and not all associations were expected to be significant in the ADNI2. Besides the sample size, there were technical differences between the two cohorts [
39,
76]. ADNI2 data was collected using older MR pulse sequence and was captured at 2.7-mm
3 resolution. ADNI3 adopted the optimized protocol established by Human Connectome Project as the standard across multiple centers and gained higher resolution at 2.0 mm
3 [
11]. There were 16 and 50 research sites involved in ADNI2 and ADNI3 studies, respectively. Four hundred nine out of 499 images in the ADNI3 dataset were acquired from 37 research sites that were not included in the ADNI2 (Additional file
2: Table S13). The results from the multi-center studies are unlikely due to biases from a few sites. As noted in the “
Methods” section, we included only imaging data of participants that were unique to ADNI3 as the ADNI3 cohort so that there was no overlap between the ADNI2 and ADNI3 cohorts in our analyses. The identified imaging-based disease associations were also consistent with known findings [
77,
78]. All these results together suggest that the associations between neuroimaging features and AD are robust to the differences between ADNI3 and ADNI2. Additionally, the replicated associations had larger effect size than the non-replicated ones, suggesting associations of smaller effect sizes require a larger sample size to validate.
There are limitations in our analyses and ADNI studies in general. Majority of the participants in the ADNI2 and ADNI3 studies were white (91.4% and 93.6%, respectively). Even though some common associations between neuroimaging features and AD were identified in ADNI2 and ADNI3 cohorts, whether the associations hold in other ethnic groups needs further studies. Additionally, there were only 17 brain regions available to construct transcriptome-based brain connectome. The limited spatial resolution of this work may increase false negatives. The Allen Human Brain Atlas has more complete coverage of the brain spatially [
79], but is limited to only 6 individuals whereas we conducted our study using 30–51 subjects depending on the brain region. Although spatially limited, our work is much better powered than the Allen Human Brain Atlas in examining correlated expression between brain regions and should better reflect the population information. Another limitation is that our study only examined gene synchronization by major white matter tracts whereas gene synchronization between two brain regions may be mediated through multiple mechanisms, including (1) direct neighbor (cis), (2) WM connected (trans), and (3) functionally connected (multi). Future works are needed to address these different gene synchronization models.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.