Perturbed functional networks in Alzheimer’s Disease reveal opposing roles for TGIF and EGR3

While Alzheimer’s disease (AD) is the most prevalent cause of dementia, complex combinations of the underlying pathologies have led to evolved concepts in clinical and neuropathological criteria in the past decade. Pathological AD can be decomposed into subsets of individuals with significantly different antemortem cognitive decline rates. Using transcriptome as a proxy for functional state, we preselected 414 expression profiles of clinically and neuropathologically confirmed AD subjects and age matched non-demented controls sampled from a large community based neuropathological study. By combining brain tissue specific protein interactome with gene network, we identify functionally distinct composite clusters of genes which reveal extensive changes in expression levels in AD. The average global expression for clusters corresponding to synaptic transmission, metabolism, cell cycle, survival and immune response were downregulated while the upregulated cluster had a large set of uncharacterized pathways and processes that may constitute an AD specific phenotypic signature. We identified four master regulators across all clusters of differentially expressed genes by enrichment analysis including TGIF1 and EGR3. These transcription factors have previously not been associated with AD and were validated in brain tissue samples from an independent AD cohort. We identify TGIF1, a transcriptional repressor as being neuroprotective in AD by activating co-repressors regulating genes critical for DNA repair, maintaining homeostasis and arresting cell cycle. In addition, we show that loss of EGR3 regulation, mediates synaptic deficits by targeting the synaptic vesicle cycle. Collectively, our results highlight the utility of integrating protein interactions with gene perturbations to generate a comprehensive framework for characterizing the alterations in molecular network as applied to AD.


Introduction
Increase in aging population with improved longevity reinforces the urgency for prevention and treatment of progressive neurodegenerative diseases including Alzheimer's disease (AD), the most common cause of dementia 11,41 . While, the predominant pathology of AD is accumulation of neuritic β-amyloid (Aβ) plaques and neurofibrillary tangles containing phosphorylated tau protein, co-occurrence of other neuropathological features is increasingly recognized to be a frequent event in brains of demented patients 1,11 . These changes including but not limited to inflammation, neuronal and synaptic loss, problems with blood circulation, and atrophy correlate with clinical symptoms of cognitive decline and have led to changes in diagnostic criteria during the last decade 6,41,78 .
Multiple causal factors underlie the complexity of sporadic AD. Primary risk factors include age, gender and family history 41 . The presence of elevated blood cholesterol, diabetes, depression, multiple lifestyle and dietary factors are also associated with increased risk of AD dementia, although not necessarily with AD pathology 6 . Genetic mutations in amyloid precursor protein (APP), presenilin 1 (PSEN1), and presenilin 2 (PSEN2) associated with autosomal dominant AD were critical in identifying pathogenic mechanisms associated with Aβ accumulation 5,83 . While many of the GWAS identified genes including the ε4 allele of apolipoprotein E (APOE) have implications of Aβ / tau processing, it is interesting to note that majority of the risk loci associate strongly with pathways involved in inflammation, lipid metabolism and endocytosis , which may partially explain the association of non-genetic risk factors 46,99 . Translation of molecular insights into predictive screening and diagnosis is important since clinical diagnosis of AD is difficult and often imprecise 41 . This is complicated by the fact that genetic mutations explain only a small proportion of autosomal dominant AD and presence of APOE ε4 is not sufficient to cause AD 6,11,41 . In addition, considering the variability in the rate of cognitive decline between individuals, transcriptomic changes at preclinical, prodromal and dementia stages of AD may be distinct and warrant subject selection refinement 11,95 .
In this study, we selected 414 clinically and neuropathologically confirmed AD subjects and cognitively normal age-matched controls, all sampled from a large well characterized community based neuropathological study 6,68 . Integration of gene perturbations with protein interactions is a powerful method to identify genes causal of a phenotype that are functionally cohesive, physically interact and share coherent biological pathways 66 . We characterize the molecular network dysregulation in AD relative to controls by integrating global gene expression profiles with a pre-calculated brain tissue specific protein-protein interactome 34 . Community detection applications to the human brain transcriptome have revealed intrinsic topological organization of functional co-expressed modules, they generally consider single interaction data 46, 99 66 . By integrating the transcriptomic changes with protein interaction networks, we identify functional composite clusters by implementing the Louvain algorithm. The distinct gene clusters reveal extensive expression changes across multiple biological and cellular pathways implicated in AD along with novel candidates with less understood biological function. By applying the strategy of gene set enrichment, we identify four master transcriptional regulators across all the clusters. These include TGFB induced factor homeobox 1 (TGIF1), and early growth response 3 (EGR3), previously not associated with AD and validated by protein analysis of brain tissue samples from an independent AD cohort. We identify transcriptional repressor TGIF which modulates the disrupted TGF-β signaling, as being neuroprotective in AD by activating co-repressors regulating genes critical for arresting cell cycle, facilitating DNA-repair and restoring homeostasis 64,89 . We show loss of regulation of EGR3 which is crucial for short term memory, to mediate synaptic deficits by targeting the synaptic vesicle cycle 69, 78 .

Refined AD phenotype reveals perturbation in the transcriptomic landscape
After adjusting for covariates and multiple testing, total of 1722 genes were significantly differentially expressed in AD compared to age-matched non-demented controls (NDC) and 57% of those genes were downregulated ( Figure 1A, Supplementary Table S1).
Gene Set Enrichment Analysis (GSEA) 82 revealed functional enrichment of specific biological processes (GO) and molecular pathways (KEGG) underlying these genes. A total of 453 upregulated and 113 downregulated MSigDB gene sets 82 were identified at 5 % FDR. Among the GO terms enriched, those related to neural development, gliogenesis, metabolism and localization of proteins and extracellular structure organization were upregulated while those related to synaptic transmission, mitochondrial and metabolic processes as well as extracellular transport were downregulated (Supplementary Tables S2, S3). These results are consistent with previous findings of multiple studies associating AD pathophysiology to genetic perturbation and demonstrate the complexity of the disease 11,46,68,95,99 .
A total of 133 differentially expressed genes were also mitochondrial genes, defined as mitochondrially encoded genes from MitoCarta 2.0 16 ( Figure 1B) (Supplementary Table   S4). Among these are nuclear-encoded genes for oxidative phosphorylation (OXPHOS) ( Figure 1C). These include NADH ubiquinone oxidoreductase core subunits S1 (NDUFS1, PEP = 2.4e-02), A5 (NDUFA5, PEP = 8.0e-05), A10 (NDUFA10, PEP = 1.3e-02) and B5 (NDUFB5, PEP = 2.4e-02), all part of complex I involved in transfer of electrons to the respiratory chain. Reduced expression of complex V genes including ATPase inhibitory factor I (ATPIF1, PEP = 2.4e-03), subunit B (ATP5F1, PEP = 1.8e-03) and F1 alpha subunit (ATP5A1, PEP = 3.7e-02) have implications for production of ATP from ADP 18 . In addition, expression of eight mitochondrial ribosomal proteins subunits (MRP) involved in translation of the mitochondrial-encoded OXPHOS genes are downregulated. Downregulated nuclear and mitochondrial genes encoding subunits involved in OXPHOS have been shown in the brain and blood of subjects with AD and mild cognitive impairment (MCI) compared to NDC 25,58 . While most mitochondrial genes were downregulated, 27 genes showed upregulation ( Figure 1B). Of these, increased expression of diazepam-binding inhibitor (DBI, PEP = 8.9e-04), known for its role as mediator in corticotropin-dependent adrenal steroidogenesis and in modulation of the action of the GABA receptor has been shown to be dose dependent on Aβ aggregates 57 .

Aggregate network of protein and genetic interactions reveals novel gene candidates in biologically distinct clusters associated with AD pathogenesis
When integrated into the GIANT brain-specific interactome 34 , the differentially expressed genes display distinct clustering structure (Figure 2A). These clusters were identified computationally using the Louvain modularity maximization algorithm 88 . This graphbased unsupervised clustering analysis does not require any explicit assumptions, and instead uses an iterative process to optimally partition nodes in a graph, by maximizing the number of edges within clusters, and minimizing the number of edges between clusters.
The largest connected subnetwork of 1391 differentially expressed genes partitions into 8 distinct clusters ranging in size from 2 to 373 genes. We restricted our analysis to clusters containing at least 5 genes and resulted in a subnetwork with 1382 genes. Many  Figure S1). Interestingly, the expression for clusters corresponding to synaptic transmission, DNA repair, immune response and metabolism were downregulated, while the cluster with overall upregulation had many uncharacterized genes ( Figure 2C).

Cluster0 -Synaptic Transmission
Loss of neurons and synapses in the hippocampus and cerebral cortex, a characteristic of AD is shown to strongly correlate with cognitive impairment, high level of Aβ production along with tau oligomers 78  In addition to neurotransmitters, neurosecretory proteins and neuropeptides are well known for their role in neuronal cell communication. Reduced expression of neurosecretory protein VGF nerve growth factor inducible (VGF, PEP = 2.9e-05) considered to be associated with synaptic plasticity and function, has recently been identified in CSF and parietal cortex of AD patients 21,38 . These results confirm and extend the trend to prefrontal cortex in AD. Based on the specificity of VGF to the central nervous system, it is an ideal candidate with the potential for blood based biomarker for AD.
Reduced cortical corticotropin-releasing factor immunoreactivity (CRH, PEP = 2.1e-04) in the face of increased hypothalamic expression) is a prominent neurochemical change in AD and is shown to mediate stress induced hyperphosphorylated tau through its receptor1 (CRFR1) with antagonist reducing Aβ levels in animal AD models 71,100 .

Cluster1 -DNA Repair and Transcription
Abnormal cell cycle reentry in response to accumulation of damaged DNA is hypothesized to precede AD pathogenesis 23,94 . Pathways in cluster1 were enriched for DNA replication and transcription ( Figure 3, Supplemental Table S7, Figure S3) with approximately 15% of genes associated with cell regulated DNA repair and chromatin remodeling 67 . One of the highly connected genes in this cluster, minichromosome maintenance complex component 7 (MCM7, PEP = 9.3e-05), was identified as an additional AD susceptible locus in large scale GWAS study 27 . Expression of DNA-dependent protein kinase (PRKDC, PEP = 9.5e-03) which is essential for the repair of the double strand breaks, a lethal form of DNA damage was downregulated and is comparable to protein expression observed in AD brains 17,23 . Reduction in the base excision repair, a multistep DNA repair pathway has been detected in early stages of the disease (amnestic MCI) with continued trend with progression to AD 94 . Reduced expression of tyrosyl-DNA phosphodiesterase 1 (TDP1, PEP = 4.0e-02), a DNA repair protein involved in base excision repair is a rational result 26 . Overexpression of DEK Proto-Oncogene (DEK, PEP = 4.1e-02) known for its role in p53 destabilization, is associated with multiple cancer phenotypes and has not been characterized in AD with potential as an early target for neuronal homeostasis 28 . PHD Finger Protein 19 (PHF19, PEP = 3.1e-05), a Polycomb-like protein is known to specifically bind to histone H3K36me3 and crucial for PRC2 recruitment to CpG islands which mainly mediate transcriptional repression 15 . Although its role in AD remain unclear, overexpression of PHF19 as observed here has been shown to correlate positively with astrocytoma grades 53 . Whether this was a consequence of cell state activity or an essential event is to be determined. Although not directly linked to AD, suppression of chromatin assembly factor I (CHAF1B, PEP = 2.0e-03) has recently been shown regulate somatic cell identity in a transcription factor-induced cell fate transition, correlated to protein levels in Down Syndrome brains and identified via whole exome sequencing to be associated with neurogenetic disorders with intellectual disability 2,19 . Interestingly, stathmin (STMN1, PEP = 3.3e-02) known to negatively correlate with neurofibrillary tangles and highly connected to genes in this cluster is shown to mediate cell cycle regulation via MELK kinase and has been investigated as a biomarker of DNA damage in AD 61,92 . This could provide an important link between dysregulation of microtubule polymerization and cell cycle dynamics as seen in AD and merits further investigation.

Cluster2 -Immune Response
Immune system pathways were enriched in cluster2 (Figure 3, Supplemental Table S7, Figure S4). Several lines of evidence suggest that chronic neuroinflammation, shown to correlate with disease progression could contribute to the pathogenesis of AD 46,62,99 .
The associated causal regulators including TYROBP as described in Zhang et al 99  to MBL2 have shown to facilitate favorable outcome in experimental models of traumatic brain and spinal injury suggesting a potential target for modulation of immune response 24,33 . Increased expression of the chemokine receptor 4 (CXCR4, PEP = 3e-0.2) is shown to be associated with higher levels of activated phosphorylated protein kinase C and with synaptic pruning in early development 65, 93 . Interestingly, this was concomitant with decreased expression of C2H2 zinc finger protein (PLAGL1, PEP = 4.2e-07) known to activate SOCS3, a suppressor of cytokine signaling 76 . Fc receptors mediated glial cell activation identified as one of the immune pathways in AD GWAS study, has been attributed in the adverse side effects associated with the failed AD immunotherapy trials reiterating the need to understand the functional roles of these receptors 32,46 . Decreased expression of bromodomain family member (BRWD1, PEP = 1.7e-06), a histone reader essential for B lymphopoiesis is in lieu with the observed loss of the cross talk between innate immune system of the brain and the adaptive immune system as facilitated by circulating B and T cells 59, 62 .

Cluster3 -Novel Gene Candidates
Genes from cluster3, the only cluster with an overall upregulation in the differential expression were not readily characterized. No enriched pathways were detected, and only a few marginally significant GO terms were found. Examination of this cluster revealed more uncharacterized genes than expected by chance (p < 0.05), suggesting genes within this cluster would be attractive candidates for follow up studies as many encode for proteins not previously characterized or associated with AD. Enriched biological processes included small GTPase mediated signal transduction (FDR = 6.21 e-05), which are known to control diverse cellular activities (Supplemental Figures, S1).
Evidence suggests that dysregulation of Rho-GTPase specifically RHOA, RAC1 and CDC42 mediated actin dynamics could be a key contributor to synaptic deficits observed in AD, yet interactions between the different signaling pathways remain unclear 40 99 . Interestingly, transcription factor EB (TFEB) a master regulator for lysosomal biogenesis was overexpressed which is associated with regulated autophagy 77 . Given that the events in AD are non-linear, whether the observed changes are pathogenic or protective is to be determined.

Cluster4 -Metabolism and Bioenergetics
Metabolic dysfunction is a well-established characteristic of AD and pathways in cluster4 were enriched for energy metabolism and vesicle mediated transport (Figure 3, Figure S5). Clinical imaging modalities have consistently shown brain glucose hypometabolism to be an early irregularity in patients with cognitive impairment and in some cases prelude memory deficits 1,11 . Increase in free radical production, a decrease in ATP/ADP ratio and increased rate of oxidant damage to lipids, proteins and mitochondrial DNA characterize the mitochondrial dysfunction in AD 4, 85 .
Whether mitochondrial dysfunction is induced by Aβ or an independent upstream process is currently unresolved. Reduction in expression of dynamin 1 like (DNM1L, PEP = 1.4e-03) is consistent with the observation that mitochondrial dynamics shift in favor of fission and is dependent on Aβ overexpression as seen in AD 4 . Increased fission could also be explained by reduced expression of hypoxia inducible domain family member 1A (HIGD1A, PEP = 1.4e-02) shown to regulate mitochondrial fusion by inhibiting the cleavage of OPA1 3 . Interestingly, it has also been attributed to regulate the mitochondrial γ-secretase, a multi-subunit protease complex known to cleave numerous transmembrane proteins including notch and APP and whose activity is dysregulated in AD 36 . Presence of amyloid precursor protein (APP) in the mitochondria is associated with reduced cytochrome c oxidase (COX10, PEP = 3.8e-05) which is essential for COX assembly and function of complex IV of the electron transport chain 25

Novel regulatory mechanisms of AD pathogenesis include TGIF1 and EGR3
We prioritized transcriptional regulators that were significantly enriched in the AD clustered subnetwork and were differentially expressed along with their gene targets in AD when compared to NDC and identified 4 master transcription factors (TF) i.e. SP1, EGR3, TGIF1 and BPTF across all the clusters (Figure 4, Supplemental Table S8, Figure   S6). Specificity protein 1 (SP1) is dysregulated in AD and can regulate key genes associated with AD pathology i.e., APP, tau and APOE 20 . Gene members of the early growth response family (EGR3, EGR1) are zinc finger transcription factors essential for synaptic plasticity and memory and are downregulated. Dysregulation of EGR1, a critical microglial homeostatic gene required for maintenance of LTP was associated with pathogenesis of APP expressing mice while EGR3 critical for short term memory has previously not been associated with AD 50, 65, 69 . BPTF/ FAC1 is a chromatin remodeler first identified in AD brain whose overexpression is associated with apoptotic cell death 81 . Transforming growth factor-βs (TGFβ) along with bone morphogenic proteins (BMP) bind via receptors (TGFBR1, TGFBR2, TGFBR3, BMPR2) and activate Smad to regulate gene expression 64  It is of interest to note that downregulation of BMPR2 observed here could be primarily mediated by downregulation of EGR1, an essential microglial transcription factor 65 .
Deficiency in BMPR2 is linked to increased sensitivity to DNA damage along with alterations in DNA repair efficiency and upregulated SP1 could reflect DNA damage control 54 . This could provide an important link between neurodegenerative microglial phenotype and DNA damage observed in AD.

Downregulated nemo like kinase (NLK) along with upregulation of NOTCH4 which is
shown to strongly couple to AD risk genes and its downstream co-activator complex RBPJ, point towards ectopic cell cycle reactivation 5,13 . Surprisingly, these observations also coincide with dysregulation of genes involved in reprogramming of cell function i.e.

TGIF1 is neuroprotective in AD
Here we propose increased TGIF1 plays a neuroprotective role by directly repressing or by activating co-repressors to regulate gene expression aimed at arresting cell cycle, enhancing DNA repair and restoring homeostasis ( TGIF1 also cooperates with the other transcription regulators to mediate the neuroprotective effect (Supplemental Figure S7). Upregulation of CDKN2C regulated by TGIF1 and EGR3 is perhaps a compensatory effect to inhibit the cell cycle progression in absence of direct regulation by TGFβ 64 . BPTF and TGIF1 regulated zinc-finger myndtype containing 8 (ZMYND8) is a chromatin reader that recruits the repressive nucleosome remodeling and histone deacetylase (NuRD) complex to transcriptionally silence gene expression as part of the DNA damage response. In conjunction with increased expression of BRD3, paralog of BRD2, ZMYND8 in AD may work to shield H4Ac during double strand repairs as part of the chromatin response to DNA damage 35 .

EGR3 affects synaptic vesicle processing in AD
Consistent with the association of synaptic loss to AD, we observe a concerted downregulation of many genes that code for proteins involved in synaptic vesicle cycle of which key genes are regulated by EGR3 ( Figure 6

Protein expression of TGIF1 and EGR3 is in concordance with transcript level expression
To substantiate the differences in transcription between AD and non-demented aged brains, we assessed the protein expression of TGIF1 and EGR1, the direct target for EGR3 in pre-frontal cortex of AD and age-matched control brains from an independent cohort. In line with changes in the transcriptome data, we observe an overall 60% decrease in EGR1 and an 81% increase in TGIF1 levels in AD brains relative to nondemented control and these trends reach statistical significance. (Figure 7). We found the EGR1 distribution across all cell types in the control brains and the protein levels were particularly diminished in pyramidal neurons in AD brains. On the other hand, upregulation of the nuclear localization of TGIF1 was a frequent feature of glial cells in AD brains. Taken together, these results demonstrate the downregulation of EGR3 and is consistent with the role of TGIF1 as a transcription factor.

Conclusions
By integrating brain tissue specific protein interactions with gene expression profiles derived from large sampling of postmortem transcriptome data, we identified changes in molecular network associated with clinical and pathological Alzheimer's disease (AD) when compared to age matched non-demented controls. Although the mechanisms for neuronal dysfunction in AD are not completely understood yet, the distinct biological clusters identified using Louvain community detection algorithm point towards associations critical for AD 9,23,46,68,85,99 . Interestingly, the average global expression for clusters corresponding to synaptic transmission, metabolism, cell cycle and survival as well as immune response were downregulated while the upregulated cluster3 had a large set of uncharacterized pathways and processes. By refining the subject selection, we identified new genes with target potential, confirmed some previous correlations and find lack of associations for other previously reported GWAS computed AD susceptible genes.
Based on gene interactions, we propose neuroprotective role for transcriptional repressor TGIF1 which activates other co-repressors to module the TGFβ signaling and acts to arrest cell cycle re-entry. We find downregulation of EGR3 to be associated to dysregulated synaptic vesicle cycle and postulate to mediate synaptic deficits as seen in AD.

Study Subjects
The Religious Orders Study and Rush Memory and Aging Project (ROSMAP) are an ongoing longitudinal clinical-pathologic cohort studies of aging and dementia with enrollment of aged individuals without known dementia at baseline 6 Table 1). Assessment of the pathological metrics as well as the characteristics of the cohort has been previously described in detail 6, 95 .

Background Interactome
The background interactome on which the network analysis was conducted was built from the brain-specific network in the GIANT database 34 . This network is composed of edges which support a strong tissue-specific functional interaction between source and target genes. This network, thresholded at an edge confidence of 0.2, contains 14,545 genes, and 1,370,174 edges, and represents genes and interactions which are likely to be present in normal brain function. The identified differentially expressed genes were mapped onto this network to create an AD subnetwork. Network visualization was accomplished using visJS2Jupyter 73 .

Clustering analysis
Clusters were identified in the AD subnetwork using a network-based modularity maximization algorithm 88 . This algorithm, commonly referred to as the Louvain clustering algorithm, identifies groups of nodes which have many connections within groups, and few connections between groups, and is efficient at extracting community structure from large networks. The algorithm iteratively maximizes the modularity, Q, defined as follows: Where is the binary adjacency matrix element representing the presence or absence of the connection between node i and node j, represents the degree of node , where degree is defined as the number of nodes directly connected to node , is the community to which node belongs, and the function ( , ) is 1 if = , and otherwise it is 0. is the total number of edges in the network, Hypergeometric test was utilized to test of statistical significance of the enriched biological process and pathways identified for the differential expressed genes in the clusters 90,97 .
Overrepresentation enrichment analysis was conducted using the full set of genes from the AD subnetwork as the reference gene set, corrected for multiple testing using the Benjamini-Hochberg procedure and FDR < 0.05 was considered significant.

Transcriptional Regulation Analysis
To identify transcription factors (TF) likely related to AD, we analyzed for enrichment of transcription factor targets in the AD clustered subnetwork 90 . This resulted in 16 significantly enriched transcription factors (Supplementary Table S8). We further filtered this list by differential expression in AD vs NDC, resulting in 4 candidate TFs. These four TFs are all significantly differentially expressed in AD as compared to NDC and have more targets than would be expected by chance in the AD subnetwork (Supplemental Figure S7). The TF subnetwork is visualized to highlight the connections between TF and targets along with connections between targets using visJS2Jupyter 73 .

Validation of TGIF1 and EGR3
We assessed the protein expression and distribution of TGIF1 and EGR3 target EGR1 in prefrontal cortex brain tissue of clinically diagnosed pathologically confirmed AD and from age matched controls using standard immunohistochemistry techniques. The subjects were selected based on diagnostic criteria as described above and brain tissue samples were obtained from University of California, San Diego Alzheimer's Disease Research Center (ADRC) brain bank. Tissue samples were obtained from two non-demented aged controls (81.5 ± 2.1 years, 12 ± 3 hours, 50% F) and three AD subjects (83 ± 4.3 years, 6.6 ± 1.1 hours, 67% F). Paraffin-embedded tissue blocks were serially sectioned and incubated with either TGIF antibody (H-1) (Santa Cruz Biotechnology, Dallas, TX) or EGR1 antibody (Cell Signaling Technology, Danvers, MA). Staining was performed with chromogen 3,3 ' -diaminobenzidine (DAB) to identify the immunoreactive structures and counterstained with hematoxylin. All images were acquired using an upright microscope (Leica DM5500B) at a resolution of 1392 x 1040 pixels and consistent aperture and gain settings. Custom designed macro was used to convert the optical images to grayscale, threshold and measure the TGIF or EGR1 positive area fraction relative to the optical field. Due to the non-normal distribution of the data, test of significance was evaluated using Wilcoxon-Mann-Whitney test and differences were considered significant when p < 0.05.