Selection and assessment of the gene expression profiles dataset

To identify neuron-specific Master Regulators (MRs) as drivers of Alzheimer’s disease (AD) initiation and progression, gene expression signatures were generated from laser-captured neurons from AD and control cases, encompassing both pathological and clinical phenotypes. Specifically, we selected the Liang et al. dataset, which comprises gene expression profiles from laser-capture microdissected (LCM) cortical neurons isolated from six anatomically and functionally distinct postmortem human brain regions representing 14 controls, 10 non-demented individuals with AD-type changes in their brains at autopsy (NDAD), and 34 demented individuals with the histopathological confirmation of AD [20,21] (S1 Table). To assemble the regulatory model (interactome), we used the full set of 193 gene expression profiles representing the entorhinal cortex (EC), hippocampus (HIP), middle temporal gyrus (MTG), posterior cingulate (PC), superior frontal gyrus (SFG) and visual cortex (VCX) of these individuals. Cluster analysis revealed that samples from regions known to be severely affected in AD (e.g., HIP, EC or MTG) show tight clustering according to diagnosis whereas samples from regions relatively less affected (i.e., VCX) do not cosegregate (Fig. 1 and S1 Fig.). Therefore, unbiased clustering analysis confirms the reproducibility of clinically relevant molecular phenotypes and suggests that patient stratification into these three diagnostic categories is biologically relevant, as supported by common molecular features. Thus, the categories “Control”, “NDAD” and “AD” were used to inform follow up analyses.

Construction of the human neuronal transcriptional interactome

The Algorithm for the Reconstruction of Accurate Cellular Networks (ARACNe) [22] was previously developed to reconstruct the transcriptional regulatory logic of specific cellular contexts. This logic comprises the set of regulatory targets, or regulon, of each transcription factor (TF) or other transcriptional regulator of gene expression. For simplicity, we will use the term TF to indicate both classes. Analysis of these 193 gene expression profiles by ARACNe yielded the first human causal transcriptional network for cortical neurons, representing 488,353 individual transcriptional interactions between 3,758 TFs expressed in this context and their transcriptional targets.

Identification of candidate Master Regulators using MARINa

We then used the previously developed Master Regulator INference algorithm (MARINa) to interrogate regulatory networks to identify candidate MR genes [14,16]. These represent regulators that are sufficient and/or necessary causal determinants of a specific phenotypic difference. MR candidates are identified by MARINa based on the enrichment of their ARACNe-predicted targets in the differentially expressed genes of a specific phenotype, similar to using a gene reporter assay approach to detect aberrant activity of a TF. The difference is that instead of using a single reporter controlled by the promoter of a TF-target gene, we use the expression of all the ARACNe-inferred targets of the TF. This allows identification of regulators whose aberrant activity is post-translationally determined, which could not be identified by differential expression, thus overcoming the limit of traditional analysis.

MARINa analysis was performed on 18 distinct gene expression profile signatures, representing three distinct phenotypic differences (Control→NDAD, NDAD→AD and Control→AD) across each of the six brain regions. Candidate MRs were first selected based on their statistical significance and then sorted by their Differentially Expressed Targets Odds Ratio (DETOR) [16]. The latter represents the density of targets in the leading edge of the GSEA analysis compared to the remaining range of expression and is a direct measure of the regulatory impact of the MR. These analyses result in relatively small ranked-lists containing TFs that are most likely to be responsible for the phenotypic difference in the brain region of interest. We also reasoned that candidate MRs for a region specific signature should be independent from the number and specific selection of the samples in the analysis. Thus, to filter out false positive MRs, we used bootstrapping on the 18 representative gene expression profile signatures, each time randomly sampling only 70% of the total samples representing the signature, with replacement. Candidate MRs that failed to be identified in all 10 bootstrap runs were then considered false-positives and were excluded from further consideration. The results of the bootstrapping step on the count of candidate MRs are summarized in Fig. 2. In the control v. NDAD phenotypic change, high numbers of candidates are found in EC, HIP and MTG, which are regions showing early pathology in AD and are also affected in aging to variable degrees. In contrast, in the control v. AD phenotypic change, we observe a positive correlation between the count of candidate MRs and the regional progression of AD. Finally, comparing NDAD and AD allows for the highlighting of MRs mostly responsible for dementia progression, as pathological hallmarks are already present in NDAD patients. Interestingly, the highest counts are found in EC, MTG, PC and SFG and a very low count is found in HIP, showing that in this region a small number of MRs are predicted to drive the phenotypic differences between NDAD and AD patients. This is a common finding for this type of analysis, such that signatures representing phenotype-relevant events produce a greater number of highly statistically significant MRs than signatures that are not the result of specific regulatory events. To this extent, PC and SFG are the two regions that differ significantly between the two phenotypic changes of interest (control v. AD and control v. NDAD). As a negative control, VCX was found to show a significant decrease in the number of candidate MRs in all three phenotypic changes, thus confirming its less representative role in AD.

Selection of candidate MRs for biochemical validation

In order to prioritize MRs for further biochemical validation, we made two assumptions. First, we reasoned that AD should be reflected across a multitude of genetic programs that are either brain region specific or common to multiple regions. However, we anticipated that focusing on region-specific programs would lead to discovery of many idiosyncratic false-positive MRs (e.g., those regulating region-specific downstream programs from the true MRs or programs activated within specific regions in response to neurodegeneration), whereas focusing on MRs conserved across multiple regions should be more specific to bona fide upstream AD drivers. Hence, for each phenotypic difference, we selected candidate MRs conserved across more than one region, thus eliminating candidate MRs identified only in one region. Fig. 2 summarizes this analysis and S2 through S7 Tables display the final lists of MRs on a region and transition-specific basis, after this step.

Second, we assumed that candidate MRs regulating hippocampus-specific signatures would be of particular relevance, as this region plays a central role in memory and is one of the first ones affected in AD [25–27]. Two strategies were thus followed in parallel: either prioritizing candidates that are most highly ranked in HIP (strategy 1), or prioritizing those that show up in the highest number of regions, including HIP (strategy 2). Through these assumptions, we reduced our original MR set to about 20 candidates per strategy (Tables 1 and 2). Two of the highest ranked genes from each strategy, for which appropriate reagents were available, were then selected for further validation in human brain tissue, by immunohistochemistry and Western blot analyses. These include the E1A binding protein p300 (EP300), Ying Yang 1 (YY1), the Zinc Finger, MYM-Type 3, (ZMYM3), and the Myocyte Enhancer Factor 2D (MEF2D).

Finally, we asked whether the expression levels of these selected candidates in the laser-captured neuron dataset correlated with that observed in mRNA obtained from whole human brain tissue (S3 Fig.). For YY1, EP300 and ZMYM3, this analysis revealed an increase of mRNA levels in AD cases compared to controls, suggesting that these two systems are comparable for these genes and therefore justifying utilization of whole tissue for further analyses. MEF2D shows no difference in the expression levels in the LCM dataset and the whole brain extracts, however, this finding is compatible with our analysis, which focuses on inferred activity rather than differences in expression levels. When investigated further using immunoblot, immunohistochemistry and binding assays, the experiments did not show consistent differences for both MEF2D and phospho-MEF2D (unpublished data). Thus, only the results for YY1, EP300 and ZMYM3 are presented below.

Experimental validation of candidate MRs in AD brain

Ethics Statement

Animal work was carried out under protocols AC-AAD9106 and AAAD4910 approved by the Columbia University Institutional Animal Care and Use Committee. Rats were sacrificed using CO2 euthanasia, and all efforts were made to minimize suffering. All human tissues used for these studies were de-identified samples obtained from the New York Brain Bank under Columbia University Institutional Review Board protocol AAAB0192 (expires Feb. 2015). This study is exempt category 4 research as it involved the study of existing data and specimens recorded by the investigator in such a manner that subjects cannot be identified, directly or through identifiers linked to the subjects under OHRP Exempt Categories 45 CFT 46.101(b).

Materials

All chemicals used were of the highest grade available. RIPA, proteases and phosphatases inhibitors were purchased from Thermo Scientific. HFIP (Hexafluoro-2-propanol) was purchased from Fluka. Poly-D-lysine, DMSO, ZMYM3 rabbit antibody (IHC, HPA003211), Actin mouse antibody (A1978) were purchased from Sigma; GAPDH mouse antibody (IMG-5019A-2) from Imgenex; Ac-Lys382-p53 rabbit antibody (2525S) and Ac-Lys14-Histone 3 rabbit antibody (4318) from Cell Signaling Technology; ZMYM3 rabbit antibody (WB, sc-130039), YY1 rabbit antibody (IHC, WB, sc-281), p300 rabbit antibody (IHC, sc-585) and p53 mouse antibody (WB, sc-126) from Santa Cruz Biotechnology; MEF2D mouse antibody (IHC, WB, 610774) from BD Biosciences; βIII-tubulin from R&D systems; phospho-Ser1834-p300 (IHC, AP3296a) from Abgent; phospho-tau (p-tau) mouse antibody (clone AT8) from Thermo Scientific. Amyloid beta (1–42) peptide was ordered from UCLA.

Patient samples

Fresh frozen and formalin fixed/paraffin-embedded autopsy brain tissue were obtained from the New York Brain Bank at Columbia University Medical Center (New York, NY, USA) for immunoblot and immunohistochemistry. Neuropathological examination was per the brain bank protocols [61]. Samples were classified as control, moderate AD (mAD) or severe AD (sAD) as presented in S8 Table. When no significant difference was observed between mAD and sAD, measurements were pooled for statistical analyses.

Dataset Processing and Normalization

The analysis was performed on the Liang et al. dataset [20,21]. The dataset was downloaded from the Gene Expression Omnibus website, reference GSE5281 and GSE9770, and contained 193 samples corresponding to six regions obtained from 14 controls, 10 NDAD individuals and 34 demented individuals with the histopathological confirmation of AD. The gene expression profiles in the dataset were collected using the Affymetrix Human Genome U133 Plus 2.0 Array GeneChip system (54,675 probe sets). Expression measurements were normalized with gcrma [62], which adjusts for background intensities in Affymetrix array data, including adjusting for optical noise and non-specific binding. The array files were processed and normalized using R version 2.15.1.

Hierarchical Clustering

Clustering analysis was performed using BRB-Array tools version 4.3.2. Out of the 54,675 probesets that were present on the array, 23,594 probesets that passed the BRB-Array tools filters were used to perform the clustering. The arrays were normalized using a modified version of RMA, which uses a random subset of the arrays to generate the normalization and probset summary values. Centered (Pearson) correlation was used as the distance metric, and the average distances between all pairs of probesets that had an associated p-value of < 0.01 were used to build the clusters.

ARACNe

Context specificity is critical to reconstruct the transcriptional regulatory logic as TF regulons are highly context dependent. The ability to accurately reconstruct such logic using ARACNe is predicated on the availability of large gene expression profiles datasets, including more than 100 samples, which represent either natural or perturbation-induced genomic variability of the context of interest. For the reconstruction of the human neuronal interactome, we assumed that different individuals represented sufficient naturally occurring genotypic variability, while the different regions represented sufficient variability associated with microenvironment related signals, thus resulting in an interactome representing an accurate estimate of the overall variability of cortical neuron gene expression. This assumption is supported by the t-SNE [23] results on all samples, which show separation between regions that are severely affected in AD, as well as high variability between regions overall (S2 Fig.). Thus, in contrast to traditional approaches, where one attempts to minimize variability by profiling several biological replicates, ARACNe benefits from inter-sample variability as this allows for a more accurate inference of transcriptional interactions by taking into account inter-sample transcriptional regulatory differences. In contrast, the signatures used to interrogate the interactome require significant phenotypic specificity and benefit from replicate samples.

To construct the neuronal transcriptional network, ARACNe was applied to the set of 193 gcrma-normalized expression profiles using the adaptive partitioning algorithm, which selects the optimal kernel width for calculating the Mutual Information (MI) threshold of a specified p-value. The MI threshold used by ARACNe (MI > = 0.2185) corresponded to the p-value threshold of 10^-7 after 100 bootstrap runs. The resulting central nervous system (CNS) Interactome contained 488,353 statistically significant MIs between the 3,758 TFs and 38,045 genes.

MARINa and Candidates Selection

MARINa was used to infer MRs that drive the transition between Control and AD samples, as defined by pathological and clinical characterization. Given a regulatory network model, MARINa requires a relatively small numbers of gene expression profiles (N ≥ 6) representing each phenotype of interest in a specific transition (e.g., 6 gene expression profiles for NDAD and 6 for AD), to identify the candidate MR genes. For each phenotype transition of interest (e.g., Control → AD in HIP), we first generate the gene expression profile signature, defined as all genes represented on the specific microarray platform, ranked from the most underexpressed to the most overexpressed in the disease phenotype compared to Control samples, as determined by a t-test. MRs are then identified as the TFs with the highest likelihood of implementing the specific signature based on the regulatory mode, i.e., those whose ARACNe-inferred targets (i.e. TF-regulon) are most enriched in differentially expressed genes in the gene expression profile signature. For instance, a positive MR (whose activation drives the transition) would have its positively regulated targets highly enriched in overexpressed genes and repressed targets highly enriched in under expressed genes in the signature of interest. The opposite would be true for negative MRs (whose inactivation drives the transition). Such enrichments can be effectively assessed by a multi-tail extension of the Gene Set Enrichment Analysis (GSEA) [24]. Thus, importantly, identification of candidate MRs is based on the expression of their ARACNe-inferred targets rather than on their own change in expression level.

Master Regulator INference analysis (MARINa) was used to analyze each TF with more than 20 targets in the human neuronal transcriptional network, with GSEA [24] used to assess the enrichment of each TF’s set of predicted targets. As a reference, we used the set of genes on the expression profile, ranked by their t-statistic calculated by comparing the candidate phenotypic transition of interest (e.g., Control → AD in HIP). P-values were computed by performing 1,000 sample-shuffling permutations and each TF was given a p-value based on its enrichment score according to GSEA. Shadow Analysis was performed as described in Lefebvre et al [16], to eliminate false positive representing TFs with substantially overlapping programs with bona fide MRs but unlikely to drive the signatures of interest. Master Regulator candidates (i.e., TFs with p < 0.01, not removed by Shadow Analysis) were sorted by their Differentially Expressed Targets Odds Ratio (DETOR) score, as given by Master Regulator Analysis [16]. The counts of candidate MRs obtained in each region at this step can be found in Fig. 2.

Transcription Factors classification

To identify transcription factors (TFs), we selected the mouse genes annotated as “transcription factor activity” in Gene Ontology and the list of TFs from TRANSFAC. This produced a final list of 1,794 TFs, which mapped to 3,758 probesets on the gcrma-normalized expression profile.

Specificity-weighted GSEA and Bootstrapping

To further classify the MRs that were most relevant for disease outcome, we applied specificity-weighted GSEA in combination with bootstrapping of the samples. Specificity-weighted GSEA is an alternative method that can be used to predict master regulator candidates. Similar to MARINa, specificity-weighted GSEA begins by measuring the enrichment of differentially expressed targets for each TF. However, in calculating the enrichment score for each TF and its target, it takes into account the specificity of interaction between TF and its target, i.e. the total number of TFs regulating the target according to the neuronal transcriptional network.

In the original GSEA paper by Subramanian [24] describes the calculation of the Enrichment Score (ES) in the following way, with ES being the maximum deviation from zero of P_hit—P_miss.

• Rank order N genes on the microarray platform to form L = g₁ … g_N according to the t-score of their differential expression between phenotype 1 and phenotype 2
• For each TF being tested by specificity-weighted GSEA, calculate the fraction of genes in S (“hit”, or targets of the TF according to the neuronal transcriptional network) weighted by their correlation and the fraction of genes not in S (“miss”) present up to a given position i in L. where

In specificity-weighted GSEA, the “hit” score of each gene is calculated in the following way, Where T_j is the number of TFs that regulate gene R_j. The application of this equation results in more conservative estimate of the enrichment score of each TF. In addition, the method rewards TFs that have specific targets.

In order to select for highly enriched MR candidates, 10 iterations of specificity-weighted GSEA was run with bootstrapping. Each bootstrap run was performed using 70% of samples in each class (random subset with replacement) and 1,000 sample-shuffling permutations. Each TF was given a p-value based on its enrichment score according to specificity-weighted GSEA, and the results were sorted by their DETOR score [16]. Eventually we only selected MR candidates that had a p-value < 0.01 according to specificity weighted GSEA in 10 out of 10 bootstrap runs. The use of bootstrapping and specificity-weighted GSEA allowed us to identify two groups of candidate MRs: those whose predicted activities were highly enriched regardless of which individual samples were used ("stable” candidates), and another group whose predicted activities showed significant fluctuations depending on which individual samples were used (“unstable” candidates). Number of stable candidate MRs in each of the regions and comparisons, as well as their effect on the final results can be found in Fig. 2.

Quantitative real-time polymerase chain reaction (QPCR)

Fresh-frozen brain tissue was pulverized in liquid nitrogen, lysed in QIAzol and homogenized using a Qiagen TissueLyser II with 5mm stainless steel beads (frequency 25/s, 2 x 2min). RNA was extracted using an RNeasy Mini kit (Qiagen) and RNA integrity was assessed using a bioanalyzer (Agilent 2100). Only samples showing a RNA Integrity Number (RIN) score above 5 were used for further experiments. cDNA synthesis was performed using a First Strand cDNA Synthesis Kit (Origene, Rockville, MD, USA) and used as template (1:4 dilution) in 20 μl reactions. Primers sequences used for the RT-PCR were the following:

p300: F: AGATGGGAATGATGAACAACC, R: ACTCACCATGTTGGGCATTC;

YY1: F: GCGGAGCCCTCAGCCATGGCCTCG, R: CAGCGGCTGCAGAGCGATCATGG;

ZMYM3: F: TGTGGATCGTAATGGCAAGA, R: TGCGGTCAACCTTGTTGTAG,

MEF2D: F: TACCCACAGCACCCAGCTT, R: TAGACTGGGAGACCCAAGG.

The FastStart Universal SYBR Green Master mix (Roche) was used for reactions on a Mastercycler ep realplex (Eppendorf, Hauppauge, NY, USA). The mRNA levels were normalized against the geometric mean of GAPDH, HPRT1, SDHA and UBC as described in Vandesompele et al. [63].

Protein extraction from whole tissue

Approximately 50 mg of fresh-frozen brain tissue were pulverized in liquid nitrogen, dounce homogenized in ice-cold PBS, diluted in RIPA buffer containing proteases and phosphatases inhibitors, sonicated, centrifuged and aliquoted to be stored at -80°C. Concentrations were measured using a BCA assay and samples were diluted to 1 mg/ml. For nuclear-cytoplasmic extractions, approximately 80 mg of fresh-frozen brain tissue were pulverized in liquid nitrogen, washed once with ice-cold PBS and centrifuged at 500 g for 5 min at 4°C. The supernatant was discarded and the pellet was resuspended in 700 μl of 1x Hypotonic buffer, containing proteases and phosphatases inhibitors, and incubated on ice for 15 min. Samples were then dounce homogenized and reincubated on ice for 15 min. 35 μl of 10% NP40 were added, samples were vortexed for 10 seconds at highest setting and incubated on ice for an additional 5 min. The homogenate was centrifuged for 10 min at 14,000 g at 4°C. The supernatant contains the cytoplasmic fraction and was aliquoted and stored at -80°C. The nuclear pellet was resuspended in 50 μl of Buffer B Working reagent from Procarta Nuclear Extraction Kit (Affymetrix, cat # AY2002) and the protocol was followed according to instructions. Nuclear extracts were aliquoted and stored at -80°C. Concentrations of nuclear and cytoplasmic extracts were measured using Bio-Rad DC Protein Assay kit.

Immunohistochemistry

The slides were deparaffinized in xylene, rehydrated through graded alcohol and antigen retrieved by pressure cooking for 30 minutes in ethylenediamine tetraacetic acid (EDTA) buffer for MEF2D; in citrate buffer for p300, phospho-Ser1834-p300 and YY1 and in Trilogy for ZMYM3. The slides were processed following the EnVision + Dual Link Kit (Dako) and incubated overnight at 4°C with rabbit polyclonal anti-p300 (Santa Cruz, 1:50), rabbit polyclonal anti-p-p300 (1:30), rabbit polyclonal anti-YY1 (1:50), rabbit polyclonal anti-ZMYM3 (1:800), mouse monoclonal anti-MEF2D (1:100) diluted in Dako antibody diluent. The slides were stained with DAB (3,3-diaminobenzidine) as described in the protocol and counterstained with Mayer’s hematoxylin, dehydrated, and mounted.

Immunocytochemistry

The slides were deparaffinized in xylene, rehydrated through graded alcohol and antigen retrieved by pressure cooking for 30 minutes in citrate buffer for 30 min. The slides were blocked 30 min using a solution of 0.1% Triton in Superblock blocking buffer (Thermo Scientific). Rabbit p-p300 (1:30) and mouse p-tau (1:100) were incubated overnight at 4°C in Dako antibody diluent. The slides were washed and goat anti-mouse and anti-rabbit antibodies diluted in Dako antibody diluent were added for 30 min. After three final washes, the slides were incubated 5 min in Sudan Black, washed again in PBS and mounted. Fluorescent images were taken with Axiovision software through Axiophot camera and a Zeiss Axioplan 2 microscope.

Hippocampal neuron cultures and treatment

Hippocampal neuron cultures from both male and female rats pups were prepared following a slightly modified version of the method of Brewer (Brewer et al., 1993). Hippocampal neurons were kept in culture at 37°C with 5% CO₂ in Neurobasal medium (Invitrogen) with B27 supplement and Glutamax (Invitrogen) and plated at a density of 2.5x10⁵ cell/ml on dishes coated with poly-D-lysine. For our experiments, neurons were treated after approximately 7–14 days in vitro (DIV) with a dodecameric preparation of Aβ_1–42 prepared according to Barghorn et al. method [64]. At the end of the treatment, cells were washed twice with ice-cold PBS, harvested with RIPA buffer containing phosphatases and proteases inhibitors, sonicated and centrifuged.

Western blotting analysis

Equal amounts, ranging from 1 to 10 μg, of cell and whole tissue lysates were resolved by SDS-polyacrylamide gel electrophoresis (4–12% Bis-Tris, Invitrogen) and electro-transferred on a polyvinylidene difluoride (PVDF) membrane (Millipore). The membranes were then incubated for 1h at RT in 5% milk and overnight at 4°C in primary antibodies (ZMYM3 1:100; YY1 1:100; Ac-p53 1:1,000; p53 1:100; Ac-Histone 3 1:1,000; Histone 3 1:1,000; Actin 1:5,000; GAPDH 1:5,000). For the competitive assay, two identical gels were run using 5 to 10 μg protein extracts from different human post-mortem brain regions (CA1, BA38, BA9) and from rat hippocampal neurons. The blots were then incubated overnight in 0.5 μg of YY1 primary antibody (sc-281) containing, or not, 80x of blocking peptide (sc-281 P). The antibody, with and without peptide, was pre-incubated 2h at RT in 500 μL of PBS before being diluted and applied onto the membrane. The immunoreactive signals were detected with enhanced chemiluminescence kit (Amersham Biosciences, Uppsala, Sweden). The procedures followed were conducted in accordance with the manufacturer’s instructions.

Pen-siZMYM3 synthesis

A 5’ thiol-modified siZMYM3 (GGAGTCTCCTCATATTGAA, Dharmacon) was reduced following Dharmacon protocol. The last precipitation step was repeated three times in order to fully eliminate the excess TCEP. The dry pellet was then reconstituted in sterile annealing buffer to 90 μM. The desired amount of reduced siZMYM3 was incubated with 1 equivalent of activated penetratin (MP biomedical, cat # 11PENA0500) for 1h at 37°C and analyzed using a 20% TBE gel electrophoresis. The conjugated pen-siZMYM3 was aliquoted and stored at -80°C.