Transcriptional Analysis of Blood Lymphocytes and Skin Fibroblasts, Keratinocytes, and Endothelial Cells as a Potential Biomarker for Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD) is a devastating and progressive form of dementia that is typically associated with a build-up of amyloid-β plaques and hyperphosphorylated and misfolded tau protein in the brain. Presently, there is no single test that confirms AD; therefore, a definitive diagnosis is only made after a comprehensive medical evaluation, which includes medical history, cognitive tests, and a neurological examination and/or brain imaging. Additionally, the protracted prodromal phase of the disease makes selection of control subjects for clinical trials challenging. In this study we have utilized a gene-expression array to screen blood and skin punch biopsy (fibroblasts, keratinocytes, and endothelial cells) for transcriptional differences that may lead to a greater understanding of AD as well as identify potential biomarkers. Our analysis identified 129 differentially expressed genes from blood of dementia cases when compared to healthy individuals, and four differentially expressed punch biopsy genes between AD subjects and controls. Additionally, we identified a set of genes in both tissue compartments that showed transcriptional variation in AD but were largely stable in controls. The translational products of these variable genes are involved in the maintenance of the Golgi structure, regulation of lipid metabolism, DNA repair, and chromatin remodeling. Our analysis potentially identifies specific genes in both tissue compartments that may ultimately lead to useful biomarkers and may provide new insight into the pathophysiology of AD.

INTRODUCTION

Alzheimer’s disease (AD) is a progressive form of dementia that affects an estimated 5.3 million Americans and this number is expected to triple by the year 2050 [1]. Additionally, the cost of medical care for those afflicted with AD and other forms of dementias are predicted to surpass $220 billion in the U.S. this year. These sobering statistics underscore the impact of AD on human health as well as the burden on the world’s health care systems. A significant part of this cost is the initial diagnosis, which includes a comprehensive medical evaluation to rule out other forms of dementia as well as neurological examinations and brain imaging. In addition to the substantial cost associated with an AD diagnosis, the lack of sensitive and specific biomarkers poses unique problems with staging and following subjects during the course of clinical trials. The protracted prodromal phase of AD compounds this issue. Indeed, it has been suggested that pathology of AD commences as early as 20 years prior to the manifestation of symptoms [2]. Therefore, it is likely that some subjects chosen as controls for clinical trials are actually asymptomatic (or pre-symptomatic) cases, underscoring the need for an accurate way to select control subjects for these studies.

Currently, six biomarkers are commonly used to make a diagnosis of AD in the research setting and for screening subjects for clinical trials. Three are imaging biomarkers and three are cerebrospinal fluid (CSF) biomarkers (reviewed by Lewczuk et al. [3]). The imaging biomarkers include hippocampal atrophy observed by magnetic resonance imaging (MRI); fluorine-18 fluorodeoxyglucose positron emission tomography (F-FDG-PET); and increased amyloid retention of Pittsburgh compound-B tracer by PET (PiB-PET). Although CSF biomarkers are helpful in staging clinical trials, the invasive nature of a spinal tap makes their use problematic when multiple sample collection is required. In light of this issue, several groups have investigated the potential utility of blood-based biomarkers. Kleinschmidt et al. recently reported that plasma levels of the Aβ_1–42 form of amyloid-β protein were significantly decreased in mild cognitive impairment (MCI) and AD subjects when compared to age-matched controls; however, levels of the Aβ_1–40 form did not differentiate cases and controls, but were observed to increases with age in healthy controls [4]. Clusterin is another potential biomarker shown to be associated with the rate of cognitive decline [5, 6]. Also, negative correlation between several members of the apolipoprotein family and grey matter volume was reported by Song et al. [7]. At this time, a single biological molecule with the required sensitivity and specificity necessary to serve as a useful AD diagnostic has not been identified suggesting that combinations of analytes may be required.

Previous studies report differential gene expression in brain tissue from postmortem AD cases. For example, disturbances in the transcription of genes involved in biochemical pathways for lipid metabolism, as well as genes for several inflammatory molecules were observed in the prefrontal cortex and hippocampus of AD cases [8, 9]. Furthermore, when comparing the brains from AD cases to those from normally aged controls, significant changes in the expression of genes related to vesicle trafficking, neurotransmitter and postsynaptic receptors have been reported [10]. Interestingly, these changes were reflective of the stage of the disease. For instance, increased transcriptional activity of genes regulating synaptic function and ATP-synthesis was detected at the early stages of AD, but declined with disease progression. Genes related to cell differentiation and proliferation, metal ion binding, and antigen processing were downregulated at the early stages of AD but were upregulated at the late stages of the disease [11].

Despite these advances, a biomarker that identifies asymptomatic (or even pre-symptomatic) subjects has yet to be described. In this pilot study, we have surveyed blood and skin punch biopsy (fibroblasts, keratinocytes, and endothelial cells) in order to expand the repertoire of potential biomarkers. Our analysis has identified a collection of candidate genes that may help identify such a biomarker in the future and further identifies novel genes that may lead to a better understanding of the pathophysiology of AD.

MATERIALS AND METHODS

Study population

Enrolled in this study were five male AD cases (mean age 74.8 y) and five female AD cases (mean age 72.2 y) diagnosed according to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) criteria for dementia [12]. Additionally, as controls, three healthy males (mean age 72.6 y) and four healthy females (mean age 71.0 y) were enrolled. This study was conducted under the guidelines of the Declaration of Helsinki and all subjects or their respective guardians provided written informed consent before participation under a protocol approved by the Institutional Review Board of the Kazan Federal University (article 20, Federal Law “Protection of Health Right of Citizens of Russian Federation” N323- FZ, 11.21.2011).

Tissue collection and Illumina BeadChip-based gene expression

Previous studies largely focus on either blood or CSF when investigating transcriptomic characteristics of AD. The invasive nature of the latter experiment limits its sample size and usefulness. Some other studies examine transcriptional differences in neurological tissue, yet these studies focus on disease pathophysiology. We speculate that a transcriptional study of two different tissue compartments, one used previously in transcriptional studies (blood), the other a novel approach (e.g., fibroblasts, keratinocytes, and endothelial cells), may provide a potential AD screening method.

From each subject, a three-millimeter skin punch biopsy (fibroblasts, keratinocytes, and endothelial cells) was collected from the forearm of the non-dominant hand and immediately dissolved in Trizol (Life Technologies, Carlsbad, CA) and stored at –80°C. Subsequent extraction was carried out according to the manufacturer’s instructions. Additionally, anti-coagulated whole blood was collected from each subject and total RNA was isolated using the LeucoLock^™ total RNA isolation system (Life Technologies) according to the manufacturer’s instructions. Global gene expression of punch biopsies and whole blood was assessed using Illumina’s HumanHT-12 Expression BeadChips v3 (Illumina, Inc, San Diego, CA). Each array targets more than 25,000 annotated genes (derived from the National Center for Biotechnology Information Reference Sequence (NCBI) RefSeq Release 38) to evaluate more than 48,000 transcripts and known splice variants across the human transcriptome. Total RNA (500 ng) from each sample was amplified and labeled using the Illumina TotalPrep RNA Amplification Kit, AMIL1791 (Ambion, Inc., Austin, TX) according to the manufacture’s instructions. Arrays were hybridized at 58°C for 16–20 h, followed by wash/stain procedures according to the Whole-Genome Gene Expression Direct Hybridization Assay Guide (Illumina). Expression arrays were analyzed with the iScan System, and data were extracted with Gene Expression Module 1.0.6 in GenomeStudio 1.0.2 (Illumina). All respective skin-sample and blood experiments were performed in a single batch.

Microarray data quality control and statistical analysis

Microarray data processing and analysis was performed using the lumi package [13] in the statistical programming language R. To filter out probes that were largely undetected across the 16 samples, detection was first defined as a detection rate (as measured by the detection p-value) with p < 0.01. Any probe that was undetected in all samples (29,736; 61%) was removed from further study. Additionally, any probe that was undetected in more than half the samples in either tissue was excluded from further study (7,167 additional probes). One sample (AD case #8) was excluded from both tissue analyses due to very low call rates (<25%). Standard quantile normalization was then applied to normalize the filtered data [13]. The remaining 11,883 probes were reduced to 10,535 unique accession numbers by averaging multiple probes per accession. A standard principal component analysis was carried out on these probes, and showed a clear separation between the two tissue types, with 85% of the total variation in the samples being explained by this difference. An additional quality control step was performed as follows: any value that was greater than two standard deviations away from the mean gene expression level of the cohort value (either case or control) was deleted. One sample (AD case #16) was excluded from the skin-sample analysis due to a high rate of these outlying values (8.5%).

Differential expression of genes with 1.5-fold or greater change between cohorts was tested for statistical significance using a Student’s t-test on quality-controlled log₂-transformed gene expression values across cohorts. A multiple correction to adjust the false discovery rate was performed on the resulting p-values [14]. Hierarchical clustering was performed on blood samples and punch biopsy samples separately, using the Manhattan distance metric and the average agglomerative method with the hclust function in R.

To examine expression data for potential biomarkers in both experiments, genes with very low variation across the control samples and relatively high variation across the patient samples were evaluated. Specifically, any gene with a Coefficient of Variation (CV) of less than 10% across the controls and a patient-CV of at least three-fold the control-CV were examined.

Enrichment of GO (Gene Ontology) biological processes in the subsets of statistically significantly differentially expressed genes was examined using the BiNGO (3.0.3) plugin for Cytoscape (3.2.1), and the April 2016 Homo sapiens gene association file (http://geneontology.org/page/download-annotations). The p-values associated to BiNGO’s hypergeometric test were adjusted to correct for the false discovery rate using Benjamini-Hochberg multiple testing correction [14].

Random Forest tree-based ensemble machine learning was utilized to identify transcript combinations that accurately identified cases and controls [15]. Subject status was used as the target variable and the gene transcript values as the predictive variables. The model was made using tree type classification mode, building 2000 decision trees and out of bag testing without replacement, using two predictors at each decision node.

RESULTS

Differential expression and BINGO analysis

Of the 25,000 annotated genes represented on the Illumina HumanHT-12 Expression BeadChip, 10,535 remained after quality control steps were implemented. Statistical analysis identified 129 differentially expressed transcripts from blood that were observed to show a more than 1.5-fold difference between AD cases and controls (adjusted p < 0.05). The 10 most significant genes are given in Table 1 and all statistically significant genes are listed in Supplementary Table 1. Distinct separation between AD cases and controls was observed in the hierarchical clustering based on these genes (Fig. 1A). Additionally, four differentially expressed transcripts from skin punch biopsy were found to be statistically significant with more than a 1.5-fold difference between AD cases and controls (adjusted p < 0.05, Table 1, and Supplementary Table 1). Hierarchical clustering of skin tissue was made using the top 17 genes with adjusted p-value p < 0.1 and raw p-value <0.05 (Fig. 1B). Based on this cluster analysis, all controls were accurately separated, while two cases were miscategorized as controls. These data suggest that blood and skin generally have the capacity to be used to stratify cases and controls.

Table 1.

Top significant genes

TRANSCRIPTS FROM BLOOD
Gene Symbol	Gene Name	Mean Ctrl	Mean Case	Ratio	p-value	Adj p val
SKAP2	Src Kinase Associated Phosphoprotein 2	8.05	8.95	1.87	1.65E-05	0.0030
BCDIN3D	BCDIN3 Domain Containing	6.53	7.17	1.55	0.000164	0.0035
CDC42EP2	CDC42 Effector Protein 2	6.92	7.69	1.70	0.000187	0.0035
CR1	Complement Component (3b/4b) Receptor 1 (Knops Blood Group)	6.70	7.41	1.64	0.000177	0.0035
FRAT1	Frequently Rearranged In Advanced T-Cell Lymphomas 1	8.02	8.63	1.53	0.000140	0.0035
MANSC1	MANSC Domain Containing 1	7.00	7.98	1.97	9.16E-05	0.0035
MAPK1	Mitogen-Activated Protein Kinase 1	8.80	9.41	1.53	0.000189	0.0035
MGAM	Maltase-Glucoamylase	6.59	7.30	1.63	0.000146	0.0035
TNFRSF1A	Tumor Necrosis Factor Receptor Superfamily, Member 1A	10.37	11.10	1.66	7.31E-05	0.0035
VPS24	Charged Multivesicular Body Protein 3	7.38	7.99	1.53	5.02E-05	0.0035
TRANSCRIPTS FROM PUNCH BIOPSIES
CPNE1	Copine I	8.63	9.25	1.53	0.004250	0.0467
ISG15	ISG15 Ubiquitin-Like Modifier	6.70	7.34	1.57	0.001278	0.0467
LDLR	Low Density Lipoprotein Receptor	10.35	11.08	1.66	0.002928	0.0467
LGALS3BP	Lectin, Galactoside-Binding, Soluble, 3 Binding Protein	7.32	7.95	1.54	0.003935	0.0467

Fig. 1.

Hierarchical clustering for blood samples (A) and punch biopsy samples (B) using the Manhattan distance metric and the average agglomerative method with the hclust function in R. Clustering of was made using statistically significant genes (adjusted p-value <0.05) and tissue from punch biopsies was made using the top seventeen genes with adjusted p-value p < 0.1 and raw p-value <0.05.

In order to explore potential functional relationships among these differentially expressed genes we used the Biological Networks Gene Ontology (BiNGO) plugin (3.0.3) for Cytoscape (3.2.1) to map transcripts to their predominant functional themes and measure enrichment of each GO theme [16]. The p-values associated to the hypothesis test of each represented GO category were adjusted to correct for the false discovery rate using Benjamini-Hochberg multiple testing correction [14]. Using the subset of 129 statistically significant genes from the blood data set, BiNGO identified 73 potential functional themes (Supplementary Table 2). The most significant themes encompassed 24 genes related to defense response (adjusted p < 1.50 × 10⁻³); 32 genes involved in cell surface receptor signaling (adjusted p < 3.50 × 10⁻³); 11 genes involved in the regulation of inflammatory responses (adjusted p < 3.51 × 10⁻³). BiNGO analysis of differentially expressed transcripts from punch biopsies primarily identified themes related to type I interferon responses and cell cycle control (adjusted p < 1.55–1.69 × 10⁻², Supplementary Table 2).

Examination of possible biomarkers by alternative analysis

Although the classic two-sample unpaired t-test is the textbook tool to test two sample means for statistical difference, it will not readily identify a gene with housekeeping-like behavior in one cohort and notably varying behavior in the second. This type of behavior, however, may be indicative of polymorphic behavior across a disease cohort. To this end, we examined any gene with a CV of less than 10% across the controls (we call this the control-CV) and a case-CV of at least three-fold the control-CV. This method identified 18 blood and 49 skin genes shown in Fig. 2A and B, respectively (complete gene list in Supplementary Table 3), that are minimally variant across control samples (control-CV <10%) and substantially variant across AD cases (case-CV >30%). In order to explore potential associations between these putative “AD-associated variant genes”, we performed a functional interaction analysis using the program STRING 10.0; a web-based program that employs a biological database of known and predicted protein–protein interactions to identify potential association among proteins [17]. Of the 18 “AD-associated variant genes” from the blood tissue analysis, 16 were annotated and recognized in the STRING database. Using a minimum required interaction score of 0.150, the identified network did not have significantly more interactions than expected; however two significant transcripts were identified in the PFAM Protein and INTERPRO Domains, (YPIF4 and YIPF6, adjusted p-values of 0.0194 and 0.0415, respectively). Of the 49 “AD-associated variant genes” from the skin study, 43 annotated genes were recognized in the STRING database. Again, a minimum required interaction score of 0.150 was used and significant number of interactions were observed (adjusted p-value: 1.41 × 10⁻⁶, Fig. 3); however, no specific biochemical pathways were identified.

Fig. 2.

Plot of variable genes with a Coefficient of Variation (CV) of less than 10% across the controls and a case-CV of at least three-fold the control-CV. This method identified 18 blood (A) and 49 skin gene (B). List of genes is given us Supplementary Table 2.

Fig. 3.

Protein-protein interaction network of 43 genes identified from variable transcripts of punch biopsies. Interaction network was identified using the web-based program STRING 10.0. A minimum required interaction score of 0.150 was used and significant number of interactions were observed (PPI enrichment p-value: 1.41 × 10⁻⁶). Lines colors indicate the interactions were determined as follows: Pink, experimentally determined; light green, textmining; purple, protein homology; dark green, gene neighborhood; black, co-expression; dark blue, gene co-occurrence.

Classification of gene transcript by importance

To further identify transcripts that might distinguish cases and controls, a Random Forest analysis (RF) was applied to the expression data. An initial analysis combined data from blood and punch biopsies to yield a model that delineated cases from controls with only 43.75% sensitivity but the specificity was observed to be 64.29% (data not shown), suggesting that the data were more useful in qualifying control subjects. An RF model using only blood expression measures yielded specificity and sensitivity 57.14% and 50.00%, respectively (Table 2). An RF model using only punch biopsy expression data returned specificity and sensitivity of 50.00% and 71.43%, respectively (Table 2). This analysis points to a possible subset of genes to select control samples for clinical trials. The ten most influential genes identified by RF analysis and their importance score for each respective tissue compartment are given in Table 3.

Table 2.

Random forest signature success rate

Actual Class	Total Cases	Percent Correct	Cases n = 7	Controls n = 8	Subjects Misclassed	Percentage Error
Case Blood	8	50.00%	4	4	4	50.00%
Control Blood	7	57.14%	3	4	3	42.86%
Case Punch Biopsy	8	50.00%	4	4	4	50.00%
Control Punch Biopsy	7	71.43%	5	2	2	28.57%

Table 3.

Random forest signature variable importance

AD-ASSOCIATED GENES FROM BLOOD
Gene	Importance	Functional Description
NKIRAS2	100	Regulator of NF-kappa-B activity by preventing the degradation of NF-kappa-B inhibitor beta
KLF12	89.58	Regulator of gene expression during vertebrate development
C9ORF91	80.99	Chromosome 9 Open Reading Frame 91 (Primarily propduced in the brain)
ARRB2	77.48	Functions in regulating agonist-mediated G-protein coupled receptor (GPCR) signaling
PKIA	76.92	cAMP-dependent protein kinase (PKA) inhibitor
NUDT16	74.88	RNA-binding and decapping enzyme
B4GALT3	72.13	Type II membrane-bound glycoprotein
DTWD1	69.93	Expressed in hematopoietic stem/progenitor cells from myelodysplastic syndromes patient
CIDEB	69.82	Required for hepatitis C virus entry into hepatocytes
IKIP	69.81	IKBKB Interacting Protein
AD-ASSOCIATED GENES FROM PUNCH BIOPIES
Gene	Importance	Functional Description
C16orf54	100	Uncharacterized protein, predected membrane protein.
C21ORF58	93.47	Uncharacterized protein
LGALS8	77.47	Beta-galactoside-binding
RFC1	75.52	Subunit of replication factor C, DNA polymerase accessory protein
IL32	75	Induces the production of TNFalpha from macrophage cells
RNF126	68.28	E3 ubiquitin-protein ligase, may play a role in the endosomal sorting
C11ORF24	66.76	Type I membrane protein. cycles between the Golgi and the plasma membrane
ARID5B	66.02	Forms a histone H3K9Me2 demethylase complex
CDK5RAP3	59.2	Function in signaling pathways governing transcriptional regulation
ZNF160	57.28	Kruppel-related zinc finger protein

DISCUSSION

The goal of this pilot study was to identify differentially expressed genes that may lead to a biomarker of AD in a larger subject cohort and provide new knowledge into the pathophysiology of this disease. While several previous studies have investigated the AD blood transcriptome, the knowledge gained by such information largely pertains to the immune system, as most mRNA in whole blood comes from leukocytes. We hypothesized that by investigating a non-canonical second tissue compartment, systemic changes that manifest in the brain may transcend into other tissues, and fibroblasts, keratinocytes, and endothelial cells from skin punch biopsies represent one of the most easily accessible tissue compartments to investigate. As expected, our analysis of blood identified a number of differentially expressed genes (129 genes) in AD cases when compared to healthy matched controls. The most significant GO themes identified by BiNGO analysis were primarily related to defense and inflammatory responses. Our analysis of punch biopsies was less rewarding, as it identified only four genes that were differentially expressed and reached statistical significance. These genes are primarily involved in interferon response and cell cycle control.

Previous studies have implicated a number of the transcripts identified in our pilot study in the pathophysiology of AD. For instance, CR1, “component (3b/4b) receptor 1 (Knops Blood Group)” was the 4th most significant transcript identified in our blood analysis and has been identified as an AD-risk factor in numerous studies [16–20]. Additionally, changes in the mRNA expression of extracellular signal-regulated kinase, including MAPK1 (mitogen-activated protein kinase 1), have associated with neurodegenerative disease including AD [21], Parkinson’s disease [22], and Huntington’s disease [23]. Although our transcriptional analysis of skin punch biopsies only produced 4 significant transcripts, 2 of these are of interest in relation to AD: ISG15 has been proposed as a general marker of acute and chronic neuronal injury [24] and alterations in LDLR has been identified in association and AD in humans [25] as well as animal models of AD [26]. The most significant transcript identified from punch biopsies, CPNE1 (Copine-1), has not been described in neurodegenerative disorders; however, it has been suggested to play an important role in neuronal differentiation [27].

As standard data analysis did not reveal biological differences strong enough to lead to potential biomarkers, we implemented a non-classical analytic approach to investigate the variability of genes across cases and across controls, as well as between the two cohorts. Any gene with a relatively small CV across the controls and a relatively large CV across the cases was identified: these genes are quite stable across controls but highly variant across AD cases. We refer to these as “AD-associated variant” genes, to represent their variability or non-predictability across the AD cases. Rather than exhibit differences between average control and average case expression values as standard differential gene analysis does, these genes reveal very notable differences in overall behavior across the two cohorts, with the behavior of these genes across the controls similar to that of standard housekeeping genes. The rational for this approach was to identify genes that normally display constitutive expression but may be irregularly up- and downregulated during the course of the disease. Indeed, such expression may go unnoticed when looking at average changes between cohorts, potentially missing important biological factors. In a very recent report, Guan and colleagues using a similar strategy of data analysis in a much larger cohort, identified potential biomarkers in autism spectrum disorder [28].

STRING analysis of these variable transcripts in blood identified two significant transcripts in the PFAM Protein and INTERPRO Domains, both of which are poorly characterized membrane spanning protein. The first, YIPF4, is believed to be involved in the maintenance of the Golgi structure [29], and has been implicated as a novel cellular binding partner of the papillomavirus E5 proteins [30]. The other, YIPF6, has been suggested as a potential biomarker in the neurodegenerative disorder Machado-Joseph disease [31] and mutations in the Yipf6 gene in mice leads to spontaneous intestinal inflammation. When take together, these data, and that of Guan and colleagues, suggest that in complex neurological disorders, the variability of gene expression may be as informative as the average expression.

In order to identify potential biomarkers that may not be obvious solely from classical statistical significance, we utilized the data-mining program RF. RF uses an ensemble of unpruned classification or regression trees produced through bootstrap sampling of the training data set and random feature selection in tree generation. Prediction is made by a majority vote of the predictions of the ensemble. The strength of the analysis was evaluated by an “out of bag” sampling without replacement of the original data. RF is an attractive method since it handles both discrete and continuous data, it accommodates and compensates for missing data, and it is invariant to monotonic transformations of the input variables. RF analysis of blood transcripts was only slightly more successful in identifying control subjects then would have been observed by random chance selection (57%); however, the RF model correctly identified control subjects with 71.4% specificity. These data suggested that this tissue compartment might be useful when qualifying controls for clinical trials; however, as this was a pilot study, the number of study subjects was small and replicating this research with a larger study cohort may be more productive. Considering the protracted prodromal phase of AD, it may be necessary to conduct studies over an extended period of time in order to reveal asymptomatic and/or pre-symptomatic individuals in the very early stages of the disease. However, the genes identified in this study provide important knowledge as to potential candidates for such investigations.