Differential and correlation analyses of microarray gene expression data in the CEPH Utah families

Abstract

The widespread microarray technology capable of analyzing global gene expression at the level of transcription is expanding its application in not only medicine but also studies on basic biology. This paper presents our analysis on microarray gene expression data in the CEPH Utah families focusing on the demographic characteristics such as age and sex on differential gene expression patterns. Our results show that the differential gene expression pattern between age groups is dominated by down-regulated transcriptional activities in the old subjects. Functional analysis on age regulated genes identifies cell-cell signaling as an important functional category implicated in human aging. Sex-dependent gene expression is characterized by genes that may escape X-inactivation and, most interestingly, such a pattern is not affected by the aging process. Analysis on sibship correlation on gene expression revealed a large number of significant genes suggesting the importance of a genetic mechanism in regulating transcriptional activities. In addition, we observe an interesting pattern of sibship correlation on gene expression that increases exponentially with the mean of gene expression reflecting the enhanced genetic control over the functionally active genes.

Introduction

The widespread microarray technology capable of analyzing global gene expression at the level of transcription is expanding its application in not only medicine but also studies on basic biology. For example, analyses of developmental and/or sex-regulated transcriptional profiles have been done in both experimental species [1, 2] and human beings [3, 4]. Recently, Morley et al [5] performed linkage analysis in the CEPH Utah pedigrees on their expression phenotypes of 3,554 genes in an attempt to map epigenetic genes that are linked to the control or regulatory network of gene expression in the immortalized lymphoblastoid cells. Their analysis has located both cis- and trans-acting regulators indicating a significant genetic contribution to transcriptomic activities. In their relatively large microarray experiment, the arrayed samples consist of, for each of the families, grandparents, parents and siblings. The rich content in their data provides valuable information for conducting statistical analysis on relating individual’s demographic variables with transcriptomic patterns. Such analyses should help us to link the global transcriptomic profiling with individual physiological parameters including age- and sex-dependent regulation on gene expression. Using the CEPH Utah family gene expression data, we apply regression analysis using the generalized estimation equations (GEEs) [6] to identify genes that are significantly regulated by the demographic variables. In addition, we also perform sibship correlation analysis on gene expression using the intra-class correlation coefficient (ICC) to validate the existence of a genetic component in global gene expression regulation reported in previous studies.

Results

We first regress the observed individual gene expression levels to the two demographic factors, i.e. generation or age group (old group: grandparents coded as 1; young group: grandchildren as 0) and sex (males coded as 1, females as 2) using GEEs with each family or pedigree as one correlated cluster. The analysis yields 898 genes that are differentially expressed in the old and the young groups with a significance level of FDR<1E-15. Figure 1 is a heatmap from the clustering analysis for the 200 topmost significant (absolute z statistic>11.728) genes (detailed information available in Supplementary Materials). Among these genes, 158 are down and only 42 are up regulated in the old group. Nearly all subjects in the two groups (old and young) are clearly distinguished by the 200 genes with grandparents clustered to the left (blue side bar, 56 subjects) and grandchildren to the right (red side bar, 110 subjects) panels. The 200 genes in Figure 1 were also submitted to EASE package for functional clustering and pathway analysis. Table 1 shows the 22 significant functional categories identified by EASE with an EASE score<0.05. After Bonferroni correction for multiple testing as provided by EASE, only one category remains significant, i.e. cell-cell signaling (adjusted p-value=6.12E-04). However, it is interesting to see that among the top categories are those involved in cell communication, signal transduction and receptor signaling pathways.

Figure 1.

Heatmap with clustering of the 200 topmost significant genes displaying age-dependent expression patterns. Nearly all subjects in the two groups are clearly distinguished with the grandparents clustered to the left and grandchildren to the right panels.

Table 1.

Significant functional categories identified by EASE for the 200 topmost significant genes regulated by age

System	Gene Catery	List Hits	Population Hits	EASE score
Biological Process	Cell-cell signaling	30	495	0.000
Biological Process	Cell communication	68	2402	0.008
Biological Process	Inorganic anion transport	6	60	0.009
Molecular Function	Channel/pore class transporter activity	14	290	0.010
Biological Process	Chloride transport	5	42	0.012
Molecular Function	Signal transducer activity	51	1733	0.014
Molecular Function	Voltage-gated ion channel activity	8	119	0.014
Molecular Function	Alpha-type channel activity	13	280	0.018
Biological Process	Anion transport	7	98	0.019
Cellular Component	Extracellular	32	979	0.022
Molecular Function	Ion channel activity	12	258	0.024
Biological Process	Ion transport	17	433	0.025
Biological Process	Muscle contraction	8	136	0.028
Biological Process	Development	42	1422	0.028
Molecular Function	Receptor binding	17	441	0.028
Biological Process	Cell surface receptor linked signal transduction	28	875	0.036
Biological Process	Sex differentiation	3	14	0.036
Biological Process	Organogenesis	26	804	0.039
Molecular Function	Voltage-gated chloride channel activity	3	15	0.040
Cellular Component	Transcription factor complex	17	458	0.043
Biological Process	Enzyme linked receptor protein signaling pathway	8	151	0.045
Biological Process	TGFbeta receptor signaling pathway	4	38	0.048

Table 2 has more details about the 30 genes involved in the cell-cell signaling pathway. In this table, the fold change is calculated as the ratio between the mean of gene expression in the old subjects and the mean of expression in the young controls. Of the 30 significant genes, only 4 genes (13%) are up-regulated (ratio>1) while the other 26 genes are all down-regulated in the old subjects. This pattern is very similar to the pattern of the 200 significant genes among which only 42 genes (21%) are up-regulated. One of the four up-regulated genes, 214637_at, is extremely highly expressed in the old group with a fold change of 9. This gene encodes a growth regulator which inhibits the proliferation of a number of tumor cell lines.

Table 2.

Details about the 30 significant genes in the cell-cell signaling pathway

Probe-set	Fold change	Gene symbol	Gene name
203757_s_at	0.846188	CEACAM6	carcinoembryonic antigenrelated cell adhesion molecule 6 (non-specific cross reacting antigen)
202668_at	0.815063	EFNB2	ephrin-B2
209409_at	0.846958	GRB10	bound protein 10
205019_s_at	0.824349	VIPR1	vasoactive intestinal peptide receptor 1
205117_at	0.801323	FGF1	fibroblast growth factor 1 (acidic)
219287_at	0.866002	KCNMB4	potassium large conductance calciumactivated channel, subfamily M, beta member 4
205110_s_at	0.781458	FGF13	fibroblast growth factor 13
205204_at	0.878151	NMB	neuromedin B
205239_at	0.795326	AREG	amphiregulin (schwannoma-derived growth factor)
207466_at	0.870967	GAL	galanin
208124_s_at	0.852245	SEMA4F	sema domain,immunoglobulin domain (Ig), transmembrane domain (TM) and short cytoplasmic domain, (semaphorin) 4F
205747_at	0.835613	CBLN1	cerebellin 1 precursor
206525_at	0.844777	GABRR1	gamma-aminobutyric acid (GABA) receptor, rho 1
210796_x_at	0.843327	SIGLEC6	sialic acid binding Ig-like lectin 6
211110_s_at	0.824972	AR	androgen receptor (dihydrotestosterone receptor; testicular feminization; spinal and bulbar muscular atrophy; Kennedy disease)
207135_at	1.162481	HTR2A	5-hydroxytryptamine
206268_at	0.841076	LEFTB	left-right determination, factor B
211222_s_at	0.795567	HAP1	huntingtin-associated protein 1 (neuroan 1)
206938_at	0.714895	SRD5A2	steroid-5-alpha-reductase, alpha polypeptide 2 (3-oxo-5 alpha-steroid delta 4-dehydrogenase alpha 2)
208251_at	1.230858	KCNC4	potassium voltage-gated channel, Shaw-related subfamily, member 4
210549_s_at	1.499597	CCL23	chemokine (C-C motif) ligand 23
208460_at	0.808868	GJA7	gap junction protein, alpha 7, 45kDa (connexin 45)
207309_at	0.856292	NOS1	nitric oxide synthase 1 (neuronal)
208052_x_at	0.898756	CEACAM3	carcinoembryonic antigenrelated cell adhesion molecule 3
214529_at	0.824701	TSHB	thyroid stimulating hormone, beta
214637_at	9.266827	OSM	oncostatin M
221330_at	0.787841	CHRM2	cholinergic receptor, muscarinic 2
211405_x_at	0.849933	IFNA17	interferon, alpha 17
221371_at	0.81324	TNFSF18	tumor necrosis factor (ligand) superfamily, member 18

In contrast to age, our analysis only identified a limited number of genes that are differentially expressed by sex. In the group consisting of young siblings (57 males and 53 females), 101 genes were found as significantly regulated (FDR<0.05). In the group of grandparents (28 males and 28 females), 66 genes were differentially expressed (FDR<0.05). Interestingly, in both groups, the genes on top of the lists are dominated by X-linked genes. In Figure 2, we show the absolute z statistic plotted against the male to female (M/F) ratio of mean expression for each of the 397 genes on the X-chromosome in the young and old groups with the names of significant genes (FDR<0.05) marked red (19 genes in the young and 15 genes in the old groups). Of the 15 sex-regulated X-linked genes in the old group, 14 overlap with the gene list of the young group (Table 3, marked as bold). Note that in both groups, the top significant genes are mainly highly expressed in females. Figure 3 is a heatmap for the 19 X-linked genes identified in the young group which displays again the predominant pattern of high expression in females together with a small number of genes up-regulated in males showing at the bottom of the figure. The 19 X-linked significant genes in the young group were further plotted on the X-chromosome (Figure 4). The genes are concentrated to the end of the short arm of the X-chromosome (pseudoautosomal genes marked blue). The distribution pattern in Figure 4 falls exactly to the distribution of genes that escape X-inactivation [7]. Table 4 has the results of EASE analysis on the 101 sex-regulated genes (7 on Y-Chromosome, 19 on X-chromosome, 75 on autosomal chromosomes; FDR<0.05) in the young group for the functional categories with EASE core<0.05. Among the 18 categories in Table 4, 4 are associated with reproduction (spermatogenesis, male gamete generation and sexual reproduction). Interestingly, all the 4 categories contain the same 5 probes with one X-linked (201099_at with a homolog on Y chromosome and gene symbol USP9X), three Y-linked (20500_at and 206624_at with homolog on X chromosome and gene symbols DDX3Y and USP9Y; 206700_s_at with gene symbol SMCY), one autosomal on chromosome 17(221326_s_at, gene symbol TUBD1). Note that all the three Y-linked genes are up-regulated while the other two genes down-regulated in males. Of the sex-regulated autosomal genes, 75 genes (20 down and 55 up regulated) were found in the young and 44 genes (30 down and 36 up regulated) in the old groups (FDR<0.05) with only 2 overlaps (gene list not shown).

Figure 2.

Absolute z statistics plotted against the male to female ratio of mean gene expression for each of the 397 genes on the X-chromosome in the young and the old groups with the names of significant genes (FDR<0.05) marked red (19 genes in the young and 15 genes in the old groups).

Table 3.

Significant sex-regulated genes on X-chromosome^*

Probe-set	Slope	FDR	Chromosomal location	Gene symbol
Young
203992_s_at	0.486	1.960E-11	chrXp11.2	UTX
201589_at	0.436	0	chrXp11.22-p11.21	SMC1L1
202383_at	0.458	0	chrXp11.22-p11.21	SMCX
200964_at	0.303	0	chrXp11.23	UBE1
201210_at	0.273	5.336E-03	chrXp11.3-p11.23	DDX3X
201099_at	0.179	8.930E-05	chrXp11.4	USP9X
208174_x_at	0.262	2.390E-07	chrXp22.1	U2AF1L2
206751_s_at	−0.186	1.497E-03	chrXp22.11	PCYT1B
201018_at	0.439	0	chrXp22.12	E1F1AX
204061_at	0.438	0	chrXp22.3	PRKX
207551_s_at	0.508	2.300E-07	chrXp22.3	SML3L1
206148_at	−0.374	6.829E-03	chrXp22.3/Yp11.3	IL3RA
203974_at	0.757	0	chrXp22.32	HDHD1A
205206_at	−0.751	0	chrXp22.32	KAL1
203767_s_at	0.718	3.090E-09	chrXp22.32	STS
201029_s_at	−0.308	2.334E-03	chrXp22.32;Yp11.3	CD99
200933_x_at	0.205	0	chrXq13.1	RPS4X
214218_s_at	4.921	0	chrXq13.2	XIST
201479_at	0.113	3.130E-02	chrXq28	DKC1
Old
203992_s_at	0.564	0	chrXp11.2	UTX
202383_at	0.552	0	chrXp11.22-p11.21	SMCX
201589_at	0.312	5.570E-08	chrXp11.22-p11.21	SMC1L1
200964_at	0.206	9.640E-05	chrXp11.23	UBE1
201210_at	0.254	2.498E-02	chrXp11.3-p11.23	DDX3X
201099_at	0.170	1.651E-02	chrXp11.4	USP9X
208174_x_at	0.451	0	chrXp22.1	U2AF1L2
201018_at	0.387	0	chrXp22.12	E1F1AX
204061_at	0.647	0	chrXp22.3	PRKX
207551_s_at	0.650	3.700E-07	chrXp22.3	SML3L1
203974_at	0.566	0	chrXp22.32	HDHD1A
205206_at	−1.048	1.710E-09	chrXp22.32	KAL1
203314_at	0.185	1.260E-02	chrXp22.33;Yp11.32	GTPBP6
200933_x_at	0.162	0	chrXq13.1	RPS4X
214218_s_at	5.397	0	chrXq13.2	XIST

^*Overlapping genes are marked as bold and unoverlapping as italic.

Figure 3.

Heatmap for the top 19 X-linked genes differentially expressed by sex in the young group. Among the 110 individuals, nearly all males are clustered to the left and females to the right panels.

Figure 4.

Location of the 19 sex regulated significant genes (marked as red circles) on the X-chromosome in the young group (pseudoautosomal genes marked blue). There is an obvious concentration on the short arm (Xp) especially to the extreme end of Xp.

Table 4.

Significant functional categories identified by EASE for the 101 significant genes regulated by sex in the young group

System	Gene Catery	List Hits	Population Hits	EASE score
Molecular Function	RNA binding	15	375	0.000
Molecular Function	Nucleic acid binding	35	1897	0.002
Cellular Component	Ribonucleoprotein complex	10	290	0.006
Biological Process	RNA processing	9	248	0.007
Cellular Component	Spliceosome complex	5	69	0.008
Biological Process	RNA metabolism	9	269	0.012
Cellular Component	cAMP-dependent protein kinase complex	3	15	0.013
Cellular Component	Obsolete cellular component	10	324	0.013
Biological Process	RNA splicing	5	92	0.021
Biological Process	Development	25	1422	0.027
Biological Process	Male gamete generation	5	102	0.029
Biological Process	Spermatogenesis	5	102	0.029
Biological Process	Sexual reproduction	6	152	0.030
Molecular Function	Pre-mRNA splicing factor activity	4	61	0.030
Biological Process	Reproduction	6	153	0.030
Cellular Component	Nucleus	34	2101	0.032
Molecular Function	Cell adhesion molecule activity	8	277	0.035
Molecular Function	Chromatin binding	3	29	0.041

Our sibship gene expression correlation analysis found that there are 4,658 genes showing ICCs with an empirical p-value<5.0E-04. Figure 5 is a scatter plot that plots the estimated ICC against the calculated coefficient of variation (CV) (left panel) and the mean of gene expression (in log scale) (right panel) with the significant genes marked in purple and insignificant gene in green. The figure shows that the correlation has nothing to do with the magnitude of variation. However, it is interesting to see that ICC increases exponentially with the mean of gene expression. Figure 5 also shows that genes displaying high sibship correlation have high expression levels although genes that are highly expressed may not necessarily exhibit high correlation.

Figure 5.

Scatter plot for the estimated ICC against the calculated CV (left panel) and the mean of gene expression (in log scale) (right panel) with significant genes marked in purple and insignificant genes in green.

Discussion

Our demographic analysis on microarray gene expression data in the CEPH Utah families has identified important differential transcriptional profiles that are associated with demographic factors such as age and sex. Results from our functional analysis on the top significant genes using EASE (Table 1) revealed that cell communication and signal transduction are important biological processes/molecular functions that are implicated in human aging. Since the age-dependent expression profiles are predominated by genes that are down regulated in the old group (158 versus 42) (Figure 1), we conclude that human aging is accompanied by the deduction or deterioration of the activities in these biological processes. Changes in signal transduction cascade and gene expression have been linked to important functional manifestations of aging such as altered regulation of physiological and behavior processes [8, 9]. It has been shown that decreases in cytokine receptor synthesis and in the activities of signal molecules are responsible for these dysregulations during aging [10, 11] leading to age-associated disorders such as depressed cardiac function [12] and decreased immune responses [13–15].

Functional analysis on top sex-regulated genes showed excessive involvement of these genes in spermatogenesis, male gamete generation and sexual reproduction (Table 4). Of the genes involved, one Y-linked gene DDX3Y (205000_at) encodes proteins that are believed to be involved in embryogenesis, spermatogenesis, and cellular growth and division. Its mutation causes male infertility, Sertoli cell-only syndrome or severe hypospermatogenesis, suggesting that this gene plays a key role in the spermatogenic process. Given the high proportion of overlapping in the X-linked significant genes in the old and the young groups, we conclude that activities of these genes are not affected by the aging process (Figure 2). Comparing the overlapping genes in Table 3, one could see that their estimated slopes are more or less similar meaning that their expression levels are maintained during aging. Although most of the significant X-linked genes are up-regulated in females than in males, there are also X-linked genes that are more expressed in males than in females. Of particular interest is probe-set 205206_at which is distinctively up-regulated in males in both the young and old groups. Chromosomal location analysis (Figure 4) shows an obvious concentration on the short arm (Xp). Such a distribution coincides with the location pattern found by Carrel et al. [7] and could suggest that the sex-regulated X-chromosome genes we have found are mainly genes that may escape X-inactivation. However, the reversed expression pattern (high expression in males) for probe-sets such as 205206_at deserves more clarification.

The results of sex-dependent gene expression from the autosomal chromosomes cast a different picture as compared with the X-linked genes. The overwhelming up-regulation pattern of the X-linked genes in females is not observed here. Importantly, the fact that nearly no top significant sex-regulated autosomal genes overlap in the old and young groups makes us to postulate that, in contrast to the X-linked genes, activities of these genes could have been re-adjusted at different developmental stages or during the aging process.

The gene expression phenotype can be regulated by both genetic (cis- or trans-effects [5] from the gene itself or from elsewhere in the genome) and non-genetic mechanisms. The genetic control over gene expression can result in correlated pattern of the phenotype in genetically related individuals due to sharing of genetic materials. There have been many recent studies on the genetic contribution to gene expression using related individuals such as twins [16–20] with predominantly positive results confirming the existence of a genetic mechanism. The important result from our correlation analysis on the large sibships is that ICC increases exponentially with the mean of gene expression. Such a pattern suggests that genes showing high sibship correlation are functionally active genes but not vice versa (Figure 5), an interesting phenomenon that deserves further analysis. Here, we emphasize that our result could reflect the enhanced genetic control over the functionally active genes. Different from twin data in which genetic and environmental components can be decomposed, our results on sibship gene expression correlation can only suggest the existence of a genetic contribution in gene expression regulation because such a correlation can well be resulted from the shared family environment. We argue that, since our data contains only large sibships (7 to 8 sibs in each family) and thus the old and the young sibs from the same family are of quite different ages, environmental sharing (in food, living condition etc) can be considerably reduced. This means that our observed correlation can actually more reflect a genetic correlation. Meanwhile, Figure 5 shows that the gene filtering procedure in microarray analysis using CV may not be a good idea since the magnitude of CV has nothing to do with the gene’s biological significance. Filtering genes according to their mean expression levels can be more meaningful.

Conclusions

Our demographic analysis of gene expression data in the CEPH Utah families has identified interesting patterns in gene expression regulation. Differential expression analysis across age groups identified the decreased cell-cell signaling as an important biological process implicated in human aging. While sex-dependent expression patterns of X-linked genes are dominated by genes that may escape X-inactivation, such a pattern is not affected by the aging process. Sibship correlation analysis suggests that genes whose transcriptomic activities show genetic control are functionally active genes but not vice versa.

Materials and methods

The CEPH Utah family gene expression data

The microarray gene expression data for the CEPH Utah families are freely available from the Gene Expression Omnibus (GEO) at http://www.ncbi.nlm.nih.gov/geo/ with an accession number GSE1485 [5]. The samples contain subjects from 14 families each covering three generations (grandparents, parents and grandchildren or offspring). In the 14 pedigrees, twelve families have 8 and two families have 7 offspring. Altogether there are 194 individuals available for the study. In order to avoid age overlap in the generations, we dropped the middle generation (parents, 28 arrays) which resulted in an old (grandparents) group (56 arrays) and a young (grandchildren) group (110 arrays) for subsequent analysis.

Microarray analysis

The gene expression data obtained by hybridizing RNA extracted from immortalized lymphoblastoid cells to the Affymetrix Focus Array containing about 8,500 genes [5]. The raw gene expression data were preprocessed by dChip software (http://biosun1.harvard.edu/complab/dchip) for data normalization using the invariant-set normalization method [21] and for summarizing the gene expression index using the model-based approach [22].

Statistical analyses

The generalized estimating equations (GEEs) [6] are an extension of the generalized linear model (GLM) to accommodate the correlated structure in observed data. Previously, we have applied the GEEs in an association analysis on twin data to account for the genetic relatedness in the twin pairs [23]. Since our aim here is to analyze the importance of individual demographic factors (i.e. age and sex) in gene expression regulation, the GEEs will be introduced to deal with the correlated patterns of gene expression within families or sibship. In our analysis, we use the exchangeable (or compound symmetry) working correlation matrix and identity link function which is equivalent to a random effect model with a random intercept for each cluster (family or sibship). Data analysis using GEEs is conducted using the free R package gee. In order to account for multiple testing, the p-values are adjusted by calculating the false discovery rate (FDR) [24]. FDR controls the expected proportion of false positives among the declared significant results and is a popular and convenient measurement in microarray studies. For genes that are significantly regulated by the demographic variables, we apply EASE software (http://david.abcc.ncifcrf.gov/ease/ease.jsp) [25] to cluster genes into biological pathways or functional categories to identify significant pathways that are involved. For a given list of genes, EASE performs over-representation analysis and report the significance of each functional category as an EASE score which is the upper bound of the distribution of the Jackknife Fisher exact probabilities. Hierarchical clustering method is applied to visualize the data for the significant genes using the free R package gplots.

To summarize an overall sibship correlation for the multiple sibs in the families, we introduce the intra-class correlation coefficient (ICC) in reliability studies which is also a method popular in use in psychology [26] and in population genetics especially twin studies [27, 28]. Recently, the method has been applied in estimating twin correlation on radiation-induced gene expression phenotypes [19]. Following the definition [26], the statistic is calculated as $ICC=\frac{MSB-MSW}{MSB+MSW}$. Here MSW is the within-sibship mean square and MSB the between-sibship mean square. An ICC is calculated for each gene on the array. A positive and large ICC implies that there is a small within-sibship variation in the expression phenotype which means high correlation while, especially, a negative or close to zero ICC indicates that the within-sibship variation exceeds or nearly equals the between-sibship variation and is equivalent to no correlation on gene expression in the siblings. In order to assess the statistical significance of the estimated ICCs, we use a computer resampling method by permuting the samples to form pseudo-sibships. The distribution of the null ICC calculated for each gene from the pseudo-sibships is used to obtain an empirical p-value for the genes.

All calculations were done using R which is a free software package for statistical computing and graphics (http://cran.r-project.org).