Systematic identification of human housekeeping genes possibly useful as references in gene expression studies

Abstract

The ideal reference, or control, gene for the study of gene expression in a given organism should be expressed at a medium-high level for easy detection, should be expressed at a constant/stable level throughout different cell types and within the same cell type undergoing different treatments, and should maintain these features through as many different tissues of the organism. From a biological point of view, these theoretical requirements of an ideal reference gene appear to be best suited to housekeeping (HK) genes. Recent advancements in the quality and completeness of human expression microarray data and in their statistical analysis may provide new clues toward the quantitative standardization of human gene expression studies in biology and medicine, both cross- and within-tissue. The systematic approach used by the present study is based on the Transcriptome Mapper tool and exploits the automated reassignment of probes to corresponding genes, intra- and inter-sample normalization, elaboration and representation of gene expression values in linear form within an indexed and searchable database with a graphical interface recording quantitative levels of expression, expression variability and cross-tissue width of expression for more than 31,000 transcripts. The present study conducted a meta-analysis of a pool of 646 expression profile data sets from 54 different human tissues and identified actin γ 1 as the HK gene that best fits the combination of all the traditional criteria to be used as a reference gene for general use; two ribosomal protein genes, RPS18 and RPS27, and one aquaporin gene, POM121 transmembrane nucleporin C, were also identified. The present study provided a list of tissue- and organ-specific genes that may be most suited for the following individual tissues/organs: Adipose tissue, bone marrow, brain, heart, kidney, liver, lung, ovary, skeletal muscle and testis; and also provides in these cases a representative, quantitative portrait of the relative, typical gene-expression profile in the form of searchable database tables.

Introduction

The quantitative study of gene expression in terms of the amount of RNA produced by a certain gene in a given biological condition is fundamental to our understanding of gene structure and function. Molecular laboratory techniques used to quantitatively measure RNA expression levels include northern blot analysis, reverse transcription-polymerase chain reaction (RT-PCR), expression microarrays and, recently, RNA sequencing (RNA-Seq). These techniques typically require a form of normalization of the measured RNA expression level of a gene to account for the potentially different RNA input quantities used in the assay. The best way to do this is to relate the transcripts to the number of templates (DNA strands) creating them. Owing to difficulties in obtaining these parameters from the same samples, several methods have been proposed over the years and are commonly used to relate the RNA amount of a given molecular species to one or more reference RNAs that are assumed to be expressed at a constant level in the cell type under consideration (1). More specifically, the ideal reference (or control) gene for the study of gene expression in a given organism should: i) Be expressed at a medium-high level so it can be easily detected; ii) be expressed at a constant/stable level in different cell types and within the same cell type undergoing different treatments; and iii) maintain these features through as many different tissues as possible within the organism (that is, ubiquitously expressed). These features would maximize the usefulness of the genes in the expression studies (2).

From a biological point of view, the theoretical requirements for an ideal reference gene appear to be best suited to the housekeeping (HK) genes, a large class of genes that are constitutively expressed, subjected to low levels of regulation in different conditions and perform biological actions that are fundamental for the basic functions of the cell (1). Their fundamental roles also mean that they tend to be expressed in high levels, confirming their suitability as reference genes.

Since the 1980s, several human genes have been widely used as ‘classic’ reference genes based on their fulfilling of the aforementioned requirements, as assessed typically by northern blot analysis (3), and this set of genes was seamlessly transferred for use in RT-PCR analyses in the 1990s (4). However, in a situation in which there was only preliminary knowledge of the human genome, the choice of these genes could be only anecdotal, among the limited pool of the genes known at the time. Following the widespread use of expression microarray techniques in the early 2000s (5), along with the initial sequencing and characterization of the human genome, it became theoretically possible to study and select HK genes by the systematic analysis of transcriptomes. This possibility was readily exploited in certain initial studies (6,7), which demonstrated that common control genes used in human studies, including the most popular glyceraldehyde-3-phosphate dehydrogenase (GAPDH), actin β (ACTB) and β₂-microglobulin (B2M) (3,8,9), actually exhibited considerable variability in expression within and across microarray data sets, and in certain cases this was confirmed by quantitative RT-PCR (RT-qPCR) analysis (10). The main conclusion was that the choice of a reference gene should be suited to the specific investigated tissue. In the following years, the problem of selecting a human HK gene by exploiting the availability of transcriptome-scale data was addressed by several studies and remains under debate (11).

The present study considered whether recent advancements in the quality and completeness of human expression microarray data, along with developments in statistical analysis, may be able to provide new clues toward the quantitative standardization of human tissue gene expression studies, cross- and within-tissue. A general framework is presented for choosing reference genes that may be useful in gene expression studies on normal human tissues and organs; the present study also addresses certain previous assumptions and provides an approach that is based on the Transcriptome Mapper (TRAM; http://apollo11.isto.unibo.it/software/TRAM) tool, which can overcome a number of problems associated with cross-platform analysis (such as, probe assignment to locus, intra- and inter-sample normalization and scaled quantile statistics) (12,13). TRAM can integrate data from hundreds or thousands of complete microarray data sets and provide unique combinations of features that are particularly suited to allow the choice of reference genes based on the three properties aforementioned. TRAM calculates a quantitative measurement for a consensus mean-expression value for tens of thousands of human transcripts expressed in a specific tissue or organ, thus allowing for a precise estimation of the intensity of its expression in terms of a percentage of the mean expression value in the pool of analyzed transcriptomes and the choice of a reference gene expressed at medium-high or high level (12,14). In addition, TRAM provides the standard deviation (SD) from the mean (normalized as the percentage of the mean value) for the mean expression value of a given locus, thus allowing for the selection of genes that may have more stable expression values in a variety of different samples and/or experimental platforms that have been investigated for a given tissue/organ. Finally, TRAM is able to integrate data from numerous sources, allowing verification of the consistency of the first two features through a wide range of different tissues within the organism studied (12).

The present study conducted a meta-analysis of 646 data sets that were obtained from different studies associated with 54 different normal human tissues and organs, using various experimental platforms. This meta-analysis produced results for 35,131 individual loci, including known genes and expressed sequence tag (EST) clusters, in the form of a database that may be extensively queried by freely combining a number of criteria, thus identifying the best intersection of moderate-high level of expression, low expression-value variability and expression in a large number of tissues. Results from the present study demonstrated that the human actin γ 1 (ACTG1) gene may potentially be used as a general reference gene for human cross-tissue studies and that specific genes are most suited for individual within-tissue studies. An enrichment analysis in functional classes for the identified HK genes is also presented.

Materials and methods

Database search

To retrieve data sets that have been derived from normal adult human tissues, a systematic search of the gene expression data repository Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo) (15) was performed for any available sample series associated with pools of normal human tissues/organs, choosing Homo sapiens as the organism. The query used was: ‘Homo sapiens [ORGANISM] AND tissue* [TI] OR organ* [TI]’. These selection criteria led to the generation of a pool of samples that included all of the main human organs and tissues, which served as a reference set already partially used by the authors of the present study (14). The searches were performed up to May 2013, and search results were then filtered using inclusion and exclusion criteria as explained in the Data set selection subsection below.

Data set selection

Inclusion criteria of data sets used in the present analysis were: i) Experiments were carried out on whole organs or tissues; ii) organs/tissues were obtained from individuals exhibiting normal phenotypes; iii) samples used in the studies were obtained from adults; and iv) the availability of the raw or pre-processed data. Exclusion criteria included: i) Studies using exon arrays (which hamper data elaboration by TRAM, owing to an exceedingly high number of data rows) or platforms using probes that are split into several different arrays for each sample, which hampers intra-sample normalization; ii) platforms that examine an atypical number of genes (that is, <5,000 or >60,000); and iii) data that is derived from cell lines, pathological or treated tissue, or children or fetal tissues.

A quantitative transcriptome map was obtained by linearizing values from each data set that were provided as logarithms. If only raw files (such as CEL files) were available in the GEO database, they were converted into pre-processed data using the AltAnalyze version 2.0 program (http://www.altanalyze.org) (16). Tables I and II summarize the data sets used to build the integrated transcriptome map from different human tissues/organs. The mean number of retrieved data sets was 22.2±6.6 for the group of the main tissues/organs (Table I) and 6.5±5.2 for the group of the minor tissues/organs (Table II) since there were a lower number of representatives of certain of them owing to the limited availability of data in the repositories.

Table I.

Summary of the samples used in the meta-analysis for 10 specific tissues/organs.

Sample	ID	n	Data points	Genes	HKa	Value	Samples % (n)	SD%
Whole poolb	1–646	646	21,840,330	35,131	27	≥150	≥65 (420)	≤45
Adipose tissue	1–20	20	943,480	25,722	363	≥300	≥80 (16)	≤30
Bone marrow	46–60	15	548,322	26,097	222	≥300	≥80 (12)	≤30
Brain	61–84	24	507,799	21,603	174	≥150	≥65 (16)	≤45
Heart	150–166	17	452,683	34,486	56	≥150	≥65 (11)	≤45
Kidney	167–183	17	405,285	33,816	65	≥150	≥65 (11)	≤45
Liver	184–216	33	971,110	35,090	49	≥150	≥65 (21)	≤45
Lung	217–249	33	817,145	34,872	49	≥150	≥65 (21)	≤45
Ovary	281–304	24	788,447	34,853	157	≥150	≥65 (16)	≤45
Skeletal muscle	403–418	16	324,382	27,160	77	≥300	≥80 (13)	≤30
Testis	495–517	23	745,430	34,853	227	≥150	≥65 (15)	≤45

^aPlease refer to the text for a description of the selection criteria used to identify HK genes according to expression values, number of samples and SD.

^bWhole pool contains all sample data sets related to the 54 different human tissues/organs listed in Tables I and II. ID, identifier used in the present study; HK, number of housekeeping genes retrieved; value, expression value; Samples % (n), the percentage of samples in each dataset in which the HK genes fulfiled the selection criteria; n, total sample number; SD%, standard deviation from the mean expression value of a given locus expressed as a percentage.

Table II.

Samples used in the meta-analysis for the whole pool in addition to all samples listed in Table I.

Sample	ID	n
Adrenal gland	21–30	10
Aorta	31–34	4
Appendix	35–38	4
Bladder	39–41	3
Blood	42–45	4
Breast	85–89	5
Bronchus	90–95	6
Ciliary ganglion	96–99	4
Colon	100–105	6
Connective	106	1
Dental pulp	107–108	2
Dorsal root ganglia	109–130	22
Esophagus	131–141	11
Fallopian tube	142–148	7
Gall bladder	149	1
Lymph node	250–266	17
Mammary gland	267–272	6
Oral mucosa	273–280	8
Pancreas	305–315	11
Parathyroid	316–318	3
Penis	319–324	6
Pericardium	325–326	2
Pharyngeal mucosa	327–334	8
Pituitary gland	335–351	17
Prostate	352–382	31
Salivary gland	383–402	20
Skin	419–431	13
Small intestine	432–437	6
Smooth muscle	438–441	4
Spinal cord	442–465	24
Spleen	466–483	18
Stomach	484–488	5
Synovial membrane	489–494	6
Thymus	518–532	15
Thyroid	533–554	22
Tongue	555–564	10
Tonsil	565–578	14
Trachea	579–592	14
Trigeminal ganglion	593–612	20
Urethra	613–620	8
Uterus	621–628	8
Vagina	629–637	9
Vena cava	638	1
Vulva	639–646	8

ID, identifier used in the present study; n, total sample number.

TRAM analysis

The TRAM tool (12) allows gene expression data to be imported in a tab-delimited text format. It also allows data integration by decoding probe-set identifiers to gene symbols using UniGene data parsing (17), normalizing data from multiple platforms using intra-sample and inter-sample normalization (that is, scaled quantile normalization) (13), and creates tables of expression values for each transcript.

A directory (folder) was created that contains the whole pool of all the sample data sets (n=646) associated with 54 different human tissues/organs that were retrieved and downloaded from the GEO database. This set included 629 samples described previously (14), which were in addition to the following GEO samples: GSM2829, GSM2856, GSM18792, GSM18793, GSM18951, GSM18952, GSM39975, GSM39980, GSM39992, GSM39994, GSM39999, GSM40001, GSM44671, GSM52565, GSM175935, GSM790822 and GSM790831. These samples were pre-processed according to the TRAM tool guide to be ready for import and processing in TRAM. During the pre-processing step, TRAM allows the linearization of data sets when provided as logarithms. All pre-processed samples of the whole pool were then imported as pool B in the TRAM table, and the entire set of analyses permitted by TRAM was performed as described in detail in the tool guide, using default parameters as previously described (12). The significance of the over-/underexpression of single genes was determined within pool B by running TRAM in ‘Map’ mode (12) with a segment window of 12,500 bp and a minimum number of one over-/underexpressed gene in that window. This window size is <20% of the 67 kb mean size of a human protein-coding gene, as determined by searching the GeneBase database (18); therefore, significant over-/underexpression of a segment [at q<0.05, where q is the P-value corrected for false discovery rate (12)] almost always corresponds to that of a single gene. When the segment window contains >1 gene, significance is maintained if the expression value of the over-/underexpressed gene prevails over the others.

To study tissues or organs individually, 10 representative biological conditions were selected (Table I) and the data were exported for the pool of samples relative to each tissue/organ from the TRAM table ‘Values B’ and reimported in TRAM as pool A to allow comparison between the specific tissue/organ (pool A) and the whole pool (pool B).

TRAM version 1.2.1, 2015 human version, was used in the present study and is freely available at http://apollo11.isto.unibo.it/software. The complete set of TRAM results from the import and analysis of the 646 data sets is not currently distributed owing to its very large size (25 GB); only the final results are available. However, the complete set of results may be regenerated locally by running the automated import and analysis of the data sets in the aforementioned pool B folder into the TRAM 1.2.1 software. Briefly, gene expression values were assigned to individual loci using UniGene, data were subjected to intra-sample normalization as a percentage of the mean value and then to inter-sample normalization by scaled quantile. The value for each locus within each biological condition is the mean of all available values for that locus. The median value of whole genome gene expression was used to determine the percentiles of expression for each gene. Only mapped genes or EST clusters with an assigned gene name or UniGene code, respectively, that have start and end genomic coordinates, and have a non-empty raw intensity value were selected for the analysis.

Further improvements were made for the present study by adding the total number of biological samples (microarrays) from which each gene expression value is derived, in addition to the number of data points, as certain experimental platforms used to assess expression levels for a sample may contain a variable number of microarray spots, each with a different probe, which generate multiple expression data points for the same gene. These improvements are available (upon request) as a dedicated script and are to be fully integrated into the next version of TRAM (TRAM 1.3), which is due to be released in 2017. Since a different number of microarray spots may be available on a platform for a given gene, these new features offer the possibility to normalize the quantity of information that is available for a gene based on the actual number of distinct biological samples that provide measured expression values for that gene.

To create transcriptome maps, TRAM does not consider probes for which expression values are not available, assuming that an expression level has not been measured. Furthermore, the software gives 95% of the minimum positive value present in a sample to those expression values ≤0 to obtain meaningful numbers when it is required to obtain a ratio between values in pool A and pool B. If it is assumed that, in these cases, the expression level is too low to be detected under the experimental conditions used, then this transformation may be useful to highlight differential gene expression.

HK gene search

The predicted genes that behave as HK genes were determined, as they are mainly involved in fundamental cellular functions and are ubiquitously and constitutively expressed in all tissues (19–21). A search for HK genes in the transcriptome maps was performed using the following parameters: i) An expression value ≥300, which in TRAM is given as a percentage of the mean expression value in a sample in order to select genes that are expressed at least threefold above the mean value and are therefore expressed at an easily appreciable level; ii) a SD ≤30, expressed as a percentage of the mean value to identify genes with a low expression variation among different samples; and iii) a sample number ≥80% of the total number of samples for each analyzed pool to select commonly expressed HK genes (for example, ≥80% indicates 16 out of 20 samples for adipose tissue). When a very low number of suitable HK genes were identified by these criteria, the parameters were relaxed to ≥150 mean expression value, ≤45% SD and ≥65% of samples in the pool with a measured value for the gene (Table I).

To select the HK genes with the best overall features to be proposed as reference genes, the genes identified as fitting the described criteria were first arranged in descending order of expression value, followed by ascending order of SD% and finally by descending order of sample number. An ascending rank number was assigned for each sorting criterion, the mean among these three ranks was calculated and the lowest mean rank was considered to correspond to the gene with the overall best fit to the three criteria. For tissue-specific analysis, a fourth criterion was added by calculating the ascending rank of the absolute deviation from 1 of the ratio between the mean expression value of the gene in the considered tissue and in the whole pool of 646 samples, respectively (lowest rank considered the best). This fourth criterion selected for genes with the most similar expression values in the whole pool and each tissue-specific pool, suggesting a particularly stable expression level. The mean rank was then calculated for all four criteria.

Functional analysis

The hypothesis that the most suitable HK genes identified in the analysis could be enriched for particular functional classes was tested using the web tool FuncAssociate version 3.0 (http://llama.mshri.on.ca/funcassociate) (22).

Results

Human HK genes for general use

Using the search criteria detailed in the Methods section of the present study, several genes that were present in ≥65% of the whole pool list of 646 samples from 54 human tissues/organs were identified. A total of eight genes were identified that best fulfilled the criteria to be proposed as reference genes (Table III), and the HK gene that best fit the combination of all the traditional criteria to be used as a general reference gene was ACTG1.

Table III.

A list of the eight genes that best fulfill the criteria proposed for use as a reference in gene expression studies across all 646 pool B samples of human tissues and organs that were examined.

Mean rank	Genea	Chromosome	Value	n	SD%
2.8	ACTG1b	17	3,453.7	513	37.8
5.5	RPS18	6	4,933.5	472	41.4
5.8	POM121C	7	349.7	425	28.0
6.5	MRPL18	6	226.6	546	39.4
6.5	TOMM5	9	273.6	426	38.2
7.0	YTHDF1	20	213.0	546	39.1
7.3	TPT1	13	5,941.7	508	43.1
8.0	RPS27	1	4,355.6	513	44.1

^aPlease see Table VIII for a complete list of gene definitions.

^bACTG1 was significantly overexpressed in the transcriptome map (q=0.03; where q is the P-value corrected for false discovery rate). Value, expression value; SD%, standard deviation expressed as percentage of the mean expression value for the locus.

Human HK genes for individual tissues

A total of 10 human tissues/organs were selected for a systematic search for the best suitable references genes within the respective biological type. By searching with the four criteria aforementioned, several genes were identified that had a mean expression value ≥300 (or ≥150, for searches performed with less stringent criteria), with an SD≤30 (or SD≤45) and with an expression value measured in ≥80% (or ≥65% with relaxed criteria) of the samples within different tissues/organs (Tables IV–VII). The eight known genes with the lowest mean rank scored for the selected criteria are listed in Table IV (adipose tissue and bone marrow), Table V (brain, heart and skeletal muscle), Table VI (kidney, liver and lung) and Table VII (ovary and testis). The complete gene name corresponding to each gene symbol listed in Tables III–VII is provided in Table VIII, along with the number of times that each gene is represented in a different pool among the 11 pools analyzed in Tables III–VII.

Table IV.

A list of the eight genes that best fulfill the selection criteria proposed for a reference in gene expression studies of human adipose tissue and bone marrow.

A, Adipose tissue
Mean rank	Genea	Chromosome	Value A	Value B	A/B	Sample count A	Sample count B	SD% A	SD% B
25.0	RPL6	12	2,207.0	2,069.0	1.1	20	592	11.6	89.5
27.5	RPS25	11	2,189.7	2,113.3	1.0	20	568	14.2	75.2
27.5	SOD1	21	1,273.0	1,231.8	1.0	20	592	12.1	67.9
33.5	RNASEK	17	912.6	881.1	1.0	20	335	11.3	41.7
35.8	GABARAP	17	1,112.7	1,185.6	0.9	20	580	12.7	72.6
41.3	ACTG1	17	3,821.3	3,453.7	1.1	20	513	15.2	37.8
43.8	GABARAPL2	16	591.6	591.9	1.0	20	592	12.7	69.8
45.0	MRFAP1	4	1,583.2	1,499.0	1.1	20	480	17.6	64.2
B, Bone marrow
13.0	RPL41	12	5,739.4	5,838.5	1.0	13	568	18.5	45.1
17.8	RPLP0	12	3,659.6	3,341.5	1.1	13	508	16.5	44.5
17.8	RPS27	1	4,494.8	4,355.6	1.0	13	513	19.5	44.1
23.3	TUBA1B	12	2,893.1	2,478.6	1.2	13	508	16.4	81.1
24.3	RPSA	3	2,016.1	2,036.8	1.0	13	575	20.6	80.1
25.5	SLC25A3	12	918.5	956.7	1.0	13	592	17.6	73.5
26.5	ACTG1	17	3,728.3	3,453.7	1.1	13	513	20.3	37.8
30.0	EEF1G	11	2,713.6	2,686.9	1.0	13	587	22.6	48.3

^aPlease see Table VIII for a complete list of gene definitions. Value A, mean expression values of the gene in the pool A, including the sample related to the specific tissue; value B, mean expression values of the gene in the whole pool B consisting of all the 646 samples; A/B, ratio between values A and B; SD%, standard deviation expressed as percentage of the mean expression value (A or B) for the locus.

Table VII.

A list of the eight genes that best fulfill the selection criteria proposed for a reference in gene expression studies of human ovary and testis.

A, Ovary
Mean rank	Genea	Chromosome	Value A	Value B	A/B	Sample count A	Sample count B	SD% A	SD% B
10.0	ACTG1	17	3,547.3	3,453.7	1.0	17	513	30.8	37.8
20.0	AP2M1	3	387.9	375.5	1.0	17	513	32.0	51.8
22.5	MIF	22	885.0	845.6	1.0	17	527	36.5	66.5
24.3	NELFCD	20	240.2	239.5	1.0	17	522	31.3	118.6
24.5	RNF181	2	331.3	356.5	0.9	16	426	24.8	52.5
28.0	PSMC1	14	367.9	421.4	0.9	17	513	29.3	48.8
30.0	TMEM147	19	295.9	282.8	1.0	22	592	35.3	47.2
30.3	TERF2IP	16	323.5	344.6	0.9	22	592	34.7	69.9
B, Testis
18.5	TUBA1B	12	2,774.3	2,478.6	1.1	18	508	30.9	81.1
21.0	FAM96B	16	339.6	325.2	1.0	19	546	28.0	47.9
22.8	RPL8	8	1,637.9	2,069.0	0.8	18	513	27.9	44.1
23.0	RPS18	6	4,434.5	4,933.5	0.9	16	472	34.9	41.4
24.5	ACTG1	17	4,320.0	3,453.7	1.3	18	513	27.7	37.8
29.3	RPS27	1	5,444.6	4,355.6	1.3	18	513	32.5	44.1
31.5	TBCB	19	343.1	290.6	1.2	21	592	29.1	67.9
32.3	RPL41	12	4,676.1	5,838.5	0.8	20	568	36.1	45.1

^aPlease see Table VIII for a complete list of gene definitions. Value A, mean expression values of the gene in the pool A, including the sample related to the specific tissue; value B, mean expression values of the gene in the whole pool B consisting of all the 646 samples; A/B, ratio between these value A and value B; SD%, standard deviation expressed as percentage of the mean expression value (A or B) for the locus.

Table V.

A list of the eight genes that best fulfill the selection criteria proposed for a reference in gene expression studies of human brain, heart and skeletal muscle.

A, Brain
Mean rank	Genea	Chromosome	Value A	Value B	A/B	Sample count A	Sample count B	SD% A	SD% B
9.3	NDUFB4	3	590.9	545.7	1.1	20	551	20.6	110.0
9.8	NDUFB1	14	546.7	586.0	0.9	22	592	22.7	59.9
17.5	GSTO1	10	345.9	370.6	0.9	22	592	22.0	75.5
28.0	AMZ2	17	299.5	276.1	1.1	20	467	29.3	99.5
28.8	POLR2I	19	302.8	254.7	1.2	22	568	23.0	79.2
29.5	NDUFA3	19	365.0	280.9	1.3	20	551	19.7	59.0
30.0	RRAGA	9	463.0	350.9	1.3	22	592	26.5	45.1
30.5	POMP	13	283.5	338.2	0.8	20	546	24.0	106.0
B, Heart
10.5	MIF	22	896.7	845.6	1.1	11	527	42.1	66.5
12.5	ECHS1	10	567.8	571.0	1.0	15	592	43.2	100.3
13.0	FAM96A	15	338.6	289.7	1.2	11	480	35.0	63.8
13.5	NOP10	15	474.8	428.5	1.1	13	546	41.4	48.0
13.5	TBCB	19	319.8	290.6	1.1	15	592	38.8	67.9
14.8	RRAGA	9	292.4	350.9	0.8	15	592	37.0	45.1
15.0	IFI27	14	510.4	441.9	1.2	15	592	42.2	103.1
15.5	MB	22	6,845.8	745.2	9.2	15	592	30.6	260.2
C, Skeletal muscle
4.3	RPL41	12	4,995.8	5,838.5	0.9	16	568	17.3	45.1
8.3	PRDX1	1	921.2	1,096.5	0.8	16	592	18.5	71.7
10.8	RPL8	8	2,065.6	2,069.0	1.0	16	513	25.2	44.1
11.3	C14orf166	14	576.0	505.7	1.1	16	546	19.1	57.0
11.8	JTB	1	714.0	606.9	1.2	16	592	20.6	61.6
11.8	RPS29	14	2,751.4	2,551.2	1.1	16	592	25.4	57.9
13.0	SNRPD2	19	516.4	524.7	1.0	16	592	21.9	57.5
14.5	NOP10	15	497.5	428.5	1.2	16	546	19.2	48.0

^aPlease see Table VIII for a complete list of gene definitions. Value A, mean expression values of the gene in the pool A, including the sample related to the specific tissue; value B, mean expression values of the gene in the whole pool B consisting of all the 646 samples; A/B, ratio between these value A and value B; SD%, standard deviation expressed as percentage of the mean expression value (A or B) for the locus.

Table VI.

A list of the eight genes that best fulfill the selection criteria proposed for a reference in gene expression studies of human kidney, liver and lung.

A, Kidney
Mean rank	Genea	Chromosome	Value A	Value B	A/B	Sample count A	Sample count B	SD% A	SD% B
6.5	PGAM1	10	688.8	741.8	0.9	14	556	29.6	64.0
7.5	NOP10	15	413.9	428.5	1.0	12	546	31.9	48.0
8.0	FIS1	7	340.7	329.0	1.0	12	546	29.6	52.9
9.0	GPX1	3	459.9	496.7	0.9	15	592	34.2	66.9
10.3	GANAB	11	266.7	260.0	1.0	15	587	35.1	49.5
11.0	NDUFB11	X	343.8	294.9	1.2	12	551	25.0	67.7
12.8	HEBP1	12	288.6	249.2	1.2	12	551	27.7	53.9
13.0	HDGF	1	444.3	432.2	1.0	14	568	40.5	58.4
B, Liver
8.5	HEBP2	6	335.1	362.5	0.9	31	592	35.8	79.2
11.8	NDUFS3	11	289.3	325.3	0.9	31	592	36.1	60.8
12.3	POLR2H	3	162.0	160.1	1.0	33	641	32.4	45.2
12.5	MRPS24	7	438.5	445.0	1.0	23	426	42.0	57.9
13.3	FAM96B	16	287.1	325.2	0.9	29	546	37.0	47.9
13.3	GTF3A	13	349.5	310.3	1.1	22	501	38.6	62.0
13.3	HIST1H2BK	6	216.3	186.6	1.2	28	568	30.3	72.7
13.8	CRELD2	22	297.3	279.9	1.1	24	467	39.5	52.7
C, Lung
7.3	RBX1	22	283.7	315.4	0.9	29	551	36.6	57.8
7.5	RRAGA	9	289.7	350.9	0.8	31	592	34.9	45.1
11.5	LAMTOR5	1	253.1	344.6	0.7	33	646	34.1	55.7
11.8	CNIH1	14	207.4	250.4	0.8	31	592	32.3	63.4
12.0	EPCAM	2	252.8	256.7	1.0	31	592	39.0	218.7
12.8	EIF4A3	17	271.2	340.0	0.8	25	563	37.8	47.1
13.5	ACTG1	17	3,067.3	3,453.7	0.9	27	513	43.9	37.8
14.0	FAM96B	16	277.0	325.2	0.9	24	546	39.2	47.9

^aPlease see Table VIII for a complete list of gene definitions. Value A, mean expression values of the gene in the pool A, including the sample related to the specific tissue; value B, mean expression values of the gene in the whole pool B consisting of all the 646 samples; A/B, ratio between these value A and value B; SD%, standard deviation expressed as percentage of the mean expression value (A or B) for the locus.

Table VIII.

Gene symbols, corresponding descriptions and the number of recurrences (n) in the Tables for the genes listed in Tables III–VII.

Gene symbol	n	Gene description
ACTG1	5	Actin γ 1
AMZ2	1	Archaelysin family metallopeptidase 2
AP2M1	1	Adaptor-related protein complex 2, µ1 subunit
C14orf166	1	Chromosome 14 open reading frame 166
CNIH1	1	Cornichon family AMPA receptor auxiliary protein 1
CRELD2	1	Cysteine rich with EGF like domains 2
ECHS1	1	Enoyl-CoA hydratase, short chain 1
EEF1G	1	Eukaryotic translation elongation factor 1γ
EIF4A3	1	Eukaryotic translation initiation factor 4A3
EPCAM	1	Epithelial cell adhesion molecule
FAM96A	1	Family with sequence similarity 96 member A
FAM96B	3	Family with sequence similarity 96 member B
FIS1	1	Fission, mitochondrial 1
GABARAP	1	GABA type A receptor-associated protein
GABARAPL2	1	GABA type A receptor associated protein like 2
GANAB	1	Glucosidase II α subunit
GPX1	1	Glutathione peroxidase 1
GSTO1	1	Glutathione S-transferase ω1
GTF3A	1	General transcription factor IIIA
HDGF	1	Heparin binding growth factor
HEBP1	1	Heme binding protein 1
HEBP2	1	Heme binding protein 2
HIST1H2BK	1	Histone cluster 1 H2B family member k
IFI27	1	Interferon α-inducible protein 27
JTB	1	Jumping translocation breakpoint
LAMTOR5	1	Late endosomal/lysosomal adaptor, MAPK and MTOR activator 5
MB	1	Myoglobin
MIF	2	Macrophage migration inhibitory factor (glycosylation-inhibiting factor)
MRFAP1	1	Morf4 family associated protein 1
MRPL18	1	Mitochondrial ribosomal protein L18
MRPS24	1	Mitochondrial ribosomal protein S24
NDUFA3	1	NADH:ubiquinone oxidoreductase subunit A3
NDUFB1	1	NADH:ubiquinone oxidoreductase subunit B1
NDUFB4	1	NADH:ubiquinone oxidoreductase subunit B4
NDUFB11	1	NADH:ubiquinone oxidoreductase subunit B11
NDUFS3	1	NADH:ubiquinone oxidoreductase core subunit S3
NELFCD	1	Negative elongation factor complex member C/D
NOP10	3	NOP10 ribonucleoprotein
PGAM1	1	Phosphoglycerate mutase 1
POLR2H	1	RNA polymerase II subunit H
POLR2I	1	RNA polymerase II subunit I
POM121C	1	POM121 transmembrane nucleoporin C
POMP	1	Proteasome maturation protein
PRDX1	1	Peroxiredoxin 1
PSMC1	1	Proteasome 26S subunit, ATPase 1
RBX1	1	Ring-box 1
RNASEK	1	Ribonuclease K
RNF181	1	Ring finger protein 181
RPL6	1	Ribosomal protein L6
RPL8	2	Ribosomal protein L8
RPL41	3	Ribosomal protein L41
RPLP0	1	Ribosomal protein lateral stalk subunit P0
RPS18	2	Ribosomal protein S18
RPS25	1	Ribosomal protein S25
RPS27	3	Ribosomal protein S27
RPS29	1	Ribosomal protein S29
RPSA	1	Ribosomal protein SA
RRAGA	3	Ras related GTP binding A
SLC25A3	1	Solute carrier family 25 member 3
SNRPD2	1	Small nuclear ribonucleoprotein D2 polypeptide
SOD1	1	Superoxide dismutase 1
TBCB	2	Tubulin folding cofactor B
TERF2IP	1	TERF2 interacting protein
TMEM147	1	Transmembrane protein 147
TOMM5	1	Translocase of outer mitochondrial membrane 5
TPT1	1	Tumor protein, translationally-controlled 1
TUBA1B	2	Tubulin α1b
YTHDF1	1	YTH N6-methyladenosine RNA binding protein 1

The genes with a number of recurrences (n) >1 are shown in bold. AMPA, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid; CoA, coenzyme A; EGF, epidermal growth factor; GTP, guanosine 5′-triphosphate; MAPK, mitogen-activated protein kinase; Morf4, mortality factor 4 (pseudogene); MTOR, mechanistic target of rapamycin; TERF2, telomeric repeat binding factor 2.

Analysis of the associated function

Using the FuncAssociate web tool, the identified HK genes were revealed to be significantly enriched in certain functional classes. Of the genes identified as suitable for human tissue-wide use listed in Table III, there were statistically significant enrichments [with adjusted P-value (P-adj)≤0.05] only in the Gene Ontology categories: Ribosome (P-adj=0.003), translation termination (P-adj=0.038), translation elongation (P-adj=0.045) and cellular protein complex disassembly (P-adj=0.05).

A significant enrichment in numerous Gene Ontology categories was also identified by pooling in a unique list all 63 genes (Table IX) identified for the 10 specific tissues/organs (Table IV–VII); all enrichments were associated with basic cellular components, molecular functions and/or biological processes. The biological process with the highest statistical significance (P-adj<0.001) was RNA catabolic processes and translation (data not shown).

Table IX.

A list of all 63 genes identified for the 10 specific tissues and organs presented in Tables IV–VII.

Number	Gene
1	ACTG1
2	AMZ2
3	AP2M1
4	C14orf166
5	CNIH1
6	CRELD2
7	ECHS1
8	EEF1G
9	EIF4A3
10	EPCAM
11	FAM96A
12	FAM96B
13	FIS1
14	GABARAP
15	GABARAPL2
16	GANAB
17	GPX1
18	GSTO1
19	GTF3A
20	HDGF
21	HEBP1
22	HEBP2
23	HIST1H2BK
24	IFI27
25	JTB
26	LAMTOR5
27	MB
28	MIF
29	MRFAP1
30	MRPS24
31	NDUFA3
32	NDUFB1
33	NDUFB11
34	NDUFB4
35	NDUFS3
36	NELFCD
37	NOP10
38	PGAM1
39	POLR2H
40	POLR2I
41	POMP
42	PRDX1
43	PSMC1
44	RBX1
45	RNASEK
46	RNF181
47	RPL41
48	RPL6
49	RPL8
50	RPLP0
51	RPS18
52	RPS25
53	RPS27
54	RPS29
55	RPSA
56	RRAGA
57	SLC25A3
58	SNRPD2
59	SOD1
60	TBCB
61	TERF2IP
62	TMEM147
63	TUBA1B

Discussion

Following the diffusion of expression microarray technology, a number of attempts have been made to use microarray data to perform a systematic analysis of the features of gene expression to identify HK genes that may be best suited as reference genes. Early attempts suffered from the limited number of samples available for analysis in addition to a lack of choice of computational biology techniques to analyze them, in particular for cross-platform analysis (6). Similar investigations were also conducted using the EST database (23) and, in recent years, RNA-Seq data (24,25). However, the generation of ESTs is subject to certain biases depending on the level of ability for an mRNA to be cloned during the generation of the cDNA EST libraries; therefore, although these data are useful to identify expressed sequences, they are less useful for quantitative analysis. RNA microarrays and RNA-Seq are the two main types of high-throughput technologies used to assess gene expression (26). Although RNA-Seq is considered to be more sensitive and has a broader dynamic range than RNA microarrays (27), in large comparative studies these two methods have produced comparable results in terms of gene expression profiling (27–29). Microarrays remain an accurate tool for measuring the levels of gene expression (29) and continue to provide useful data-mining resources. The approach of the present study takes advantage of the large number of previous transcriptomic studies that were performed with microarray technology and stored in publicly available databases, and also of the results provided in the form of a list of genes and the corresponding expression values.

The diverse origin of the data, in terms of different investigated individuals, different experimenters and different experimental platforms in the field of microarray analysis, provided a richness in the context of an analysis such as in the present study. That is, following data integration, the final results were not affected by systematic biases that may be linked to the particular samples or experimenters/platforms involved in the generation of the data, and they are likely to best represent the actual ‘mean’ status for a gene (12), compared with works based only on the original data obtained through a single platform (6). In addition, the approach of the present study exploited the combination of: i) Automated reassignment of probes to the corresponding genes by the updated UniGene data embedded in TRAM; ii) intra- and inter-sample normalization, including the scaled-quantile method that allows for comparison among platforms with a highly different number of probes; and iii) elaboration and representation of gene expression values in linear form within an indexed, searchable database, with a graphical interface recording quantitative levels of expression (mean expression values), expression variability (SD) and cross-tissue expression of more than 31,000 transcripts. These features represent a clear advancement in comparison with other meta-analyses that were based on published microarray data and were also aimed at identifying human reference genes (30,31), particularly considering that several studies on the subject were conducted in years when there was a reduced availability of samples and/or the experimental platforms were less complete (32).

The meta-analysis in the present study was performed on a pool of 646 data sets from 54 different human whole tissues/organs, and excluded analyses of individual cell types as the whole organ/tissue includes a vast number of cell types in its structure [as discussed by Fagerberg et al (33)], and also due to the requirement of selecting a representative set of samples as a result of the very long elaboration time for each analysis. The ACTG1 gene was identified as the HK gene that best fit the selection criteria for use as a general reference gene in the study of human gene expression. This gene proved to be statistically significantly over-expressed in the transcriptome map, according to the described criteria (34), with the following features: A very high mean expression value (3,453.7, indicating an expression level of ~35-fold in comparison with the mean expression value of all the genes in each sample, set as equal to 100) and an SD% of 37.8%; the number of samples in which a measure for this gene was available was 513 out of 646 (79.4%). Notably, this well-characterized gene, whose coding sequence appears to be completely characterized (35), encodes for a cytoplasmic form of actin that is known to be ubiquitously expressed in human cells, but is different from the actin β (ACTB) that is routinely used as a reference gene. According to the data of the present study, the commonly used reference genes ACTB and GAPDH had excellent features in terms of high expression value and diffuse expression in human cells. However, they have an SD% almost double that of ACTG1. Owing to the high similarity between ACTB and ACTG1 (91% identity with no gaps between their coding sequences, as determined by standard BLASTN analysis; data not shown), probes and primers need to be accurately selected to specifically identify the desired form of RNA.

Among the HK genes identified to be best suited as general reference genes for human studies, two ribosomal protein genes, RPS18 and RPS27, and one aquaporin gene, POM121 transmembrane nucleporin C (POM121C), were identified. ACTG1 and RPS27 were also included in the top 20 HK human genes across 42 human tissues in a previous study based on a single platform (36), however, the data sets from the present study were not included in the present meta-analysis since it was not possible to derive the expression values as linear numbers for each microarray channel from the deposited data. Notably, the eight genes listed in Table III were also classified at the transcript and protein level as ‘expressed in all tissues’ in the Human Protein Atlas (33), further supporting the results of the present study regarding them as the most generally suitable reference genes.

In accordance with the relevance of the ACTG1 gene in cross-tissue analysis, this gene is also present in the greatest number of lists (n=5) of the 10 tissue-specific genes best fulfilling criteria to be used as reference genes (Tables IV–VII). Ribosomal proteins were another notable example of known classes of general HK genes that are well represented in several human tissues. Although there is a clear, expected prevalence among the identified loci of genes encoding for basic cell structure (such as genes encoding for cytoskeletal components) and function (such as genes encoding for transcription and translation, reduction-oxidation metabolism and signaling proteins), it is worth noting that specific members of the same gene family involved in these processes may be identified in one particular tissue/organ and not in the others. The vast majority of genes that may be more suitable as reference genes for individual tissues (Tables IV–VII) are still typical HK genes, with the clear exception of the tissue-specific myoglobin gene in the heart (Table V).

The approach used in the present study allows for the systematic search for ideal reference genes and, at the same time, made available a ‘consensus’ reference gene-expression profile for 10 human tissues/organs. From this particular point of view, the presented results are less systematic than other previous attempts conducted in the case of the brain (34) and heart (14). Only the samples belonging to experiments in which a series of normal human tissues were analyzed have been included here, without searching for any single samples recorded for a given tissue in any type of available experiment (for example, comparisons between normal and pathological samples). However, the present data have been obtained through an improvement of the search algorithm for HK genes and may still offer interesting hints to their biological specificity in the transcriptome of these tissues. The same approach may be applied to data sets deriving from cell lines or pathological samples.

Systematic analysis aimed to evaluate if the genes identified as possible reference genes (Tables IV–VII) were significantly enriched in a particular class of genes confirmed to be involved in the most basic biological processes; in particular, in the metabolism of the informational macromolecules (nucleic acids and proteins). This therefore justifies their tendency to constitutive, stable and almost universal expression, which was also observed in a previous analysis that ranked genes by combining the average expression level and its SD in a single score (37). The biological peculiarity of HK genes was also highlighted by a significant difference in complexity between HK and tissue-specific gene promoters, as revealed by DNA entropy analysis (38).

While the present study was in progress, an article on the topic was published that suggested that a ‘universal’ human HK gene does not exist and provided a list of suitable reference genes for individual tissues/organs (11). However, the method employed by that study was different from the approach of the present study in a number of relevant aspects. In particular, the results of the previous study were originally obtained by combining the lists of genes retrieved from studies performed using heterogeneous techniques (such as microarray, EST or RNA-Seq analysis), were analyzed using the logic of classification, previous judgments concerning the suitability of certain genes as HK/reference genes were accepted and then these lists were combined. This approach of combining the lists of results was also used by Chang et al (39) and, following ranking, by Shaw et al (40). By contrast, the present study re-elaborated and normalized original raw data, and generated a fully quantitative analysis of human gene expression, which may explain certain differences in the results obtained by the algorithm used in the current study. Conversely, certain shared general conclusions were highlighted, including the necessity to calibrate the search criteria for HK genes according to cell/tissue type; however, the general analysis of the present study can still identify certain general-purpose genes with acceptable criteria that may be proposed as reference genes. Several previous studies have demonstrated that the results provided by the TRAM tool were highly reliable, having been confirmed by RT-qPCR experiments for several diverse human tissues, demonstrating a correlation coefficient (r) between TRAM and RT-qPCR data of r=0.98 for brain (34), r=0.99 for hippocampus (41) and r=0.98 for heart (14). However, it is commonly accepted that additional experimental studies may be required to verify that the identified candidate reference gene is suitable for the actual biological condition investigated (42). Additional studies are in progress to verify if the HK profile identified in normal tissues may be applicable to aneuploid cells, in particular for systematic analysis of trisomy 21 cells (43), in light of the fact that all of the best reference genes identified by that study are not located on chromosome 21 and that a very small portion of this chromosome appears to be associated with the basic features of Down syndrome (DS) (44). In this regard, consulting the published differential transcriptome map comparing acute megakaryoblastic leukemia (AMKL) cells from children with DS (DS AMKL) and euploid megakaryocyte cells (euploid MK) (45) reveals promising values for ACTG1 (DS AMKL/euploid MK gene expression ratio=1.08) and POM121C (DS AMKL/euploid MK gene expression ratio=1.01); the ratios range between 0.47 and 3.55 (45) for the other genes listed in Table III.

A previous study using RNA-Seq data to identify reference genes across multiple human tissues focused mainly on low SD and so proposed a list of 11 genes with exceptionally low variability (1). The corresponding values have been checked by TRAM analysis in the present study, which confirmed their expression across multiple tissues, with a generally low SD (although not exceptional in the data of the current study). However, these identified genes had low expression values, all in the range of 150–400 in terms of a percentage of the mean value. This was recognized by Kwon et al (46), who selected low variability as the leading parameter, as have other studies, whereas the approach of the present study was aimed at finding genes with the best combination of criteria, considering that a high expression value may be advantageous in practice for the usability of a reference gene. Finally, it should be noted that other studies have frequently used very different and original approaches to the problem, using computational classifiers (47), the controlled vocabulary of Medical Subject Headings (48) and Gene Ontology classifications (29).

It may finally be noted that several of the available analysis tools are aimed at determining the best reference genes for normalization of gene expression data, and are largely based on the evaluation of the minimal variation of the expression level of a given gene among different conditions. BestKeeper (49) and GeNorm (50) are commonly used for screening the reference genes to perform an accurate normalization of RT-qPCR data, whereas NormFinder (51) may be used to evaluate reference genes for normalization of RT-qPCR and microarray experiments. Regarding NormFinder, a direct comparison with the tool employed in the present study is not possible, since NormFinder requires the same number of measured values for each gene and the TRAM algorithm does not have this limitation. In addition, NormFinder evaluation is based only on the analysis of variation among different samples, whereas the TRAM approach integrates this parameter with the level of expression (at least medium-high) and with the highest possible number of samples in which each gene is measured; all are essential parameters by which to search for a suitable reference gene, thus allowing the identification of genes with the overall best fit to the three criteria. Finally, although NormFinder has the ability to identify the ideal combination of biologically independent genes for each tissue, this analysis requires the creation of a matrix with two groups of data deriving from two different conditions, which in the case of TRAM can be either one of the ten tissues analyzed or the whole pool. In the whole pool sample, the difference in the number of measures for each gene increases, so the exclusion of a large number of measures to create a matrix of data does not make the direct comparison between NormFinder and TRAM results possible.

In conclusion, the present study provided, to the best of our knowledge, the first systematic analysis to quantitatively combine all of the traditional criteria aimed at identifying the HK genes that are best suited to be reference genes for the study of human gene expression. Several genes were identified and proposed to be suitable in cross-tissue studies, and certain genes were proposed as references for tissue/organ-specific studies. The wealth of data generated by this approach may also provide a representative portrait of typical gene-expression profiles for several human tissues and organs in the form of searchable database tables and suggested that currently uncharacterized transcripts, even EST clusters, may be worthy of further investigation as strong candidates to represent HK genes, or tissue-specific genes, expressed in high levels in human cells.