Detecting transcription of ribosomal protein pseudogenes in diverse human tissues from RNA-seq data

Background Ribosomal proteins (RPs) have about 2000 pseudogenes in the human genome. While anecdotal reports for RP pseudogene transcription exists, it is unclear to what extent these pseudogenes are transcribed. The RP pseudogene transcription is difficult to identify in microarrays due to potential cross-hybridization between transcripts from the parent genes and pseudogenes. Recently, transcriptome sequencing (RNA-seq) provides an opportunity to ascertain the transcription of pseudogenes. A challenge for pseudogene expression discovery in RNA-seq data lies in the difficulty to uniquely identify reads mapped to pseudogene regions, which are typically also similar to the parent genes. Results Here we developed a specialized pipeline for pseudogene transcription discovery. We first construct a “composite genome” that includes the entire human genome sequence as well as mRNA sequences of real ribosomal protein genes. We then map all sequence reads to the composite genome, and only exact matches were retained. Moreover, we restrict our analysis to strictly defined mappable regions and calculate the RPKM values as measurement of pseudogene transcription levels. We report evidences for the transcription of RP pseudogenes in 16 human tissues. By analyzing the Human Body Map 2.0 study RNA-sequencing data using our pipeline, we identified that one ribosomal protein (RP) pseudogene (PGOHUM-249508) is transcribed with RPKM 170 in thyroid. Moreover, three other RP pseudogenes are transcribed with RPKM > 10, a level similar to that of the normal RP genes, in white blood cell, kidney, and testes, respectively. Furthermore, an additional thirteen RP pseudogenes are of RPKM > 5, corresponding to the 20–30 percentile among all genes. Unlike ribosomal protein genes that are constitutively expressed in almost all tissues, RP pseudogenes are differentially expressed, suggesting that they may contribute to tissue-specific biological processes. Conclusions Using a specialized bioinformatics method, we identified the transcription of ribosomal protein pseudogenes in human tissues using RNA-seq data.


Background
Pseudogenes are "fossil" copies of functional genes that have lost their potential as DNA templates for functional products [1][2][3][4][5][6]. While the definition of pseudogenes is still somewhat fuzzy, most of them are defined operationally by bioinformatics criteria, e.g., genomic scans of signatures of homology to known genes. Ribosomal protein (RP) pseudogenes represent the largest class of pseudogenes found in the human genome: over 2000 ribosomal protein pseudogenes are identified by bioinformatics scan of genomic sequence [5].
These pseudogenes are commonly thought to be nonfunctional due to the lack of promoters and/or the presence of loss of function mutations. Indeed, the vast majority of these pseudogenes either carry dysfunctional mutations such as in-frame stop codons, or lack of proper regulatory sequences, such as promoters, mTOP signals, and first introns [7]. Interestingly, three RP pseudogenes, with 89%-95% sequence identity to their parent (progenitor) RP genes, were found to be transcribed and seem to be functional, by a bioinformatics scan of cDNA and expression sequence tag (EST) databases and confirmation by PCR and Northern blot [8]. A genome-wide bioinformatics scan identified over 2000 potential pseudogenes [5]. Moreover, it was found [9] that the six RP pseudogenes shared at syntenic loci between the human and the mouse genomes are more conserved than other RP pseudogenes.
However, data were lacking to experimentally validate pseudogene expression. It is unclear from the literature whether the reported cases are merely anecdotal or that pseudogenes do play some cellular roles. This is largely hindered by the lack of methods for the identification of pseudogenes transcription. The traditional method of transcriptome profiling, gene expression microarray, is not sensitive in distinguishing transcripts among very similar gene sequences.
Recent advancements of next-generation sequencing allow for direct massive transcriptome sequencing (RNA-seq), and thus providing unprecedented insights into all transcribed sequences. For example, RNA-seq has been applied to detect complex transcriptional activities such as alternative splicing [10,11] and allelicspecific expression [12]. Recently, RNA-seq has been applied to reveal RNA editing events [13]. However, to the best of our knowledge, there were yet no attempts to detect the transcription of pseudogenes in RNA-seq data. The main challenge for pseudogene identification in RNA-seq data is the difficulty of high fidelity read mapping. Because sequences of pseudogenes are highly similar to the sequences of the mRNAs of the parent genes, specialized read mapping methods are required to detect reads unambiguously generated from pseudogenes.
In this study, we conduct a bioinformatics analysis of pseudogene expression using RNA-sequencing data of 16 human tissues of the Illumina Human Body Map 2.0 project. We first describe our new computational pipeline for detecting pseudogene expression that disentangles sequencing reads of pseudogenes from those of the parent genes, with consideration of possible mismatches due to SNPs and RNA-editing. This is followed by a description of our findings and a discussion of their implications.

Illumina Human Body Map 2.0 RNA-seq data
The Human Body Map 2.0 Project by Illumina generated RNA-seq data for 16 different human tissues (adipose, adrenal, brain, breast, colon, heart, kidney, liver, lung, lymph node, ovary, prostate, skeletal muscle, testes, thyroid, and white blood cells). In our analysis we used the 75 bps single read data, with one lane of HiSeq 2000 data per tissue. Standard mRNA-seq library preps were used to extract poly-A selected mRNAs.

Discovery of pseudogene transcription in RNA-seq data
Our primary goal is to detect transcriptional activities of any of the 1709 processed RP pseudogenes. In addition, we also want to provide a preliminary quantification of their level of transcription.
We developed a novel bioinformatics approach for detecting the subtle signals of pseudogene expression. Briefly, we first compiled a "composite genome" consisting of known RP gene spliced mRNA sequences and the human genome (hg18) [14]. We then mapped RNA-seq reads to the composite genome using Bowtie [15], allowing no mismatches and discarding reads mapped to more than one locus. Thus we ensured that the reads mapped to RP pseudogenes are neither from repetitive regions nor from normal RP genes. On average 89% of the reads aligning to RP pseudogenes can also be mapped to real RP spliced mRNA sequences and are removed when mapped to the composite genome (see Table 1). Furthermore, to remove mapped reads that may be caused by SNPs and RNAediting, we extended the concept of the mappability (the mappable regions of human genome is called the uniqueome) [16] and masked regions in RP pseudogenes that are duplicated in the composite genome within 4 mismatches over the 75 bps read length. The number of reads we removed from non-unique locations in both the composite genome and hg18 genome can be seen in Table 2. The mappability regions only correspond to the unaltered human genome locations, so all reads mapped to RP gene mRNA sequences in the composite genome are removed during this step. Additionally, the composite genome alignment lacks the reads that mapped both to the unaltered human genome locations and spliced RP gene mRNA sequences as we only retained reads aligning to a single location. With both of these groups of reads removed, the number of reads mapped uniquely in the composite genome is always less than that in the unaltered human genome. Finally, we calculated the transcription levels, as measured by the Reads Per Kilobase per Million reads (RPKM) [11] of all pseudogenes according to the mapped reads in their mappable regions. As a benchmark for normal expression levels, we also aligned reads to an unaltered genome using TopHat and measured FPKM of all RefSeq genes using Cufflinks [17]. The alignment information of reads to the composite genome, and to the unaltered genome (hg18), can be seen in Table 2. Please see Methods for details.

Prevalent transcription of RP pseudogenes
The expression levels of the top seventeen highly expressed ribosomal protein pseudogenes in 16 human tissues are shown in Figure 1 and Table 3 (See Table S1  The number of reads mapped to the composite genome (which includes spliced ribosomal protein gene sequences) and to the unaltered human genome (hg18), and the number of reads overlapped with uniqueome ("mapped uniquely") for both are shown. For the composite genome, the number of reads aligning to the entire composite genome and the unaltered hg18 human genome are shown.
in Additional file 1 for complete list for all RP pseudogenes). As expected the majority of pseudogenes have no reads aligning to their sequence. Interestingly, there were some pseudogenes with high expression levels.
One RP pseudogene (PGOHUM-249508) is transcribed with RPKM 170 in thyroid. Moreover, three additional RP pseudogenes are transcribed with RPKM > 10. Furthermore, thirteen more RP pseudogenes are of RPKM > 5. We describe pseudogenes with an RPKM > 10 as "highly expressed", with the understanding that they may be only representing the top 10-15 percentile of all 37,681 RefSeq genes in the Human Body Map 2.0 data set, while RPKM > 5 represents the top 20-30 percentile (see Table 4). Below we discuss these cases in detail. PGOHUM-249508, an RPL21 pseudogene, is expressed with RPKM = 170 in thyroid (Figure 2). This highest expressed RP pseudogene is located in an intron of the Thyroglobulin (TG) gene. The TG gene is highly and specifically expressed in the same tissue, thyroid, and the gene encodes a glycoprotein that acts as a substrate for the synthesis of thyroxine and triiodothyronine as well as the storage of the inactive forms of thyroid hormone and iodine [18]. The transcription of this pseudogene goes beyond the annotated pseudogene region, but is less than the entire intron region. Although the pseudogene is specifically expressed in the same tissue as TG, the RP coding frame is on the reverse strand of the TG gene. Therefore, it is possible that this pseudogene is on a distinct transcript other than the TG gene. Moreover, according to UCSC genome browser [19], this pseudogene region is only conserved within the primates (between human and the Rhesus monkey), but not in other mammalian and vertebrate lineages. As a side note, the genome browser shows a peculiar conservation pattern between human and the stickleback fish, but it is likely to be an artifact of matching human genomic sequence with the RPL21 gene of stickleback fish.
Three additional pseudogenes are highly transcribed (RPKM > 10). PGOHUM-237215, an RPL7A pseudogene, is expressed RPKM = 17 in white blood cells. This  Table S1 in Additional file 1 for the complete list.
pseudogene is located in an intergenic region. Also, the transcription unit seems to span a longer region ( Figure 3). It is transcribed in a white blood cell specific fashion. PGOHUM-249146, an RPS24 pseudogene, is expressed in kidney. This pseudogene is located in the intronic region of gene SLC12A3 (Figure 4). This gene encodes a renal thiazide-sensitive sodium-chloride cotransporter that is important for electrolyte homeostasis. PGOHUM-239833, an RPS11 pseudogene, is expressed in testes. This pseudogene is located in an intergenic region ( Figure 5). The comparison of read coverage with or without uniqueome filtering for these four RP pseudogenes can be been in Figures S1-S4 in Additional file 2. Expression levels of pseudogenes with their pseudogene ID (pg-id, prefix 'PGOHUM00000' omitted) are measured in terms of RPKM. Only pseudogenes with maximum RPKM > 5 are shown. Tissue specificities are measured by the JS divergence [20]. Read coverage is the ratio of pseudogene exon length covered by uniquely mapped reads to the total pseudogene exon length. Percentile columns represent the percentage of RefSeq genes in all tissues with FPKM above a given value.

Tissue-specificity of pseudogene transcription
Many genes are expressed in a tissue-specific fashion. The Human Body Map 2.0 data allow us to study the tissue-specificity of transcriptions of these pseudogenes. We adopt the entropy-based Jensen-Shannon (JS) divergence measure used in [20]. The distributions of tissuespecificity JS divergences of RP pseudogenes versus RP genes are shown in Figure 6. In the Human Body Map 2.0 data set, all RP genes are not transcribed in a tissue specific fashion (JS divergence <0.5 for all RP genes) ( Table S2 in Additional file 3). Unlike ribosomal protein genes that are constitutively expressed in almost all tissues, many RP pseudogenes are differentially expressed (Table S1 in Additional file 1). Among the seventeen pseudogenes with RPKM > 5 at some tissue, 8 of them have a JS divergence > 0.5. In fact, all of the top 4 pseudogenes with RPKM > 10 are transcribed in a highly tissue specific fashion (JS divergence > 0.8). These results suggest that these highly expressed RP pseudogenes may contribute to tissue-specific biological processes.

Discussion and conclusions
In this work, we conducted a bioinformatics analysis of the pseudogenes of ribosomal protein genes in diverse human tissues. Using our specialized pipeline, we identified several cases of pseudogene expression. Most notably, one pseudogene in an intron of the TG gene is extremely highly expressed in thyroid. Moreover, several other pseudogenes are also highly expressed. We found that many pseudogenes are expressed in a tissue-specific fashion. Also, the expression of pseudogenes seems to often go beyond the boundaries of the annotated pseudogenes. Apparently, further experimental investigations will be needed to reveal the biological relevance of these expressions. Transcriptome sequencing, RNA-seq, provides an unprecedented opportunity to uncover many complexities of cellular gene expression. A key computational challenge in RNA-seq data analysis is to discern reads among multiple potential sources with similar sequences. In this work we focused on the detection of evidences of  pseudogene expression. We used extremely strict read mapping criteria to minimize potential false positives. Admittedly we did not utilize all potential reads, especially at regions with low uniqueness. Further research may consider using looser mapping criteria combined with sophisticated statistical algorithms to take into account the mapping ambiguity.
The bioinformatics methods presented here may find application in other RNA-seq studies dealing with high similarity in reference sequences. In particular, the same methodology may be able to identify differential expression between other homologous genome regions. Studies in other fields, such as metagenomics, dealing with high similarity DNA sequences may also find benefits from strict alignment and intersection with uniquely mappable locations.

Human tissue samples
The Human Body Map 2.0 RNA-seq data for 16 human tissue samples were provided by Gary Schroth at Illumina and can be accessible from ArrayExpress (accession no. E-MTAB-513). Reads were 75 base pairs long and came from the following samples: adipose, adrenal, brain, breast, colon, heart, kidney, liver, lung, lymph, muscle, ovary, prostate, testes, thyroid, and white blood cells. The samples were prepared using the Illumina mRNA-seq kit. They were made with a random priming process and are not stranded.

Software and datasets
Bowtie version 0.12.7 [15] and TopHat version 1.2.0 [21] were used for the mapping. Cufflinks version 1.0.3 [17] was used for differential expression calculation for RefSeq genes. BEDTools version 2.12.0 [22] was used to analyze alignments. The uniqueome dataset was collected from the Uniqueome supplementary database [16] for human genome (hg18, NCBI Build 36.1) marking genome locations where reads of length 75 bps are unique within 4 mismatches (hg18_uniqueome. unique_starts.base-space.75.4.positive.BED). The 75 bps read length matches the RNA-seq data provided by Illumina. RefSeq genes and DNA sequences of spliced ribosomal protein genes were collected from NCBI (RefSeq database D32-6) [14]. Pseudogene annotations and sequences were downloaded from pseudogene.org [23] database (human pseudogenes build 58). Pseudogenes whose parent genes are ribosomal protein genes were selected, totaling 1788. Among them, 79 were annotated 'Duplicated'. As we are only interested in processed pseudogenes, our analysis focuses on the remaining 1709 pseudogenes. The human genome sequence (hg18) was collected from NCBI build 36.1.

Composite genome
A composite genome index was constructed with Bowtiebuild using the sequences of the human genome (hg18, NCBI build 36.1) and NCBI spliced RP gene sequences.

Alignment
RNA-seq data for each tissue was aligned using two distinct methodologiesone for pseudogenes and one for real genes. Pseudogene alignment protocol consists of strict alignment (Bowtie, no mismatches, report reads with only one alignment location only) to the composite genome. Real gene alignment protocol consists of strict alignment (Bowtie, no mismatches, single alignment location) to the human genome (hg18, NCBI Build 36.1).

Uniqueome
A uniqueome data set [16] was obtained for Build 36.1 marking genome locations where reads of length 75 bps are unique within 4 mismatches. Alignments for all tissues for both real genes and pseudogenes were intersected with the uniqueome dataset for all genome locations (intersectBed from BEDTools [22]). The total number of remaining reads in each alignment was counted. The uniqueome dataset was used to filter out ambiguously mapped reads.

Comparative expression analysis
Gene expression values were calculated as reads aligned to gene per kilobase of exon per million mapped reads (RPKM) [11]. The number of reads aligned to all gene exons and additionally aligning in unique locations was counted for each gene. Exon length for genes was calculated as the sum of unique positions as marked by the uniqueome across all gene exons. It is worth noting that RP pseudogenes appear spliced in the human genome and therefore have only  Figure 6 Distribution of tissue specificity, as measured by the JS divergence [20].