Background

Massively parallel whole transcriptome sequencing, commonly referred to as RNA-Seq, and its ability to generate full transcriptome data at the single transcript level, provides a powerful tool with multiple interrelated applications, including transcriptome assembly [1–4], gene and transcript expression level estimation [5–8], also known as transcriptome quantification, studying trans- and cis-regulatory effects [9], studying parent-of-origin effects [9–11], and calling expressed variants [12].

RNA-Seq has become the technology of choice for performing transcriptome analysis, rapidly replacing array-based technologies [13]. The Ion Torrent technology offers the fastest sequencing protocol for RNA-Seq experiments able to sequence whole transcriptome in few hours [14]. Most current research using RNA-Seq employs methods that depend on existing transcriptome annotations. Unfortunately, as shown by recent targeted RNA-Seq studies [15], existing transcript libraries still miss large numbers of transcripts. The incompleteness of annotation libraries poses a serious limitation to using this powerful technology since accurate normalization of RNA-Seq data critically requires knowledge of expressed transcript sequences [5–8]. Another challenge in transcriptomic analysis comes from the ambiguities in read/tag mapping to transcripts. Ubiquitous regulatory mechanisms such as the use of alternative transcription start and polyadenylation sites, alternative splicing, and RNA editing result in multiple messenger RNA (mRNA) isoforms being generated from a single genomic locus. Most prevalently, alternative splicing is estimated to take place for over 90% of the multi-exon human genes across diverse cell types [8], with as much as 68% of multi-exon genes expressing multiple isoforms in a clonal cell line of colorectal cancer origin [16]. The ability to reconstruct full length transcript sequences and accurately estimate their expression levels is widely believed to be critical for unraveling gene functions and transcription regulation mechanisms [17].

Here, we focus on two main problems in transcriptome analysis, namely, transcriptome assembly and quantification. Transcriptome assembly, also known as novel transcript discovery or reconstruction, is the problem of assembling the full length transcript sequences from the RNA sequencing data. Assembly can be done de novo or it can be assisted by existing genome and transcriptome annotations. Transcriptome quantification is the problem of estimating the expression level of each transcript. In the remainder of this section we give a brief description of the common protocols used for mRNA sequencing.

RNA-Seq protocol

RNA-Seq uses next generation sequencing technologies, such as SOLiD [18], 454 [19], Illumina [20], or Ion Torrent [21]. Figure 1 depicts the main steps in an RNA-Seq experiment, ending with the first step of analysis which is typically mapping the data to a reference. The mRNA extracted from a sample is converted to cDNA using reverse transcription and sheared into fragments. Fragments with lengths within a certain range are selected, and ligated with sequencing adapters. This is usually followed by an amplification step after which one or both ends of the cDNA fragments are sequenced to produce either single or paired-end reads. cDNA synthesis and adapter ligation can be done in a strand-specific manner, in which case the strand of each read is known; this is commonly referred to as directional sequencing. In the more common non-directional RNA-Seq protocols strand specificity is not maintained. The specifics of the sequencing protocols vary from one technology to the other. In particular, the length of produced reads varies depending on the technology, with newer high-throughput technologies typically producing longer reads.

Figure 1
figure 1

A schematic representation of the RNA sequencing protocol.

Full size image

Related work

Transcriptome assembly and quantification from RNA-Seq data has been the focus of much research in recent years. The sequences of novel transcripts together with their expression levels can be inferred from deep RNA-Seq data, but this is computationally challenging due to the short length of the reads, high percentage of sequencing errors, uneven coverage of expressed transcripts, and the need to distinguish between highly similar transcripts produced by alternative splicing. A number of methods address the problem of transcriptome assembly and quantification from RNA sequencing data. Methods for transcriptome assembly fall into three categories: "genome-guided", "genome-independent" and "annotation-guided" methods [22]. Genome-independent methods such as Trinity [23] or transAbyss [24] directly assemble reads into transcripts. A commonly used approach for such methods is the de Brujin graph [25] utilizing "k-mers". The use of genome-independent methods becomes essential when there is no trusted genome reference that can be used to guide assembly. On the other end of the spectrum, annotation guided methods [26–28] make use of available information in existing transcript annotations to aid in the discovery of novel transcripts. RNA-Seq reads can be mapped onto the reference genome, reference annotations, exon-exon junction libraries, or combinations thereof, and the resulting alignments are used to assemble transcripts.

Many transcriptome reconstruction methods fall in the genome-guided category. They typically start by mapping sequencing reads onto the reference genome, using spliced alignment tools, such as TopHat [29] or SpliceMap [30]. The spliced alignments are used to identify putative exons, splice junctions and transcripts that explain the alignments. While some methods aim to achieve the highest sensitivity, others work to predict the smallest set of transcripts explaining the given input reads. Furthermore, some methods aim to reconstruct the set of transcripts that would insure the highest quantification accuracy. Scripture [1] construct a splice graph from the mapped reads and reconstructs transcripts corresponding to all possible paths in this graph. It then uses paired-end information to filter out some transcripts. Although Scripture achieves very high sensitivity, it may predict a lot of incorrect isoforms. The method of Trapnell et al. [4, 31], referred to as Cufflinks, constructs a read overlap graph and reconstructs transcripts using a minimal size path cover via a reduction to maximum matching in a weighted bipartite graph. TRIP [3] uses an integer programming model where the objective is to select the smallest set of putative transcripts that yields a good statistical fit between the fragment length distribution empirically determined during library preparation and fragment lengths implied by mapping read pairs to selected transcripts. IsoLasso [32] uses the LASSO [33] algorithm, and it aims to achieve a balance between quantification accuracy and predicting the minimum number of transcripts. It formulates the problem as a quadratic program, with additional constraints to ensure that all exons and junctions supported by the reads are included in the predicted isoforms. CLIIQ [34] uses an integer linear programming solution that minimizes the number of predicted isoforms explaining the RNA-Seq reads while minimizing the difference between estimated and observed expression levels of exons and junctions within the predicted isoforms. Traph [35] proposed a method based on network flows for a multiassembly problem arising from transcript identification and quantification with RNA-Seq. Another method, CLASS [36] uses local read coverage patterns of RNA-seq reads and contiguity constraints from read pairs and spliced reads to predict transcripts from RNA-Seq data. iReckon [37] is a method for simultaneous determination of the transcripts and estimation of their abundances. This probabilistic approach incorporates multiple biological and technical phenomena, including novel isoforms, intron retention, unspliced pre-mRNA, PCR amplification biases, and multi-mapped reads. iReckon utilizes regularized Expectation-Maximization to accurately estimate the abundances of known and novel transcripts.

Methods

Spliced alignment

Alignment of RNA-Seq reads onto the reference genome, reference annotations, exon-exon junction libraries, or combinations thereof is the first step of RNA-Seq analyses, unless none of these are available in which case it is recommended to use de novo assembly methods [23, 24]. The best mapping strategy depends on the purpose of RNA-Seq analysis. If the focus of the study is to estimate transcripts and gene expression levels rather then discover new transcripts then it is recommended to map reads directly onto the set of annotated transcripts using a fast tool for ungapped read alignment. To be able to discover new transcriptional variants one should map the reads onto the reference genome. Recently, many bioinformatics tools, called spliced read aligners, have been developed to map RNA-Seq reads onto a reference genome [29, 30]. Alternatively, RNA-Seq reads can be mapped onto the genome using a local alignment tool such as the Ion Torrent mapper, TMAP. Both spliced alignments and local alignments can be used to detect novel transcriptional and splicing events including exon boundaries, exon-exon junctions, gene boundaries, transcriptional start (TSS) and transcription end sites (TES).

In our experiments we used TopHat [29] with default parameters. For assessing transcriptome quantification accuracy Ion Torrent reads from cancer datasets were mapped on the External RNA Controls Consortium (ERCC) RNA spike-in controls reference [38] with added polyA tails of 200 bp using TMAP. Reads for the MAQC datasets were mapped onto Ensembl known transcripts with added polyA tails of 200 bp, also using TMAP.

Splice graph and putative transcripts

Typically, a gene can express multiple mRNA transcripts due to alternative transcriptional or splicing events including alternative first exon, alternative last exon, exon skipping, intron retention, alternative 5' splice site (A5SS), and alternative 3' splice site (A3SS) [39]. To represent such alternative transcripts, a gene is processed as a set of so called 'pseudo-exons' based on alternative variants obtained from aligned RNA-Seq reads. A pseudo-exon is a region of a gene between consecutive transcriptional or splicing events, i.e., starting or ending of an exon, as shown in Figure 2. Hence, every gene consists of a set of non-overlapping pseudo-exons. This gene representation lets us easily enumerate all possible transcripts of a gene. To generate the set of putative transcripts, we first create a splice graph based on pseudo-exon boundaries and splice junctions.

Figure 2
figure 2

Pseudo-exons. An example of three transcripts, Tr 1 , Tr 2 and Tr 3 . Each transcript is represented as a set of exons. Pseudo-exons are regions of a gene between consecutive transcriptional or splicing events. S psej and E psej represent the starting and ending position of pseudo-exon j, respectively.

Full size image

The splice graph is a directed acyclic graph (Figure 3) whose vertices represent pseudo-exons and edges represent pairs of pseudo-exons immediately following one another in at least one transcript (which is witnessed by at least one spliced read). Both splice junctions and pseudo-exon boundaries are inferred from read alignments. To construct the splice graph, MaLTA infers splice junctions from gapped alignments of RNA-seq reads. Next, inferred splice junctions are used to partition the reference genome into a set of non-overlapping segments, which are classified as (a) intron, (b) pseudo-exon, or (c) combination of both. It is easy to classify a segment as pseudo-exon if it is entirely covered, and as intron in case it is entirely uncovered. In case of partial coverage we require 80% of the segment to be covered to be classified as pseudo-exon, otherwise it is classified as (a) or (c). Segments containing a combination of introns and exons most likely contain gene boundaries. In this case we identify islands of coverage inside the segment. A segment may contain several coverage islands which correspond to single exon genes.

Figure 3
figure 3

Splice graph. The red horizontal lines represent single reads. Reads interrupted by dashed lines are spliced reads. Each vertex of the splice graph corresponds to a pseudo-exon and each directed edge corresponds to a (splice) junction between two pseudo-exons. Red vertices of the slice graph serve as transcription start sites (TSS). Blue vertices - transcription end sites (TES).

Full size image

After constructing the splice graph, MaLTA enumerates all maximal paths using a depth-first-search algorithm. These paths correspond to putative transcripts. Note that a gene with n pseudo-exons may have as many as 2n − 1 possible candidate transcripts, each composed of a subset of the n pseudo-exons. The next subsection presents a maximum likehood transcriptome assembly and quantification algorithm that selects a minimal subset of candidate transcripts that best fits the observed RNA-Seq reads. The key ingredient is an expectation-maximization algorithm for estimating expression levels of candidate transcripts.

Maximum likehood transcriptome assembly

Existing transcriptome assembly methods [3, 4] use read pairing information and fragment length distribution to accurately assemble the set of transcripts expressed in a sample. This information is not available for current Ion Torrent technology, which can make it challenging to assemble transcripts. The Ion Torrent PGM platform is able to produce single reads with read length in 50-300 bp range. Our approach is to simultaneously explore the transcriptome structure and perform transcriptome quantification using a maximum likelihood model. MaLTA starts from the set of putative transcripts and selects the subset of this transcripts with the highest support from the RNA-Seq data. Maximum likelihood estimates of putative transcripts are computed using an Expectation Maximization (EM) algorithm which takes into account alternative splicing and read mapping ambiguities. EM algorithms are currently the state-of-the-art approach to transcriptome quantification from RNA-Seq read, and have been proven to outperform count-based approaches. Several independent implementations of EM algorithm exist in the literature [7, 40].

We developed a new version of IsoEM [7] suitable for Ion Torrent RNA-Seq reads. IsoEM is an expectation-maximization algorithm for transcript frequency estimation. It overcomes the problem of reads mapping to multiple transcripts using iterative framework which simultaneously estimates transcript frequencies and imputes the missing origin of the reads. A key feature of IsoEM, is that it exploits information provided by the distribution of insert sizes, which is tightly controlled during sequencing library preparation under current RNA-Seq protocols. In [7], we showed that modeling insert sizes is highly beneficial for transcript expression level estimation even for RNA-Seq data consisting of single reads, as in the case of Ion Torrent. Modeling insert sizes contributes to increased estimation accuracy by disambiguating the transcript of origin for the reads. In IsoEM, insert lengths are combined with base quality scores, and, if available, strand information to probabilistically allocate reads to transcripts during the expectation step of the algorithm. Since most Ion Torrent sequencing errors are insertions and deletions, we developed a version of IsoEM capable of handling insertions and deletions in read alignments.

The main idea of the MaLTA approach is to cover all trancriptional and splicing variants presented in the sample with the minimum set of putative transcripts. We use the new version of the IsoEM algorithm described above to estimate expression levels of putative transcripts. Since IsoEM is run with all possible candidate transcripts, the number of transcripts that are predicted to have non-zero frequency can still be very large. Instead of selecting all transcripts with non zero frequency, we would like to select a small set of transcripts that explain all observed splicing events and have highest support from the sequencing data. To realize this idea we use a greedy algorithm which traverses the list of candidate transcripts sorted in descending order by expression level, and selects a candidate transcript only if it contains a transcriptional or splicing event not explained by the previously selected transcripts.

Results and discussions

We evaluated the accuracy of the MaLTA-IsoEM approach on both simulated and real human RNA-Seq data. The human genome sequence (hg18, NCBI build 36) was downloaded from UCSC together with the KnownGenes transcripts annotation table. Genes were defined as clusters of known transcripts defined by the GNFAtlas2 table. In our simulation experiments, we simulate reads together with spliced alignments to the genome; these alignments are provided to all compared methods. We varied the length of single-end reads, which were randomly generated per gene by sampling fragments from known transcripts. All the methods were compared on datasets with various read length, i.e., 50 bp, 100 bp, 200 bp, and 400 bp. Expression levels of transcripts inside each gene cluster followed uniform and geometric distributions. To address library preparation process of RNA-Seq experiment we simulated fragment lengths from a normal probability distribution with different means and 10% standard deviation.

All reconstructed transcripts were matched against annotated transcripts. As in [4] and [32], two transcripts were assumed to match if and only if internal exon boundaries coordinates (i.e. all exons coordinates except the beginning of the first exon and the end of the last exon) were identical. We use sensitivity and positive predictive value (PPV) to evaluate the performance of different assembly methods. Sensitivity is defined as the proportion of assembled transcripts that match annotated transcripts, i.e., sensitivity = TP/(TP + FN). Positive predictive value (PPV) is defined as the proportion of annotated transcript sequences among assembled sequences, i.e., PPV = TP/(TP + FP).

Transcriptome quantification accuracy was evaluated by comparing RNA-Seq estimates with TaqMan qRT-PCR measurements [41] or External RNA Controls Consortium (ERCC) RNA spike-in controls [38]. The coefficient of determination (R2) between the (qRT-PCR/ERCC) and Fragment Per Kilobase of exon length per Million reads (FPKM) estimates was used as accuracy measure.

Comparison on simulated RNA-Seq data

In this section, we use sensitivity and PPV defined above to compare the MaLTA to other transcriptome assembly tools. The most recent versions of Cufflinks (version 2.1.1) [4] and IsoLasso (v 2.6.0) [2] with the default parameters are used for performance comparison. We explore the influence of read and fragment length on performance of assembly methods.

Table 1 reports sensitivity and PPV of transcriptome assembly for reads of length 400 bp, simulated assuming both uniform and geometric expression of transcripts. MaLTA significantly outperforms the other methods, achieving both sensitivity and PPV of over 75% for all datasets. For all methods the difference in accuracy between datasets generated assuming uniform and geometric distribution is small, with the latter one typically having a slightly worse accuracy. Thus, in the interest of space we present remaining results for datasets generated using uniform distribution.

Table 1 Sensitivity and PPV comparison between methods on datasets simulated assuming uniform, respectively geometric expression of transcripts, with reads length 400 bp, mean fragment length 450 bp and 10% standard deviation.
Full size table

There is a strong correlation between the number of splicing events within the gene and the number of annotated transcripts. A high number of splicing events leads to increased number of candidate transcripts, which makes the selection process more difficult. To explore the behavior of the methods depending on number of transcripts per gene we divided all genes into categories according to the number of annotated transcripts and calculated the sensitivity and PPV within each such category.

Figures 4(a)-4(b) compare the performance of three methods (Cufflinks, IsoLasso, MaLTA) on simulated data with respect to the number of transcripts per gene. Note that sensitivity and PPV (Figure 4) for single-transcript genes is 100% for all methods and is excluded from consideration. MaLTA achieves equivalent or better results in both sensitivity and PPV for all categories.

Figure 4
figure 4

Sensitivity and PPV comparison between methods for groups of genes with n transcripts on simulated datasets with mean fragment length 250 bp, 10% standard deviation, and read length of 100 bp.

Full size image

Table 2 compares assembly accuracy of Cufflinks, IsoLasso, and MaLTA for different combinations of read and fragment lengths: (50 bp,250 bp), (100 bp,250 bp), (100 bp,500 bp), (200 bp,250 bp), (400 bp,450 bp). The results show that MaLTA provide 5-15% improvement in sensitivity and 1-10% improvement in PPV.

Table 2 Sensitivity and PPV comparison between methods for different combinations of read and fragment lengths: (50 bp, 250 bp), (100 bp, 250 bp), (100 bp, 500 bp), (200 bp, 250 bp), (400 bp, 450 bp).
Full size table

Comparison on Ion Torrent cancer and MAQC RNA-Seq datasets

For this study, we compared MaLTA and Cufflinks on 3 cancer datasets downloaded from the Ion Community website: GOG-382 (HepG2 - hepatocellular carcinoma), DID-416 (K562 - myelogenous leukemia) and DID-413 (MCF-7 - breast ductal carcinoma). Comparison with IsoLasso on the real datasets is omitted due to technical problems (IsoLasso results were consistently incomparable to other methods). Reads were mapped to the hg18 reference genome using TopHat2 (with default parameters) which is able to produce spliced alignment used by transcriptome assembly tools (Table 3).

Table 3 Read mapping statistics and read length for Ion Torrent HeLa datasets.
Full size table

Although UCSC annotations are known to be incomplete, we expect a significant proportion of assembled transcripts to be consistent with these annotations. Thus, the performance of transcriptome assembly methods was evaluated by the total number of assembled transcripts matching UCSC annotations. Table 4 gives the results obtained by MaLTA and Cufflinks on DID-413, DID-416 and GOG-382 datasets. Both methods assemble highest number of transcripts confirmed by reference annotations for GOG-382 dataset. Cufflinks and MaLTA respectively were able to assemble 13,887 and 16,143 transcripts, of which 1,557 and 4,395 are known annotated transcripts. A large number of identified annotated transcripts were confirmed by both methods (Figure 5). The GOG-382 dataset contains the highest number of annotated transcripts confirmed by both methods; among identified annotated transcripts 1,291 transcripts were confirmed by both methods.

Table 4 Performance comparison of transcriptome assembly between Cufflinks and MaLTA for Ion Torrent HeLa datasets.
Full size table
Figure 5
figure 5

Consistency of transcriptome assembly. Number of identified annotated transcripts confirmed by both methods for GOG-382,DID-416 and DID-413 datasets.

Full size image

To evaluate transcriptome quantification accuracy of the methods we ran IsoEM and Cufflinks on Ion Torrent RNA-Seq data generated from two commercially available reference RNA samples that have been well-characterized by quantitative real time PCR (qRT-PCR) as part of the MicroArray Quality Control Consortium (MAQC); namely the Ambion Human Brain Reference RNA, Catalog #6050), henceforth referred to as HBRR and the Stratagene Universal Human Reference RNA (Catalog #740000), henceforth referred to as UHRR. We used five HBRR datasets and five UHRR datasets for the comparison. To assess accuracy, gene expression estimates obtained from RNA-Seq data were compared against those obtained from TaqMan qRT-PCR measurements (GEO accession GPL4097) collected as part of the MAQC project. As described in [41], each TaqMan Assay was run in four replicates for each measured gene. POLR2A (ENSEMBL id ENSG00000181222) was chosen as the reference gene and each replicate CT was subtracted from the average POLR2A CT to give the log2 difference (delta CT). For delta CT calculations, a CT value of 35 was used for any replicate that had CT >35. The normalized expression value for gene g was computed as 2(CT of POLR2A)-(CT of g), and the average of the qPCR expression values in the four replicates was used as the ground truth. Mapping gene names to Ensembl gene IDs using the HUGO Gene Nomenclature Committee (HGNC) database resulted in TaqMan qPCR expression levels for 832 Ensembl genes. Tables 5 and 6 show statistics for the size, number of mapped reads, and accuracy of gene expression levels estimated by IsoEM for each of the 10 datasets as well as the combined reads for each sample. Figure 6 presents a comparison between IsoEM and Cufflinks results. IsoEM estimates correlate better with qPCR measurements compared to Cufflinks. Additionally, IsoEM estimates have less variability across different Ion Torrent runs.

Table 5 Read mapping statistics and correlation between gene expression levels estimated by IsoEM and qPCR measurement for Ion Torrent UHRR dataset.
Full size table
Table 6 Read mapping statistics and correlation between gene expression levels estimated by IsoEM and qPCR measurement for Ion Torrent HBRR dataset.
Full size table
Figure 6
figure 6

Correlation of estimates obtained by both IsoEM and Cufflinks with qPCR measurments for HBRR and UHRR datasets. The red color represents the 2nd quartile and the green color represents the 3rd quartile.

Full size image

We also compared IsoEM and Cufflinks on two of the cancer Ion Torrent datasets, GOG-382 and DID-413. Methods were evaluated by calculating correlation between estimated FPKMs for External RNA Controls Consortium (ERCC) spike-in controls [38] with the known frequencies of these RNA controls in the samples (ERCC mix1 was spiked in for both runs). Table 7 presents the results of this comparison, showing higher R2 for IsoEM in both cases.

Table 7 Correlation (R2) between known frequencies of spiked in ERCC controls and gene expression levels estimated by IsoEM and Cufflinks for Ion Torrent HeLa datasets.
Full size table

Conclusion

In this paper we described the MaLTA-IsoEM method for simultaneous transcriptome assembly and quantification from Ion Torrent RNA-Seq data. Our approach explores transcriptome structure and incorporates a maximum likelihood model into the assembly and quantification procedure. Results on real cancer and MAQC RNA-Seq datasets show that Ion Torrent RNA-Seq data can be successfully used for transcriptome analysis. Transcriptome assembly and quantification accuracy was confirmed by comparison to annotated transcripts and TaqMan qRT-PCR measurements and External RNA Controls Consortium RNA spike-in controls. Experimental results on both real and synthetic datasets generated with various sequencing parameters and distribution assumptions suggest increased transcriptome assembly and quantification accuracy of MaLTA-IsoEM compared to existing state-of-the-art approaches.