Understanding the molecular processes responsible for normal development or tumorigenesis necessitates both identifying functionally important mutations and exploring the transcriptomic diversity of various tissues. RNA sequencing (RNA-seq) provides genome-scale access to the RNA complement of a cell with unprecedented depth, and has therefore proven useful in unraveling the complexity of transcriptomes 12. The analyses of RNA-seq reads aim at detecting a variety of targets: from transcribed exons and classical splice junctions with canonical splice sites, to alternatively spliced RNAs, RNAs with non-standard splice sites, read-through and even non-colinear chimeric transcripts 3. Moreover, RNA-seq also gives access to those somatic mutations and genetic polymorphisms that are transcribed. Chimeric RNAs result from the transcription of genes fused together by chromosomal rearrangements 4, especially in cancer 5, and they can also be induced by trans-splicing between mature messenger RNAs (mRNAs) 6. RNA-seq can also capture these complex, non-colinear transcripts, whose molecular importance is still poorly assessed and which may provide new diagnostic or therapeutic targets 78.

As next generation sequencing (NGS) improves and becomes cheaper, bioinformatic analyses become more critical and time consuming. They still follow the same paradigm as in the first days of NGS technologies: a multiple step workflow - mapping, coverage computation, and inference - where each step is heuristic, concerned with only a part of the necessary information, and is optimized independently from the others. Consequently analyses suffer from the drawbacks inherent to this paradigm: (a) pervasive erroneous information, (b) lack of integration, and (c) information loss, which induces re-computation at subsequent steps and prevents cross-verification. An example of the lack of integration is that the mapping step cannot use coverage information, which prevents it from distinguishing biological mutations from sequencing errors early in the analysis.

Here, we design a novel and integrated strategy to analyze reads when a reference genome is available. Our approach extracts information solely from the genome and read sequences, and is independent of any annotation; we implemented it in a program named CRAC. The rationale behind it is that an integrated analysis avoids re-computation, minimizes false inferences, and provides precise information on the biological events carried by a read. A peculiarity of CRAC is that it can deliver computational predictions for point mutations, indels, sequence errors, normal and chimeric splice junctions, in a single run. CRAC is compared with state-of-the-art tools for mapping (BWA, SOAP2, Bowtie, and GASSST) 910111213, and both normal (GSNAP, TopHat, and MapSplice) 31415 and chimeric (TopHat-fusion) 16 splice junction predictions. The results show the relevance of the approach in terms of efficiency, sensitivity, and precision (which is also termed specificity in the literature). We also provide true assessments of the sensitivity of each method by analyzing complex simulated data.

Availability: CRAC is distributed under the GPL-compliant CeCILL-V2 license and is available as source code archive or a ready-to-install Linux package from the CRAC project website 17 or the ATGC bioinformatics platform 18. It includes two programs: crac-index to generate the index of the genome, and crac for analyzing the reads.

We evaluated CRAC for mapping reads, predicting candidate SNVs, indels, splice junctions, and chimeric junctions, and compared it to other tools. Simulated data are needed to compute exact sensitivity and accuracy levels, while real data enable us to study predictions with biologically validated RNAs. For simulating RNA-seq, we first altered a reference genome with random substitutions, indels, and translocations to derive a mutated genome, then reads were sequenced in silico using FluxSimulator 22, the annotated RefSeq transcripts, and a realistic distribution of random expression levels (Additional file 2). As read lengths will increase, we used two simulated datasets to assess different strategies: one (hs75) with a typical read length of 75, another (hs200) with reads of 200 nt representing the future.

CRAC is a multi-purpose tool for analyzing RNA-seq data. In a single run it can predict sequencing errors, small mutations, and normal and chimeric splice junctions (collectively termed events). CRAC is not a pipeline, but a single program that can replace a combination of Bowtie, SAMtools, and TopHat/TopHat-fusion, and can be viewed as an effort to simplify NGS analysis. CRAC is not simply a mapper, since it uses local coverage information (in the support profile) before computing the genomic position of a read. In contrast to the current paradigm, mapping and post inferences are not disjoint steps in CRAC. Instead, it implements a novel, integrated approach that draws inferences by simultaneously analyzing both the genomic locations and the support of all k-mers along the read. The support of a k-mer, defined as the number of reads sharing it, approximates the local read coverage without having the reads mapped. The combined k-mers location and support profiles enable CRAC to infer precisely the read and genomic positions of an event, its structure, as well as to distinguish errors from biological events. Integration is not only the key to an accurate classification of reads (Additional file 1), but it avoids information loss and saves re-computation, and is thereby crucial for efficiency. Indeed, CRAC takes more time than state-of-the-art mappers, but is considerably faster than splice junction prediction tools (for example, Bowtie plus TopHat). The other key to efficiency is the double-indexing strategy: a classical FM-index (where FM stands for Ferragina - Manzini) for the genome and the Gk arrays for the reads 21. This makes CRAC's memory requirement higher than that of other tools, but fortunately computers equipped with 64 gigabytes of memory are widespread nowadays. Experiments conducted on simulated data (where all answers are known), which are necessary for assessing a method's sensitivity, have shown that for each type of prediction CRAC is at least competitive or surpasses current tools in terms of sensitivity, while it generally achieves better precision. Moreover, CRAC's performances further improve when processing longer reads: for example on 200 nt reads, it has 85% sensitivity and 99.3% precision for predicting splice junctions.

CRAC analyzes how the location and support profiles vary and concord along the read. Hence k-mers serve as seeds (in the genome and in the read set), and k is thus a key parameter. Its choice depends on the genome length 19, and quite conservative values - k = 22 for the human genome - have been used in our experiments. Smaller k values are possible with smaller genomes (like bacterial ones). k affects the number of false genomic locations (FLs) that occur in the profile; a FL indicates a wrong location for a k-mer, which differs from the location of origin of the sequenced molecule. This tends to induce a false location for the read (mapping) or a false location for a junction border (normal and chimeric junction prediction). However, CRAC uses two criteria to avoid these pitfalls: the coherence of locations for adjacent k-mers over a range and the concordance of locations for the k-mers around the break (especially in the break verification and fusion procedures; see Additional File 2). When k-mers surrounding the break have a few, but several, locations, CRAC examines all possible combinations, and as FL occurrences are governed mainly by randomness, this eliminates discordant positions. FLs have a larger effect on the prediction of chimeras. Overall, the results on both simulated and real data, like the improved mapping sensitivity (+15 points compared to Bowtie, BWA, and SOAP2), show that CRAC makes accurate predictions with conservative values. k controls the balance between sensitivity (shorter seeds) and precision. The breast cancer libraries we used have 50 nt reads, but CRAC could still find 74% of the chimeric RNAs validated by Edgren et al. 27. Of course, the k value has two limitations: first, the minimal exon size detectable in a read is ≥k, second, reads must be long enough (>40 nt with k = 20 for the human genome). However, NGS is progressing towards longer reads, which should become standard, and Figure 4c illustrates well CRAC's ability to detect short exons within single reads. The k-mer profiling approach detects events located near the read extremities, but cannot exactly determine their position in the read. Hence the inference rules cannot be fully applied, and CRAC classifies such reads as incompletely determined (Undetermined and BioUndetermined files). However, the position of an event in a read is random, and thus, the high coverage delivered by NGS nearly ensures that the same event occurs in the middle of other reads covering it. Consequently, border cases do not hinder CRAC from detecting mutations, splice junctions, etc. Only errors escape this rule since they are mostly read specific. A more complex drawback of k-mer profiling is when two events are located <k positions apart on the genome (see the BCAS4-BCAS3 chimera); again such cases even with a high support are not fully resolved and end up in the BioUndetermined file. A post-processing of reads in this file, for example by an alignment program, could clearly save such cases. Obviously, such cases are rare, and we keep this as future work. As briefly mentioned, k-mer profiling also detects when reads span a repeat border region, which should help in inferring the locations of mobile genetic elements, duplications, or copy number variations; this suggests further developments and CRAC's usefulness for analyzing genomic data.

Determining the correct genomic location of reads is crucial information for any NGS data analysis and especially for cataloging all transcripts of a cell with RNA-seq. Generally, a mapping step computes this information using efficient, well-known tools (BWA, Bowtie, and SOAP2), but the mapping sensitivity is rarely questioned. We performed extensive mapping tests on simulated data, which showed that sensitivity can truly be improved and that CRAC makes a significant step in this direction. Of course by considering discontinuous alignments (as do CRAC and GSNAP) many reads covering splice junctions can be mapped, which BWA, Bowtie/Bowtie2, and SOAP2 cannot detect. However, the mapping results for categories of reads carrying one mutation, a short indel, or even errors indicate that classical mappers missed between 15 to 20 points in sensitivity, thereby confirming that the difference due to splice junction reads is critical even for other events, while CRAC performs equally well (>90%) whatever the category (Figure 2). The other way around, those tools are able to map 10% to 35% of reads containing a splice junction. This can negatively affect downstream analyses depending on the type of events under investigation. For instance to predict splice junctions, in the current strategy (TopHat, MapSplice, or TopHat-fusion), reads are first mapped with Bowtie to divide the collection into: (a) reads having a continuous alignment on the genome and (b) unmapped reads. The former serve further to delimit exons, and the latter are then processed again to search for spliced alignments. If a read that requires a discontinuous alignment is mapped by Bowtie, it is not considered by TopHat, MapSplice, or TopHat-fusion as potentially containing a junction, and they will not find a spliced alignment for it. In contrast, CRAC's k-mer profiling approach is flexible, reliable in this respect (Figure 3), and importantly, adapts well to longer reads (for example, 200 nt). This last point is key since longer reads will be available soon. They will much more likely incorporate not one, but several events - errors, mutations, splice junctions, etc. - and thus be harder to map. Whatever the class of required predictions, CRAC's sensitivity is always improved with longer reads. This is crucial for detecting multiple exons within single reads, and CRAC exhibits a better ability in this as exemplified by a transcript of TIMM50 gene (Figure 4c).

An issue in transcriptomics is to reliably extract the complete set of splice junctions with a minimal number of false positives 24. In this regard, our results (Table 2) demonstrate that k-mer profiling approaches (MapSplice and CRAC) profit greatly in sensitivity from longer reads, and that CRAC is the tool with the highest precision whatever the read length. They also indicate that CRAC handles difficult cases with higher sensitivity, like long-distance splices, multi-exon reads, or RNA expressed at a low level. The analysis of a multi-tissue library shows that CRAC, GSNAP, and MapSplice have a very large (>90%) agreement on the set of reported known junctions (>140,000 distinct junctions), RefSeq transcripts, and genes, thereby providing evidence of their ability to extract splice junctions of well-annotated transcripts (Figure 4b and 4a). In contrast, TopHat misses 21% of these known RefSeq junctions. Comparatively, CRAC reports fewer novel or unknown junctions than other tools, and tends to be more conservative, which likely reflects its precision. Altogether, CRAC is a solution for exploring qualitatively the transcriptome of a sample with high sensitivity and precision, and thus provides the primary material for determining all transcript structures, which is indispensable for estimating the expression levels of all RNA isoforms 326.

Recent investigations have suggested that non-colinear RNAs are quantitatively more abundant in human transcriptomes than previously thought, underlining the structural diversity of these chimeric RNAs and their occurrence in cancers 827283536. Predicting chimeric RNAs (chRNAs) is the most difficult and error-prone computation when analyzing RNA-seq. The combinatorial possibilities of aligning a read partly to two distinct regions on the same or different chromosomes 4 increase the likeliness of predicting FPs. It explains why filtering for suboptimal but colinear alignments of an apparent chimeric read may still help, and also partly why TopHat-fusion per se yields so many more chRNA candidates compared to CRAC (Table 4). Paired end reads are not sufficient: analyzing single reads by splitting them is inevitable for predicting the chimeric junction point; hence k-mer profiling also suits this purpose. Nevertheless, paired end reads are useful for performing a complementary consolidation of chRNA candidates, which we may develop in the future. However, chRNAs can occur at low expression levels and be much less expressed than their parental genes; this impels CRAC to rely less on the support profile than for mutation prediction. In addition, transcriptional noise or template switching during library preparation may generate true chimeric reads from biologically irrelevant chRNAs. Thus, subsequent criteria are definitely needed to prioritize chRNA candidates: the consistent finding of the same junction point has been suggested as an important one 273637. Notably, CRAC predicted for the four breast cancer libraries 36 recurrent chRNAs that were not reported previously 1627, and 35/36 always harbor the same junction point in the different libraries and among the distinct reads predicting them. Several of these involve genes known to be implicated in tumorigenesis or tumor maintenance, like GNAS 31 or HSPD1 30. As CRAC outputs also included 74% of validated chRNAs with a single clear false negative, this shows that CRAC consistently reports interesting chRNA candidates based on the read data. As already mentioned, CRAC distinguishes between five chRNA classes, included those exhibiting small-scale sequence inversions, as illustrated by a chRNA within the LONP1 gene, which recurs in normal and tumoral libraries. We also reported cases of chRNAs, which although validated, do not constitute good candidates for the computational inference step, since not enough reads in the data support their existence. The latter point is critical and strengthens how difficult chimeric RNA prediction is.

Here, the in silico experiments focus on transcriptomic data, but the method is also applicable to genomic sequencing. For instance, the counterparts of splice junctions and chimeras in RNA-seq are large deletions and rearrangements (translocation, inversion, and displacement of a mobile element) in DNA. Thus, CRAC may also prove useful for genomic analyses.

chRNA: chimeric RNA; EST: expressed sequence tag; FL: false location; indel: insertion or deletion; FP: false positive; KJ: known splice junction; NGS: next generation sequencing; NJ: new splice junction; nt: nucleotide; OJ: other splice junction; RNA-seq: RNA sequencing; SJ: splice junction; SNP: single nucleotide polymorphism; SNV: single nucleotide variant; TP: true positive.

The authors declare that they have no competing interests.

NP, MS and ER devised the algorithm. NP and MS developed the source code. All authors devised and analyzed the software comparisons and evaluations, as well as the analysis of real datasets. NP and MS performed the experiments. NP and MS prepared the figures. ER wrote the manuscript with contributions from all authors. ER and TC supervised the research. All authors read and approved the final manuscript.

^a TopHat-fusion without the extra post-filtering script.

^b If TopHat-fusion-post is applied to TopHat-fusion's results with default parameters, it reports 27 chimera, 11 of them being validated chimeras, which is about half those reported by TopHat-fusion alone.

^c Only intergenic chRNAs are counted here.