Quantitative bacterial transcriptomics with RNA-seq

Abstract

RNA sequencing has emerged as the premier approach to study bacterial transcriptomes. While the earliest published studies analyzed the data qualitatively, the data are readily digitized and lend themselves to quantitative analysis. High-resolution RNA sequence (RNA-seq) data allows transcriptional features (promoters, terminators, operons, etc.) to be pinpointed on any bacterial transcriptome. Once the transcriptome is mapped, the activity of transcriptional features can be quantified. Here we highlight how quantitative transcriptome analysis can reveal biological insights and briefly discuss some of the challenges to be faced by the field of bacterial transcriptomics in the near future.

RNA-seq comes of age

Advances in RNA sequencing technology have revolutionized the study of bacterial transcriptomes [1,2]. At its core, RNA-seq generates digital information that allows transcriptional features to be located with single-nucleotide precision in a strand specific manner. Since the data are digital, RNA-seq facilitates quantitative computational analysis of any selected region of the transcriptome, but the transcriptome must first be annotated properly. Since bacterial genomes are organized in operons, it is logical that RNA-seq data should be annotated with the operon architecture in mind. In practice, only three transcriptional features need to be defined: 5′ transcript ends (promoters), 3′ ends (terminators), and RNA sequence read coverage to connect the ends, which together define operons [3-5].

The true power of RNA-seq resides in its potential as an analytical tool for quantifying promoter activity, terminator efficiency, and differential expression of transcripts, including operons, transcription units within operons (e.g. generated by promoters internal to operons), and antisense RNAs. As described in more detail below, RNA-seq datasets consist of tens of millions of sequence reads and typically the reads are 50 bases in length. The raw sequence reads are aligned to a reference genome and only high quality reads are retained and mapped. Conversion of sequence data into digital format is accomplished by employing freely available computer scripts that count the number of times each transcribed base was sequenced in a read-aligned dataset, thereby converting aligned sequence reads to base count data. Normalization of the base count data is necessary to quantify the differential expression (i.e., relative base counts) of each transcriptional feature within a sample or between different samples. The normalized base count data can be quantified by averaging the base count across a selected region of the genome. Since the average of the base counts is used, the relative expression of any given transcription feature, regardless of its length, can be expressed in this way. Here we focus on the analysis of an E. coli RNA-seq dataset to demonstrate the strategy we developed to quantify the expression of the transcriptional features that define operons in bacteria.

Single-nucleotide resolved RNA-seq dataset

To obtain an RNA-seq dataset suitable for quantitative analysis, we prepared RNA from a culture of E. coli K-12 strain BW38028 during logarithmic- and stationary-phase growth on glucose limited minimal medium, as described previously [4]. In addition, we starved E. coli BW38028 and its isogenic rpoS mutant BW39452 for nitrogen by decreasing by three-fold the amount of ammonium chloride in the growth medium [6]. The RNA was extracted by using the hot-phenol method [7] and DNase I treated to remove contaminating DNA. The RNA samples were not depleted for rRNA prior to sequencing, which tends to eliminate some experimental biases [8]. The RNA samples were shipped on dry ice to vertis Biotechnologie AG (Germany) for library preparation and Illumina HiSeq2000 sequencing, as described by others [7,9]. For library preparation the RNA samples were split and subjected to differential RNA-seq (dRNA-seq) as described [2,10]. Briefly, one portion of the RNA was fragmented by ultrasound and then the fragments were poly(A)-tailed and an RNA adapter was ligated to the 5′ phosphate of the RNA. First strand cDNA synthesis was with a poly(dT) primer and reverse transcriptase. Second strand cDNA synthesis incorporated a barcoded 3′ TruSeq adapter. The other portion of the RNA samples were fragmented and treated with terminator exonuclease (TEX), which enriches for 5′ triphosphate containing transcripts that are generated by transcription initiation at promoters. The TEX treated samples then were tailed and ligated, and cDNA was prepared as described above. The cDNAs were sequenced on an Illumina HiSeq2000 system using 50 bp read length, with each library yielding approximately 20 million reads.

Datasets consisting of 10 million reads per sample are sufficient for transcriptional feature mapping and differential gene expression analysis without ribo-depletion for a transcriptome the size of E. coli [9,11]. For quantification the genome-aligned, strand-specific RNA-seq data should be converted from aligned reads to base counts. Our RNA-seq data analysis pipeline involves alignment of the raw data to the reference genome by using Bowtie2 to generate the sequence read alignment file (SAM) [12]. SAMTOOLS [13] were used to convert the SAM file to a binary alignment file (BAM). The BAM file was converted to a BigWig file (base count file), which contains the count of the base at each base location and is the standard for visualization in genome browsers such as J-Browse [14]. Conversion of BAM to BigWig formatted files can be accomplished by using tools available in the Galaxy Toolshed [15] or at UCSC Genome Browser [16].

Alternatively, users can analyze their datasets by using pipelines such as Galaxy [17] or READemption [18], which outputs normalized wiggle files (base count files). A simple and straightforward way to normalize base count data is by using a strategy analogous to the total count approach [19] for normalizing gene-specific read alignments, which expresses each value as the base count per billion bases counted [4]. Because the BigWig file represents the base count at each nucleotide position, all downstream analysis begins with this file. The advantages of the base count approach are: a) the digital base count data are inherently computable because of their format and smaller size, b) the average base counts of individual transcriptional features can be computed and queried at any desired resolution, from a single nucleotide to an entire operon, to quantify the expression level or activity, c) normalization of base count data makes all samples directly comparable, and d) the use of average base count values eliminates the length bias when comparing transcriptional features of different length [19].

Identification of transcription start sites

Several published RNA-seq studies have focused on transcription start site (TSS) identification [7,9,10,20-28]. The annotation of TSSs is essential for analyzing promoters, 5′ UTRs, operon architecture, and for discovering novel transcripts. To assure accuracy, a set of “best practices” for TSS identification is emerging. Enrichment of the 5′ RNA ends that are generated by transcription initiation is critical for accurate TSS identification. The many advantages of dRNA-seq were recently reviewed [2]. The initiating nucleotide in bacteria is a nucleotide triphosphate, which can be distinguished from 5′-monophosphate and 5′-OH containing RNAs that are generated by RNA processing or RppH pyrophosphohydrolase activity [29]. The enrichment strategy preferred by many researchers makes use of 5′ monophosphate-dependent terminator exonuclease (TEX), which degrades RNA with 5′ monophosphate ends to enrich for primary transcripts that contain 5′ triphosphate ends and hence represent the product of transcription initiation [10]. dRNA-seq works by enumerating differences in base counts between TEX-enriched and unenriched sequencing libraries. Experimental replication is critical for accurate TSS identification. Since dRNA-seq is remarkably reproducible, comparison of datasets generated by using the same protocols yet different growth conditions adds confidence to the process and the use of different growth conditions also increases the number of mapped TSSs. RNA samples from many growth conditions can be pooled for dRNA-seq identification of thousands of promoters [9]. For example, a recent dRNA-seq analysis of Salmonella using RNA pooled from 22 different growth conditions led to mapping of 96% of the TSSs that could be identified by independently analyzing the 22 samples [9].

When annotating transcriptome data, it is convenient to use widely available computer programs to search dRNA-seq datasets for TSSs [20,30,31]. The advantages of the computational process compared to manual annotation are the speed and precision of recording transcription feature locations. However, like all bioinformatics approaches, some features will be missed and there will be false positives. In the end, human supervision of the results is critical and the state-of-the-art in transcriptome annotation remains a manual process [9]. Manual annotation of TSSs is made more efficient by plotting the count of only the first base at the 5′ end of each TEX-enriched read (Fig. 1A) [32]. In practice this allows visualization of the 5′ triphosphate nucleotide at the TSS.

Figure 1.

Transcriptional feature map and analysis of the cysK-ptsHI-crr operon. The dRNA-seq data are available at GEO, GSE58556. (A) The genes and feature locations are drawn to scale and annotated to the positive strand of the E. coli MG1655 U00096.3 reference genome. Promoters (P) are indicated by an arrow and are numbered in order from left to right on the positive strand. Terminators (T) are indicated by a diamond. The base count data, consisting of TEX-treated samples pointing up and unenriched coverage data (fragmented RNA not treated with TEX) pointing down, are visualized in J-Browse [14], as described previously [4]. Only positive strand data are shown. Tracks: wild type (WT), glucose-grown E. coli K-12 in logarithmic phase (blue track); WT in stationary phase, 30 min after exhaustion of glucose (red track); WT starved for nitrogen (green track); and an isogenic rpoS mutant starved for nitrogen (tan track). The base count scale (on the left) is from 0 to 100, with values exceeding 100 indicated by dark red. (B) The relative activities of the nine promoters is plotted in the graphs as log2 average counts of the first 10 transcribed bases under the four different growth conditions, which are colorized as above. (c) The decrease in average counts of the 25 bases before and after the terminator T-A are shown by light green and pink arrows. (D) Time series analysis of the relative expression levels of three transcripts within the complex cysK-ptsHI-crr operon is plotted as the log2 average counts of bases from the indicated promoters to terminators, as described previously [4]. Time point 1 is during middle logarithmic phase, time point 2 is immediately prior to entry into stationary phase, time point 3 is 15 min after entry into stationary phase, time point 4 is 30 min after entry into stationary phase, and time point 5 is 180 min after entry into stationary phase. Additional details of the analysis are described in the text.

Subsequent to identification of TSSs by dRNA-seq, bioinformatics and functional analyses can add weight to promoter identification. For example, the DNA sequences immediately upstream of putative TSSs can be analyzed by using a bioinformatics approach to score sigma factor specific RNA polymerase binding sequence motifs [4,33]. ChIP determination of RNA polymerase binding provides a robust and comprehensive validation of putative promoters [23]. When used in combination, dRNA-seq, consensus amongst experimental replicates, promoter sequence analysis, and RNA polymerase binding assays are a powerful set of tools for the identification of promoters.

Annotation of 3′ ends

To obtain the full analytical value of RNA-seq data it is essential to map the 3′ transcript ends. Annotating 3′ ends is a notably more difficult endeavor than mapping TSSs because there currently is no method of enriching for them. The 3′ ends are the primary sites of exonuclease-dependent RNA decay, which may be the reason that RNA base counts decline at the 3′ ends of operons, and few reads extend into the stem loop structures of intrinsic terminators (Fig. 1C). Further complicating 3′ end analysis is that termination is typically inefficient [34], which allows read-through transcription. Currently, the best method for annotating 3′ ends is to search for correlation between replicates of the furthermost downstream bases transcribed, keeping in mind that the base counts near the 3′ end will be low even for highly expressed transcripts. Comparison of the 3′ ends to terminator predictions adds confidence to the analysis. For example, the TransTermHP software package works very well for finding intrinsic terminators [35]. In addition, a ChIP-chip analysis of the distribution of RNA polymerase after treatment with the Rho-specific inhibitor bicyclomycin led to identification of 200 Rho-dependent terminators [36]. Once both the 5′ and 3′ transcript ends are mapped, it is possible to annotate operons.

Annotation of operons

The transcriptome is a map of the activities of promoters and terminators. These activities are located on both strands of the genome [37] and depending on their arrangement, can give rise to antisense transcription and overlapping, divergent [38,39] and convergent operons [40,41]. To accommodate this naturally occurring complexity it is necessary to annotate the operon architecture. Three transcriptional features are necessary to define operons: 5′ ends (promoters), 3′ ends (terminators), and sufficient RNA-seq read coverage to connect the ends. If sequence reads cover 90% of the bases, this is a sensible indicator that the operon is real [4,32]. While there are computer algorithms that can find operons [5,42,43], just as for TSS mapping, the state-of-the-art remains a manual process [9]. Once the operons have been mapped, it is a straightforward task to annotate additional promoters and terminators within operons, which add complexity to the transcriptome. Mapping of internal promoters can be done manually or by bioinformatics analysis of mapped promoters that fall within the base locations of annotated operons. The transcriptional feature locations can be formatted as a GenBank feature file by using “promoter”, “terminator” and “operon” as feature keys (see for example, GSE52059 [4]). This format accommodates incremental annotation of condition specific regulatory information and is an accepted standard for disseminating genome annotation data [44]. Once the transcriptional feature locations are annotated, it is reasonably straightforward to calculate the average base count value for each feature, from each dataset, as described below.

Computing the activities of transcriptional features

Analysis of RNA-seq reads at the base count level permits normalized base counts to be readily averaged across any range of base locations to calculate the relative expression level, activity, or efficiency of individual transcriptional features [4]. We determined empirically that computing the average count of the first 10 transcribed bases accurately represents promoter activity and allows closely spaced promoters to be discriminated [4]. Likewise, the efficiency of transcription termination can be calculated as the relative decline in average base counts in 25-base windows before and after terminators (Fig. 1C). The relative transcript levels of operons can be calculated by averaging the base counts from the promoter to the terminator locations. Likewise, the expression levels of alternative transcripts generated by promoter and terminator activities within operons can be calculated. These applications of single-nucleotide-resolution analysis are exemplified in Fig. 1, for wild type E. coli K-12 during logarithmic growth on glucose minimal medium and during starvation for carbon (stationary phase) or nitrogen, as well as an rpoS mutant during nitrogen starvation.

The cysK-ptsHI-crr operon contains 4 genes and multiple transcription units (Fig. 1A). Conservatively, more than 40% of E. coli operons contain multiple transcription units that are differentially expressed, underscoring the need for an annotation system that accommodates operon architecture [4]. In addition to the primary promoter (P-1) and terminator (T-B) that define the operon, there are 8 additional promoters and one terminator within the operon (Fig. 1A). The activities of the promoters range from 12 to more than 10,000 average base counts (calculated from +1 to +10 at each promoter) and their relative activities under the four growth conditions are plotted in Fig. 1B.

There are two promoters (P-1 and P-2), separated by 33 base pairs, which drive transcription of cysK (Fig. 1B). Comparison of the average counts of the first 10 transcribed bases indicates that P-2 is greater than 30-fold more active than P-1. Inefficient termination (approximately 40% of cysK transcripts are not terminated, as indicated by the ratio of average base counts) at the internal terminator (T-A) suggests that cysK and ptsHI-crr are co-transcribed (Fig. 1C). Nevertheless, the T-A terminator segments the operon into cysK and ptsHI-crr specific transcripts, which makes sense because CysK is a cysteine biosynthetic enzyme and the remaining genes encode components of the phosphotransferase system (PTS) involved in sugar uptake [45]. In the current annotation these genes are thought to comprise two operons (cysK and ptsHI-crr) [46], but the data in Fig. 1 show a low but significant number of RNA-seq reads across the terminator T-A, most clearly in the log phase sample. There is also a promoter (P-3) internal to cysK that under all four conditions is relatively active compared to the other promoters and could contribute to transcription across the cysK-ptsH intergenic region (Fig. 1B), yet P-3 activity does not appear to correlate with the base counts in the corresponding unenriched samples and therefore is unlikely to contribute to operon function (Fig. 1A). Given its location at the end of a transcript and immediately upstream of an inefficient terminator, this could be an example of a pervasive transcript, which is discussed below.

Two promoters, P-4 and P-5, which are located within the cysK-ptsH intergenic region, drive transcription of ptsHI-crr. P-4 is approximately 15 times more active in logarithmic phase than it is under the other three conditions (Fig. 1B). On the other hand, P-5 is induced (2.5-fold) in stationary phase and nitrogen-starved conditions by comparison to logarithmic phase and its activity is rpoS-dependent, as indicated by a 40-fold reduction in promoter activity by comparison to the wild type under the same conditions (Fig. 1B). The transcripts originating from these two promoters apparently are terminated at T-B, downstream of crr (Fig. 1C). The collective activities of P-4 and P-5 correlate well with the modest decline in average base counts of the P-4:T-B (ptsHI-crr) transcript upon entry into stationary phase (Fig. 1D). Within the ptsI gene are three closely spaced promoters (P-6, P-7, and P-8) that are of relatively low activity compared with the others (Fig. 1B). P-6 is expressed approximately equally in the four conditions, P-7 is induced in stationary phase and nitrogen-starved conditions and is RpoS-dependent, and the least active of the three, P-8, is also dependent RpoS. It does not appear that these three promoters contribute to transcription of the downstream crr gene, as indicated by a lack of change in the unenriched base counts visualized in Fig. 1A, and so these promoters could also generate pervasive transcripts. On the other hand, P-9 is highly active in stationary phase and nitrogen-starved conditions, is RpoS-dependent, and is located near the 3′ end of ptsI (Fig. 1B), where it apparently drives expression of a crr specific transcript (Fig. 1A).

Time series analysis shows that the three major transcripts within the operon are differentially expressed during growth and entry into stationary phase (Fig. 1D). The cysK-specific transcript is expressed at high levels during logarithmic phase and its level declines rapidly during stationary phase. Hence expression of cysK reflects the decline in P-1 and P-2 promoter activity in stationary phase and nitrogen-starved conditions. The ptsHI-crr transcript level declines little during the first 30 min of stationary phase and then declines modestly 3 hours into stationary phase (Fig. 1D), probably because P-4 is less active and P-5 is induced upon entry into stationary phase (Fig. 1B). Expression of the crr transcript is partially dependent on read-through from promoters within ptsH and ptsI, and there is no evidence from the base counts to indicate that there is termination within the ptsI-crr intergenic region. The crr-specific transcript level increases upon entry into stationary phase in the wild type, yet declines in an RpoS-dependent manner in the rpoS mutant (Fig. 1D). Indeed, P-9 is RpoS dependent, as indicated by 16-fold higher expression in the wild type starved for nitrogen compared to the rpoS mutant, and it has a -10 promoter element with the base sequence (CTAnnnTTAA) that is characteristic of RpoS promoters [47].

The primary goal of many RNA-seq experiments is to determine differential gene expression between growth conditions and treatments [9,19,27,32,48-52]. Typically these experiments involve calculating for control and test conditions the number of reads that map to the genome between the start and stop codons of individual genes. Similarly, differential expression of operons can be determined by calculating the average base counts between the promoters and terminators. Since the average operon contains 2 genes, plus intragenic sequences, and 5′ and 3′ UTRs, there is significantly more information used (more bases) to compute the operon expression level than what is available to represent expression of individual genes. So, the statistical significance of differential expression can be greatly enhanced by using normalized base count data to measure relative operon or transcript expression levels. Differential transcription of operons is readily accomplished by employing algorithms such as DEseq [48] to compute the differential expression and statistics.

Challenges

Massive amounts of RNA sequencing data can now be readily obtained. Precise mapping of transcriptional features, logical organization of the annotated data, and meaningful feature quantitation are key to maximizing the value of the resulting transcriptomes. Critical analysis of dRNA-seq data is needed to minimize the number of false positive promoters annotated. Thus it is necessary not only to properly replicate dRNA-seq experiments, but also to augment the analysis with information to corroborate that a predicted TSS is indeed a functional promoter, such as by promoter motif analysis and RNA polymerase binding assays. It would be useful if future advances in TSS mapping technology include methods to directly label the nucleotides corresponding to TSSs, rather than simply enriching for them. Mapping of 3′ transcript ends is an even larger issue and there is a real need for technology that directly labels the 3′ ends generated by transcription termination. Perhaps in vitro poly(A) tailing of the 3′ ends of RNA prior to fragmentation, followed by sequencing from that end would be helpful. However, it appears from existing RNA-seq data that termination is not a precise biological process and transcripts do not stop at a single nucleotide. For the time being, the state-of-the-art for 3′ transcript end mapping remains consensus between replicates.

Lastly, it is important to determine whether “pervasive transcription”, defined as TSSs in non-canonical locations [53], is real and if such transcripts have a functional role. Pervasive transcription is seen in yeast, mammals, and fruit flies [54,55] and is frequently observed in viruses and bacteria [32,56,57]. So, there seems to be little doubt that pervasive transcription is real. As to whether pervasive transcripts are functional, that topic was recently reviewed, but it is too early to be sure [53]. The finding that some pervasive transcripts in herpesvirus decreased viral protein production [56] suggests that the functional role of such transcripts should be investigated in bacteria. It is becoming apparent that H-NS and NusG suppress some pervasive transcripts [57,58]. Several potential examples of pervasive transcription can be seen in Fig. 1. Using a conservative approach we previously mapped 4 promoters to the cysK-ptsHI-crr operon [4]. However, dRNA-seq revealed 9 promoters that map to the operon (Fig. 1A), only 4 of which appear to drive transcription of the corresponding genes (P-2, P-4, P-5, and P-9). The other 5 include a weak promoter upstream of the major promoter in front of cysK and a relatively strong promoter located within the cysK coding region and just upstream of the terminator that is intergenic to cysK-ptsH. Neither of these promoters appears to contribute to transcript expression levels. The remaining 3 putative pervasive promoters are located within the ptsI gene, have relatively low activity levels, and yet all have reasonably well conserved -10 promoter sequence elements, including two that have RpoS promoter motifs and appear to be RpoS-dependent. If these turn out to be real promoters, and there is no reason to think they are not, then the number of promoters on bacterial genomes is being underestimated by perhaps two-fold [9,32].

Highlights.

• An annotation schema that accommodates operon architecture is ideal for genome-wide bacterial transcriptome analysis.

• RNA sequencing data can pinpoint transcriptional features with single-nucleotide-resolution.

• Single-nucleotide-resolved data essentially are digital and facilitate quantitative analysis of promoter activity and terminator efficiency, as well as differential expression of transcripts.

• Challenges facing the transcriptomics field are the lack of technologies to directly label transcript ends, need to validate promoters identified by dRNA-seq, and resolving the issue of “pervasive transcription”.