To assess the extent of artifacts generated by reverse transcription, we performed five replicates (biological) of standard reverse transcription reactions lacking ActD (ActD−) on RNA samples from yeast grown in rich media and analyzed cDNA targets by hybridization to tiling arrays. For sense transcripts, hybridization signals were concordant between all replicates. However, two types of signals were registered on the antisense strands opposite to expressed genomic regions. In one class, the signal intensities corresponded proportionally to the intensities of the sense counterparts (A, upper panel); furthermore, substantial variability existed across replicates (B and C, upper panel). In the other class, antisense signals were highly reproducible across all replicates and did not correlate with sense strand expression levels (D, upper panel). These differences suggested that the first class of antisense signals might be artifacts, potentially triggered by spurious second-strand synthesis during reverse transcription as proposed by the models (). Besides growth in rich media, we have seen the same pattern for several other conditions (data not shown). We postulated that the putative artifacts can be resolved using ActD. Indeed, three replicates of reverse transcription reactions performed with ActD (ActD+) resulted in expression signals below background for the first class, but not for the second class of antisense regions (A–D, lower panels). In addition to the examples shown in , profiles for all genomic regions are available online (http://www.ebi.ac.uk/huber-srv/actinomycinD
Figure 2. Genome-wide analysis of sense and antisense signals from hybridizations performed in either the absence (ActD−) or presence (ActD+) of ActD during reverse transcription. (A) Dependence of antisense signal intensity on the sense expression level. (more ...)
A genome-wide comparison of expression signals with or without ActD over all coding genes [as defined by ORFs in the Saccharomyces
Genome Database (SGD, http://www.yeastgenome.org
)] and their opposite regions supported the interpretation that ActD reduces artifactual antisense signals. The number of antisense regions detected as expressed above background declined from 1046 in ActD− to 325 in ActD+. Moreover, only 25% (260/1046) of the cases observed upon standard first-strand cDNA synthesis are still detectable in ActD+ (Supplementary Table 1). Furthermore, consistent with its specific role in second-strand synthesis, ActD did not affect the number of sense transcripts detected above background. Out of 5703 coding-genes (ORFs), the majority is expressed above background in both conditions: 5214 ORFs in ActD+ and 5186 ORFs in ActD−; in addition, there is a nearly complete overlap between the two gene lists (5172 ORFs) (Supplementary Table 1).
Tiling arrays can be used for de novo
identification of transcripts (7
). Therefore, the hybridization signals of probes were examined along their chromosomal positions irrespective of previous annotation and separately for each strand. The profiles were partitioned into segments of constant hybridization intensity using a segmentation algorithm (7
). We analyzed the effect of ActD for sense and antisense transcripts defined by segmentation (Supplementary Table 2). Antisense segments were defined based on three criteria: (i) the segments are expressed and overlap annotated genes located on the opposite strand, but not on the same strand, (ii) the segment lengths are longer than 48 bp and (iii) they are flanked by segments with reduced hybridization signal on both sides. A comparison shows that ActD has no quantitative effect on first-strand cDNA synthesis, since there is concordance between the expression levels of sense segments measured in ActD+ and ActD− reactions (). However, the addition of ActD reduces the expression level below background for more than half of the antisense segments: among a total of 553 antisense segments, 347 give signal above background only in ActD− and 14 only in ActD+, while 192 antisense segments are detected in both conditions (Supplementary Table 3). Therefore, strikingly, the number of antisense segments observed by using a standard reverse transcription protocol is decreased by 64% (347/539) with the inclusion of ActD.
Figure 3. Scatter plot of segment expression levels between arrays generated with and without ActD. Green dots, antisense segments detected above background in ActD− samples; Red line, identical expression in both conditions; dashed lines, background thresholds. (more ...)
Three independent lines of evidence are in agreement with the array hybridization results generated in the presence of ActD. First, we compared our array hybridization data to the stringent set of antisense transcripts obtained by full-length cDNAs sequencing (8
). Despite complementarities between these two methods, as well as the comparison being between different experimental growth conditions (rich media versus minimal media), a better overlap was achieved for the ActD+ than the ActD− dataset (P
value of 1.6e−9 versus 2.5e−6, Fisher's exact test). Second, we evaluated the effect of a stringent computational filter on removing putative antisense artifacts from hybridization results. This filter was previously developed to computationally remove putative antisense artifacts by requiring segments to have higher expression signal than seen on the opposite strand for at least part of their length (7
). On ActD− data, the computational filter reduced the number of antisense segments by 63% (337/539). In contrast, the filter had only a mild effect on the number of antisense segments detected in ActD+: of 206 antisense segments, only 39 were filtered out. In addition, a comparison of antisense segments detected in ActD− after filtering, and antisense segments detected in ActD+, shows good (but imperfect) concordance (158/250) (Supplementary Table 4)
. Since computational filters always face the dilemma of compromising between false positives and false negatives, the experimental improvement using ActD is more advantageous. This is clearly the case for antisense detection as the majority of antisense transcripts are weakly expressed (Supplementary Table 3) and thus hard to distinguish from noise. Notably, several antisense segments filtered out in ActD− (48 segments, A orange dots) are detectable in ActD+, which suggests that these cases were erroneously filtered (manual inspection confirms this). In addition, ActD+ also resolves artifactual antisense segments that erroneously passed the computational filter (44 segments, A green dots). Third, we conducted semi-quantitative strand-specific RT-PCR analysis for 10 antisense segments (B): of the three antisense transcripts detected both in ActD+ and ActD− (MBR1, EPL1, MRK1
), all three yielded a signal by strand-specific RT-PCR. In contrast, of seven antisense segments detected exclusively in ActD− (CYS4, EMP24, PNC1, MDH3, HAC1, TRR1, PFK1
), all seven yielded negative results by strand-specific RT-PCR.
Figure 4. Validation of antisense transcripts. (A) Comparison of ActD and computational filtering. Expression level of antisense segments in reactions with (ActD+) and without ActD (ActD−). For ActD− data, only antisense segments which passed the (more ...)