To accurately determine RNA levels in the parasite, any biases that are inadvertently introduced during the preparation of cDNA libraries must be kept at a minimum while strand-specificity, coverage and sequencing depth should be maximized. Previously, a thorough comparison of different strand-specific RNA-Seq methods indicated that the least amount of ‘false’ antisense RNA was generated when libraries were prepared by using the ‘RNA-ligation method’ involving the sequential ligation of 3′-preadenylated and 5′-adapters to RNA followed by reverse transcription using a primer complementary to the 3′-adapter 17 (Figure  1A). Therefore, we adopted this method for use in Plasmodium although alternative strategies to generate strand-specific libraries may also be well suited 20 .

To increase the sequencing coverage of non-ribosomal RNA, we generated libraries from polyA-enriched RNA, unless indicated otherwise. We found polyA-enriched libraries to contain less rRNA than libraries prepared from RNA treated with the 5′-phosphate-dependent exonuclease (Tex) (Figure  2), an enzyme that specifically digests processed RNAs with a 5′-monophosphate end (e.g. rRNA). Furthermore, we noticed that the genome-wide coverage was substantially higher for libraries prepared from polyA-enriched libraries than for libraries prepared from Tex-treated RNA. This difference is probably a consequence of the lower percentage of rRNA found in libraries prepared from polyA-enriched RNA (Figure  2). Next we compared the distribution of sequence reads across CDSs for libraries prepared from polyA-enriched and Tex-treated RNA and noticed a strong bias towards the 5′-end in the Tex-treated library (Figure  1B). While this bias was not apparent for all genes, it was highly reproducible for individual genes and distinctly more pronounced in large genes. Therefore, to obtain more uniform sequencing coverage, we prepared all subsequent libraries from polyA-enriched RNA.

Using the RNA-ligation-based protocol, we prepared sequencing libraries from RNA extracted 10 h, 20 h, 30 h and 40 h post infection (p.i.) (n = 4), as well as from cytoplasmic and nucleic RNA of parasites harvested at 20 h and 30 h p.i. (n = 4). Combining the data from these libraries with the data from libraries prepared for protocol development (n = 3), our dataset consists of ~600 million mapped strand-specific reads derived from 11 libraries (Figure  2).

Although the existence of NATs has been documented in P. falciparum based on SAGE, microarray and strand-specific high-throughput sequencing data 15 21 , the understanding of their genome-wide distribution is still incomplete due to the limited depth or uncertainties in the level of strand-specificity of the published datasets. Using strand-specific RNA-Seq data from ~30 million mapped reads, López-Barragán et al. observed NAT transcription in 312 coding genes 7 . To gain more insight into the genome-wide distribution of NATs, we took advantage of the unprecedented depth of our dataset (~600 million mapped reads, Figure  2) and combined the mapping results of 11 libraries to generate a coverage map. We also re-analyzed the López-Barragán et al. dataset (pooled from 4 strand-specific libraries in 7 ) in parallel for comparison of coverage and level of strand-specificity.

Using the combined data from 11 libraries, we detected transcription of 78.3% of the Plasmodium genome (≥5-fold coverage, Additional file 1: Figure S1). Keeping all parameters constant, re-analysis of the López-Barragán et al. dataset indicated transcription of ~39.5% of the Plasmodium genome. Most of the transcribed genomic positions detected in the López-Barragán et al. dataset were also detected in this study (Additional file 2: Figure S2A), suggesting both datasets are consistent and that our dataset represents a substantially deeper coverage.

Next, we tried to determine the level of strand-specificity of our datasets compared to that generated by López-Barragán et al. Previously, a globally positive correlation between sense and antisense transcripts has been used as an indicator for potential presence of artifactual antisense transcripts 22 23 . This approach is based on the fact that the observed antisense transcripts could be derived from the sense transcript during library preparation, e.g. incomplete second-strand cDNA digestion in dUTP method 7 , spurious synthesis of second-strand cDNA 22 , or other unidentified sources of artifacts. While the correlation in our dataset was close to 0 (Pearson’s correlation = 0.057, P < 0.01), we observed a strong positive correlation for the López-Barragán et al. dataset (Pearson’s correlation = 0.82, P < 0.01, Additional file 2: Figure S2B), suggesting a relatively higher level of strand-specificity of our dataset. Likewise, the ratios of sense to antisense reads differ significantly between our (1 antisense to 328 sense reads) and the López-Barragán et al. dataset (1 to 11.25). The relatively lower level of strand-specificity in López-Barragán et al. dataset maybe due to the imperfect second-strand cDNA digestion in dUTP method as the authors mentioned 7 , comparing to the RNA ligation method used in this study. López-Barragán et al. thus avoided false positives in NAT identification by applying stringent cutoffs on both proportion of antisense read (>70%) and antisense read number (>150) 7 , which may have underestimated the prevalence of NATs.

To estimate the prevalence of NATs in Plasmodium, we established two sets of thresholds with different stringency (see Methods for details). Our data indicate that between ~24% (n = 1247, stringent thresholds) and ~45% (n = 2389, relaxed thresholds) of all coding genes (n = 5284) are overlapping with NATs (Additional file 3: Table S1). In the following sections, unless mentioned otherwise, we restricted our analyses to the genes with NATs defined at stringent thresholds (n = 1247).

Numerous mechanisms have been described by which antisense RNA can regulate gene expression even in the absence of a functional RNAi-machinery. Common to these mechanisms is the ability of the respective organism to regulate antisense transcription independently of the complementary sense transcription.

Previously, the distinct regulation of mRNA levels during the IDC has been described 2 . The IDC of P. falciparum takes 48 h to complete, thus, to evaluate developmental regulation of NATs across the IDC, we compared patterns of sense and antisense expression in parasites at evenly spaced intervals (10 h, 20 h, 30 h and 40 h p.i.).

To determine whether transcription of NATs is developmentally regulated, we searched for changes in NAT levels across the IDC. Using the software package EdgeR, we observed significant changes (FDR <0.05) in antisense transcript levels for 357 of the 1247 genes with NATs in at least 1 of the 6 non-redundant time point pair comparisons. For 125 genes with NATs we observed significant changes in antisense transcript levels in at least 2 of the 6 non-redundant time point pair comparisons (see Methods for details). These findings suggest that at least ~10% of NATs are regulated during the IDC (Additional file 3: Table S1 and Additional file 4: Tables S2).

Subsequently, we investigated whether the changes in sense and antisense transcript levels between times points are correlated and observed no global correlation (Additional file 5: Figure S3). Thus, our data suggest that transcription of NATs does not exert a globally positive or negative effect on the transcription of its sense counterpart. To further investigate this observation visually, we sorted the genes based on patterns of antisense expression levels throughout the IDC and plotted the antisense expression heat map (n = 2389, relaxed threshold), and observed a distinct life cycle-specific expression pattern of antisense transcripts. Then, we re-plotted the heat map using sense expression levels while keeping the gene order sorted using antisense expression levels. In agreement with the above observation, no apparent pattern was observed (Figure  3A, top panel). It should be noted that sorting the genes based on sense transcript levels confirmed the previously observed regulation of sense transcript levels 2 (Figure  3A, bottom panel). Thus, both sense and antisense transcript levels are regulated during the IDC but regulation is not synchronous. An example of developmentally regulated antisense transcript levels is shown in Figure  3B.

To further characterize the newly identified NATs, we mapped antisense transcript levels across the CDSs of all genes and observed a striking enrichment of antisense transcripts at the 3′-ends of genes (Figure  4A). It has been suggested that NATs could be a consequence of run-through transcription or transcription initiation of bi-directional promoters from adjacent genes (for example see Figures  4B-C) 24 25 . To determine if P. falciparum contains NATs whose expression is regulated independently from that of adjacent genes, we investigated the correlation between the changes in mRNA levels of the downstream genes and the changes in antisense transcript levels of the corresponding upstream genes during the IDC. We reasoned that, should antisense transcription be a consequence of active transcription of the downstream gene, changes in antisense transcript level across the IDC should be positively correlated to those of the downstream mRNA. To this end, we performed 6 non-redundant time-point pair comparisons for each gene (see Methods and Additional file 6: Table S3) and observed a positive correlation between the change of antisense transcript levels of one gene and that of mRNA levels of the gene located downstream (Figure  5). The observed correlation was independent of the orientation of the downstream gene, suggesting that both run-through transcription and bi-directional promoter activity contribute to the observed antisense transcription in Plasmodium. However, while we observed a positive correlation for 473 of 1247 genes with overlapping NATs (37.9%), for 66 genes (5.3%) we observed a negative correlation and for 519 genes (41.6%) no clear correlation (Figure  6). For 189 (15.2%) genes we did not have enough data points or no gene was located downstream. Among the NATs whose transcription did not correlate with the transcription of adjacent genes (n = 519 of 1247), i.e. NATs that appear to be independently regulated, we found ~13% (n = 67 of 519) to be developmentally regulated (for details see Methods, Additional file 3: Table S1 and Additional file 6: Tables S3). Taken together, these findings indicate that while run-through transcription and bi-directional promoter activity are likely to contribute to the observed antisense transcription, a significant number of NATs seems to be regulated independently of mRNA transcription.

To understand the role of antisense transcription in Plasmodium, we sought to determine whether our data provide evidence to support the known mechanisms for antisense-mediated regulation of gene expression.

Collisions between RNA polymerases transcribing the sense and antisense strands have been described to interfere with the transcription of mRNAs as shown in Escherichia coli 26 and S. cerevisiae 27 . If antisense transcription has a negative impact on sense transcription in Plasmodium, an increase in antisense levels from one time point to the next should correlate with a decrease in sense levels and vice versa. However, as mentioned above, comparison of changes in sense and antisense levels during four time-points showed a correlation close to 0 (Pearson’s correlation = 0.07, Additional file 5: Figure S3).

Alternatively, duplex formation between sense and antisense RNA molecules can result in nuclear retention of antisense transcripts and modulate gene expression. If antisense transcripts played a role in retaining sense transcripts in the nucleus, an enrichment of antisense transcripts in the nucleus would be observed. Several studies have reported nuclear localization of ncRNA in P. falciparum 28 29 30 . Thus, even though polyadenylated NATs are not commonly enriched in the nucleus, and our sequencing libraries were enriched for polyadenylated RNA, we examined the possible nuclear enrichment of antisense reads in genes that showed significant levels of antisense transcription but are not likely to be affected by run-through transcription from neighboring genes (listed in Additional file 3: Table S1). We did not observe a significant enrichment of antisense reads in these genes (n = 198) in the nucleus against both the total and the cytosolic fractions at 20 h p.i., while ncRNA from telomere-associated repeated elements were enriched in the nuclear fraction (Additional file 7: Figure S4).

A final mechanism we considered was antisense-mediated translational inhibition in which case overexpression of antisense transcripts leads to a reduction of proteins but not to reduction of sense RNA levels 31 . To this end, we correlated our RNA-Seq data with previously published stage-specific proteomic datasets 32 . As observed by Le Roch et al., we saw a positive correlation between mRNA abundance (in RPKM (reads per kilobase per million)) and protein abundance (in peptide counts per kilobase), thus an increase in sense RPKM correlates with an increase in peptide counts per kilobase for a given gene (Pearson’s correlation = 0.59 to 0.66 in 4 time points; Additional file 8: Figure S5A). To evaluate the global effect of antisense transcripts on translation, we investigated the correlation between antisense transcript levels (in antisense RPKM) and translation efficiency (in peptide counts per sense RPKM). However, for all four time points the correlation was close to 0 (Pearson’s correlation = -0.02 to 0.16 in 4 time points; Additional file 8: Figure S5B), suggesting that the level of antisense transcription has no global effect on translation efficiency.

The sequencing depth of our dataset combined with the use of a polymerase better suited for the amplification of extremely AT-rich regions than previously used enzymes (Additional file 9: Figure S6 and 19 ), allowed us to map for the first time polyadenylation sites (PAS) for P. falciparum. Despite the importance of 3′-UTRs in the regulation of gene expression, the genome-wide characterization of 3′-UTRs in Plasmodium lags far behind that of coding sequences. To determine PASs and to address the question of whether the observed antisense transcripts are polyadenylated, we used the combined RNA-Seq data from all 11 libraries to map PASs on a genome-wide scale. Due to the high abundance of coding mRNAs compared to antisense RNA (Additional file 3: Table S1), we expected most of the identified PASs to correspond to coding mRNAs and only a small fraction to correspond to NATs.

To map genuine PASs and to avoid false positive assignments derived from internal polyA stretches or sequence reads of low quality, we followed a set of previously published criteria 33 (see Methods). To assess whether the identified PASs could represent genuine PASs, we plotted the occurrence of PASs around stop codons of all genes (Figure  7A). Most of the identified PASs fell within 1000 nt downstream of the stop codon of genes, corresponding to the expected sites of polyadenylation and therefore validating our approach. As microheterogeneity in PAS selection is well documented 34 , we grouped individual PASs located in close proximity to each other into so-called ‘PAS-clusters’ with a maximum width of 20 nt. Setting the maximum 3′-UTR length to 2000 nt, we were able to assign 6678 PAS-clusters to 3443 coding mRNAs (i.e. 1.94 PAS-clusters per coding mRNA), yielding an average 3′-UTR length of 523 nt (median = 451 nt) (Additional file 10: Table S4), for an example see Figure  7B. Genome-wide we observed that 51.8% genes (n = 1785) contained multiple PAS-clusters compared to 72.1% genes with multiple PASs found in S. cerevisiae 35 .

Previously, some NATs in Plasmodium were shown to be transcribed by RNA polymerase II and it was suggested that NAT stability may be regulated at the level of polyadenylation 36 but no polyadenylation of NATs has been described for P. falciparum. Thus, we decided to search for PASs of NATs, i.e. antisense PAS-clusters. Antisense PAS-clusters were defined as those that 1) are located on the antisense strand within the coding region of an annotated gene, 2) contain at least two PASs within the cluster, 3) are located at least 2000 nt upstream of an annotated stop codon (to avoid including polyA tails of neighboring coding mRNA). Because most of the antisense transcripts exist in low abundance, the likelihood of identifying their PASs are much lower than for coding mRNAs. Nonetheless, based on the above criteria, we identified 154 antisense PAS-clusters for 135 genes (for an example see Figure  7C). These data indicate that at least some of the observed NATs are polyadenylated.

Based on the splicing junctions identified from our datasets using HMMSplicer 37 , we found evidence for antisense spliced junctions in 123 of the 1247 genes with NATs (Additional file 11: Table S5). Thus, at least a portion of the observed NATs (~10%) is spliced. Furthermore, we found 55 of the 123 antisense junction clusters to overlap with sense junctions on mRNA (Additional file 11: Table S5). These findings are consistent with a previous study, which has identified antisense introns and antisense junctions overlapping with sense junctions, using a similar approach 6 . These observations further validate antisense transcription in P. falciparum to be a widespread phenomenon.

Parasites from highly synchronous cultures were harvested 10 h, 20 h, 30 h or 40 h post-infection, red bloods cells lysed (0.15% Saponin in PBS) and total RNA (including small RNA) isolated using a miRNeasy Mini Kit (Qiagen). Separation of nuclear and cytosolic RNA fractions was performed as published previously 48 . Subsequently, genomic DNA was removed by on-column DNase treatment according to the manufacturer’s instructions (Qiagen) and mRNA was enriched by subjecting the total RNA to one round of polyA-selection using oligo(dT)-coated Dynabeads (Invitrogen) or treatment with 5′-phosphate-dependent exonuclease (Tex). For polyA-selection (libraries 1-9, Figure  2), we followed the instruction provided by Invitrogen. Alternatively (libraries 10 and 11, Figure  2), 25 μg of total RNA was treated with 10U of Tex (Epicentre), 1x Buffer A (Epicentre), 200U RNase Out (Invitrogen) in a final volume of 200 μl for 60 min at 30°C.

For the preparation of each strand-specific RNA library we used between ~50 ng (ring stage parasites) and ~150 ng (trophozoites) mRNA. mRNA was fragmented to approximately 100-200 nt in length. Reproducible fragmentation was obtained by mixing RNA with a RNA fragmentation reagent (Ambion) and heating it to 70°C for exactly 5 min. To remove the 5′-terminal 7-methylguanylate cap, the fragmented RNA was treated with 10U of Tobacco Acid Pyrophosphatase (Epicentre) for 2 hours at 37°C. All subsequent steps were performed according to an Illumina application Note (Note: directional mRNA-Seq sample preparation) with the following exceptions: Custom-made 5′- adapter (5′-GUUCAGAGUUCUACAGUCCGACGAUCNNNN-3′, conc. 20 μM) was used instead of the RNA adapter provided by Illumina, the PCR amplification was performed for 12-16 cycles using the KAPA HiFi DNA polymerase (Kapabiosystems) and KAPA HiFi Fidelity Buffer according the manufacturer’s instructions (Mg² concentration was adjusted to 2.5 mM). The PCR product was purified and concentrated using AMPure XP beads (Beckman Coulter). Quality and concentration of all libraries was determined using a Bioanalyzer 2100 (Agilent) and high throughput sequencing was performed on a HiSeq2000 (Illumina) except for library 11, Figure  2, which was sequenced on a Genome Analyzer (Illumina). For read statics see Figure  2.

For the preparation of gDNA libraries we fragmented 2.5 μg gDNA to a size of 200-300 bp using Bioruptor (Diagenode). Subsequently, the DNA was blunted using the End-It DNA end-repair kit from Epicentre according to the manufacturer’s instruction and a single dA was added to the 3′ end by mixing the DNA with 5 μl of 10x Buffer 2 (100 mM Tris-HCl, pH 7.9, 500 mM NaCl, 100 mM MgCl₂, 10 mM DTT, (New England Biolabs)), 10 μl of 1 mM dATP, 15U Klenow fragment (3′-5′ exo^-) in a final volume of 50 μl followed by an incubation for 30 min at 37°C. Next, a ‘forked DNA adapter’ consisting of DNA strands 5′ P-GAT CGG AAG AGC GGT TC AGCAGGAATGCCGAG-3′ and 5′-ACACTCTTTCCCTACACG ACGCTCTTCCGATct-3′ (phosphodiester bond between c and t) was ligated to the DNA by mixing the DNA with 2.5 μl of 40 μM forked adapter, 25 μl of 2× ligation buffer (Enzymatics), 5 μl of highly purified T4 ligase (600U/μl, Enzymatics) in a total volume of 50 μl followed by incubation of 15 min at 25°C. Ligated DNA ranging in size between 300 and 500 bp was size-selected on an agarose gel and PCR amplified for 12 cycles using the KAPA HiFi DNA polymerase (Kapabiosystems) and KAPA HiFi Fidelity Buffer according to the manufacturer’s instructions (Mg² concentration was adjusted to 2.5 mM). The PCR product was purified and concentrated using AMPure XP beads (Beckman Coulter). Quality and concentration was determined using a Bioanalyzer 2100 (Agilent) and high-throughput sequencing was performed on a HiSeq2000 (Illumina). Between enzymatic reaction the DNA was purified using NucleoSpin Extract II columns (Macherey-Nagel).

First, reads longer than 50 nt flagged with Illumina’s low quality flag “B” were removed from all datasets. Then, we removed the 4 custom index nucleotides at the 5′-end and the low quality nucleotides at the 3′-end of the reads. The P. falciparum 3D7 genome and its gene model annotations (version 3 on January 2012) were downloaded from Sanger FTP (ftp://ftp.sanger.ac.uk/pub/pathogens/Plasmodium/falciparum/). Unless specified otherwise, the reads in all datasets were mapped onto the reference genome using Bowtie 0.12.8 (parameters “-n 2 -k 1 -m 50 --best”), with maximum 2 mismatches and multiple hits (maximum 50) distributed to the best-aligned location 49 . To map the splicing junctions, the unaligned reads from Bowtie were mapped using HMMSplicer (parameters “-w 4 -j 10 -k 3000 -ja 10 -e 2 -m 500 -n 700 -d True”). It should be noted that HMMSplicer was specifically developed to work with P. falciparum RNA-Seq data 37 . Read pileups were generated for each library and all read pileups (n = 11) were pooled using custom scripts to generate a coverage map for defining the genomic distribution of NATs. Strand-specific fastq data of López-Barragán et al. 7 were downloaded from Short Read Archive of NCBI (http://www.ncbi.nlm.nih.gov/sra/) under the accession of SRR364836, SRR364841, SRR364842 and SRR364846. Second reads of the pairs in these datasets were reverse complemented and all reads were processed in the same way as our datasets (mentioned above).

The expression level of a transcript was expressed as number of reads per kilobase per million (RPKM) 50 . Briefly, we counted the number of reads mapped to all annotated transcriptomic features (e.g. mRNA) on the same strand (i.e. sense) and opposite strand (i.e. antisense). Both the sense and antisense read numbers were normalized by the length of the feature (in kilobase) and the total number of reads (in millions) mapped to non-structural RNAs in the corresponding library (i.e. number of mappable reads excluding rRNA and tRNA reads). To visualize the changes of sense and antisense expression levels across the 4 time points across the IDC, we generated expression heat maps based on both sense and antisense RKPM using Genesis (http://genome.tugraz.at/). Briefly, the order of genes was sorted based on the changes of sense RKPM (in log2 scale) across the 4 time points, and a heat map was generated using both sense and antisense RKPM while keeping the same gene order. The process was then repeated using antisense RKPM for gene sorting. The patterns of these heat maps were then visually inspected as described in the main text. To identify transcripts that are developmentally regulated, we used EdgeR 51 to screen for genes that are differentially expressed across the IDC. Briefly, each gene was assigned to have a sense and antisense transcript of the same length (i.e. coding region), which were then treated as independent transcripts in the EdgeR analyses. The read counts on these transcripts among the 4 time points were paired into 6 non-redundant pairs and compared using EdgeR with the biological coefficient of variation set to 0.6. We then used the exact test for determining differential expression and transcripts at false discovery rate (FDR) < 0.05 were considered to be differentially expressed. A transcript is considered developmentally regulated if it is differentially expressed in at least 2 of the 6 non-redundant time point pair comparisons.

To determine the prevalence of natural antisense transcripts, we pooled the reads from all 4 time-points (i.e. library 1 to 4) and generated a strand-specific read coverage map. Based on this map, we scanned for transcribed fragments (transfrags) longer than 150 nt covered by at least 2 fold-coverage per nt with a maximum 10nt of coverage gap. A transfrag will be split into shorter transfrags if a dramatic difference in coverage is detected within a sliding window (≥100 fold difference between two halves of a 20 nt window). An antisense transfrag is then defined as a transfrag that is overlapping with the CDS of a gene in the opposite strand. A significant antisense transfrag is defined based on the following 4 criteria 1) percentage of ORF covered by the transfrag: ≥10% (stringent criteria) or ≥5% (relaxed criteria), 2) average coverage depth of the transfrag: ≥10 (stringent) or ≥3 (relaxed), 3) antisense read count: ≥50 (stringent) or ≥15 (relaxed), and 4) sense to antisense read ratio: <200 (stringent) or <2000 (relaxed). A significant sense transfrag was defined in a similar way except for criteria 4). To investigate the overall coverage pattern of antisense (or sense) transfrags (at stringent cutoffs), we divided the CDSs into 5 equal bin regions and recorded the overlap of antisense (or sense) transfrags within these bin regions. The percentage of genes being covered by antisense (or sense) transfrags in these bin regions was plotted (as shown in Figure  4A).

We reasoned if an antisense transcript observed in a gene was the consequence of the transcription activities from its downstream gene through a bi-directional promoter (CDSs are in ‘Tail-to-Head’ orientation) or run-through transcription (CDSs are in ‘Tail-to-Tail’ orientation), changes in expression level of the upstream antisense transcript across IDC should be positively correlated to that of the downstream mRNA. We therefore investigated the correlation between the changes in antisense transcript level of the upstream gene (upStrmA RPKM) and the changes in sense transcript (i.e. mRNA) level of the downstream gene (dnStrmS RPKM) across the 4 IDC time-points. Briefly, we calculated the fold changes of both sense and antisense RPKM for each gene among the 4 time points (i.e. 6 non-redundant pairs of time points). A valid comparison requires either 1) both RPKM values are ≥0.2 in both time points (i.e. quantifiable), or 2) one RPKM value is ≥2 if the other RPKM value is <0.2 (i.e. unquantifiably large). In a comparison, if both upStrmA RPKM and dnStrmS RPKM were changed in the same direction at ≥1.5 fold, this comparison is scored as “positive”, and alternatively, as “negative” if the change was in opposite direction, or otherwise, as “not correlated”. A comparison could also be scored as “no data” if the RPKM values were lower than the mentioned cutoff. Based on the scores of these 6 comparisons, the correlation of upStrmA RPKM and dnStrmS RPKM in a gene pair is said to be “positive (strong)”, “positive (medium)” or “positive (weak)” if “positive” is scored in ≥3, ≥2 and ≥1 of the 6 comparisons and no “negative” was scored. Negative correlations were defined in the same manner but in the opposite direction. Correlations of a gene pair were defined as “not correlated” in other scenarios, e.g. both “positive” and “negative” were scored, or as “no data” if all 6 scores were “no data”. Detailed results can be found in Additional file 6: Table S3. Finally, strength of evidence supporting a bi-directional promoter affecting its upstream antisense transcript was defined as strong, medium and weak when the correlation of a Tail-to-Head gene pair is “positive (strong)”, “positive (medium)” and “positive (weak)”, respectively. Strength of evidence supporting run-through transcription affecting the upstream antisense transcript is defined in the same manner except that Tail-to-Tail gene pairs were considered (Additional file 6: Table S3). It should be noted that 181 coding genes were not analyzed because they do not have an annotated coding gene at their 3′-end, e.g. located at chromosome end, or next to pseudogene, tRNA or rRNA, etc. Totally, 2774 Tail-to-Tail and 2329 Tail-to-Head pairs were analyzed.

Interpretation of differences in alignment patterns observed for uniquely and non-uniquely mapped reads to repetitive regions of the genome is difficult, if the repetitiveness of these regions was unknown. We therefore assessed the repetitiveness of various genomic regions by simulation. Briefly, we randomly extracted 10 million ‘fragments’ from both strands of the whole genome (mean 200 nt with standard deviation 50 nt). Then, we extracted the first 60 nt from these simulated ‘fragments’ as simulated ‘reads’. These simulated reads were mapped to the genome using Bowtie with zero mismatch allowed and maximum alignment hit of 100. Then, we defined the repetitiveness of a position in the genome as the averaged number of alignment hits (i.e. NH attribute of SAM file specification) of all the reads mapped to this position. By definition, repetitiveness of 1 means all reads mapped to this position are uniquely mapped reads. This ‘repetitiveness’ value was used in Figure  3.

The criteria used here are primarily based on Lee et al. 33 . Briefly, reads containing 15 or more consecutive “A” at their 3′-end were selected from all datasets and redundant reads within the same library were discarded. These non-redundant reads were pooled. These reads potentially contain the sequence of polyA tails. The A stretches at the end were trimmed and the reads with minimal 18 nt after trimming were mapped to the reference genome using Bowtie with parameters “-n 2 -k 1 -m 50 -1 30”. To distinguish polyA tracks of true polyadenylation from polyA tracks of internal polyA stretches on the mRNAs themselves (i.e. false positives), we analyzed the base compositions surrounding the end of the mapped reads and discarded those that might not represent true polyadenylation. Reads with the following properties were regarded as false positives and removed. 1) Reads with ≥ 5 nt immediate downstream of the end site are As; 2) Depending on the actual length of the polyA stretch of the read (e.g. N nt), reads with 70% of N nt downstream of the end site are As; 3) Reads with ≥ 6nt within 10 nt immediate upstream of the end site are As. The polyadenylation sites were then defined as the immediate downstream base of the reads. To ensure the identified polyadenylation site are not false positives derived from low quality base calls, reads with quality scores in any of the upstream and downstream 5 nt flanking the polyadenylation site less than 20 were further removed. These procedures should be able to remove false positives derived from internal polyA stretches and low quality base calls.

As most of the observed polyadenylation sites appear as clusters 34 , we grouped the polyA sites into clusters by allowing an optimal maximum intra-cluster distance (at 20 nt) between sites. A polyA cluster was then represented by the polyA site with the highest number of supporting reads (i.e. peak), and these peak positions were used in all downstream analyses. A polyA cluster is defined as valid when the number of reads at the peak position is ≥ 2. To assign polyA tails to mRNAs, we searched for polyA clusters within 2000 nt downstream of their stop codons on the same strand and recorded the size of the coverage gap between the polyA clusters and the stop codon. A polyA tail is defined as valid when the coverage gap is ≤ 30 nt.

All splicing junctions identified by HMMSplicer 37 were clustered as mentioned in a previous study 52 . A junction cluster is considered to be ‘antisense’ when its representative junction is located within the coding region of a gene on the opposite strand.

All sequence reads from this study have been submitted to the EMBL Nucleotide Sequence Database (EMBL-Bank) (http://www.ebi.ac.uk/embl/) under accession no. ERP001849.