Leaders in Pharmaceutical Business Intelligence (LPBI) Group

Similarity between alternative exons

We were interested in determining the feasibility of using the MinION nanopore sequencer to characterize the connectivity of distantly located exons in the mRNAs expressed from genes with complex splicing patterns. For the purposes of these experiments, we have focused on fourDrosophila genes with increasingly complex patterns of alternative splicing (Fig. 1). Resistant to dieldrin (Rdl) contains two clusters, each containing two mutually exclusive exons and therefore has the potential to generate four different isoforms (Fig. 1a). Multidrug-Resistance like Protein 1(MRP) contains two mutually exclusive exons in cluster 1 and eight mutually exclusive exons in cluster 2, and can generate 16 possible isoforms (Fig. 1b). Myosin heavy chain (Mhc) can potentially generate 180 isoforms due to five clusters of mutually exclusive exons – clusters 1 and 5 contain two exons, clusters 2 and 3 each contain three exons, and cluster 4 contains five exons. Finally, Dscam1 contains 12 exon 4 variants, 48 exon 6 variants, 33 exon 9 variants (Fig. 1d), and two exon 17 variants (not shown) and can potentially express 38,016 isoforms. For this study, however, we have focused only on the exon 3 through exon 10 region of Dscam1, which encompasses the 93 exon 4, 6, and 9 variants, and 19,008 potential isoforms (Fig. 1d).

Because our nanopore sequence analysis pipeline uses LAST to perform alignments [8], we aligned all of the Rdl, MRP, Mhc, and Dscam1 exons within each cluster to one another using LAST to determine the extent of discrimination needed to accurately assign nanopore reads to a specific exon variant. For Rdl, each variable exon was only aligned to itself, and not to the other exon in the same cluster (data not shown). For MRP, the two exons within cluster 1 only align to themselves, and though the eight variable exons in cluster 2 do align to other exons, there is sufficient specificity to accurately assign nanopore reads to individual exons (Fig. 2a). For Mhc, the variable exons in cluster 1 and cluster 5 do not align to other exons, and the variable exons in cluster 2, cluster 3, and cluster 4 again align with sufficient discrimination to identify the precise exon present in the nanopore reads (Fig. 2b). Finally, for Dscam1, the difference in the LAST alignment scores between the best alignment (each exon to itself) and the second, third, and fourth best alignments are sufficient to identify the Dscam1 exon variant (Fig. 2c). This analysis indicates that for each gene in this study, LAST alignment scores are sufficiently distinct to identify the variable exons present in each nanopore read.

Optimizing template switching in Dscam1 cDNA libraries

Template switching can occur frequently when libraries are prepared by PCR and can confound the interpretation of results [9], [10]. For example, CAM-Seq [11] and a similar method we independently developed called Triple-Read sequencing [12] to characterize Dscam1 isoforms, were found to have excessive template switching due to amplification during the library prep protocols. To assess template switching in our current study, we generated a spike-in mixture of in vitro transcribed RNAs representing six unique Dscam1 isoforms – Dscam1 ^{4.2,6.32,9.31} , Dscam1^{4.1,6.46,9.30} , Dscam1 ^4.3,6.33,9.9 , Dscam1 ^{4.12,6.44,9.32} , Dscam1 ^4.7,6.8,9.15 , and Dscam1 ^4.5,6.4,9.4. We used 10 pg of this control spike-in mixture and prepared libraries for MinION sequencing by amplifying the exon 3 through exon 10 region for 20, 25, or 30 cycles of RT-PCR. We then end-repaired and dA-tailed the fragments, ligated adapters, and sequenced the samples on a MinION (7.3) for 12 h each. We obtained 33,736, 8,961, and 7,511 base-called reads from the 20, 25, and 30 cycle libraries, respectively. Consistent with the size of the exon 3 to 10 cDNA fragment being 1,806–1,860 bp in length, depending on the precise combination of exons it contains, most reads we observed were in this size range (Fig. 3a). We used Poretools [13] to convert the raw output files into fasta format and then used LAST to align the reads to a LAST database containing each variable exon. From these alignments, we identified reads that mapped to all three exon clusters, as well as the exon with the best alignment score within each cluster. When examining the alignments to each cluster independently, we found that for these spike-in libraries, all reads mapped uniquely to the exons present in the input isoforms. Therefore, any observed isoforms that were not present in the input pool were a result of template switching during the RT-PCR and library prep protocol and not due to false alignments or sequencing errors.

When comparing the combinations of exons within each read to the input isoforms, we observed that 32 % of the reads from the 30 cycle library corresponded to isoforms generated by template switching (Fig. 3b). The template-switched isoforms observed by the greatest number of reads in the 30 cycle library were due to template switching between the two most frequently sequenced input isoforms. In most cases, template switching occurred somewhere within exon 7 or 8 and resulted in a change in exon 9. However, the extent of template switching was reduced to only 1 % in the libraries prepared using 25 cycles, and to 0.2 % in the libraries prepared using 20 cycles of PCR (Fig. 3b). Again, for these two libraries the most frequently sequenced template-switched isoforms involved the input isoforms that were also the most frequently sequenced. These experiments demonstrate that the MinION nanopore sequencer can be used to sequence ‘full length’ Dscam1 cDNAs with sufficient accuracy to identify isoforms and that the cDNA libraries can be prepared in a manner that results in a very small amount of template switching.

Dscam1 isoforms observed in adult heads

To explore the diversity of Dscam1 isoforms expressed in a biological sample, we prepared aDscam1 library from RNA isolated from D. melanogaster heads prepared from mixed male and female adults using 25 cycles of PCR and sequenced it for 12 h on the MinION nanopore sequencer obtaining a total of 159,948 reads of which 78,097 were template reads, 48,474 were complement reads, and 33,377 were 2D reads (Fig. 4a). We aligned the reads individually to the exon 4, 6, and 9 variants using LAST. A total of 28,971 reads could be uniquely or preferentially aligned to a single variant in all three clusters. For further analysis, we used all 16,419 2D read alignments and 31 1D reads when both template and complement aligned to same variant exons (not all reads with both a template and complement yield a 2D read). The remaining 12,521 aligned reads were 1D reads where there was either only a template or complement read, or when the template and complement reads disagreed with one another and were therefore not used further. We observed 92 of the 93 potential exon 4, 6, or 9 variants – only exon 6.11 was not observed in any read (Fig. 4f). To assess the accuracy of the results we performed RT-PCR using primers in the flanking constitutive exons that contained Illumina sequencing primers to separately amplify the Dscam1exon 4, 6, and 9 clusters from the same RNA used to prepare the MinION libraries, and sequenced the amplicons on an Illumina MiSeq. The frequency of variable exon use in each cluster was extremely consistent between the two methods (R ²  = 0.95, Fig. 5a).

Over their entire lengths, the 2D reads that map specifically to one exon 4, 6, and 9 variants map with an average 90.37 % identity and an average LAST score of approximately 1,200 (Fig. 5b). The 16,450 full length reads correspond to 7,874 unique isoforms, or 42 % of the 18,612 possible isoforms given the exon 4, 6, and 9 variants observed. We note, however, that while 4,385 isoforms were represented by more than one read, 3,516 of isoforms were represented by only one read indicating that the depth of sequencing has not reached saturation (Fig. 4b and c). This was further confirmed by performing a bootstrapped subsampling analysis (Fig. 4d) and by using the capture-recapture method to attempt to assess the complexity of isoforms present in the library (Fig. 4e), which suggests that over 11,000 isoforms are likely to be present, though even this analysis has not yet reached saturation. The most frequently observed isoforms were Dscam1^{4.1,6.12,9.30} and Dscam1 ^4.1,6.1,9.30 which were observed with 30 and 25 reads, respectively (Fig. 4e). In conclusion, these results demonstrate the practical application of using the MinION nanopore sequencer to identify thousands of distinct Dscam1 isoforms in a single biological sample.

Nanopore sequencing of ‘full-length’ Rdl, MRP, and Mhc isoforms

To extend this approach to other genes with complex splicing patterns, we focused on Rdl, MRP, and Mhc which have the potential to generate four, 16, and 180 isoforms, respectively. We prepared libraries for each of these genes by RT-PCR using primers in the constitutive exons flanking the most distal alternative exons using 25 cycles of PCR, pooled the three libraries and sequenced them together on the MinION nanopore sequencer for 12 h obtaining a total of 22,962 reads. The input libraries for Rdl, MRP, and Mhc were 567 bp, 1,769-1,772 bp, and 3,824 bp, respectively. The raw reads were aligned independently to LAST indexes of each cluster of variable exons. The alignment results were then used to assign reads to their respective libraries, identify reads that mapped to all variable exon clusters for each gene, and the exon with the best alignment score within each cluster. In total, we obtained 301, 337, and 112 full length reads forRdl (Fig. 6), MRP (Fig. 7), and Mhc (Fig. 8), respectively. For Rdl, both variable exons in each cluster was observed, and accordingly all four possible isoforms were observed, though in each case the first exon was observed at a much higher frequency than the second exon (Fig. 6d). Interestingly, the ratio of isoforms containing the first versus second exon in the second cluster is similar for isoforms containing either the first exon or the second exon in the first cluster indicating that the splicing of these two clusters may be independent. For MRP, both exons in the first cluster were observed and all but one of the exons in the second cluster (exon B) were observed, though the frequency at which the exons in both clusters were used varied dramatically (Fig. 7d). For example, within the first cluster, exon B was observed 333 times while exon A was observed only four times. Similarly, in the second cluster, exon A was observed 157 times whereas exons B, E, F, and G were observed 0 times, thrice, once, and twice, respectively, and exons D, E, and H were observed between 40 and 76 times. As a result, we observed only nine MRP isoforms. For Mhc, we again observed strong biases in the exons observed in each of the five clusters (Fig. 8d). In the first cluster, exon B was observed more frequently than exon A. In the second cluster, 109 of the reads corresponded to exon A, while exons B and C were observed by only two and one read, respectively. In the third cluster, exon A was not observed at all while exons B and C were observed in roughly 80 % and 20 % of reads, respectively. In the fourth cluster, exon A was observed only once, exons B and C were not observed at all, exon E was observed 13 times while exon D was present in all of the remaining reads. Finally, in the fifth cluster, only exon B was observed. As with MRP, these strong biases and near or complete absences of exons in some of the clusters severely reduces the number of possible isoforms that can be observed. In fact, of the 180 potential isoforms encoded by Mhc, we observed only 12 isoforms. Various Mhc isoforms are known to be expressed in striking spatial and temporally restricted patterns [14] and thus it is likely that other Mhc isoforms that we did not observe, could be observed by sequencing other tissue samples.

Here we have demonstrated that nanopore sequencing with the Oxford Nanopore MinION can be used to easily determine the connectivity of exons in a single transcript, including Dscam1, the most complicated alternatively spliced gene known in nature. This is an important advance for several reasons. First, because short-read sequence data cannot be used to conclusively determine which exons are present in the same RNA molecule, especially for complex alternatively spliced genes, long-read sequence data are necessary to fully characterize the transcript structure and exon connectivity of eukaryotic transcriptomes. Second, although the Pacific Bioscience platform can perform long-read sequencing, there are several differences between it and the Oxford Nanopore MinION that could cause users to choose one platform over the other. In general, the quality of the sequence generated by the Pacific Bioscience is higher than that currently generated by the Oxford Nanopore MinION. This is largely due to the fact that each molecule is sequenced multiple times on the Pacific Bioscience platform yielding a high quality consensus sequence whereas on the Oxford Nanopore MinION, each molecule is sequenced at most twice (in the template and complement). We have previously used the Pacific Bioscience platform to characterize Dscam1 isoforms and found that it works well, though due to the large amount of cDNA needed to generate the libraries, many cycles of PCR are necessary and we observed an extensive amount of template switching, making it impractical to use for these experiments (BRG, unpublished data). However, over the past year that we have been involved in the MAP, the quality of sequence has steadily increased. As this trend is likely to continue, the difference in sequence quality between these two platforms is almost certain to shrink. Nonetheless, as we demonstrate, the current quality of the data is more than sufficient to allow us to accurately distinguish between highly similar alternatively spliced isoforms of the most complex gene in nature. Third, the ability to accurately characterize alternatively spliced transcripts with the Oxford Nanopore MinION makes this technology accessible to a much broader range of researchers than was previously possible. This is in part due to the fact that, in contrast to all other sequencing platforms, very little capital expense is needed to acquire the sequencer. Moreover, the MinION is truly a portable sequencer that could literally be used in the field (provided one has access to an Internet connection), and due to its size, almost no laboratory space is required for its use.

Although nanopore sequencing has many exciting and potentially disruptive advantages, there are several areas in which improvement is needed. First, although we were able to accurately identify over 7,000 Dscam1 isoforms with an average identity of full-length alignments >90 %, there are several situations in which this level of accuracy will be insufficient to determine transcript structure. For instance, there are many micro-exons in the human genome [15], and these exons would be difficult to identify if they overlapped a portion of a read that contained errors. Additionally, small unannotated exons could be difficult to identify for similar reasons. Second, the current number of usable reads is lower than that which will be required to perform whole transcriptome analysis. One issue that plagues transcriptome studies is that the majority of the sequence generated comes from the most abundant transcripts. Thus, with the current throughput, numerous runs would be needed to generate a sufficient number of reads necessary to sample transcripts expressed at a low level. In fact, this is one reason that we chose in this study, to begin by targeting specific genes rather than attempting to sequence the entire transcriptome. We do note, however, that over the past year of our participation in the MAP, the throughput of the Oxford Nanopore MinION has increased, and it is reasonable to expect additional improvements in throughput that should make it possible to generate a sufficient number of long reads to deeply interrogate even the most complex transcriptome.

In conclusion, we anticipate that nanopore sequencing of whole transcriptomes, rather than targeted genes as we have performed here, will be a rapid and powerful approach for characterizing isoforms, especially with improvements in the throughput and accuracy of the technology, and the simplification and/or elimination of the time-consuming library preparations.