[sequencingfan]

I also have a question.what about non template adding nucleotides by reverse transcriptase.is there possibility that Rt add no 3 C but for example CCG.Is there possibility to use degenerate TSO?

[HeidelbergScience]

This is a very nice protocol, thanks for sharing.
I have two questions about the 5' ends of RNA libraries:
1. Why is there a bias in the 5' towards G (and not A)
2. Are there always 3 G's at the 5' end or there might be more which will add additional G's to the 5' end of the read? (this can be an answer to Q1). If I need to know exactly where the read (or RNA fragment) begins, is this protocol sensitive enough?
Thanks

The bias to G-ending RNA templates is likely due to the fact that the template switching is facilitated when RT-product has terminal C. However, DNA templates have no 5-end bias, and, therefore, there could be other explanations (e.g. Mg2+ fragmentation and RNAses digestion produce more 5'-G... RNA fragments and less 5-A.... RNA).

The number of G is always 3 (evident from Sanger-seq of the libraries from synthetic control RNA in the publication), and it is dictated by the 3xrG in the 3'-end of the TSO. When TSO has 4xrG, the products have 4 G (we did not put it in the paper though). Therefore, the protocol gives exact information when the RNA template starts.

[Asaf]

The bias to G-ending RNA templates is likely due to the fact that the template switching is facilitated when RT-product has terminal C. However, DNA templates have no 5-end bias, and, therefore, there could be other explanations (e.g. Mg2+ fragmentation and RNAses digestion produce more 5'-G... RNA fragments and less 5-A.... RNA).

The number of G is always 3 (evident from Sanger-seq of the libraries from synthetic control RNA in the publication), and it is dictated by the 3xrG in the 3'-end of the TSO. When TSO has 4xrG, the products have 4 G (we did not put it in the paper though). Therefore, the protocol gives exact information when the RNA template starts.

Thanks! We'll use this protocol for our next RNA library construction.

[HeidelbergScience]

I also have a question.what about non template adding nucleotides by reverse transcriptase.is there possibility that Rt add no 3 C but for example CCG.Is there possibility to use degenerate TSO?

Under conditions of the described protocol (e.g. no Mn2+ ions), the terminal transferase activity of the RT enzyme is very limited. Most likely, it only adds 1 nucleotide to the 3'-terminus of the first cDNA strand before template switch occurs. Enhancing TT activity of the RT (e.g. by adding Mn2+) produces mostly "empty" cDNAs libraries, because RT starts to tail also the poly(dT) primer.

We are planning to test degenerative TSO, but they are likely to be inefficient, since RT adds predominantly Cs, and with much lower probability G, A,and Ts.

[HeidelbergScience]

An important update added to the protocol.
Incubation with TSO for 2 hours (instead of 15 min) increases the yield of the library 5-10 times.

[Asaf]

Some insights from our test nextseq run.
We took some RNA (8 RNAseq libraries), prepared the libraries after RiboZero, fragmentation and size selection. The process was indeed shorter (1.5 days) and required less RNA.
The nextseq results were disappointing, the multiplexing barcodes (i7) weren't read well for some reason, most of the barcodes were just poly-A. In addition, most of the reads contained poly-A which in some point in the read were changed to poly-G. I think that these issues are related to the fact that nextseq uses 2-colors system, A is both and G is neither and using some sort of normalization the software should determine which signal is too low to be an A and calls it as a G. Another issue that might arise is that there is a bias towards G in the first base of the read, this means that nextseq will have difficulties determining clusters (again - no color, and variability) which might lead to a small number of clusters (we got 100M reads instead of 400M, maybe it's this but maybe error in quantification).
We are planning on doing another round, this time we will have the barcodes on the adapter before the GGG, we will use barcodes in different lengths so there won't be a G read on the entire flowcell, and use standard Illumina primers.
Any more suggestions? Had anyone else tried this method with nextseq?

[Simone78]

Even though there are some original and interesting ideas, I want to make several comments on things that are either not convincing or plainly wrong. Sorry, there is no order in that. I just wrote them down while reading the paper

1 - you mention that Superscript II, SMARTscribe and SMART RT are the only ones giving detectable amounts of cDNA. Well, Superscript II and SMARTscribe are probably the same thing as we already showed in the Smart-seq2 paper (Picelli et al., Nat Methods 2013, Suppl Info).
2 - When using Superscript III, 4 more pre-ampl cycles were needed. Also this is well documented in the literature (and also described in the Smart-seq2 paper). Mutations in Superscript II and III are different and the III has a negligible strand-switch activity. Nothing new here.
3 - the method is not of any use for short circulating RNA given the huge amount of unmappable reads, as reported in the “Result” section.
4 - Interesting observation about the bias in the template switching efficiency. You speculate that a TSO with different or modified 3´terminal could solve the problem. We already tested LNA-based TSO with “N” in 1, 2 or all of the 3´terminal bases (again, see Suppl Info). It turned out that 3 LNA-G bases are the best solution. If you don´t use LNA bases, then 3 rG are the best. Clontech had already thought about it, obviously!
5 - Throughout the protocol it is repeatedly stated that one should cut the agarose and extract the samples ready for seq. This is a terrible (terrible!) troublesome, time-consuming, inefficient and non-scalable way of doing library prep. If you really want to do size selection why not using E-gels? Or PippinPrep if available?
6 - Even though this method claims that only picograms of material are needed, you need hundreds ng/few ug of tot RNA to start with, due to losses in extraction, column purification, fragmentation, purification again…And then it can sequence all the RNA species…yes, but if I would simply do a Ribozero treatment (or a “home-brewed” version of it. There are some around) on as little as 10 ng + SMARTer (or SMART-seq2) I would achieve the same result with the same or less effort!
7 - You always compared your method to others (for single cell seq) present on the market repeatedly saying that they are expensive and that they rely on an inefficient adaptor ligation step and so on (page 824). There are several inaccuracies in this statement. First, you can´t compare your method to those because yours is not designed for single cells. Second, it´s absolutely not true that all the methods rely on the (classic, ligase-based) adaptor ligation. Nextera (Illumina) and our recent method (Picelli et al., Genome Res 2014) don´t and they are efficient with ng (Nextera) or sub-pg (ours but also Adey et al., 2010) amounts of input DNA. And in the “Discussion” at the end of page 825 you refer to Ramsköld et al. (Nat Biotech 2012, ref #30) which is NOT based on any ligase but on the first Nextera kit from Illumina! Besides the cost of Smart-seq2+home-made Tn5 is 10-15 euros, comparable to the cost for your library prep.
8 - in the discussion on page 825 in the same sentence there are several inaccuracies (a record!). It says that “(Nextera) requires at least 50 ng of DNA and, apparently, is restricted to the long DNA molecules. The full capacity of the tagmentation technique for DNA library prep is yet to be tested and compared with other methods”. First: the DNA Nextera XT kit is specifically designed to start from 1 ng input DNA (the old Nextera kit was using 50 ng). We have also shown in Picelli et al., Gen Res 2014 that you need as little as 0.1 pg DNA, but you should have just read Adey et al. 2010. Second: the tagmentation is NOT restricted to long molecules. In fact Adey and Shendure (the original Genome Biol 2010 paper where they describe the method) say that molecules as short as ca. 35 bp can be tagmented. Third, in the same paper they also compare the method to other standard fragmentation methods, showing that the Tn5 has just a weak preference for cutting the DNA at specific sites. Additionally there are also other papers on bisulfite-converted DNA prepped with the Tn5 (Adey et al. Genome Res 2012)…and even one from your Institution (!!!), Wang et al (Nat Prot 2013). So the Tn5-based approach is a viable option for exactly everything you claim not be good for.
9 - You also state (claim 2 at the end of page 824) that there are no reports of “strand-specific mRNA transcriptome from 1 ng of polyA enriched RNA”, which is obviously inaccurate. Clontech has a protocol for FFPE samples where it couples Ribozero to a stranded SMARTer protocol and starts from as little as 10 ng of degraded TOTAL RNA (designed for FFPE samples), with no column purifications, fragmentation or other preparation steps needed. And 10 ng of tot RNA are in the same order of magnitude as 1 ng of polyA RNA you use in the paper.
10 - Regarding the circulating DNA in the “Discussion” section. It is stated that the Thruplex kit (Rubicon Gen) is not capable to generate libraries from lower quantity of DNA. In principle this is correct, but you forgot to mention that the Picoplex (same company) allows the sequencing of even single CTCs.
11 - You also claim that your method is better for sequencing circulating DNA compared to Tam-Seq because you sequence the whole genome and not only selected loci. However, the sotry is not that simple: I cite from the original Tam-Seq paper (Forshew et al., Sci Transl Med, 2012): “This generates a large amount of data on genomic regions that do not, at present, inform clinical decisions. Moreover, the depth of coverage for clinically significant loci is not sufficient to detect changes that occur at low frequency (<5%). Such approaches have recently been complemented by methods for examination of individual amplicons at great depth”. If just sequencing the whole genome would be that informative and cheap don´t you think it would have already been done?
12 - It says in a previous post that “Under conditions of the described protocol (e.g. no Mn2+ ions), the terminal transferase activity of the RT enzyme is very limited. Most likely, it only adds 1 nucleotide to the 3'-terminus of the first cDNA strand before template switch occurs”. Sorry again, but we have reported (Nat Meth 2013, Suppl Info) that manganese chloride is NOT necessary for the template switch reaction to occur. Besides, even in the original SMARTer paper (Zhu et al., Biotechniques 2001) manganese chloride was not even mentioned.

[luc]

Hello Simone,
thanks for pointing me to your new transposase paper and for the interesting details in your post. Obviously the transposase will allow to start library prep from even lower RNA amounts and thus should likely be the method of choice for ultra-low input applications. Nevertheless the biases of the enzyme ( e.g. http://omicfrontiers.com/2013/07/04/...biased-genome/ [ please note the "pcr-free" in this comparison which is not realistic for low input RNA-seq]) and the usually wide insert size range of the libraries lead me to avoid it whenever possible. Thus refinements to the template-switching protocols remain of great interest to me.

[Simone78]

Hello Simone,
thanks for pointing me to your new transposase paper and for the interesting details in your post. Obviously the transposase will allow to start library prep from even lower RNA amounts and thus should likely be the method of choice for ultra-low input applications. Nevertheless the biases of the enzyme ( e.g. http://omicfrontiers.com/2013/07/04/...biased-genome/ [ please note the "pcr-free" in this comparison which is not realistic for low input RNA-seq]) and the usually wide insert size range of the libraries lead me to avoid it whenever possible. Thus refinements to the template-switching protocols remain of great interest to me.

thanks for the link, very interesting read! Unfortunately, the bias introduces by the transposase is not the only one. We used KAPA HiFi DNA Pol both in the Smart-seq2 and the transposase papers in order to reduce the PCR bias (based on what reported by Quail et al., Nat Meth 2012), but when you work with single cells we need 2 rounds of PCR, which means that the bias just gets bigger.
One thing I wanted to point out is that in the post you linked it says "The flaw in our analysis (of course) is that we should have used the Illumina Nextera XT kit for these samples as these are optimized for smaller genomes to minimise excessive numbers of smaller fragments". I believe, as I already said elsewhere in Seqanswer, that the enzyme is exactly the same for the standard and the XT kits (as I joke I always say that Illumina just bought Epicentre, repacked their enzyme and started selling for a few thousand USD but they didn´t make any effort to improve/change it) . What is different is just the buffer, as we also showed in our paper. Therefore, you need a two-buffer system to work with inputs that span from sub-pg to tens of ng DNA. Besides, varying the amount of PEG and the amount of enzyme will enable a better control of the size of the fragments (thus avoiding over-fragmentation). If you believe in coincidences (I don´t) the "long-fragments" and "short fragments" buffers that were used in the old Epicentre kit are the same used now in the Nextera and Nextera XT, respectively.

[HeidelbergScience]

Some insights from our test nextseq run.
We took some RNA (8 RNAseq libraries), prepared the libraries after RiboZero, fragmentation and size selection. The process was indeed shorter (1.5 days) and required less RNA.
The nextseq results were disappointing, the multiplexing barcodes (i7) weren't read well for some reason, most of the barcodes were just poly-A. In addition, most of the reads contained poly-A which in some point in the read were changed to poly-G. I think that these issues are related to the fact that nextseq uses 2-colors system, A is both and G is neither and using some sort of normalization the software should determine which signal is too low to be an A and calls it as a G. Another issue that might arise is that there is a bias towards G in the first base of the read, this means that nextseq will have difficulties determining clusters (again - no color, and variability) which might lead to a small number of clusters (we got 100M reads instead of 400M, maybe it's this but maybe error in quantification).
We are planning on doing another round, this time we will have the barcodes on the adapter before the GGG, we will use barcodes in different lengths so there won't be a G read on the entire flowcell, and use standard Illumina primers.
Any more suggestions? Had anyone else tried this method with nextseq?

Many thanks for your feedback.
Albeit we never run HiSeq after CATS (only MiSeqs), we have observed the similar problem with the P7 barcode reads. So far we made only 2x MiSeq runs using P7 barcodes (one with 8 and one with 4 different barcodes). So, the one with 8 had barcodes messed up, however, the run with 4 worked well. Also, we noticed that the quality of the Index read was much lower as compared to Read1. We first though this is the problem of our primers or MiSeq, but after your post we suspect that the proximity of 30xdA-tail to the P7 index might have caused it. However, we can not find any feasible explanation of how the polyA-tail could interfere. The best option would be to introduce the bar-codes into TSO oligo before the GGG together with making them of different length. However, one can also try shorter 20xdA tails in the reverse primer, or "dilute" the tail with dG and dC nucleotides between each 10xdT (so RT primer wold be XXXXXttttttttttgttttttttttcttttttttttV). We will test those strategies and update the protocol.

The fact that most of the reads contained poly(A) is normal, as long as they are not “empty”. You need to use the trimming algorithm to remove the polyA before mapping of course. The only way to evade the poly(A) tails after each read is either by using longer RNA templates or shorter Read1 length. We never cared about it, since the presence of polyA does not bring any trouble. We did not observe changing of A to G though (at least at the majority of the reads), but again we never run HiSeq.

Regarding clusters. Despite the fact that in RNA runs 80% of the reads starts with G (due to the bias), we have only a slight drop in number of clusters passing the filter as compared to DNA runs (which does not have G bias). For example with DNA runs we usually see 90% of clusters passed the filter, while for Mg2+ fragmented RNA it is about 80%. It might happen that for HiSeq the G bias creates bigger problem. Actually, 400M reads is a “theoretical” maximum which you can achieve on Hi200 and it is for paired-end reads (for single read it it 200 respectively). I guess you used single reads, right? So, 100M is definitely ok, if you did not have too many clusters. Could you specify what was the cluster density and how many clusters has passed the filter? This info would help to determine whether G bias interfered in the quality of Read1.

Also, you mentioned that the process took 1,5 days. Did it include the initial preparation of all solutions, primers, Ampipure/Gel isolation and final QC? What was the hands-on time of the library prep? Also, how much DNA did you use?
Again many thanks for the feedback, it was very helpful.

[HeidelbergScience]

Even though there are some original and interesting ideas, I want to make several comments on things that are either not convincing or plainly wrong. Sorry, there is no order in that. I just wrote them down while reading the paper

Thank you for investing your time and commenting on our article. We appreciate your effort and think that it would be fair from our side to answer to your comments here. We apologize for not responding earlier since your comment was posted during the Christmas holidays and we have been on vacation until today.
Please find below a point-by-point response. We hope that this information would be helpful for the readers to better perceive the important advantages and fundamental differences between current NGS library prep methods.

[HeidelbergScience]

1 - you mention that Superscript II, SMARTscribe and SMART RT are the only ones giving detectable amounts of cDNA. Well, Superscript II and SMARTscribe are probably the same thing as we already showed in the Smart-seq2 paper (Picelli et al., Nat Methods 2013, Suppl Info).
2 - When using Superscript III, 4 more pre-ampl cycles were needed. Also this is well documented in the literature (and also described in the Smart-seq2 paper). Mutations in Superscript II and III are different and the III has a negligible strand-switch activity. Nothing new here.

The goal of this manuscript and the described experiments was not to claim “a new discovery” of different template switching (TS) capacities of MMLV RT mutants. The experiment was important to demonstrate which commercial MMLV RTs can be used for the TS in CATS protocol. We tested the 6 most widely used commercial RT enzymes, but it was neither the goal nor a topic of our publication to extensively research on all of their properties. We also cannot agree that the phenomenon of different TS capacity is well documented in the literature. For most commercial RT enzymes this information is simply not present.

Secondly, in the main text and supplementary material in [Picelli et al, Nat Met // 2013] we did not find any written comments on the different TS capacities of various RT enzymes. Albeit from the supplementary table it is indeed possible to infer that both Smartscribe and SSRTII secures much higher yield of final DNA library as compared to SSRTIII, there is no information about other three enzymes tested in our paper.

[HeidelbergScience]

3 - the method is not of any use for short circulating RNA given the huge amount of unmappable reads, as reported in the “Result” section.

Thanks for giving us an opportunity to comment on that. The library prepared from 100pg of plasma RNA which were shown in the paper represents the true circulating RNA content in the sample. Relatively high % of un-mappable reads in this run could be caused by significant % of non-human RNA in plasma or/and using sub-optimal RNA mapper.

However, we have extensively addressed the problem in the meantime and currently our plasma RNA-seq runs yield much better mapping statistics towards human genome/transcriptome (e.g. >80% can be unambiguously mapped to human genome using an alternative RNA mapping database) and indicates that the plasma RNA runs described in the manuscript were not yet an accurate representation of the full capacity of the technique.

The CATS protocol itself does not create any “visible” amounts of irrelevant libraries/by-products. Thus, libraries prepared from only 5pg of synthetic cel-miR-39 consisted only of fragments carrying cel-miR-39 as was evident from “clean” Sanger-seq chromatograms (Fig.2 in the paper). Moreover, there was no signal in negative controls prepared without adding nucleic acids templates.

[HeidelbergScience]

4 - Interesting observation about the bias in the template switching efficiency. You speculate that a TSO with different or modified 3´terminal could solve the problem. We already tested LNA-based TSO with “N” in 1, 2 or all of the 3´terminal bases (again, see Suppl Info). It turned out that 3 LNA-G bases are the best solution. If you don´t use LNA bases, then 3 rG are the best. Clontech had already thought about it, obviously!

The bias occurred only in RNA, but not DNA templates. Therefore we cannot be 100% sure whether it derived from the template switching preference to 5’-rG, or a bias towards generation of 5’-rG templates after Mg2+ RNA fragmentation. We will address it in our follow-up work, and also test if degenerative bases reduce this bias.

Also, in [Picelli et al, 2013] maintext we found one solid statement that “exchanging only a single guanylate for a locked nucleic acid (LNA)11 guanylate at the TSO 3′ end (rGrG+G) led to a twofold increase in cDNA yield relative to that obtained with the SMARTer IIA oligo”. Could you tell how big was the drop in library yield when you used a N-base at the end, as you stated above?

[HeidelbergScience]

5 - Throughout the protocol it is repeatedly stated that one should cut the agarose and extract the samples ready for seq. This is a terrible (terrible!) troublesome, time-consuming, inefficient and non-scalable way of doing library prep. If you really want to do size selection why not using E-gels? Or PippinPrep if available?

In fact, gel purification step is completely optional. Thus, in Fig2 of our paper Sanger-seq and Bioanalyser demonstrate that the purity of CATS libraries after the column purification and gel extraction step are equal (except the remaining of the pre-amp primers which Qiaquick columns cannot efficiently remove). So you need only to remove pre-amp primers by magnetic beads (e.g. Ampipure) before NGS and can completely skip the gel-extraction step.

We usually cut our libraries from E-gels, as it is - in our opinion - more convenient than magnetic bead or column purification, and also gives us the information about the library peaks distribution. This also enables to skip the – in our opinion - inconvenient Bioanalyser step and only use Qubit after library purification from E-gel. Of course, any researcher should decide on his own whether to purify or which purification method is best suited to his experimental needs or preferences concerning time-efficiency, scalability, etc.

In fact, CATS is the only method which allows NGS of small (<5ng) amounts of short RNA(DNA) without gel-purification. All other methods that we are aware of are associated with ligation of adaptors and thus inevitably yield high percentages of “empty libraries” and self-ligated adaptors which cannot be separated by magnetic beads from template containing libraries.