I agree with the triplicate approach. We must consider that one library with 3x depth is not same as 3 library with 1x depth. Yet, going very deep without the biological replicates can risk the results. As in transcriptomics, expression is very critical and any artefact could cause flaws in results. Medium depth sequencing in triplicate give us more confidence in results.

I work in a Translation Molecular Genetics lab, where we have patients with synonymous/splicing affecting variants. Not exactly your case but the similarity is in terms of 'quest of rare transcripts'. We search truncated or alternatively spliced transcripts. Here we rely on high depth, rather than having replicates. Since you are looking for rare transcripts, high depth can discover them, while less depth might miss them. In our case 20M total reads serves the purpose. I would advice to try initially with 1X depth and see if you observe your transcripts of interest and then decide your approach. Read depth would vary depending on how infrequent your transcript is!!

As per the suggestions provided by Abhijeet & Sushree previously, it would be vital to strike a balance between the 2 methods. An optimal approach would be to do a triplicate (as Abhijeet mentioned) for sample comparison & statistical validation, but at a moderately high depth (20M - 25M reads) per sample. However, what is more crucial here is the interpretation of the results - sometimes, a rare transcript might be incorrectly dismissed as an artefact (simply because it occurs in only 1 sample out of the three), when (in fact) this is not the case.

Thank you all for your responses to my question (@Abhijeet Singh, @Sushree Sangita Sahoo, @Shiraz Sk).

I do see the value in replication, as well as the caution to not consider rare transcripts detected only once in three replicates as artefacts (ht @Shiraz SK).

However, I still think there can be further explanation that can help this issue.

Specifically, (@Abhijeet Singh), you say "We must consider that one library with 3x depth is not same as 3 library with 1x depth", and I am not sure why this is the case, aside from the benefit of technical replication.

This gets to the core of my question, is there a theoretical reason why

"one library with 3x depth is not the same as 3 libraries with 1x depth" ?

I remember hearing that within a library, one has to 'sequence through the high copy reads to get to the low copy genes', and so am I making an error to think that I can just pile three replicate libraries together in silico and get the same as one library sequenced at the 3x depth?

Thank you!

"Am I making an error to think that I can just pile three replicate libraries together in silico and get the same as one library sequenced at the 3x depth?"

Yes, you're making an error. The core issue is the small number of copies of the rare transcript present in each sample. This becomes a mathematical probability problem: if there is 1 red ball in a bag with 1000 green balls, what are the odds of grabbing the red ball if you pull out 10 balls? 100? 500?

This is analogous to sequencing depth. Say there are only 10 copies of your rare transcript present in each library. You would need to sequence a given library at great depth to be statistically likely to find at least one of those copies. If you do "medium depth" sequencing on three libraries, you will probably miss those 10 copies in all three of your samples.

@Jennifer Hardee: This is not necessarily so. Consider the following analogy of your example, where there is 1 rare transcript in a library with 9999 "non-rare" transcripts - if the original sample size (a single library containing 10,000 transcripts) was used, you would likely need a sequencing depth of ~10k reads (assuming 1 read/transcript) before you could sequence the rare transcript (on a probabilistic basis), which would imply a "greater depth" of sequencing (as you were referring to).

However, if the original library of 10,000 transcripts was split into 10 separate libraries (each library containing 1,000 transcripts) and each library is then sequenced individually, the maximum depth of "sequencing" in this case would be only 1,000 reads for each library (assuming 1 read/transcript as previously); any additional depth would only be redundant in this respect. Thus, performing a lower depth of sequencing on multiple mutuallly exclusive libraries (derived from the same principal library) would yield a similar accuracy as doing high depth sequencing on the primary library alone.

I agree with what @Jennifer Hardee. This is a sampling problem. By sampling less (i.e. using low seq-depth) you'll grab rare transcripts less likely because most of the reads will be "taken away" from the non-rare ones. The sampling depth VS replicate number issue has been described in this nice paper.

Liu, Yuwen, Jie Zhou, and Kevin P. White. "RNA-seq differential expression studies: more sequence or more replication?." Bioinformatics 30.3 (2013): 301-304.

It really boils down to what is your end goal. If you just want to understand if a transcript is present or not (presence/absence) having more depth is the best strategy. On the other end, if this is just an intermediate step and you end goal is to perform a differential expression analysis you'll need to have more biological replicates.

@Marco Moretto re your statement "By sampling less (i.e. using low seq-depth) you'll grab rare transcripts less likely because most of the reads will be "taken away" from the non-rare ones": This was why I suggested (in my previous post) that you'd need to split the original (primary) library into multiple sub-libraries before performing the sequencing - to circumvent the issue of rare transcripts being missed without using an excessive sequencing depth. In fact, this approach can be easily proven to be more effective in detecting rare transcripts than using a high sequencing depth on the primary library alone - the mathematical principles and logic governing this idea are relatively elementary. The paper you cited has also indicated the advantage of doing replicates over high sequencing depth, though the authors have resorted to empirical means to prove their findings, although (in reality) this can also be easily proven mathematically.

Thank you all for the great responses to my question.

@jennifer_hardee that is a great example, and I am very satisfied w/ the answer.

In the simplest case: if you have a jar w/ 1 red balls and 1001 green balls,

-if you prep your library three times separately and sequence each at 334 reads, there is a reasonable chance that you will not get the red ball.

-if you prep your library once and sequence at a depth of (334 * 3 = 1002 reads), you will get the red ball.

@Shiraz_Sk I also agree if we were to prep the library once, split it in three afterwards, then sequence the three parts of the split library each w/ 334 reads it would be the same as sequencing the unsplit library once w/ 1002 reads. In fact however, this was not what I was considering doing - I did want to prep the library three separate times.

The reason I was planning on prepping the library three separate times is one part of this issue that is still slightly perplexing, but may be out of the scope of this question and can be posed as a new question:

the benefit of prepping three libraries separately in terms of rare transcripts being 'missed' by initial PCR steps. i.e. if the rare transcript is not caught in the first rounds of PCR, then at subsequent rounds it would be lost in the continuous amplification of output previous rounds of PCR. In this case, starting the PCR separately three times may in fact bring a benefit to having three separate libraries. Whether this would be outweighed by the benefit of the probability discussed in this question is unknown to me.

This practice of multiple PCR replicates seems to be applied to metabarcoding for detection of rare species.

Again, thanks all. I will consider posting this ^^ follow-up Q as a new question.

@Ben J G Sutherland: Thanks for your input. Yes, the method you proposed initially (to do 3 replicates, or triplicates, with a lower sequencing depth) suffers from the possibility that the rare transcript would be missed, as Jennifer Hardee & Sushree Sangita Sahoo mentioned. However, if you sequence 1 single library at great depth, you will likely find the transcript, although the possibility of read errors (due to declining read accuracy with increased read depth as a consequence of the polymerase activity) may also cause you to mis-read the transcript. Splitting the primary library into multiple sub-libraries will enable you to minimise the possibility of read errors occurring, since you may choose to use a freshly-prepared polymerase mixture for each round of sequencing. This is in addition to the fact that subsampling would statistically enhance your chances of obtaining the rare transcript.