2500 vs 4000

[kwalton]

We are in a position to buy an Illumina. I have a lot of experience with the 2500 at my previous job, but have not run Illumina technology in ~2yrs. At my current position we are debating the 2500 vs. 4000. We predominantly run RNA-seq (with Kapa lib build), ChIP-seq, PCR-seq (low diversity), targeted seq, "weird" libraries, some exome seq and with the Illumina technology we will be running some WGS. Who has run samples on both platforms? What is your experience? How do the above mentioned libraries work on the 4000? I have read there are problems with the 4000 (AD, not tolerant of differing library sizes, does not tolerate low diversity, etc). Are these actual concerns in the field with anyone running the 4000? How many million reads per lane are people actually getting with the 4000 with the above library types, or any library that you are running? How are you spiking in phiX on the 4000? Do you see a problem with spiking in phiX? I realize this is a lot of questions and I appreciate any help you can give.

[nucacidhunter]

The short answer is that with HiSeq 2500 you can run all of those libraries but not with 4000 and currently there is no long read (2x250) option either. For those limitation one gets 25% increased raw read but if library has not been prepared carefully, that 25% can be lost after processing and removing low quality reads and primer-dimers.

[Brian Bushnell]

The 2500 is certainly a safer bet. The 4000 *may* reach the quality of the 2500 some day, but I think that's hardly a valid justification for committing >$1m in some direction. The 2500 is good *now*.

[luc]

With libraries made by us, we are getting 370 million reads on average on the 4000. The quality suffers with longer insert sizes and the system is less tolerant to low complexity issues (it requires more spike in of high complexity libraries), but you can run basically anything and it is more tolerant to library concentration variations. If you are operating the sequencer you will certainly prefer the 4000.

[nucacidhunter]

@luc
I wonder if you have done any bisulfite converted library either RRBS or WGBS on 4000.

[Brian Bushnell]

With libraries made by us, we are getting 370 million reads on average on the 4000. The quality suffers with longer insert sizes and the system is less tolerant to low complexity issues (it requires more spike in of high complexity libraries), but you can run basically anything and it is more tolerant to library concentration variations. If you are operating the sequencer you will certainly prefer the 4000.

That's quite interesting. Have you done any analysis of the quality of output reads?

[GenoMax]

With libraries made by us

This is probably the most critical aspect.

@kwalton: If you are going to make your own libraries then 4000 would work but if you expect to run libraries made by someone else then you may find 2500 a safer bet (especially if you have no control over the library quality/characteristics).

@luc: Have you run any libraries that were not made by you?

[luc]

Hi,
we are running libraries made by customers, which often have very wide insert size ranges (e.g. the ugly Nextera libraries). For these the yields (while still better than 2500) and the quality scores tend to be lower. We have run epignome WGBS as well as RRBS libraries. For these it seems to be difficult to dose the PhiX - the amount of spike in has not a good correlation to the % of resulting PhiX reads - requiring a test lane. Likely due to some algorithmic problem the reverse reads of bisulfite converted lanes have low quality scores, but the reverse read sequences seem not to contain notably more errors (e.g. alignment rates are similar).
Problematic are libraries that have sudden changes between very low complexity and high complexity regions.

[bilyl]

Dumb question, but how is it possible that the HiSeq 4000 is not robust to low diversity? Isn't the whole point of patterned flow cells to get good registration of clusters, and then from there get reliable base calls? Low diversity was primarily a problem because of bad base calling when some channels get flooded with signal and you can't see any defined clusters.

[luc]

Sorry, I am as dumb. I did post this question several times to Illumina people but never got an answer.

Dumb question, but how is it possible that the HiSeq 4000 is not robust to low diversity? Isn't the whole point of patterned flow cells to get good registration of clusters, and then from there get reliable base calls? Low diversity was primarily a problem because of bad base calling when some channels get flooded with signal and you can't see any defined clusters.

[GenoMax]

Speculating about a cause: it is possible that 4000 can't tolerate low diversity in the first 25 cycles when chastity filtering is active for each cluster. Even though the features are defined, having majority of them light up at the same time may interfere with RTA/chastity filtering.

[Genohub]

Highlights of current HiSeq 3000/4000 limitations: https://genohub.com/services/sequenc...seq-3000-4000/

[luc]

With regards to the Genohub page info:

As mentioned before - low complexity sequencing is possible with the HiSeq 4000 (although not recommended by Illumina).

Point 6 is misleading. What is this supposed to mean? - without any quantification or details? Starting in autumn there were (are?) reagent quality problems that showed up only the last 40 bases of PE150 reads (reagent quality problems tend to show up on all illumina platforms). For us these reagent problems fortunately seem to have disappeared. HiSeq 3000s and 4000s are identical with regards to parts and read qualities. How big is the install base of the HiSeq 3000 ? (perhaps 3 or 4 machines? - the vast majority are HiSeq 4000 sequencers).

[huguesparri]

As far as I know, some facility running a HS3000 is still facing reagent issues in february 2016.
And from what Illumina told them, they are still investigating on the source of this issue.

[bmangeat]

Does anyone have some information regarding how the HiSeq 3000/4000 compares with the NextSeq when sequencing low complexity libraries? Is it as bad, or intermediate between the NextSeq and HiSeq 2500?