It’s taken us a while, but at this point we’ve now submitted two multiplex NextSeq Illumina kits through CWRU genomics. We’ve performed the multiplexing by appending dual indices during targeted amplification of the barcodes. In order to mix everything together, we’ve taken a rather simple approach of gel extracting the amplified bands and quantitating the amount of extracted DNA using Qubit, and mixing every extracted amplicon to equimolar amounts. This mixture is then submitted to the CWRU genomics core for qPCR-based quantification, and loaded onto the kit for sequencing.
Since everything has been mixed to equimolar amounts so far, it’s quite simple to see how well our samples were quantitated and mixed together using the Qubit readings. This is done by taking the number of paired reads associated with each pair of indices, and seeing how their counts / frequencies compare to each other. The distributions tended to be a log-normal distribution, as seen in the below plots:
The second kit seemed to do a little better than the first one. Each kit ended up having a singular poorly sequenced outlier, with the sample in the first kit lower than the median by about ~300-fold, and the sample in the second kit lower by about 10x. The first kit also had about 4 samples that were below the median by about 10-fold.
Regardless, this log-normal distribution is showing what kind of mixing precision we can expect to consistent get with this scheme, even when its working well. Both distributions had a sd(log) of ~0.5. The coefficients of variation were ~ 0.04.
What are the red lines you ask? Well, those are the idealized / hypothetical number of reads we should have gotten if we got all of the reads from the kit (130 million for the mid kit used in kit 1, and 400 million for the high kit used in kit 2) divided by the total number of samples. Clearly, the real life read numbers aren’t hitting that idealized number, which is as expected.
To frame it another way though, sure having read a sample with too much read-depth in inefficient, but what is arguably more annoying is if end up under-reading a sample because we were off in out estimates. What seems like a reasonable approach is to choose a target read-count that will ensure that ~95% of our log-normal distribution hits that minimum read-number needed to get interpretable data. Looking at these n = 2 results, it seems reasonable that we would want to give each sample ~ 5 to 10-fold more reads than one would expect based on idealized numbers.