2
$\begingroup$

How do the Oxford Nanopore Flongle flow cells perform compared to the standard MinION flow cells, when it comes to reliability and error rates? I found this comparison article from last year, from where the below graph is taken, but they used the old chemistry. How is the situation today with the new R10.4.1 nanopores and Kit 14 chemistry? Does anyone have first-hand experience, or a good reference?

My use case would mainly be metabarcoding, i.e. amplicon sequencing. If anyone has experience of Flongle flow cells for specifically metabarcoding, I would be happy to hear your thoughts.

enter image description here

$\endgroup$

1 Answer 1

4
$\begingroup$

For us, R10.4.1 Flongle flow cells are running at reduced yield due to lower numbers of working pores (something like 10-20 pores instead of 30-60 pores for R9.4.1 flow cells). This doesn't really affect the stuff we're using Flongles for, because I make sure we've got plenty of headroom built in when designing experiments.

Here are the yield assumptions I use for flow cells, which are basically the lower ranges that we've seen when using them:

  • Flongle - 200 Mb
  • MinION - 5 Gb
  • PromethION - 50 Gb

For Flongle, this has been fine for us to do sequencing and assembly of 12 plasmids on one flow cell using the rapid barcoding kit, and the improved accuracy from the Kit14 + R10.4.1 combination makes reliable assembly easier (especially where homopolymers are involved). I expect it will work similarly well for amplicon sequencing.

Error rates are pretty similar across platforms, because it's the pore that matters, not the package. Here's a comparison plot of mapped accuracy vs predicted accuracy (from the basecaller) that I did recently on a TrackIt 1kb Plus DNA ladder, with bead cleanup using LFB, run on a PromethION flow cell with the V14 ligation kit (sequenced on a P2 Solo), then called using the new bacterial model. The reads represented in this plot have been filtered to only keep (at most) the longest 2000 matches for each target. The run produced 1,128,671 reads (1.95 Gb) over 1.5 hours:

Predicted vs mapped quality distribution(PromethION), demonstrating that the mapped quality is usually higher than the predicted quality from the base calling model

Here's a comparison plot of mapped accuracy vs predicted accuracy (from the basecaller) for the same DNA ladder sample run on a Flongle flow cell (sequenced on a MinION Mk1b), with bead cleanup using SFB, also taking at most 2000 longest sequences from each mapping. The run produced 465,144 reads (191 Mb) over 24 hours:

Predicted vs mapped quality distribution (Flongle), demonstrating that the mapped quality is usually higher than the predicted quality from the base calling model

The predicted accuracy is taken from the basecaller (i.e. fastq quality string), and is calculated by summing up the quality values after converting to a linear representation, then dividing by the number of bases and converting back to a log representation:

sub getMeanQual {
  my ($qual) = @_;
  if(length($qual) == 0){
    return(0);
  }
  my $qBase = 33;
  my @qspl = split("", $qual);
  my $qualTotal = 0;
  my $qualCount = 0;
  foreach my $qual (@qspl) {
    $qualTotal += 10 ** (-(ord($qual) - $qBase) / 10);
    $qualCount++;
  }
  return(int(-10 * log($qualTotal / $qualCount) / log(10)));
}

According to the trained models that LAST generated, there's not much different between the called reads between Flongle and PromethION flow cells. Here's the PromethION base mismatch matrix (q45 SNPs):

# substitution percent identity: 99.9971
#last -Q 1
#last -t4.38507
#last -a 23
#last -A 19
#last -b 2
#last -B 4
#last -S 1
# score matrix (query letters = columns, reference letters = rows):
       A      C      G      T
A      6    -45    -35    -46
C    -45      6    -45    -44
G    -33    -45      6    -46
T    -46    -45    -46      6

And here's the Flongle base mismatch matrix (q41 SNPs):

# substitution percent identity: 99.9922
#last -Q 1
#last -t4.36483
#last -a 20
#last -A 20
#last -b 4
#last -B 5
#last -S 1
# score matrix (query letters = columns, reference letters = rows):
       A      C      G      T
A      6    -45    -36    -46
C    -45      6    -44    -45
G    -32    -45      6    -46
T    -46    -45    -46      6

So I guess you could say that the Flongle has slightly lower accuracy, but I wouldn't call that difference significant for any practical purposes.

$\endgroup$
1
  • $\begingroup$ Thank you for the update, I have marked your answer as a solution! $\endgroup$
    – Joel
    Commented Oct 25, 2023 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.