[poisson200]

This is misinformation. An error in colorspace does mess with the rest of the read momentarily, but colorspace errors are easily corrected thanks to the same phenomenon because there will be an inconsistency across the different ligations, and colorspace error trends are known and modeled. In actuality, colorspace reads are a lot more accurate in theory than base space (but in chemistry, they may get messier at the ends).

Thanks for the correction; I am interested to know where the error trends are modelled? Is that in a publication somewhere and/or do some mapping packages correct/know about these errors when mapping reads and correct for it?

To clarify; you say that colourspace reads are more accurate than base space "theoretically", but because of chemistry issues, in reality they are not. Is that what you mean?

[dsidote]

Bacdirector,

What did you use to trim your SOLiD reads?

Dave

[drio]

I am looking forward to see what's the number of uniquely mapped reads when trimming compared to the non trimmed version. I am with MJC, if you use bwa, NOvoalignCS or bfast (particularly the last two) you'll see an increase in mapped reads without trimming. It will be nice if you can recompute your alignments with those and show the numbers.

[westerman]

The ABI/LifeTech Bioscope software does 'progressive' mapping where it starts considering reads that map at 50 bases (and with various mismatches), then at 49 bases, etc. This approach is probably superior to one of simply chopping off the reads to 35 bases.

As an example, a partial statistics file from one of my recent SOLiD runs shows:

Read Length 50 0 mismatches 19,901,946 (64.86%)
Read Length 50 1 mismatches 1,336,272 ( 4.35%)
...
Read Length 35 0 mismatches 137,601 ( 0.45%)
Read Length 35 1 mismatches 40,989 ( 0.13%)
...
Down to a read length of 25.

[Michael.James.Clark]

Thanks for the correction; I am interested to know where the error trends are modelled? Is that in a publication somewhere and/or do some mapping packages correct/know about these errors when mapping reads and correct for it?

To clarify; you say that colourspace reads are more accurate than base space "theoretically", but because of chemistry issues, in reality they are not. Is that what you mean?

Not exactly. Sorry, that may have been a little unclear. You can see any of the tech specs for SOLiD to help understand how the base correction works. Check the ABI site--many publications listed here: http://www.appliedbiosystems.com/abs...equencing.html

Basically, colorspace reads are more accurate than base space reads because of the ability to correct colorspace errors. There are (IIRC) five repeat resets during SOLiD sequencing, so a single read is observed five separate times at different ligation starting positions. Due to the increased number of observations, one can correct away errors thanks to knowing how the colorspace-to-basespace translation would be affected by a specific mismatching colorspace read.

The chemistry gets messy at the 3' end, so the accuracy start falling off, but it doesn't get as bad as single base sequencing after 35-bases--more like the last five bases or so. So 2-base encoded colorspace reads are more accurate than base space reads generally, but along the length of the read, they may be less accurate at the very 3' end. However, that inaccuracy at the end has zero impact on gapped alignment anchored by masking at the 5' end.

Nils Homer has a nice article about two-base encoding and how it works to improve accuracy here: http://www.biomedcentral.com/1471-2105/10/175

[bacdirector]

thanks, everyone

well, a lot of great feedback and perspective here; looks like a lot of people are doing different things; some use a masking tool, which should allow reads to map in spite of non-calls; that makes sense; we have heard of progressive mapping, but not done it yet, and will be studying that in comparison to what we've found. also, re: # uniquely mapped reads, will be sure to report back on that after we look at that closely (over the various trimmed lengths).

with the various strategies for data representation, filtering on qvals; trimming back based on qvals; and given the various algorithms, i.e., bfast, and bowtie, and right now we're doing nextgene and bioscope, there seems, as usual, to be a combinatorial number of strategies. we've been careful to be to look at the number of valid adjacent errors as well to be sure we're not just mapping short reads anywhere and everywhere.

yeah, the dibase encoding = higher accuracy is real.

looks like we have a lot of careful comparative evaluations to do!

any additional thoughts would be appreciated.

[drio]

For your evaluations (simulations), take a look to this.
The author of bfast talks about simulating reads and then plotting the results using ROC curves to find the best tool for the problem.

[bacdirector]

Ok, here it is - Percent of Reads that Map that Also Map Uniquely

In our study that started this thread, I originally reported a doubling of reads mapped from 25 million to 50 million achieved by trimming all reads >35 to 35. At 50b, we get only 25 million reads mapped. This was based on empirical comparisons of reported valid adj errors.

The answer to the question of what percent of reads map uniquely is--- it barely matters. The range is 85.5% to 86.5% Percent of Reads that Map that Also Map Uniquely, with an increasing trend from 85.5% to 86.5% going from 25b trimmed reads to 50b reads.

So far, we've seen the optimal overall response to trimming at 35b. This is for Bioscope, not Nextgene (my bad!!!), but without any masking of -1's..... so in response to feedback from the thread, we'll be expanding our repetoire to include progressive mapping as well as non-call masking. And bfast. And bowtie. We have all the trimmed data sets ready to go, it's just a matter of punching buttons...

thanks, everyone for your interest and feedback!

James Lyons-Weiler
Director
Bioinformatics Analysis Core
University of Pittsburgh

[Chirag]

Hey,

Could anyone help me on how to trim the color space data.
I am using bowtie/tophat to map the data, and i map only around 40% of the reads. Reads are 75 bp long and i had allowed 3 mismatches, where allowing 2 mismatch had given me only 35% mapping.

Could someone help me on how to trim the end of the reads, for eg, last 10/15 bases are trimmed, so could probably map more. But haven't found yet any S/W that does trimming for SoliD ?

regards
Chirag

[Chirag]

Try this:

http://hts.rutgers.edu/filter

[colindaven]

The best aligners for CS data use iterative trimming, see for example NovoalignCS and Lifescope.