How to confirm the best parameter for assembling?

[anyone1985]

Some assemblers have less parameter, such as Edena, Velvet; Some other are not. I'd like to use vcake or ssake to assemble some solexa data. However, there are so many parameters that it is not possible to try them all. How can I find out the best parameter?

[Torst]

Some assemblers have less parameter, such as Edena, Velvet; Some other are not. I'd like to use vcake or ssake to assemble some solexa data. However, there are so many parameters that it is not possible to try them all. How can I find out the best parameter?

Your first problem is to strictly define what you mean by "best". This is called your "objective function".

Once you have that, you can use various methods to navigate the parameter space to find those parameters which maximize (or minimize) your objective function.

Torst.

[anyone1985]

I defined the best parameter was that the least number of contigs with the high accurate. Because I wanted to use solexa data to finish a bacteria genome with a draft reference genome.

Your first problem is to strictly define what you mean by "best". This is called your "objective function".

Once you have that, you can use various methods to navigate the parameter space to find those parameters which maximize (or minimize) your objective function.

Torst.

[lmrodriguezr]

While "the least number of contigs" is a good variable (measurable and meaningful), "high accurate" is hard to measure. What does "high accuracy" means in terms of assemblage?

The Salzberg's group at the UMD have developed some tools for dealing with this question. You can give a look at http://www.cbcb.umd.edu/research/ass...lidation.shtml.

I hope to be useful,
Luis M. Rodriguez-R

[anyone1985]

I am finishing a bacteria genome with the solexa data. It is said that Solexa is not suitable for de novol assembling. The contigs that I assembled by velvet maybe exist mistakes which I will never know. This would lead me never to finish it . This is what I worried about.

While "the least number of contigs" is a good variable (measurable and meaningful), "high accurate" is hard to measure. What does "high accuracy" means in terms of assemblage?

The Salzberg's group at the UMD have developed some tools for dealing with this question. You can give a look at http://www.cbcb.umd.edu/research/ass...lidation.shtml.

I hope to be useful,
Luis M. Rodriguez-R

[dan]

Any papers comparing assemblies?

I'd like to know which assemblies are best ... I don't care *how* best is defined, just so long as it *is* defined in the paper...

I think as a community we really need to think about infrastructure to allow rigorous comparison of different assembly methods over different datasets... I know this is hard, and efforts like the "Genome Assembly Validation" project are very welcome, but I think there is still a lot to do.

For example, people are rushing to perform hybrid assembly - where can we find guidelines about how best to do this? Is there any theoretical study of what mix of technologies gives the 'best' assembly?

Sorry for ranting, but the sooner we can get these questions sorted, the sooner we can turn to more interesting analysis.

[mchaisso]

I think it would be difficult to judge assembly methods based on papers right now -- the publication rate is slower than the rate of improvement in these methods, so you might be stuck trying out the latest of different methods. Perhaps the best way to "publish" assembly results is to simply post them in a blog format.

Within the realm of assembly development, the common objectives are: <br>

First: Correctly order read overlaps so as to have the largest N50, without creating any mis-assemblies. With next-gen data, coverage is so high that one typically doesn't have to "bet" on any overlaps as being correct, whereas in Sanger sequencing there are regions with only two reads overlapping with possible ambiguity, so decisions are made to consider overlaps that are more likely to cause mis-assemblies.

Second: Get base calling correct. This is typically post-processing (or as in euler-sr, only planned for post-processing).

Regarding the hybrid assembly, there are some tools that I started writing for euler-sr that would indicate what experiments are necessary to finish the genome, but development in this has been handed off to a new crop of students, so it may take some time to figure out. I'm not sure what Birney and Zerbino plan for Velvet, but it would be pretty easy to do. The best way to judge what is going on in an assembly is to look at the repeat graph, but typically it's difficult to make any sense of it unless you have looked at a lot of repeat graphs before.

-mark

Any papers comparing assemblies?

I'd like to know which assemblies are best ... I don't care *how* best is defined, just so long as it *is* defined in the paper...

I think as a community we really need to think about infrastructure to allow rigorous comparison of different assembly methods over different datasets... I know this is hard, and efforts like the "Genome Assembly Validation" project are very welcome, but I think there is still a lot to do.

For example, people are rushing to perform hybrid assembly - where can we find guidelines about how best to do this? Is there any theoretical study of what mix of technologies gives the 'best' assembly?

Sorry for ranting, but the sooner we can get these questions sorted, the sooner we can turn to more interesting analysis.