Sbastien Boisvert, the developer of Ray assembler, wrote a long comment that we believe would be helpful for our readers. So, we are posting it here with slight modifications (=hyperlinks moved around text):
homolog.us admin: The total length produced by SPAdes is 4927558, but E. coli K-12 MG1655 has 4.6 Mb. This strongly suggests duplicated genomic regions caused by misassemblies !
Rick Westerman: You should try Ray Meta for your metagenome !
I am the developer and maintainer of the Ray assembler.
Like Rayan Chikki, assemblers is what I do. And like him, I believe that graph traversal algorithms are better for genome assembly (better than all those assembly algorithms based on simplification and/or concatenation and/or erosion).
Anyway, single bacterial genome assembly is pretty much a solved problem if you want the perfect assembly and if you take the time to look at ALLPATHS.
ALLPATHS is probably the best assembler for the correctness of the output, but its usability and scalability are poor for what I hear (I dont know, I am not a end user). It produces 1 scaffold, but is not of broad utility because it needs an huge array of fancy sequencing libraries.
For E. coli with just two insert sizes of 215 +/- 10 and 487 +/- 18 (the famous SRA001125 benchmark), the Ray assembler generates an assembly in minutes if you throw 32 Ray processes on the problem on a good interconnect.
The memory usage is below 200 MB per Ray process (virtual memory, this includes the mapped software library and the Ray executable size too !). So globally Ray gives the same result produced by these rectangle graphs in SPAdes, but with less than 6 GB of physical memory and in minutes. And Ray produces almost no misassemblies, if not 0.
Ray is so easy to use that you wont believe it !!! ;-)
mpiexec -n 32 Ray \
Output of Ray:
Scaffolds >= 500 nt
Total length: 4583532
Our three paradigms for Ray and Ray Meta (and Ray Communities) are:
- outstanding usability
- best-in-class correctness
- infinite scalability
You can actually run Ray on a ultrabook with 8 GB of memory if you dont like super computers for a bacterial genome. Ultrabooks are quite the things these days.
The issue with SPAdes and its Paired de Bruijn Graph (PDBG) is combinatorial explosion. The authors of SPAdes do not seems to see this as a problem though. In the paper about it, no running times are provided. This fact and the fact provided by Rick Westerman above about 300 GB for 8 GB of input data strongly indicate that SPAdes is more a prototype to test theories and models than a software tool ready for the genomics community.
EULER, also from the group of Pavel Pevzner, was also a prototype with similar usability features.
All the best !
1. Our blog did not write much on three assemblers - Ray, ABySS and ALLPATHS, because they were already known to be very good assemblers, when we started writing, and we never had time to read their papers in detail and present with some new insight (yes, bad excuse). If you like to explore Ray assembler further, here are three links you can start with -
2. Regarding SPAdes, we believe its primary appeal is in proposing an algorithmic approach to use paired end reads within the de Bruijn graph framework. For that reason, we are not very concerned about the RAM usage and other performance metrics (except assembly quality) of the assembler. Many others (link, link) are also looking into the same problem and we do not know, whether SPAdes algorithm will be the best in the long run, but it is unusual enough to start thinking about.
paired short-read assembly will remain a hot topic for at least a few years.
If he is right, we will have heated arguments for at least a few years :)