[ewingad]

Thanks for the reply Ben, and thanks for bowtie! I was introduced to it via a talk by Lior Pachter, he speaks highly of you and Cole. In my short experience with it so far it has been extremely fast and accurate when re-mapping reads from the 1000 genomes project.

I agree with lh3 that it may be vital to know whether a given alignment is to a repetitive sequence i.e. if other equally good alignments exist - I suppose that if there are multiple _identical_ alignments they should all be coming from the same BW rows? If this is provable then it is possible to reassure users about lh3's point #2.

I also agree that something like lh3's point #5 is useful - when aligning short reads it is often a useful quality measure to know how many alignments would result from a 1-bp change in the query sequence. So, if for a given read, if I know that even if I changed one base, the reported alignment would not change, I feel more confident about that alignment.

Thanks again!

-Adam

[lh3]

To Ben:

First thank you for Bowtie. It is really amazing. I have not read Bowtie codes, but from README, I think all the BWT based aligners (together with SOAP2 and mine) share a lot in common. Some comments:

1. I forget how much time is spent on building BWT index, but my impression is BWT-SW (also used in soap2 and bwa) is much faster and more light-weighted. Maybe it is worth having a look at the publications by Tam's group. Of course, building index is an once-for-all process. Just remind you of a possibly better algorithm.

2. My main concern about bowtie is actually related to the column 7. I think by default (no --best), bowtie just outputs the first group of hits it meets. Users would not know whether it is the best or whether it is a repeat or not. I think (maybe wrong) this behaviour is only useful for screening human contaminations. With "--best", user would know the output is the best hit, but whether it is a repeat is still unknown in some cases. I know the "unknown" cases should be rare, but it would be necessary to convince users that the rare cases would not affect accuracy. Only with "--best -k 2", a user may know whether it is a repeat or not, although he/she would not know the number of occurrences. I think the "--best -k2" is the most desired behaviour and should become the default. Bowtie is fast enough. Slowing it down by a factor 3 will still make most users quite happy (see also below). Also quoting the speed under the default option would be unfair to others.

3. I think it is worth metioning that the speed of soap/bowtie/novoalign is sensitive to error rate, while eland/zoom/rmap not. On bowtie's home page, it is claimed that bowtie is 35X faster than maq and 350X than soap. However, to my experiences, on high quality data, maq and soap is about the same speed. We can only see this big difference when base quality is low. Actually on my small experiment, bowtie is 170X faster than maq (cmd option: -f -v 2; I know using quality may be a bit slower), but I guess the difference would be smaller if the error rate is higher.

4. Comparison of BWT based aligners. So far as I know, there are three BWT based short read aligners: bowtie, soap2 and bwa. Soap2 gives the number of occurences of the best hits and bwa also reports the number of hits with additional mismatches if the best hit is unique (like what chipper is asking for). bwa is the only one that finds short indels at present, although I am sure this is not hard for soap2/bowtie. In addition, soap2 only finds hits with up to 2 mismatches. I think they have to do so according to a brief description of its algorithm. On speed, bowtie-f-v2 is 3X faster than soap2 (bowtie--best-v2-f-k2 is similar to soap2 in speed) and soap2 is 2.5X faster than bwa. On memory, both bwa and bowtie use 2.3GB while soap2 uses 5.3GB.

5. back to how the alignments are reported. I think the bwa behaviour is useful if people do not care too much about speed. Knowing the number of suboptimal hits would help us to decide which alignments are reliable. I know this is important to some (not all) SV detection algorithms. If you think the bwa behaviour is costly (possibly it is), I would recommend the soap2's one. Frequently, we may want to know the exact number of occurences (no need to output the detailed aligments). I am sure having the soap2 behaviour would make bowtie more popular.

[PS: soap2 has binary available; bwa is released under GPL and so source codes are available.]

To swbarnes2:

I do think referees found by Nature thought maq indel caller is rubish. As for the maq output, I wonder whether mapview (not maqview) suits your goal?

To Dmamartin:

I think your benchmark may not be very fair. On one hand, you use chr1 only, but BWT-based method only shines when the reference genome is longer. To this end, the comparison unfair to bowtie and vmatch. On the other hand, ZOOM is only efficient when you feed several million reads in a batch. You will find that the speed on aligning 20k or 40k reads is about the same for ZOOM.

[Chipper]

Hi Ben,

is column #7 the (minimum) number of equally good hits? Would it be possible to report also the number of hits with an additional mismatch?

[Ben Langmead]

Hi ewingad,

That's *almost* what field #7 is. What this field really tells you is how many other rows there were in the range of suffix-array rows (equivalently: Burrows-Wheeler matrix rows) from which the alignment was randomly selected and reported. To put it another way, that field tells you how many other alignments there are with the same reference string. To put it yet another way: that field gives an estimate of how many other places the read aligns, but it can be an *underestimate* (in rare cases, it can be a large underestimate). If you want to know exactly how many places it aligns then, yes, giving bowtie the -a option and counting up the alignments is the best thing to do for now.

Everything's a tradeoff! The current Bowtie default maximizes performance but doesn't expend any of the additional effort needed to get a more meaningful result for field #7. Bowtie with the -a options minimizes performance but tells you exactly how many alignments there are and what they are. In the future, it's possible for me to add an option that doesn't necessarily *report* all the alignments, but which expends the effort needed to count them up and reports the right number in field #7. This will be faster than bowtie -a, but probably only a little faster. Is that a useful feature for you, or are the current options sufficient?

Thanks,
Ben

[ewingad]

Hi, is there a way to find out how many equally good alignments exist for a given read using bowtie (other than using the -a option and counting them up)? Is this what field #7 ('Reserved') in the bowtie aligner output is?

[Ben Langmead]

Hello Ryan,

FYI, I made multiple improvements in Bowtie 0.9.8 that will help prevent such problems for people in the future. First of all, the Mac binary available from sourceforge is now "Universal" for i386 and x86_64, whereas it was previously i386 only, which forced 64-Mac users to start from source. Also, by default bowtie-build will now automatically look for values for the --bmax, --dcv and --packed parameters that fit into the memory of the computer it's running on. This obviates the tedious trial-and-error of trying larger and larger --bmaxdivn. Also, bowtie-build and bowtie-build-packed are now both "in" bowtie-build - they are no longer separate binaries. Packed mode is activated by passing -p/--packed to bowtie-build.

I hope these improvements help you and others.

Thanks,
Ben

[myrna]

Bowtie on a Mac

I found a way to fix the memory issue I mentioned in this thread on a Mac. It seems that the binary was run as a 32-bit intel process, which forces it to use 32-bit memory addressing. This meant that as soon as the process hit the 32-bit memory ceiling, it choked. I edited the Makefile and recompiled, and it runs as a 64-bit process now. I no longer get any complaints about memory, and don't have to tweak any of the runtime parameters.

Makefile modification:
old:
EXTRA_FLAGS =
new:
EXTRA_FLAGS = -arch x86_64

Ryan

[Ben Langmead]

Hello myrna,

As soon as the *next* version of Bowtie comes out, this pain will go away because there will be a "-a/--auto" option that automatically follows the suggestion printed in the error message. As 0.9.7.1 stands, you'll have to do what it says yourself, i.e., just try larger values of --bmaxdivn until it fits in memory. Again - I promise this will be easier in the next version.

Thanks,
Ben

[myrna]

memory issues when creating index file

Hi Ben.
Thanks for your prompt reply. This time around I see this error (after quite awhile):

bowtie-build-packed(14780) malloc: *** mmap(size=2860744704) failed (error code=12)
*** error: can't allocate region
*** set a breakpoint in malloc_error_break to debug
Could not allocate a suffix-array block of 2860744708 bytes
Please try using a larger number of blocks by specifying a smaller --bmax or
--bmaxmultsqrt or a larger --bmaxdivn

I will play with bmaxdivn and bmaxmultsqrt to see if I can get a successful build. Any suggestions?

Regards,

Ryan

[Ben Langmead]

Hello myrna,

Yes, sorry, other users have seen that problem too. It seems that even if your machine has plenty of RAM in total, the memory allocator may not be able to dole it out in large enough chunks to satisfy Bowtie (due to memory fragmentation within the allocator). I'm working on a solution for the 0.9.8 release. For now, you can usually work around the problem by using bowtie-build-packed, which uses 2-bit-per-base encoding to save memory.

BTW, a good place to report issues is the sourceforge bug tracker: (https://sourceforge.net/tracker/?fun...7&atid=1101606). It leaves a better paper trail.

Thanks!
Ben

[myrna]

memory issues when creating index file

I am unable to index the human genome on my MacPro (16G RAM). I have the same problem when using the provided Mac binary or compling from source. I have posted the error output below.

Any ideas?

Thanks

./bowtie-build -f ../../genomes/all_human_build_36.fa human_all
Settings:
Output files: "human_all.*.ebwt"
Line rate: 6 (line is 64 bytes)
Lines per side: 1 (side is 64 bytes)
Offset rate: 5 (one in 32)
FTable chars: 10
Max bucket size: default
Max bucket size, sqrt multiplier: default
Max bucket size, len divisor: 4
Difference-cover sample period: 1024
Reference base cutoff: none
Endianness: little
Actual local endianness: little
Sanity checking: disabled
Assertions: disabled
Random seed: 0
Sizeofs: void*:4, int:4, long:4, size_t:4
Input files DNA, FASTA:
../../genomes/all_human_build_36.fa
Reading reference sizes
Choose best chunkRate: 15
Time reading reference sizes: 00:01:09
Calculating joined length
= 2860744704 (5384364 characters of padding)
Writing header
Reserving space for joined string
bowtie-build(6713) malloc: *** mmap(size=2860744704) failed (error code=12)
*** error: can't allocate region
*** set a breakpoint in malloc_error_break to debug
Out of memory creating joined string in Ebwt::initFromVector() at ebwt.h:586

[Ben Langmead]

Hi jyli,

The memory footprint of the whole-human index is about 2.2 GB without the -z ("phased") option. With the -z option it's closer to 1.3 GB (last I checked). If your machine has 3 GB of RAM or more and you'd like to align to human, the default mode should be fine. If your machine has 2 gigabytes of RAM and you'd like to align to human, you'll need to use the -z option.

(The unfriendly error message is my fault! - I'm going to fix that for the next release.)

Thanks,
Ben

[jyli]

Memory requirement on a window 32x

I tried to test human index downloaded from recommended site with the command

bowtie -c h_sapiens_asm ATTCAGTAGGTACTATAAATGGCCGAT

then, I got error:

Out of memory allocating ebwt[] in Ebwt::read() at ebwt.h:2811
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc

This application has requested the Runtime to terminate it in an unusual way.
Please contact the application's support team for more information.

So, my question is about the memory allocation or whether I did anything wrong?

Thank you for your attention.

[swbarnes2]

They seem to be. But I'm only doing bacteria, and that's easier to align correctly to. Reference genomes are rarely what they are cracked up to be, but when aligning I look across the genome at what aligned where, I see mostly 48-mers, but also 40-mers, 32-mers and occasionally 24-mers, when I trim by 8's. And I know that when I compare the two output files of my test, the reads that show up in SOAP that didn't show up in Bowtie are all ones that SOAP trimmed.

[Chipper]

Are the extra million reads aligned after truncatinon really correctly placed?