[maker-devel] Couple quick questions about Maker

Mon Jul 7 10:26:43 MDT 2014

I don't think RepeatMasker produces GFF3.  I believe it is GFF2 with the
-gff option (which is pretty different). Also If you provide GFF# files for
repeats, you will still need to turn of repeat masking in the control files
by blanking out the options.  Also MAKER uses a step called RepeatRunner
against an internal transposable element protein databases which is probably
still running (and is slow because it's a search in translated protein
space).

For performance, you may want to give a larger max_dna_len for the MAKER run
given that you have a large RAM machine. Also set all the depth_blast in
maker_bopts.ctl to 15 or 20.

CEGMA is convenient for training predictors because it finds genes that will
always be in every eukaryote (I.e. high confidence).  You can combine these
with est2genome/protein2genome results from MAKER if you want.  You can then
use the resulting HMM for a larger MAKER run with experimental evidence, and
then train again on those results.  But beware than there is rarely any
benefit from training beyond that second round.  More training actually
tends to makes things worse (the overtraining paradox).

--Carson

From:  Daniel Ence <dence at genetics.utah.edu>
Date:  Monday, July 7, 2014 at 10:00 AM
To:  Nathaniel Jue <n.jue at uconn.edu>
Cc:  "<maker-devel at yandell-lab.org>" <maker-devel at yandell-lab.org>
Subject:  Re: [maker-devel] Couple quick questions about Maker

Hi Nathaniel, 

1) We'll need to see the error messages that MAKER was giving to understand
what might have gone wrong with the Repeat Masker gff3 file. If you could
run maker on one of your scaffolds with your current settings and send us
the complete output, we can start to figure out what happened.

2) MAKER interacts with its gene predictors (augustus, snap, and the other
ones listed in the control files) in a way that improves their performance
(with the hints and such). When you supply predictions through the pred_gff
parameter, MAKER can't give that performance improvement, so there's
something of a tradeoff there. I think the performance improvement is a key
part of MAKER's success, so I would definitely recommend running the
ab-initio tools internally.

MAKER tries to save you time by saving results from run to run and only
rerunning tools (usually blast tools) that had their parameters changed in
the control files. Taking advantage of that will probably be the biggest
time saver for you. Something else that could save you almost as much time
would be to set a reasonable lower-bound on the size of contigs that maker
will try to annotate (usually <5kbp or <10kbp depending on your genome).
This parameter is set with the min_contig parameter.

I'll have to check with my lab mates about the Repeat ORF searching and how
they use CEGMA results. I think you can probably just put them all in there
at once though. 

~Daniel

Daniel Ence
Graduate Student
dence at genetics.utah.edu
Eccles Institute of Human Genetics
University of Utah
15 North 2030 East, Room 2100
Salt Lake City, UT 84112-5330

On Jul 7, 2014, at 9:26 AM, Nathaniel Jue <n.jue at uconn.edu>
 wrote:

> Hi, 
> 
> I'm trying to run maker on a couple genomes right now and was wondering if
> folks had any thoughts on way to speed it up a bit. I'm running it on a
> 48-processor supercomputer (lots of RAM, usually use it for genome assembly).
> Both these genomes are a little fragmented, so there are lots of contigs,
> which slows down the whole process. I am looking for ways to speed things up
> and was wondering about a couple things:
> 
> 1) I'm currently just at the first round of maker predictions using EST and
> protein evidence to build models. Had already done RepeatMasking so thought
> I'd just input subsequent GFF to speed it up. Didn't seem to like the GFF, so
> two questions: i) any thoughts on why that GFF wasn't acceptable? It's the one
> that repeatmasker outputs if you ask it to; and ii) Providing this GFF, should
> generally allow the program to bypass the RepeatMasking step, correct? Does it
> also make it bypass the Repeat ORF searching step?
> 
> 2) I plan to run both SNAP and Augustus on these genomes as well. The two-step
> SNAP training from the tutorials seems straightforward, but I was wondering
> about the Augustus step. From what I can tell, simply providing an Augustus
> "trained" species name should turn on Augustus and blast/blat-like hints
> generated within Maker are then used in gene prediction. Any thoughts on if
> it's either more accurate or faster to do the Augustus predictions outside of
> the Maker pipeline and then import them using the pred_gff parameter in the
> maker_opts file?
> 
> 3) Finally, I noticed that you had a script for converting cegma gff files to
> zff file for snap training? Currently, I am using predicted transcript for
> this species and protein sequences from related species to training. Does
> anyone have any insight into using CEGMA results as well? Do you work
> iteratively with them? For instance, start with the using hints from more
> distant taxa (i.e. CEGMA) and then work your way closer? Just throw everything
> in at once and retrain after that?
> 
> Thanks in advance for any advice and insight.
> 
> Cheers,
> Nate
> 
> 
> Nathaniel Jue, Ph.D.
> Department of Molecular and Cell Biology
> University of Connecticut
> Storrs, CT 06269
> 
>  
> <http://s.wisestamp.com/links?url=http%3A%2F%2Fwww.linkedin.com%2Fpub%2Fnathan
> iel-jue%2F1%2F531%2F176%2F&sn=>
> _______________________________________________
> maker-devel mailing list
> maker-devel at box290.bluehost.com
> http://box290.bluehost.com/mailman/listinfo/maker-devel_yandell-lab.org

_______________________________________________ maker-devel mailing list
maker-devel at box290.bluehost.com
http://box290.bluehost.com/mailman/listinfo/maker-devel_yandell-lab.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://yandell-lab.org/pipermail/maker-devel_yandell-lab.org/attachments/20140707/103b0aa3/attachment-0003.html>