[maker-devel] Filtering of ab initio gene models
Carson Holt
carsonhh at gmail.com
Fri Jun 6 12:51:18 MDT 2014
There can be overlapping meddles if you have multiple gene predictors. Also
the hint based models will overlap the ab initio models, but you never get
to see them (they are not kept in the evidence because they are confusing
and really not useful unless they are chosen as the best model). So they
will overlap the ab initio models, but you may never get top see them. All
models regardless of location and overlap get sorted by their AED score.
The best model is then kept from the list. Then the next, then the next. If
the next best model overlaps a model that has already come off the list
(which means the other model has a better AED score), then it gets skipped,
and the next best one in the list gets added to the non-overlapping space.
The result is that the final models will be non-redundant and
non-overlapping, but if you look at the evidence aligments you will find ab
initio models different than the MAKER models that were rejected and do not
overlap the final models.
model_gff competes just like any other model with AED. Ties always go to
model_gff, and if there is a region where no model gets chosen (they all
have AED of 1) and a model_gff entry will fit (even with an AED score of 1),
then it will be chosen, because model_gff do not need evidence support to
end up in the final annotations.
--Carson
From: Daniel Standage <daniel.standage at gmail.com>
Date: Friday, June 6, 2014 at 12:38 PM
To: Carson Holt <carsonhh at gmail.com>
Cc: Maker Mailing List <maker-devel at yandell-lab.org>, Volker Brendel
<vbrendel at indiana.edu>
Subject: Re: [maker-devel] Filtering of ab initio gene models
Carson et al,
Your feedback so far has been very helpful, and we are grateful for the time
you have taken to respond!
We're still trying to understand the precise procedure by which competing
models are chosen. You mentioned that a single model must be chosen (via
AED) from a pool of available models: are these pools constructed by
overlap? It is not uncommon in our experience to see overlapping genes
reported by Maker, although for the most part it appears these overlapping
genes don't have CDS overlap.
Looking more closely at the Maker 2.10 output from the test data we sent
yesterday, we also noted that exclusion of the transcript in question also
had an effect on the interval exon structure (exon 7717-7776 becomes exon
7737-7776) of a downstream model with which it overlaps 3 nucleotides.
And still unclear to us is how the model_gff data fits in with all this.
>From my previous searching of the list archives I was under the impression
that these models would be given substantial weight in the prediction
process, and would only be altered if a considerably better model could be
identified. Our experience with this small data set, though, is that which
overlapping gene is reported, and which corresponding exon structure is
selected, is dependent on very slight changes in the evidence.
Thanks,
Daniel
--
Daniel S. Standage
Ph.D. Candidate
Computational Genome Science Laboratory
Indiana University
On Fri, Jun 6, 2014 at 12:59 PM, Daniel Standage <daniel.standage at gmail.com>
wrote:
> This helps, thanks.
>
>
> --
> Daniel S. Standage
> Ph.D. Candidate
> Computational Genome Science Laboratory
> Indiana University
>
>
> On Fri, Jun 6, 2014 at 12:56 PM, Carson Holt <carsonhh at gmail.com> wrote:
>> I got the e-mail. Thanks for the test set.
>>
>> Multiple ab initio predictors don't inform a single annotation, rather one
>> must be chosen from the pool of available models (I.e. it has to be SNAP or
>> Augustus, or GeneMark). They all supply their own ab initio as well as hint
>> based prediction, and then the one with best evidence match (measured by AED)
>> is kept (it's like a competition that only one predictor can win).
>>
>> If you want a consensus model instead, you can take MAKER results in GFF3
>> format and give them to Evidence Modeler (EVM). The upcoming MAKER 3.0 is a
>> collaboration with the EVM group and will have this option, but for now users
>> can just split the MAKER GFF3 by evidence types and give it to EVM. EVM then
>> produces consensus models based on the GFF3 content.
>>
>> --Carson
>>
>> From: Daniel Standage <daniel.standage at gmail.com>
>> Date: Friday, June 6, 2014 at 10:46 AM
>>
>> To: Carson Holt <carsonhh at gmail.com>
>> Cc: Maker Mailing List <maker-devel at yandell-lab.org>, Volker Brendel
>> <vbrendel at indiana.edu>
>> Subject: Re: [maker-devel] Filtering of ab initio gene models
>>
>> Good to know, thanks. If multiple ab initio predictors inform a single
>> annotation, how does Maker decide which one will be included in the gene's
>> ID?
>>
>> Given your quick response just now, I wanted to confirm that you got the
>> message and data set I sent yesterday. I received an email saying the size of
>> my message required list admin approval to be distributed, but since you were
>> also a direct recipient of the email I didn't worry about it too much.
>>
>> Thanks again!
>>
>>
>> --
>> Daniel S. Standage
>> Ph.D. Candidate
>> Computational Genome Science Laboratory
>> Indiana University
>>
>>
>> On Fri, Jun 6, 2014 at 12:39 PM, Carson Holt <carsonhh at gmail.com> wrote:
>>> snap_masked-$seqid-processed-gene was produced by SNAP on the repeat masked
>>> sequence without hints (i.e. the ab initio call).
>>> maker-$seqid-snap-gene was produced by SNAP after receiving hints from
>>> MAKER.
>>>
>>> In both cases MAKER is allowed to add UTR to the model (hence the
>>> 'processed' tag).
>>>
>>> --Carson
>>>
>>>
>>> From: Daniel Standage <daniel.standage at gmail.com>
>>> Date: Friday, June 6, 2014 at 10:33 AM
>>> To: Carson Holt <carsonhh at gmail.com>
>>> Cc: Maker Mailing List <maker-devel at yandell-lab.org>, Volker Brendel
>>> <vbrendel at indiana.edu>
>>>
>>> Subject: Re: [maker-devel] Filtering of ab initio gene models
>>>
>>> Another question: is there documentation anywhere for the naming conventions
>>> of the genes annotated by Maker? Of course it's easy to spot genes based on
>>> a particular ab initio gene predictor, as the names are in the IDs. But what
>>> is the significance of, say, "snap_masked-$seqid-processed-gene" in a gene
>>> ID vs "maker-$seqid-snap-gene"?
>>>
>>> Thanks,
>>> Daniel
>>>
>>>
>>> --
>>> Daniel S. Standage
>>> Ph.D. Candidate
>>> Computational Genome Science Laboratory
>>> Indiana University
>>>
>>>
>>> On Thu, Jun 5, 2014 at 2:05 PM, Daniel Standage <daniel.standage at gmail.com>
>>> wrote:
>>>> I have attached data for a small 18kb region with a handful of genes, as
>>>> well as the corresponding maker_opts.ctl file. (This is a smaller and
>>>> different data set than what I was looking at yesterday, with a more
>>>> well-defined problem).
>>>>
>>>> With the data files as is, Maker 2.31.3 reports a model from 4125 to 6400
>>>> with an AED of 0.23. If you exclude transcript TSA024184, Maker reports a
>>>> different gene from 6111 to 8345 with an AED of 0.01. Both of these genes
>>>> have transcript support: will Maker report overlapping genes under any
>>>> conditions? And even if Maker is forced to choose only a single gene to
>>>> report, why would the model from 4125 to 6400 ever be reported in place of
>>>> the one from 6111 to 8345, especially since this is provided in the
>>>> model_gff file?
>>>>
>>>> Even when transcript TSA024184 is included, Maker 2.10 reports the
>>>> high-confidence gene from 611 to 8345.
>>>>
>>>> Any light you could shed would be helpful. Thanks!
>>>>
>>>>
>>>> --
>>>> Daniel S. Standage
>>>> Ph.D. Candidate
>>>> Computational Genome Science Laboratory
>>>> Indiana University
>>>>
>>>>
>>>> On Wed, Jun 4, 2014 at 3:17 PM, Carson Holt <carsonhh at gmail.com> wrote:
>>>>> Just eAED, but eAED can affects selection of ab initio results. For
>>>>> example reading frame match of protein evidence which also affects whether
>>>>> evidence from single_exon=1 and genes with single_exon protein evidence
>>>>> get kept. There is also the assumption that your alignments in GFF3 are
>>>>> are correctly spliced (like BLAT does). So giving blastn results as
>>>>> precomputed est_gff would create a lot of noise, since maker ignores
>>>>> blastn and is using it only to seed the polished exonerate alignments.
>>>>>
>>>>> --Carson
>>>>>
>>>>>
>>>>> From: Daniel Standage <daniel.standage at gmail.com>
>>>>> Date: Wednesday, June 4, 2014 at 1:11 PM
>>>>> To: Carson Holt <carsonhh at gmail.com>
>>>>> Cc: Maker Mailing List <maker-devel at yandell-lab.org>
>>>>> Subject: Re: [maker-devel] Filtering of ab initio gene models
>>>>>
>>>>> I do not provide Gap or Target attributes in the GFF3. Will this affect
>>>>> the AED as well, or just the eAED?
>>>>>
>>>>>
>>>>> --
>>>>> Daniel S. Standage
>>>>> Ph.D. Candidate
>>>>> Computational Genome Science Laboratory
>>>>> Indiana University
>>>>>
>>>>>
>>>>> On Wed, Jun 4, 2014 at 3:09 PM, Carson Holt <carsonhh at gmail.com> wrote:
>>>>>> Sure. that would be helpful. One question. Do you provide the Gap
>>>>>> attribute in your precomputed alignments? Having or not having that
>>>>>> attribute affects the eAED score which takes reading frame into account,
>>>>>> and may cause some things to be kept that normally would be dropped,
>>>>>> because MAKER won't be able to take the points of mismatch of the
>>>>>> alignment into account (it just assumes match everywhere).
>>>>>>
>>>>>> --Carson
>>>>>>
>>>>>>
>>>>>> From: Daniel Standage <daniel.standage at gmail.com>
>>>>>> Date: Wednesday, June 4, 2014 at 1:03 PM
>>>>>> To: Maker Mailing List <maker-devel at yandell-lab.org>
>>>>>> Subject: [maker-devel] Filtering of ab initio gene models
>>>>>>
>>>>>> Thanks everyone for your responses recently!
>>>>>>
>>>>>> The reason for my recent flurry of email activity is that I'm seeing some
>>>>>> unexpected trends when running the new version of Maker with precomputed
>>>>>> alignments. Compared with an annotation I did a while ago (Maker 2.10,
>>>>>> Maker-computed alignments), this new annotation has a substantial number
>>>>>> of new genes annotated. If I compare distributions of AED scores between
>>>>>> the old and new annotation, it's clear that the new annotation has a lot
>>>>>> more low-quality models. If I look at new gene models that do not overlap
>>>>>> with any gene model from the old annotation, the likelihood that it's a
>>>>>> low-quality model is much higher.
>>>>>>
>>>>>> I decided to run a little experiment. I annotated a scaffold first using
>>>>>> Maker 2.10 and then using Maker 2.31.3. I both cases, I used the same
>>>>>> pre-computed transcript and protein alignments and the same (latest)
>>>>>> version of SNAP as the only ab initio predictor. Maker 2.10 predicted 44
>>>>>> genes while Maker 2.31.3 predicted 63. If we group gene models into loci
>>>>>> by overlap, there are 33 loci with gene models from both 2.10 and 2.31.3,
>>>>>> 1 locus with only models from 2.10, and 28 loci with only models from
>>>>>> 2.31.3.
>>>>>>
>>>>>> Before this experiment, I assumed the issue was related to providing
>>>>>> pre-computed alignments in GFF3 format and perhaps violating some
>>>>>> important assumption. However, this experiment makes me wonder whether
>>>>>> there have been changes to how Maker filters ab initio gene models
>>>>>> between version 2.10 and version 2.31.3? Do you have any ideas? If it
>>>>>> would help, I could put together a small data set that reproduces the
>>>>>> behavior I just described.
>>>>>>
>>>>>> Thanks!
>>>>>>
>>>>>> --
>>>>>> Daniel S. Standage
>>>>>> Ph.D. Candidate
>>>>>> Computational Genome Science Laboratory
>>>>>> Indiana University
>>>>>> _______________________________________________ maker-devel mailing list
>>>>>> maker-devel at box290.bluehost.comhttp://box290.bluehost.com/mailman/listinf
>>>>>> o/maker-devel_yandell-lab.org
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://yandell-lab.org/pipermail/maker-devel_yandell-lab.org/attachments/20140606/d48d28ac/attachment-0003.html>
More information about the maker-devel
mailing list