Originally posted by Damian Kao in the old forum on 08.07.2015 - 10:44
At the end of section5 of the readme, there are some instructions for retraining hint parameters. The warning for the section said it is for internal use.
Is this not recommended for end users to run? I've tried running the command given in the readme with my training gb and hint files, but it kept giving me an error that has something to do with the extrinsic.cfg file. I've tried specifying the extrinsic.cfg file via --extrinsicCfgFile, but it still seems to be using the extrinsic.cfg from the software's config folder.
Is this feature worth using? And how do I make it work?
retraining hint parameters
Moderator: bioinf
Re: retraining hint parameters
Originally posted by mario in the old forum on 08.07.2015 - 11:14
I had implemented procedure a long time ago and changed my viewpoint on this in the meantime.
If you just want to annotate a singe or a few genomes I rather recommend to vary the extrinsic.cfg file manually. For example you can create copies that have 1e2,1e4,1e6 as intron bonus, similarly for other parameters. Then run augustus with the different extrinsic.cfg variants on a browser and visually inspect the predictions that differ against the evidence and maybe closer look at some examples, for example by BLASTing the predicted protein sequence.
The problem is that a training set is likely to be biased, e.g. towards expressed genes. And the automatic procedure could therefore chose parameters to reproduce the bias, e.g. to not predict genes without hints.
I had implemented procedure a long time ago and changed my viewpoint on this in the meantime.
If you just want to annotate a singe or a few genomes I rather recommend to vary the extrinsic.cfg file manually. For example you can create copies that have 1e2,1e4,1e6 as intron bonus, similarly for other parameters. Then run augustus with the different extrinsic.cfg variants on a browser and visually inspect the predictions that differ against the evidence and maybe closer look at some examples, for example by BLASTing the predicted protein sequence.
The problem is that a training set is likely to be biased, e.g. towards expressed genes. And the automatic procedure could therefore chose parameters to reproduce the bias, e.g. to not predict genes without hints.