UTR training sensitivity not improving

[katharina]

Originally posted by PeterM in the old forum on 14.02.2014 - 15:30

Hi everyone,

I am using optimize_augustus.pl (Augustus V2.7) to train parameters. The option --UTR=on was set. For training I use 1000 manually curated genes with UTR.

I am surprised to see that on a test set the sensitivity of the UTRs after training is always worse than to the pre-computed A. thaliana parameters.
The specifictiy however did improved. For testing I always run augustus with "--UTR=on".

Also surprisingly there was no difference when running optimize_augustus.pl with or without the option "--UTR=on".

If I ran optimize_augustus.pl with "--UTR=on --trainOnlyUtr=1" I got an error message.
IntronModel::readProbabilities: Error reading file Tr1_intron_probs.pbl
..Could not read the accuracy values out of predictions.txt when processing bucket 1. at optimize_augustus.pl line 795.

How can I improve the UTR sensitivity?
Is "--UTR=on" set per default or why there is no difference between setting it or not in optimize_augustus.pl.

Many thanks,
PeterM

[katharina]

Originally posted by PeterM in the old forum on 28.02.2014 - 12:18

The error message probably was because I was accidentally optimizing one parameter set twice.

I still have no Idea why the UTR sensitivity never improves after training and why despite not training the UTR (by not setting "--UTR=on"), UTR sensitivity is still worse compared to arabidposis parameters. I am highly confident that my training genes have correct UTRs since they have strong support from expression data. To me is seems that there is something wrong with the UTR training algorithm.
Since UTRs are generally poorly conserved I would expect an increase of UTR sensitivity when training with genes from the same species. Has someone any idea why this might not be the case?

[katharina]

Originally posted by Anand KS Rao in the old forum on 19.08.2014 - 06:16

Hi Peter,
I just came across your post.
I have had similar issues, see my post with values from pre- and post-training (recent post, with same user name).
Have you had any resolution to your problem? If yes, could you share here on this forum?
Thanks!