Dear mGene User,

This is the first public release of mGene version 0.1 beta.  This
software is the result of further development of mGene as described in
[1] and is similar to the implementation used for mGene.web described
in [2]. 

The main focus of the improvements relative to [1] is to make it
easier to use, the software more stable and applicable to a broad
range of organisms. The current version of mGene does not support:

1. use of whole genome alignments (features used in
   mGene.multi in [1])

2. use of alignment of ESTs, cDNAs or proteins (features used in
   mGene.seq in [1])

3. Trans-splicing, operons and polyA sites. These feature are
   primarily important for nematode genomes and are often not necessary
   for other organisms.

4. Model selection: All models are trained with preset
   hyper-parameters. Model-selection can significantly improve the
   results of the predictions.

5. Prediction tuning: There are several ways to trade-off between
   sensitivity and specificity in gene prediction. However, right now
   these expert options are not supported in this version.

There are some additional limitations in this release:

1. Currently, we implemented some inherent settings for sub-sampling
   examples to reduce the computation time. Enabling the use of all
   examples (as in [1]) leads to better results.

2. Reading genes for training from GFF3 files is tricky. We only
   accept a gene's annotation, if it satisfies certain, quite
   stringent conditions, leading to smaller sets for training.
   In some cases we try to infer missing information, like the open
   reading frame (which can be easily wrong). Training on poor-quality
   input data generally leads to a poor prediction performance.

3. Currently, intermediate files a relatively big and can easily addup
   to a few gigabytes for a genome like the one of
   C. elegans. Moreover, depending on the length of the
   chromosomes/contigs mGene may need considerable amounts of main
   memory.

4. Currently, mGene uses regions around annotated genes for training the gene 
   predictor, where we use some heuristic to determine where to cut
   intergenic regions. (These regions are assembled into blocks for
   training.) This strategies always works to a certain extend (even
   for missing genes in the annotation), but is inferior to using fully
   annotated regions (as done in [1]). There will be more options to
   control this behaviour in future versions.

5. Right now we use two versions of the shogun toolbox, leading to
   problems running the monolithic workflows in octave. This will be
   solvedx in the next release.

We are working towards solving these issues, to make the software more
stable and to include more features into mGene.

If you are interested in using mGene for annotating an organism's
genome or for comparing it with other predictions, please let us know
(support@mgene.org or Gunnar.Raetsch@tuebingen.mpg.de). We would be
glad to assist you to obtain the optimal results using mGene.


The mGene Development Team