readme

C-package of "Long Short-Term Memory" for Protein classification

Here one can assemble an individual experiment. To do this the following
steps are necessary:

1.) Determining sequences for positive training and test as well negative sequences
    for training and test. This yields to four files. The files must be in FASTA format.
    There are already four files in the dataset folder which can be filled accordingly.
    Also there are ready examples in the quichstartExperiment datasets folder.

2.) Copying the four files in to the dataset folder. The files should be named
    pos-train.fasta, pos-test.fasta, neg-train.fasta and neg-test.fasta for each
    of the four sequence types from step 1. However the names and locations of the four
    files can be configured also in the parameter file (see step 4).

    IMPORTANT: Currently the parameter file (see step 4) is configured for a maximum
    of 50000 trainings sequences and 10000 test sequences with a maximum length of
    1200 amino acids. One can change the parameters by editing the parameter file
    in the last four lines.
 
3.) Compiling the sources by running 'make' in the src folder.

4.) Choosing an architecture for the LSTM network by choosing a parameter file.
    E.g. 'lstmpars.mem12.ws9.txt' for a network with 12 memory cells and a window size
    of 9 amino acids.

5.) Starting the experiment by running e.g. choosing the parameter file from step 4:

    src/lstm -c lstmpars.mem12.ws9.txt


By default the parameter files are configured to stop the LSTM after 100 epochs. See
the line beginning with: 'stop learning after n epochs: 100'. The LSTM constantly gives
out the performance of the experiment for every epoch. The performance metrics are 
Mean Square Error (mse), false negatives (fn) and false postives (fp) in percent.
The performance is also written in to the file 'out.txt'.