Recent Changes - Search:

Data provide to benchmark a method

The current verion of BRASERO contains 5 benchamrks, for 5 RNA families.

See Benchmarking protocol for details on the structure of a benchmarking data set.

Noise sources

To generate sequences of set F, we use various sources:

  • random genome segments from viral genomes
  • random genome segments from encode sequence (hg18_dna range=chr5:131284314-132284313)
  • GenRGenS: a generator of random sequences or structures

Adding new RNA families

This is a short how-to explaining how we generated data sets for each family.

The R set

First we looked for 6 representants of each family, carefully chosen to be as representative as possible (i.e. sequence/structure are not very close within a family).

The T set

Then we selected a hundred sequences of RNA known to belong to the family.

The T2 set

To fold each sequence of the set T we used a python script build_benchmark_1.0.zip. With our sequences in stored in distinct files, we did fold them as follows:

for i in *.fasta
do 
        build_benchmark.py --mfold  --mfold-opt " P=10 W=10 MAX=5 " \
              -d "TRUE %f: %s => %e (%x) % %m %o %N" -f ../true_events.srp.dp $i
done

In this example, only mfold is used to fold the sequence but rnashapes and rnasubopt are also available. Please use build_benchmark.py -h to obtain information on how to use this script.

Note: the mfold parameter depends on the input

The F2 set:

We directly build the set F2 from the noise source (encode or viral genome) using the build_benchmark.py script. First extracted the lengths of each reference structure (for example 120,121,140,90,120,130).

If the noise source (sequence) is in a file named noise.fasta and we want to generate noise in the file noise.dp, we use the following command:

build_benchmark.py --mfold  --mfold-opt " P=10 W=10 MAX=5 " -d "%f: %s => %e (%x) % %m %o %N" -f noise.dp \
 --sequence-name noise.fasta noise.fasta  {::50}/90,{::100}/120,{::50}/121,{::50}/130,{::50}/140

This line tell build_benchmark.py to use noise.fasta as input and randomly choose 50 sequences of length 90, 100 of length 120, 50 of length 121, ... (see build_benchmark.py -h for a description of the parameters) This give 300 sequences for F (the size of F2 depends on the folders options).

Note: again, the mfold parameter depends on the input

Assembling the data together

Finally, we created two files for each benchmark: the first usually called reference_set contains the paths to references structures and the second, true_events_with_noise, contains the paths to T2 and F2 structures files.

Edit - History - Print - Recent Changes - Search
Page last modified on December 09, 2010, at 03:17 PM EST