instructions.md

MiRduplexSVM instructions

Train - Test MiRduplexSVM

1. Download miRBase data (miRNA.dat) from mirbase.org/ftp.shtml. We have test the code on versions 17 (included in the download) and 19.
2. Put the .dat file to MiRduplexSVM/code/input/data folder
3. Go in folder MiRduplexSVM/code, and run script init.m. The performed steps are printed in matlab's command window. Note:

• The fold of the cross validation are set in the second cell.
• Only human and mouse hairpins are selected to train – test the algorithm. You can change this in cell 6, line 76.

4. Run script runexpCrossVal.m to optimize hyper parameters employing 5 fold cross validation.

• In the first cell, the user can set SVMs hyperparameters. Only the polynomial kernel can be used. The default parameters are the ones used in the MiRduplexSVM publication Karathanasis et al., 2015
• Second cell trains the SVMs
• In the third cell the user should re-set SVMs hyperparameters.
• The forth cell tests the models which were produced from the second cell.
• The fifth cell generate figures with several metrics to evaluate performance.

5. Run script runexpHoldOut.m to train and test the final model using a hold out set.

• In the second cell the user should provide the desired parameters. The default parameters are the ones used in the MiRduplexSVM publication Karathanasis et al., 2015
• The last cell generate figures with several metrics to evaluate performance.

Results

The .mat file with the actual numbers of the performance metrics can be found in the MiRduplexSVM/code/Results folder.

The cumulative distributions of the errors can be found by following the steps below: - Load a _CumFreq_10.mat file - Duplex errors, (similar to figure 3), are included in the meanAbsErrorMeanCumRelFreq double. - k55, k53, k35, k33 errors (similar to figure 4) are included in the f5p5pMeanAbsErrorCumRelFreq, f5p3pMeanAbsErrorCumRelFreq, f3p5pMeanAbsErrorCumRelFreq, f3p3pMeanAbsErrorCumRelFreq doubles, respectively.

Thank you for using MiRduplexSVM!