Prediction of bacterial virulent protein sequences has implications for identification and characterization of novel virulence-associated factors, finding novel drugs/vaccine targets against pathogens, and understanding the complex virulence mechanism in pathogens.

The prediction of virulent proteins will aid studies aimed at knowing more about bacterial virulence and annotation of (unknown) virulent genes for the identification of novel antimicrobial targets. Hence, in the present study, an attempt has been made to develop reliable SVM-based method for the automated prediction of virulent proteins. The training was carried out using non-redundant dataset of virulent and non-virulent proteins and evaluation with 5-fold cross-validation technique. In addition independent datsets were also used for evaluating the unbiassed performace of different SVM module. User can download the sequences of different datasets (Training and Independent datasets) by clicking the link DATA
The different individual modules were based on features such as -Compositions (amino acid, dipeptide and higher order dipeptides), evolutionary information in the form of multiple sequence alignment and similarity search. Finally, the performance of cascade SVM module, developed using individual feature modules was found to be much more efficient, hence used as default option for the VirulentPred server.

  Details about CASCADE SVM

Sometimes, machine learning techniques are unable to handle the noise produced due to the large/complex number of input units/patterns, which further, effects their classification efficiency. However, this problem can be overcome by the construction of very important module i.e. cascade SVM. In the present study, two-layered cascade approach-based SVM module was constructed. The brief description about each layer is as follows

 First layer

In the first layer 5 modules based on protein features such as amino acid compositions, dipeptide composition, amino acids composition of divided protein sequences, PSSM profiles and similarity-search were developed. These modules gave SVM predicted scores and similarity-search based information for each sequence.

 Second layer

The second layer received the binary scores output generated by 5 best modules constructed in the first layer to train second layer SVM model. Here, SVM was provided with a vector of 7 dimensions (1 for amino acid composition, 1 for dipeptide composition, 1 amino acids composition of divided protein sequences, 1 for PSSM and 3 for similarity-search based results). Hence, second layer correlates the predicted information of the first layer models to provide final output as shown in the picture below