Exploiting the sequence and evolutionary information for the identification of virulence factors

The pathogenicity of a bacterium depends on its virulence factors (VF); and predicting the VF allows to understand the pathogenesis of infection. It can be used to identify a possible point of interference for disease treatment or vaccination. In present work, a machine-learning based prediction model for (VF’s was developed, using two different types of feature extraction: classical sequence-based features and evolutionary information-based features for the representation of VFs and non-VFs protein sequences. For splitting data into training and testing sets, uniform sampling based approach was used (Kennard- Stone algorithm) to create diversified and representative training/testing sets. For accurate prediction of VF’s, different machine learning algorithms (Random forest and Gradient boosting machine) and deep learning algorithms were used and further analysed using model agnostic interpretation methods. The highest accuracy of 84.5% was obtained by RF with PSSM feature set, followed by GBM with PSSM feature set (accuracy 83.9%). With feature fusion, the best performance evaluation metric was obtained by GBM with 92.0% sensitivity, 86.6% specificity, 89.3% accuracy, 0.788 mcc and 0.953 AUC on 10- fold cross validation. The feature importance was captured by Variable importance plots (VIP) and Shapely plots. Evolutionary information based features i.e. PSSM and PSE-PSSM are the two most important features in discriminating the VFs from non VFs.