Keratin protein property based classification of mammals and non-mammals using machine learning techniques

Keratin protein is ubiquitous in most vertebrates and invertebrates, and has several important cellular and extracellular functions that are related to survival and protection. Keratin function has played a significant role in the natural selection of an organism. Hence, it acts as a marker of evolution. Much information about an organism and its evolution can therefore be obtained by investigating this important protein. In the present study, Keratin sequences were extracted from public data repositories and various important sequential, structural and physicochemical properties were computed and used for preparing the dataset. The dataset containing two classes, namely mammals (Class-1) and non-mammals (Class-0), was prepared, and rigorous classification analysis was performed. To reduce the complexity of the dataset containing 56 parameters and to achieve improved accuracy, feature selection was done using the t-statistic. The 20 best features (parameters) were selected for further classification analysis using computational algorithms which included SVM, KNN, Neural Network, Logistic regression, Meta-modeling, Tree Induction, Rule Induction, Discriminant analysis and Bayesian Modeling. Statistical methods were used to evaluate the output. Logistic regression was found to be the most effective algorithm for classification, with greater than 96% accuracy using a 10-fold cross validation analysis. KNN, SVM and Rule Induction algorithms also were found to be efficacious for classification.