SVM-based prediction of the calpain degradome using Bayes Feature Extraction

Calpains belong to a family of calcium-dependent cysteine proteases which are implicated in a myriad of pathologies such as cancer and neurodegeneration. Despite extensive experimental studies on these proteases, our knowledge of the calpain degradome is still limited. Using a dataset of 341 unique, experimentally verified calpain cleavage sites, we conducted extensive sequence analyses and discovered novel residue propensities in the region flanking the cleavage site which could be modeled for prediction using machine learning algorithms. We have developed a series of computational models incorporating support vector machines and Bayes Feature Extraction for the prediction of calpain cleavage sites. The best models achieved AROC and accuracy scores ranging from 0.79 to 0.93 and 71% to 86% respectively when tested on independent test sets. We predicted calpain cleavage sites on proteins from the receptor tyrosine kinase family and discovered potential sites of cleavage at critical regulatory domains. The results suggest a novel role of calpains as a direct regulator of receptor tyrosine kinase activity in cell survival and cell death pathways.