Consistently predicting protein function based on MKL

Using Multiple Kernels Learning(MKL) to integrate heterogeneous data sources to train Support Vector Machine(SVM) classifier is becoming popular. For the protein function prediction problem, all the function categories form a directed acyclic graph(DAG), that is, Gene Ontology(GO). Given a protein to be predicted, after applying a trained SVM to output probabilistic prediction at each function category node, we use a cost-based model to consistently adjust function assignment on GO, which is called PredConsist/MKL. Experiments show PredConsist/MKL has higher ROC score than unadjusted method SDP/SVM, and better prediction performance. This adjustment is necessary.