False Positives Reduction in Top-down Protein Informatics using Support Vector Machines

The scarce but consistent chance of getting falsepositive matches [1], [2] in protein database search [3] hasalways casted a shadow over the reliability of results. Thesituation can be helped by viewing the protein data froma descriptive and the probabilistic framework, together.Using the conventional approach as the first stage, top downprotein data is descriptively searched for proteins and theresults are scored and ranked, using a top down proteinsearch engine. We then suggest applying Support VectorMachine, (SVM) as a second stage probabilistic scoringsystem, to the first stage protein database search results so asto further enhance protein classification. For SVM scoring,features are extracted from the top down data and a featuretable is constructed. An SVM using Radial Basis Functionis trained with this feature table. Later classification isperformed on the test data using this SVM. The classificationcan then be viewed together with the previously calculatedsearch engine score and a reordering of top ranked proteinsmay be done.