A data fusion approach in protein homology detection

ABSTRACTA DATA FUSION APPROACH IN PROTEIN HOMOLOGY DETECTIONAydın Can POLATKANBaskent UniversityComputer EngineeringMachine learning techniques are frequently and extensively used for classifyingproblems in the field of computational biology. These techniques require constantlength feature vectors as inputs. As far as it is known that proteins are in differentlengths, therefore all proteins are needed to be represented with a constant numberof features.One of the effective methods developed for this goal is n-peptite combinations of theprotein strings. These methods are represented with the availability percentage ofeach of the n-length substrings inside the sequence. To reduce the spacecomplexity, for increasing values of n, amino acid alphabet is reduced regularly forthe resulting feature vectors to conform available memory resources today.In this solution, all feature inputs were given to a single classifier. In this thesis, thesefeature inputs are classified into specific significant groups, according to the n-peptitecompositions and reduced amino alphabets. These groups are given to severaldifferent classifiers to achieve a data fusion approach with a few techniques that arewandering in the narrowed search space by abstraction. Aim is to have better resultswith techniques that are converging in exact and leading to different regions of asolution. In that approach, to evaluate the output values of different classifiers,various cases like averaging, weighted averaging and choosing the most successfulone in the training set are compared.Each of these methods was tested on remote homology detection problem which isone of the major and actual problems of computational biology and results arepresented relatively.As the results are considered, the case in which the output of the most successfultraining set is granted, observed as the more accurate one. To explore the statisticalsignificance of differences between results, paired samples T-tests were carried outbetween all methods. Furthermore, all data fusion approaches tested, through out thethesis has more efficient memory usage according to the single classifier case. Thedata fusion approach which has been tested with support vector machines is alsothought to be efficient for not only protein homology detection problems but alsoother problems of classification.Keywords: Protein Homology Detection, N-peptite Compositions, Support VectorMachines, Classification, Data Fusion.Supervisor: Hayri SEVER, Prof. Dr., Çankaya University, Department of ComputerEngineering