Applying Back Propagation Algorithm for classification of fragile genome sequence

Abstract

Most frequently occurring recurrent chromosomal translocation allied with all subtype of leukemia are available in Mitel Mann Data base. We have retrieved about 55 such genome sequence from TIC dB data base with 100% similarity score and got noncoding sequence of chromosome 9 and 22 as positive example of fragile site. Another 55 housekeeping genome sequence is taken for classification purpose. For content based analysis we have extracted 20 features of frequency density of mono nucleotide and dinucleotide. The network is designed by determining hyper parameters like number of hidden layer, hidden neurons and input features. First we took 20 input features and there after 16 for reducing number of free parameters (i.e. weight space). Network is also pruned for succeeding experiments. The training strategy was also exhaustively explored, based on literature study and trial and error heuristic methods to achieve more and more accuracy. Regularization is also employed by cross validation and early stopping. We have achieved 95% accuracy for training data and 70% to test data in first experiment. To avoid this over fitting at last we could achieve 93% over all accuracy and outlier detection, too. We could be able to show that dinucleotide frequency density is important statistical feature for classifying genome sequence. This classifier can show the probability of fragility to occur in genome sequence at very early stage so as to deal with the diesis at prognosis phase.