Computational Models For DNA Sequence Analysis And Prediction Of Anticancer Drug Sensitivities

It has been produced massive genomics data and anti-cancer drugomics data by high-throughput sequencing technology.Establishing an effective computing model is an impor-tant tool for quickly mining the information contained in these big data.The thesis constructed a 24-D feature vector which is composed of base transition prob-abilities,base contents and base position ratios.The vector was applied to compare complete coding sequences of ?-globin genes of 11 species and whole mitochondrial genomes of 18 eutherian mammals respectively.The derived phylogenetic trees were quite agreement with the evolutionary relationship.In addition,the essential genes of 28 bacteria were success-fully identified by support vector machine method.The average AUC value is 0.808,much higher than some other methods.The results of experiments demonstrated that the proposed three characteristics are alternative classifiers in related bioinformatics research.Accurate computational prediction of anticancer drug responses in cell lines can signif-icantly contribute to fulfilling the precision medicine in oncology.Many popular computa-tional models are powerless for predicting the sensitivity of 'new'(untested)drug versus'new'(untested)cell line.Using the CCLE and the GDSC as a benchmark data set the the-sis first demonstrated a general assumption,that is,genetically more similar cell lines always exhibit higher response correlations to structurally more related drugs.Next,constructed a cell line-drug complex network based on similar cell lines and similar drugs.It is natural that the prediction of new cell line-new drug could be made by the responses of their neighbor cell lines to neighbor drugs.Relying on this important result we constructed a simple com-putational model with intuitive interpretability,named CDCN model.Finally,we verified the prediction performance of the proposed model by leave one out method,derived good general anticancer drug response prediction and the highest Pearson correlation coefficient between predicted and observed response values reached 0.88,also obtained acceptable spe-cial prediction,i.e.,'new cell line-new drug' response prediction,and the highest Pearson correlation coefficient between predicted and observed response values is 0.61.Moreover,the CDCN model could correctly predict that MEK1/2 inhibitors are more sensitive to BRAF mutant cell lines than BRAF wild-type cell lines.The EGFR gene mutation inhibitor,La-patinib,can also be correctly predicted.All the results showed that CDCN model can be used as a tool for predicting anticancer drug responses,it may potentially save the cost of cell line-drug screening and contribute to optimizing exploration of new drugs.