Classification Of Gene Expression Data Of Tumor Microarray Based On Intelligent Optimization Algorithm

The conventional methods of tumor classification present highly subjective experience as they are based on the clinical observation, anatomy and pathematology. With the development of biology and biologic technology, the tumor research increasingly relay on the application of molecular biology, in which the microarray gene expression data is employed as a potent platform for study. In order to classify those tumor expression data, some conventional and intelligent optimization algorithms have been proposed. However, there are short of stable classification algorithm with higher accuracy for the tumor samples which limits their application.In this thesis, the TLPSO classification algorithm based on the concept of different layers as well as swarms and the MOA classification algorithm based on the iterative search of global as well as local optimization atoms are established to overcome the faultiness of the conventional and intelligent optimization classification algorithms applying in the classification of the microarray gene expression data. In order to evaluate the performance of the TLPSO and the MOA, they are employed to classify the testing data came from American UCI machine learning database which contains Heart, Vote, Iris and Wine dataset. After that, they are applied to the tumor testing data derived from the broad institute of MIT and Harvard containing leukemia, DLBCL and multi-tumor gene expression data. Lung, breast, prostate and color tumor samples consist of the multi-tumor gene expression data. The classification algorithm based on the GA and the PSO is also established. The classification results of the tumor testing data obtained by the TLPSO and the MOA are compared with those based on the GA and the PSO.It can be seen from the experimental results that the TLPSO and MOA classification algorithm expresses good classification performance for UCI datasets. The TLPSO is surpassing the MOA in classification stability and convergence as the results of TLPSO differ slightly during the alteration of dataset and the convergence process of the TLPSO is quick than that of the MOA. Although the experimental results vary slightly for different tumor gene expression datasets, both TLPSO and MOA could get favorable classification accuracy, stability and convergence in different experiment dataset. There is no significant impact of the parameters on the performance in leukemia dataset for TLPSO and MOA, while in DLBCL there is some influence.Comparing with GA and PSO, the TLPSO and MOA achieves satisfying results in tumor expression data with small number of training samples. When the number of training samples increase, the MOA could get high accuracy while the accuracy of TLPSO descents. GA gets bad results for each subgroup. The accuracy of PSO is also declined with the raising of training samples.In conclusion, the effectiveness of the TLPSO is proved as it solve the problem that particles easily trap in local optimal and the MOA is also proved by introduction of the global and local search structure. The TLPSO and MOA could get good classification results in tumor gene expression samples which could provide objective results to tumor classification.