| As an important experimental technology, DNA microarray provides an effective way to measure the expression levels of tens of thousands of genes simultaneously under different conditions, which makes it possible to investigate the gene activities of the whole genome. However, computational challenges have to be faced as a result of the large volume of generated data. In this dissertation, two important applications of microarray data, i.e., genetic regulatory networks inference and cancer classification, are addressed with machine learning and neural networks.; Genetic regulatory network (GRN) inference is important for investigating gene functions and understanding their relations. In this dissertation, a recurrent neural network (RNN) and particle swarm optimization (PSO) approach is presented to infer GRNs from time series gene expression data. The proposed method aims to deal with the insufficiency of previous computational methods in dealing with the temporal behavior of the data and capturing the complex nonlinear dynamics of the gene regulation. The results demonstrate that the RNN/PSO can effectively capture the nonlinear dynamics of the gene expression time series and is capable of revealing regulatory interactions between genes. Different from traditional morphological appearance-based methods, a cancer identification system consisting of semi-supervised ellipsoid ARTMAP (ssEAM) and PSO is presented in this dissertation for cancer identification based on the corresponding gene expression profiles. The effectiveness of ssEAM/PSO for multi-class cancer diagnosis is demonstrated by testing it on three benchmark cancer data sets. Furthermore, gene expression profiles are extended for patient survival analysis, with a hybrid system combining probabilistic neural networks (PNN) and PSO. |
