Gene expression temporal patterns classification with hierarchical Bayesian neural networks and time lagged recurrent neural networks

DNA microarray technology allows simultaneous measurement of thousands of mRNA concentrations from a single cell across different conditions or over time. By systematically investigating thousands of genes in parallel and by monitoring the time dependence of expression levels we can study various patterns of gene expression profiles. Often cell functions and status can be determined from their patterns of expression. Highly accurate classification of gene patterns is crucial for revealing relationships among the genes and the genes with diseases in the presence of environmental hazards. Identification of subsets of relevant information (experimental conditions or genes), which include high correlations and overwhelming interactions is difficult, but critical. Handling the high level noise involved in the measurements and the presence of uncertainties in the modeling process pose further challenges for the task at hand. Moreover, microarray gene expressions typically follow multiple complicated dynamic patterns. These interesting challenges inspirited this thesis to investigate and explore automated learning systems with the combination of advanced statistical techniques to facilitate the characterization of large scale gene temporal patterns according to known functions of the relative gene expression.; In this work, Hierarchical Bayesian Neural Networks and Time Lagged Recurrent Neural Networks with appropriate data preprocessing for information selection, noise estimation and reduction are presented for characterizing the multiple gene expression temporal patterns. We investigate Automatic Relevance Determination with Bayesian regularization algorithm and an algorithm, which employs dynamic trajectory learning with back-propagation through time to deal with dynamic data and other complicated features in order to avoid overtraining and improve the generalization performance. A new Hierarchical Bayesian Neural Network with correlated weight structure is developed and implemented to model the correlation of multidimensional gene data. With the hierarchical Bayesian setting, the network parameters and hyperparameters were simultaneously optimized. By optimizing regularized performance functions and statistical criteria, such as Bayesian Information Criteria, the optimal network architecture for modeling gene expressions is learned. The model performance of the proposed methods was compared to other popular machine learning methods such as Nearest Neighbor, Support Vector Machine, and Self Organized Map.