A machine learning approach to query time-series microarray data sets for functionally related genes using hidden markov models

Microarray technology captures the rate of expression of genes under varying experimental conditions. Genes encode the information necessary to build proteins; proteins used by cellular functions exhibit higher rates of expression for the associated genes. If multiple proteins are required for a particular function then their genes show a pattern of coexpression during time periods when the function is active within a cell. Cellular functions are generally complex and require groups of genes to cooperate; these groups of genes are called functional modules. Modular organization of genetic functions has been evident since 1999. Detecting functionally related genes in a genome and detecting all genes belonging to particular functional modules are current research topics in this field. The number of microarray gene expression datasets available in public repositories increases rapidly, and advances in technology have now made it feasible to routinely perform whole-genome studies where the behavior of every gene in a genome is captured. This promises a wealth of biological and medical information, but making the amount of data accessible to researchers requires intelligent and efficient computational algorithms. Researchers working on specific cellular functions would benefit from this data if it was possible to quickly extract information useful to their area of research. This dissertation develops a machine learning algorithm that allows one or multiple microarray data sets to be queried with a set of known and functionally related input genes in order to detect additional genes participating in the same or closely related functions. The focus is on time-series microarray datasets where gene expression values are obtained from the same experiment over a period of time from a series of sequential measurements. A feature selection algorithm selects relevant time steps where the provided input genes exhibit correlated expression behavior. Time steps are the columns in microarray data sets, rows list individual genes. A specific linear Hidden Markov Model (HMM) is then constructed to contain one hidden state for each of the selected experiments and is trained using the expression values of the input genes from the microarray. Given the trained HMM the probability that a sequence of gene expression values was generated by that particular HMM can be calculated. This allows for the assignment of a probability score for each gene in the microarray. High-scoring genes are included in the result set (of genes with functional similarities to the input genes.) P-values can be calculated by repeating this algorithm to train multiple individual HMMs using randomly selected genes as input genes and calculating a Parzen Density Function (PDF) from the probability scores of all HMMs for each gene. A feedback loop uses the result generated from one algorithm run as input set for another iteration of the algorithm. This iterated HMM algorithm allows for the characterization of functional modules from very small input sets and for weak similarity signals. This algorithm also allows for the integration of multiple microarray data sets; two approaches are studied: Meta-Analysis (combination of the results from individual data set runs) and the extension of the linear HMM across multiple individual data sets. Results indicate that Meta-Analysis works best for integration of closely related microarrays and a spanning HMM works best for the integration of multiple heterogeneous datasets. The performance of this approach is demonstrated relative to the published literature on a number of widely used synthetic data sets. Biological application is verified by analyzing biological data sets of the Fruit Fly D. Melanogaster and Baker.s Yeast S. Cerevisiae. The algorithm developed in this dissertation is better able to detect functionally related genes in common data sets than currently available algorithms in the published literature.