Evolutionary Machine Learning of Higher Order Relationships in Genome-wide Sequence Analysis

Abstract

One of the basic research goals in life science is to understand the complex relationships between biological factors and phenotypes, and to identify the various factors affecting the phenotype. In particular, genomic sequences play a significant role in determining the phenotype, such as gene expression and a susceptibility to disease, so the studies for the fundamental information stored in genome is essential to understanding biological processes. Previous genomic sequence analyses mainly focused on identification of a single associated factor or pairwise relationships with significant effects. Recent development of high-throughput technologies has made it possible to identify the causal factors by carrying out genome-wide analysis. However, it still remains as a challenge to discover higher-order interactions of multiple factors because this involves huge search spaces and computational costs.

In this dissertation, we develop effective methods for identifying the higher-order relationships of sequence elements affecting the phenotype, by combining statistical learning with evolutionary computation. The methods are applied to finding the associated combinatorial factors and dysfunctional modules in various genome-wide sequence analysis problems. Firstly, we show statistical learning-based methods to detect co-regulatory sequence motifs and to investigate combinatorial effects of DNA methylation, affecting on downstream gene expression. Next, to examine the sequence datasets with a huge number of attributes on human genome, we apply evolutionary computation approaches. Our methods search the problem feature space based on machine learning techniques using training datasets in evolutionary computation processes and are able to find candidate solution well in computationally expensive optimization problems. The experimental results show that the approaches are useful to find the higher-order relationships associated to disease using genomic and epigenomic datasets. In conclusion, our studies would provide practical methods to analyze complex interactions among sequence elements in genomic/epigenomic studies.