Optimization and Machine Learning Algorithms for Condition-Specific Biological Network Construction and Analyses

Abstract

Network bioinformatics has been successfully used to help reveal the complex mechanism<br/> of cells in research and industrial domains of diverse fields such as medical<br/> science, the pharmaceutical industry, biological science, and agriculture. In network<br/> bioinformatics, myriad of studies relied on the static association between interacting<br/> nodes to construct the network or to discover novel knowledge from networks. However,<br/> biological associations are condition-specific, which means that they can change<br/> dynamically depending on the status of the cell. Thus, the needs for algorithms to<br/> address these technical challenges have been increased. In this context, I developed<br/> stochastic optimization (SO) and machine learning (ML) algorithms with sequencing<br/> data that can capture status of cells. The main focus of my work is on ensemble SO<br/> and ML approaches with network bioinformatics.<br/> The first work is miRNA-mRNA regulation inference algorithm called PlantMirnaT<br/> to construct condition-specific miRNA-mRNA bipartite network. This is a challenging<br/> problem since it needs to consider expression data of two different types,<br/> miRNA and mRNA, and target relationship between miRNA and mRNA is not clear,<br/> especially when microarray data is used. Fortunately, due to the low sequencing cost,<br/> small RNA and RNA sequencing are routinely processed and it may be able to infer<br/> regulation relationships more accurately. To fully leverage the power of sequencing<br/> data, I proposed a parameterized miRNA-mRNA expression model based on a novel<br/> idea named split-ratio, and utilized genetic algorithm and the quasi-newton method<br/> to determine optimal model parameters.<br/> The second work is to discover functional subnetworks of miRNA-mRNA regulation<br/> network that work in specific biological condition. These subnetworks are named<br/> functional miRNA-mRNA regulatory module (MRM). Mining functional MRM has<br/> exponential time complexity, thus heuristic algorithm is needed to discover optimal<br/> MRM set. My algorithm operates in two steps: 1) grouping and ordering the miRNAs<br/> and mRNAs to build per sample matrices representing miRNA-mRNA regulations,<br/> and 2) determining maximum sized modules from structured miRNA-mRNA matrices.<br/> The third work is to use deep learning method for network bioinformatics. Deep<br/> learning has shown a great potential to address the various learning problems. However,<br/> deep learning technologies conventionally use grid-like structured data, thus<br/> application of deep learning technologies to the classification of human disease subtypes<br/> is yet to be explored. Recently, graph based deep learning techniques have<br/> emerged, which becomes an opportunity to leverage analyses in network biology. I<br/> propose a hybrid model, which integrates two key components 1) graph convolution<br/> neural network (graph CNN) and 2) relation network (RN). I utilize graph CNN as<br/> a component to learn expression patterns of cooperative gene community, and RN as<br/> a component to learn associations between learned patterns. The proposed model is<br/> applied to the synthetic dataset and PAM50 breast cancer subtype classification task,<br/> the standard breast cancer subtype classification of clinical utility. In experiments of<br/> both subtype classification and patient survival analysis, my algorithm achieve the<br/> better result in both quantitative and qualitative perspectives than previous methods.