Denoising and Interaction Learning of Biological Data

Abstract

Since the Human Genome Project was completed, enormous biological data have been accumulated as an attempt to understand the biological mechanisms of human. However, errors induced during the sequencing procedures and unrevealed inherent features of biological data for inferring their interactions arouse the necessity of large-scale data-driven applications. In this regard, this dissertation exploits the recent advances in machine learning and artificial intelligence techniques that have shown their success in time series sequence learning, including natural language processing and neural machine translation, to improve the reliability and computational performance of investigating biological data. This dissertation discusses three issues in sequence analysis and proposes methodologies to overcome them. First, to alleviate the error-prone nature of sequence reads from next-generation sequencing (NGS), we present an information theoretic approach for correcting sequence errors from various sequencers. Next, we show a generalized multi-graphics processing units (GPUs) accelerated sequence denoiser to address the computational challenges of denoising high-throughput sequences. Finally, we describe an end-to-end machine learning framework for robust sequence (e.g., miRNA) target prediction to boost the sensitivity without the laborious manual feature extraction procedure. In summary, this dissertation proposes a set of methodologies on the basis of machine learning algorithms to handle biological sequences that can boost the reliability of downstream analysis.