The high-throughput sequencing technologies have been developed and applied to the human genome studies for nearly 20 years. These technologies have provided numerous research applications and have significantly expanded our knowledge about the human genome. In this thesis, computational methods that utilize sequence data to study human genomic variants and transcripts were evaluated and developed.

Indel represents insertion and deletion, which are two types of common genomic variants that are widespread in the human genome. Detecting indels from human genomes is the crucial step for diagnosing indel related genomic disorders and may potentially identify novel indel makers for studying certain diseases. Compared with previous techniques, the high-throughput sequencing technologies, especially the next- generation sequencing (NGS) technology, enable to detect indels accurately and efficiently in wide ranges of genome. In the first part of the thesis, tools with indel calling abilities are evaluated with an assortment of indels and different NGS settings. The results show that the selection of tools and NGS settings impact on indel detection significantly, which provide suggestions for tool selection and future developments.

In bioinformatics analysis, an indel’s position can be marked inconsistently on the reference genome, which may result in an indel having different but equivalent representations and cause troubles for downstream. This problem is related to the complex sequence context of the indels, for example, short tandem repeats (STRs), where the same short stretch of nucleotides is amplified. In the second part of the thesis, a novel computational tool VarSCAT was described, which has various functions for annotating the sequence context of variants, including ambiguous positions, STRs, and other sequence context features. Analysis of several high- confidence human variant sets with VarSCAT reveals that a large number of genomic variants, especially indels, have sequence features associated with STRs.

In the human genome, not all genes and their transcripts are translated into proteins. Long non-coding ribonucleic acid (lncRNA) is a typical example. Sequence recognition built with machine learning models have improved significantly in recent years. In the last part of the thesis, several machine learning-based lncRNA prediction tools were evaluated on their predictions for coding potentiality of transcripts. The results suggest that tools based on deep learning identify lncRNAs best.