Computational methods to improve genome assembly and gene prediction

Abstract

DNA sequencing is used to read the nucleotides composing the genetic material that forms individual organisms. As 2nd generation sequencing technologies offering high throughput at a feasible cost have matured, sequencing has permeated nearly all areas of biological research. By a combination of large-scale projects led by consortiums and smaller endeavors led by individual labs, the flood of sequencing data will continue, which should provide major insights into how genomes produce physical characteristics, including disease, and evolve. To realize this potential, computer science is required to develop the bioinformatics pipelines to efficiently and accurately process and analyze the data from large and noisy datasets. Here, I focus on two crucial bioinformatics applications: the assembly of a genome from sequencing reads and protein-coding gene prediction.

In genome assembly, we form large contiguous genomic sequences from the short sequence fragments generated by current machines. Starting from the raw sequences, we developed software called Quake that corrects sequencing errors more accurately than previous programs by using coverage of k-mers and probabilistic modeling of sequencing errors. My experiments show correcting errors with Quake improves genome assembly and leads to the detection of more polymorphisms in re-sequencing studies. For post-assembly analysis, we designed a method to detect a particular type of mis-assembly where the two copies of each chromosome in diploid genomes diverge. We found thousands of examples in each of the chimpanzee, cow, and chicken public genome assemblies that created false segmental duplications.

Shotgun sequencing of environmental DNA (often called metagenomics) has shown tremendous potential to both discover unknown microbes and explore complex environments. We developed software called Scimm that clusters metagenomic sequences based on composition in an unsupervised fashion more accurately than previous approaches. Finally, we extended an approach for predicting protein-coding genes on whole genomes to metagenomic sequences by adding new discriminative features and augmenting the task with taxonomic classification and clustering of the sequences. The program, called Glimmer-MG, predicts genes more accurately than all previous methods. By adding a model for sequencing errors that also allows the program to predict insertions and deletions, accuracy significantly improves on error-prone sequences.