Quality-Control of Next-Generation Sequencing Data

点击翻译按钮获取中文摘要

Next-generation DNA sequencing (NGS) allows the entire genome of an organism to be sequenced quickly (within days) and affordably (less than $10,000). Continued technological advances are reducing both time and cost, and soon this approach will become viable for routine commercial applications. NGS has allowed researchers to understand evolutionary history, improve and optimize agricultural practices and better understand fundamental properties of the cell. In the future, creative entrepreneurs will develop clever new commercial applications that leverage cost-effective NGS.******But today each individual sample requires hands-on analysis. Future production use of NGS will require automated pipelines to facilitate processing of large sample numbers and improved quality-control to ensure outputs remain within acceptable tolerances. As in other areas of industry, this will likely involve application and extension of techniques from the fields of control theory and statistical process control. To date the quality-control and optimization of NGS experiments is under-studied.******To address these issues we developed SeqControl, a framework for evaluating NGS data quality. Across 53 human genomes SeqControl was nearly perfect in identifying low-quality experiments early, before they incurred significant costs. Our long-term goal is to create a robust quality-control system for NGS data. To reach that goal we have three aims, focused on the major gaps in the NGS data analysis.******First, we will create a global standard for optimizing NGS data analysis. We will create a well-understood "gold standard" dataset and release it to groups around the world in a "Challenge", with an public leaderboard to identify the best methods. We have already successfully used this approach for other problems, attracting hundreds of participants.******Our second objective will enhance the SeqControl statistical model. Our initial formulation employs supervised machine-learning. While successful, this approach does not incorporate non-numerical information (e.g. technician or protocol), require large training datasets and are computationally intense. We will create on-line machine learning techniques, where predictive models are continually updated based on new information.******Finally we will develop new ways to optimize and merge different analysis methods using the "wisdom of the crowds" -- the idea that a group of analysis methods are superior to any single method. We will exploit this in NGS analysis for the first time.******Rapidly declining costs are bringing NGS closer to routine commercial use. A robust statistical process-control framework is urgently needed to predict quality, optimize experimental design and rapidly identify low-quality data. We will create global standards to improve the quality and reliability of hundreds of millions of dollars worth of NGS studies in-progress in Canada and world-wide.