Next-Generation Algorithms in Statistical Genetics Based on Modern Machine Learning

项目摘要/摘要 技术的进步使收集数以百万计的人类基因组的海量数据集成为可能。这个 这一提议背后的长期愿景是利用现代数据集和机器学习(ML)来 说明基因和环境如何决定对改善人类健康至关重要的特征和结果。 现代ML在数百万个非结构化数据点(基因组、临床笔记、图像)的海量数据集上蓬勃发展， 这将对统计遗传学产生重大影响，统计遗传学是研究基因-表型联系的领域。改善- 统计遗传学反过来有可能阐明疾病的遗传基础，并支持每一种疾病。 声学医疗疗法。 该建议通过两个方面推进了上述设想：(1)开发新颖的机器学习(ML)算法。 (2)为科学家创建开放源码软件系统 以及基于上述算法的临床医生。具体地说，我们描述了一个发展计算的计划-- 统计遗传学中三个广泛领域的传统方法：模型、连锁不平衡和结构 遗传变异，分析全基因组关联研究数据，并预测遗传和环境风险- 心理因素。在每个领域内，我们的目标是为关键的应用问题开发开源软件，包括- ING基因归因、单倍型、低通测序、因果变异鉴定和风险评分。 我们的研究旨在为基于现代ML的统计遗传学奠定基础，并推动 ML在其他应用程序领域中可能无法追求的方向。我们的方法将支持技术 在医疗保健方面有直接应用，并有助于揭示影响疾病的新的遗传因素。 简化；提高从预防医学到药物基因组学领域的基因组预测的准确性； 显著降低基因组测序分析的成本，并最终改善人类健康。

PROJECT SUMMARY/ABSTRACT Advances in technology are enabling the collection of massive datasets of millions of human genomes. The long-term vision underlying this proposal is to leverage modern datasets and machine learning (ML) to under- stand how genetics and the environment determine traits and outcomes important to improving human health. Modern ML thrives on vast datasets of millions of unstructured datapoints (genomes, clinical notes, images), and stands to greatly impact statistical genetics, the field which studies the genotype-phenotype link. Improve- ments in statistical genetics have in turn the potential to elucidate the genetic basis of disease and support per- sonalized medical therapies. This proposal advances the above vision via two thrusts: (1) developing novel machine learning (ML) algo- rithms motivated by problems in statistical genetics; (2) creating open-source software systems for scientists and clinicians based on the above algorithms. Specifically, we describe a plan for the development of computa- tional methods in three broad areas in statistical genetics: modeling linkage disequilibrium and the structure of genetic variation, analyzing genome-wide association study data, and predicting risk from genetic and environ- mental factors. Within in each area, we aim to develop open-source software for key applied problems includ- ing genetic imputation, haplotyping, low-pass sequencing, causal variant identification, and risk scoring. Our research seeks to establish a foundation for statistical genetics based on modern ML and also advance ML in directions that may not be pursued in other application domains. Our methods will support technologies that have immediate applications in healthcare and that help reveal novel genetic factors that influence dis- ease; improve the accuracy of genomic prediction in domains from preventive medicine to pharmacogenomics; significantly reduce the cost of genomic sequencing assays, and ultimately improve human health.