Methods and Software to Enhance Genomic Privacy and Sharing of RNA-Seq Data

摘要 随着生物医学的广度和深度前所未有地增加，隐私正受到越来越多的关注 数据集，特别是个人基因组数据集。大多数关于基因组隐私的研究都集中在保护上 个人基因组中的变异。然而，分子表型数据集也可以包含大量 敏感信息。虽然没有明确的基因类型信息，但微妙的基因-表型 相关性可以用来在统计上将表型和基因数据集联系起来。我们会研究 从表型数据集中分析敏感信息泄漏的方法。我们将重点关注 RNA-seq数据集及其相关敏感信息泄漏源。这些泄漏是通过 通过数量性状基因座的表达。我们将在三个目标下进行隐私分析。我们首先要瞄准的是 提出可用于量化敏感信息泄露的统计指标 表型数据集。这些量化可以用来评估侵犯隐私的风险。在 第二个目标，我们将重点系统地分析如何实例化和分析链接攻击。我们 我将研究如何推广链接攻击，使隐私研究人员能够研究其风险 更有系统地与这些攻击相关联。然后我们将评估不同的基因模型 预测并评估如何将其用于链接攻击。我们将特别关注离群值基因 表达水平并评估如何将离群值用于基因预测和链接攻击。 在第三个目标中，我们将开发实现量化、风险评估和风险管理的工具 方法，并将这些方法集成到一个连贯的软件套件中，以进行全面的隐私分析， 能够在不同级别的数据集摘要中保护RNA-seq数据集，例如读取、基因 和文字记录量化。我们将致力于增加基因组隐私软件工具的数量 分析。我们将研究不同的算法方法来处理高计算复杂性的 文献中的匿名化技术。我们将研究敏感信息泄露的其他来源 基因表达水平，例如剪接和非编码转录。这些信息来源将被研究 在先前目标中提出的风险量化和管理战略的背景下。我们最终会 使用这些工具量化来自大型测序的公开可用数据集中的敏感信息 项目，例如ENCODE、1000基因组、TCGA、GEUVADIS和GTEx。

Abstract Privacy is receiving much attention with the unprecedented increase in the breadth and depth of biomedical datasets, particularly personal genomics datasets. Most studies on genomic privacy are focused on protection of variants in personal genomes. Molecular phenotype datasets, however, can also contain substantial amount of sensitive information. Although there is no explicit genotypic information in them, subtle genotype-phenotype correlations can be used to statistically link the phenotype and genotype datasets. We will study the methodologies for analysis of sensitive information leakage from phenotype datasets. We will focus on the RNA-seq datasets and the associated sources of sensitive information leakage. These leakages are mediated by the expression quantitative trait loci. We will approach the privacy analysis under 3 aims. We will first aim at proposing statistical metrics that can be used for quantification of the sensitive information leakage from phenotype datasets. These quantifications can be used to evaluate the risks of privacy breaches. In the second aim, we will focus systematical analysis of how linking attacks can be instantiated and analyzed. We will study how one can generalize linking attacks that enables the privacy researchers study the risks associated with these attacks more systematically. We will then evaluate different models of genotype prediction and assess how these can be used in linking attacks. We will focus, specifically, on the outlier gene expression levels and evaluate how the outliers can be used for genotype prediction and in the linking attacks. In the third aim, we will develop tools that implement the quantification, risk estimation, and risk management methodologies and integrate these in a coherent software suite for a comprehensive privacy analysis, which enables protecting RNA-seq datasets at different levels of summarizations of the datasets, e.g., reads, gene and transcript quantifications. We will aim at increasing the number of software tools for genomic privacy analysis. We will study different algorithmic approaches to tackle with the high computational complexity of anonymization techniques in the literature. We will study sources of sensitive information leakage other than gene expression levels, e.g. splicing and non-coding transcription. These sources of information will be studied in the context of risk quantification and management strategies presented in the previous aims. We will finally use the tools to quantify the sensitive information in the publicly available datasets from large sequencing projects, for example ENCODE, 1000 Genomes, TCGA, GEUVADIS, and GTex.