New Statistical Methods for Medical Signals and Images

对计算生物学和医学的大型数据集的分析代表了一个重要的挑战 对于统计学家来说是很困难的。这些数据通常具有大量相关特征，但相关特征相对较弱 用于预测感兴趣的表型的信号。此类数据的示例包括 DNA 序列和 GWAS， 质谱、MRI 和 EEG 图像、RNAseq 和蛋白质阵列等等。此次活动的总体目标是 正在进行的三名研究者资助是为了开发和研究增强分析的统计技术 以及这些数据的解释。我们的团队结合了统计建模、算法开发方面的经验 并对这些技术进行理论分析。在新项目中，我们的重点是开发 最先进的方法来利用已知或隐含的结构，以便从中提取有用的信息 高维数据。 此次更新将通过四个具体目标来实现这些目标。研究人员将研究： 1. 染色质结构建模的主曲线。我们提出新的统计方法 基于 Hi-C 测定得出的接触图对 DNA 染色质结构进行建模。我们使用 受主曲线启发的技术，但应用于度量缩放的背景下，考虑到 考虑沿染色体的局部结构。 2. 将稀疏模型拟合到大数据和汇总数据。许多现代数据集（例如 GWAS 1M SNP 和 500K 受试者）在计算上具有挑战性。我们提出计算进步 使套索能够扩展到此类场景。已发表的 GWAS 研究的作者通常会这样做 出于隐私和其他原因，不共享原始数据。我们提出了近似拟合的技术 这些模型的多变量版本仅给出通常报告的单变量汇总分数。 3. 估计病毒学和遗传学中的高维特征结构。我们将利用低阶结构 序列数据的真实性，以比较推断病毒蛋白部分的不同方法。对于 数量遗传学，我们将开发用于特征分析的统计理论、方法和软件 多层次的变异，特别是遗传协方差矩阵。 4. 利用辅助信息进行预测。许多研究都在寻找可预测的生物标志物特征 结果，例如各种治疗下的疾病状况。我们提出了一种统计方法 利用辅助信息，例如基因途径的成员资格或每个途径的定量测量 生物标记，以提高在这些具有挑战性的领域中发现特征的能力。 研究人员和他们的学生将共同努力，将新的统计工具应用到公共领域。 非常可用的软件，遵循本次资助的早期周期中建立的模式，其中我们的软件包 已在斯坦福大学和世界各地的医学研究人员中广泛使用。

The analysis of large datasets from computational biology and medicine represents an important chal- lenge for Statisticians. These data typically have a large number of correlated features with relatively weak signals for predicting phenotypes of interest. Examples of such data includes DNA sequences and GWAS, mass-spectra, MRI and EEG images, RNAseq and protein arrays, to name a few. The broad goal of this ongoing three-investigator grant is to develop and study statistical techniques that enhance the analysis and interpretation of these data. Our team combines experience in statistical modeling, algorithmic devel- opment, and theoretical analysis of these techniques. In the new projects, our focus is the development of state-of-the art methods to exploit known or implied structure in order to extract useful information from high-dimensional data. The renewal will address these goals through four Specific Aims. The investigators will study: 1. Principal curves for modeling chromatin architecture. We propose new statistical methodology for modeling the chromatin structure of DNA based on contact maps derived from Hi-C assays. We use techniques inspired by principal curves, but applied in the context of metric scaling, that take into account local structure along the chromosome. 2. Fitting sparse models to large data and to summary data. Many modern datasets (e.g. GWAS with 1M SNPs and 500K subjects) are computationally challenging. We propose computational advances that enable the lasso to scale to such scenarios. Often the authors of published GWAS studies do not share the raw data for privacy and other reasons. We propose techniques for approximately fitting multivariate versions of these models given only the univariate summary scores typically reported. 3. Estimating high-dimensional eigenstructure in virology and genetics. We will exploit low rank struc- ture in sequence data to compare different methods for inference about sectors in viral proteins. For quantitative genetics, we will develop statistical theory, methods and software for eigenanalysis of multiple levels of variation, and specifically for genetic covariance matrices. 4. Prediction with side information. Many studies seek biomarker signatures that are predictive of outcomes such as disease status under various treatments. We propose a statistical approach for exploiting side information such as membership in gene pathways or quantitative measures for each biomarker in order to increase the power for discovering signatures in these challenging domains. Working together, the investigators and their students will implement the new statistical tools into publi- cally available software, following a pattern established in earlier cycles of this grant, in which our packages have found wide use among medical researchers both at Stanford and around the world.