Managing and Modeling Time in Genomics Data

了解与观察到的生物表型相关的机制是许多基因组研究的共同目标。然而，通常情况下，在单个时间点测量分子实体不足以捕捉许多生物系统的复杂性；真正系统的测量需要考虑跨越时间和空间的动态变化。随着成本和样本需求的降低，纵向基因组数据正在迅速积累。然而，帮助研究人员在基因组学研究中建模和理解时间变化的工具在很大程度上是缺失的。该研究计划的长期目标是开发工具和算法来处理涵盖广泛时间尺度或频率的时间变化。具体来说，我们将要开发的工具和算法分为三个“层”：(a)数据层；(b) a建模层；(c)解释层。数据层侧重于数据质量和离群值管理；建模层以可扩展的方式对基因组数据进行建模和分析；解释层通过将用户与在线出版物和讨论中嵌入的知识联系起来，帮助研究人员解释他们的数据和分析结果。关于数据层，“垃圾输入，垃圾输出”。如果手头的数据不是高质量的，那么后续的建模和解释就不太可能有成果。我们建议开发方法来评估质量，并从新基因组技术和平台（包括下一代测序）产生的数据中去除随机或系统噪声。这里的一个关键思想是，可以利用数据点之间的时空关系来评估和提高数据质量。识别异常样本也很重要。可以利用数据的时间序列性质来获得高质量的模型，从中可以标记出偏差样本以供进一步检查。我们还计划找出可能的原因，解释为什么某些样本被认定为偏远地区。对于建模层，出于建模和解释的原因，时间序列经常需要增加潜在变量。此外，观测结果通常是从混合分布中获得的，这些分布通常是先验未知的。我们提出发展非参数贝叶斯模型来学习复杂时间序列的潜在结构。一个关键的挑战是建模多组数据的大量特征。因此，我们将重点关注可伸缩或可并行的方案。基因组学数据可能基于细胞的异质亚群。因此，为了更有效地建模，重要的是开发方法来反卷积底层组成，以更好地捕捉时变动力学。从基因组学数据中有效地发现知识需要领域专家在解释数据和模型方面的大量参与。解释层侧重于将用户链接到在线出版物中嵌入的知识的工具。例如，使用全基因组技术鉴定的涉及基因/肽的关系对研究人员解释其结果很有价值。基本的google风格搜索不足以取代更复杂的自然语言处理，从文本中提取关系，包括时间表达式，以及实体之间的时间或顺序关系（例如，基因）。随着社交网络在过去十年中得到广泛使用，研究博客也在基因组学和医学研究社区中得到了发展。我们建议开发方法来总结研究人员之间的在线讨论。一个新颖的想法是探索如何生成抽象摘要来响应用户给出的查询。

Understanding the mechanisms associated with observed biological phenotypes is a common goal of many genomic studies. However, it is often the case that measuring molecular entities at a single time-point is insufficient to capture the complexity of many biological systems; truly systematic measurement needs to consider dynamic changes across time and space. With decreasing cost and sample requirements, longitudinal genomic data are accumulating quickly. Yet tools that help researchers to model and understand temporal changes in genomics studies are largely missing. The long-term objective of the proposed research program is to develop tools and algorithms to handle temporal changes covering a wide range of time scales or frequencies. Specifically, we categorize the tools and algorithms to be developed into three “layers”: (a) a Data layer; (b) a Modeling layer; and (c) an Interpretation layer. The Data layer focuses on data quality and outlier management; the Modeling layer focuses on modeling time and analyzing genomics data in scalable ways; and the Interpretation layer helps researchers interpret their data and the outcomes of their analyses by linking users to knowledge embedded in online publications and discussions.Regarding the Data layer, “Garbage in, garbage out.” If the data at hand are not of high quality, it is unlikely that subsequent modeling and interpretation will be fruitful. We propose to develop methods for assessing the quality of, and removing random or systematic noise from, data generated from newer genomics technologies and platforms, including next-generation sequencing. A key idea here is that the space-time relationships between the data points can be exploited to assess and enhance data quality. It is also important to identify samples that are outliers. The time series nature of the data can be exploited to derive quality models, from which deviant samples can be flagged for further examination. We also plan to identify possible reasons that explain why certain samples are identified as outlying. Regarding the Modeling layer, time series often need to be augmented with latent variables for modeling and interpretation reasons. Moreover, the observations are typically obtained from a mixture of distributions, which are often unknown a priori. We propose to develop non-parametric Bayesian models to learn the latent structure of complex time series. One key challenge is the large number of features for modeling multi-omic data. Thus, we will focus on scalable or parallelizable schemes. Genomics data may be based on heterogeneous sub-populations of cells. Thus, for more effective modeling, it is important to develop methods to deconvolve the underlying composition to better capture the time-varying dynamics. Effective knowledge discovery from genomics data requires significant involvement of domain experts in interpreting the data and models. The Interpretation layer focuses on tools for linking users to knowledge embedded in online publications. For example, relationships involving genes/peptides identified using whole genome technologies are valuable to researchers in interpreting their results. Basic Google-style searches are not sufficient to replace more sophisticated natural language processing to extract relations from text, including temporal expressions, and temporal or sequential relationships among entities (e.g., genes). As social networking has gained widespread use in the past decade, research blogging has also been growing in the genomics and medical research communities. We propose to develop methods for summarizing online discussions among researchers. One novel idea is to explore how to produce abstractive summaries in response to user-given queries.