Integrative Probabilistic Models for Identifying Transcriptional Modules

描述（由申请人提供）： 转录模块（transscriptional modules，TM）是一组共调控基因，沿着调控其表达的转录因子。基于实验数据和基因组调控序列的TM识别是生物医学中的一个重要而困难的问题。可用于重建TM的数据来自全基因组基因表达谱实验、全基因组转录因子结合实验、实验建立的DNA调控基序序列和基因调控区序列。 在鉴定和表征TM的过程中使用所有可用类型的数据的益处已在许多研究中得到证实。虽然精确的概率模型通常确实存在用于单独分析不同数据类型，但用于所有可用数据类型的统一模型很少。生物医学研究人员目前可用的计算方法是不够的，要么是由于缺乏适当的计算工具，要么是由于基础数学框架的不足。此外，用于建立不同数据类型的联合建模的不同策略的相对效益的协议是不存在的。这使得生物医学研究人员在选择最佳数据分析方法时无法做出明智的决定。 我们建议开发无限转录模块（ITM）的框架组成的一个新的概率模型和相关的计算工具，用于识别转录模块，通过联合建模基因表达和调控数据。统一的概率模型将利用无限混合模型机制来对具有不同模块数量的模型进行平均，从而避免估计“正确”模块数量的问题。每种不同的数据类型将在特定于上下文的无限混合模型的不同上下文中单独建模。这种模块化方法将有助于使用最适当的概率模型来表示不同类型的数据。我们的目的不是为不同的数据类型开发新的模型和分析方法。相反，我们将专注于开发一个原则性的概率框架，用于集成当前可用的单个数据类型的最新模型。我们假设，我们的统一建模方法将导致显着更高的精度识别的转录模块比它将通过单独分析不同的数据类型，或通过应用目前可用的算法进行联合分析。我们还期望，后验分布的共同成员在TM，基于我们的模型，将提供可靠的评估确定TM的统计意义。使用真实的世界数据;我们将构建数据集和协议，客观地比较TM重建的不同方法的关键性能方面。

DESCRIPTION (provided by applicant): Transcriptional modules (TM) are groups of co-regulated genes along with transcriptional factors regulating their expression. Identifying TMs based on experimental data and genomic regulatory sequences is an important and difficult problem in biomedicine. The data that can be used in reconstructing TMs comes from genome-wide gene expression profiling experiments, whole genome transcription factor binding experiments, sequences of experimentally established DNA regulatory motifs and sequences of gene regulatory regions. Benefits of using all available types of data in the process of identifying and characterizing TMs have been demonstrated in numerous studies. While precise probabilistic models generally do exist for analyzing different data types separately, unifying models for all available data types are scarce. Computational methods currently available to biomedical researchers are inadequate either due to the lack of appropriate computational tools, or due to inadequacies of underlying mathematical framework. Furthermore, protocols for establishing relative benefits of different strategies for joint modeling of different data types are non-existent. This leaves biomedical researchers without means to make an informed decision when choosing the optimal data analysis approach. We propose to develop Infinite Transcriptional Modules (ITM) framework consisting of a novel probabilistic model and related computational tools for identifying transcriptional modules by jointly modeling gene expression and regulatory data. The unifying probabilistic model will utilize the Infinite Mixtures Model mechanism for averaging over models with different number of modules and thus circumvent the problem of estimating the "correct" number of modules. Each different data type will be modeled separately within different context of a Context Specific Infinite Mixture Model. Such modular approach will facilitate the use of the most appropriate probabilistic models for representing different types of data. Our intention is not to develop new models and analytical approaches for different data types. Instead, we will focus on developing a principled probabilistic framework for integrating currently available state of the art models for individual data types. We hypothesize that our unifying modeling approach will result in significantly higher precision of identified transcriptional modules than it would be achieved by either separately analyzing different data types, or by applying currently available algorithms for joint analysis. We also expect that the posterior distribution of co-membership in a TM, based on our model, will offer credible assessment of statistical significance of identified TMs. Using real world data; we will construct datasets and protocols for objectively comparing key performance aspects of different methods for TM reconstruction.