转录因子结合位点序列基元挖掘的计算方法研究

Transcription factors (TFs) can modulate gene expression patterns and hence are key components of cellular regulatory networks. TFs bind to DNA in a sequence-specific manner. The relative preferences of TFs to various nucleotide sequences are often referred to as TF binding site (TFBS) motifs. These motifs are of considerable interest to biological study, as they are central to understanding the mechanisms of gene expression. In this project, we systematically study the computational methods for the motifs discovery of transcription factor binding sites. Firstly, we propose a discriminative motif finder for discovering high quality initial motifs between two sequence datasets, which uses area under receiver-operating characteristic curve (AUC) as a measure of the discriminating power of motifs and incorporate novel search strategies. Secondly, we propose a new framework for estimating generative probabilistic motif models via a contrasting process, which can provably learn the optimal motif parameter by discriminating the observed binding data from samples from an adaptive noise distribution. Finally, we reformulate the discriminative motif finding problem into a multiple-instance learning framework, thereby more properly modeling the underlying inference problem and facilitating the incorporation of advanced machine learning and optimization tools. The development of this project will promote the understanding of the underlying mechanisms of regulation. It will also help to understand the cells from the system level and explain the pathogenesis of the disease.

转录因子能够调控基因表达的模式，因而是细胞调控网络的关键组成部分之一。转录因子和DNA序列间的结合关系是具有序列偏好性的，转录因子对于不同核苷酸序列的相对结合偏好通常被称为转录因子结合位点基元，由于它们在基因表达机制理解中的核心地位，对于生物学的研究具有极其重要的意义。在本项目中，我们将系统地研究转录因子结合位点基元的计算挖掘方法。首先，提出一种新的判别方法用于在两组序列间寻找高质量的基元初始解，这种方法采用受试者工作特征曲线下面积来判别解的判别能力，并引入了新颖的搜索策略。然后，我们将基元生成模型参数学习问题转化为对比训练过程，从而可以通过将观测数据和人工数据对比，最优地学习模型参数。最后，我们将判别基元模型训练转化为多示例学习问题，从而可以更适当地建模其潜在的推断问题，并方便引入新的机器学习和优化计算工具。本项目的成功实施将有助于认识调控的内在机制,并帮助进一步从系统层次来理解细胞活动