Discovering Novel Structural Genomic Rearrangements Using Deep Neural Networks

摘要 准确检测基因组中的结构变异是一项具有挑战性的任务。许多方法都是 在过去的几十年里发展起来的，但据估计，仍有数万个变种被遗漏 在给定的样本中。由于使用短读测序的限制，许多这些变体被遗漏 确定较大的变种。尽管许多这些缺失的变异体位于 基因组方面，已经表明一些仍然具有临床意义，这使得他们的发现变得重要。新的 已经开发了使用长阅读对基因组进行测序的平台，并展示了克服这些问题的希望 许多这些限制创造了识别简单和复杂结构变体的全谱的能力。 由于这项技术相对年轻，新的计算方法支持长期阅读的分析 测序数据可以帮助发现这些仍然被遗漏的变异。除了检测 在样本中具有长时间读取测序数据的新变异，可以开发计算方法来 利用这些新颖的变体调用重新分析当前数十万个短读数据集 可用。在这项提案中，我们计划开发新的计算方法来识别新的结构变异 在基因组中。在目标1中，我们将应用递归方法来分析长读取序列数据集 深度神经网络。在目标2中，我们将开发一个工具来推导长期预测的结构变体的轮廓。 可用于识别短读取数据集中的结构变体调用并对其进行基因分型的读取。一起， 这些方法将使研究人员能够准确地表征多头和短头的结构变化。 读取数据集。

Abstract Accurately detecting structural variation in the genome is a challenging task. Many approaches have been developed over the last few decades, yet it is estimated that tens of thousands of variants are still being missed in a given sample. Many of these variants are missed due to the limitations of using short-read sequencing to identify large variants. Although many of these missed variants are located within complex regions of the genome, it has been shown that some still have clinical relevance making their discovery important. New platforms have been developed for sequencing the genome using long-reads and show promise for overcoming many of these limitations creating the ability to identify the full spectrum of simple and complex structural variants. Because this technology is relatively young, new computational approaches to support the analysis of long-read sequencing data can aid in the discovery of these variants which are still being missed. In addition to detecting novel variation in samples with long-read sequencing data, computational approaches can be developed to leverage these novel variant calls to reanalyze the hundreds of thousands of short-read datasets currently available. In this proposal, we plan to develop new computational approaches to identify novel structural variation in the genome. In Aim 1, we will apply a recurrence approach to analyze long read sequencing datasets utilizing deep neural networks. In Aim 2, we will develop a tool to derive profiles of structural variants predicted in long- reads which can be used to identify and genotype structural variants calls in short read data-sets. Together, these approaches will allow researchers to accurately characterize structural variation in both long and short- read datasets.