Comparative Analyses and Genomic Sequence-Only Prediction of DNA Breakpoints via Gradient Boosting Machines and Deep Learning

基因组插入/缺失改变通过形成DNA链断裂而发生，是仅次于点突变的第二大DNA修饰。然而，与点突变不同的是，DNA链断裂点的长程和短程序列背景依赖性，以及基因组中断裂倾向的详细区域变异，还没有通过尖端的计算手段进行广泛的研究。这种情况普遍存在，因为关于这种DNA改变的现有数据相对稀少，再加上DNA链断裂的复杂多病因性质，这将使本已稀少的数据进一步分层。然而，理解DNA断裂的重要性导致了许多计算工作概述了CNV断裂点与非B DNA构象、频繁的SCNA断裂点与癌症基因、诱导性DNA断裂与染色质堆积和基本位点的潜在关联。基于CNV数据开发的离散值双变量隐马尔可夫模型可以对人类基因组中约400个断裂点达到~300BP的预测分辨率，显示出至少对CNV的DNA断裂点进行更高分辨率的基于序列的预测的前景。在过去的十年里，见证了用于实验性基因组范围DNA断点图谱的特定测序方法的出现，尖端的机器学习技术开始变得容易获得，最近在生物大数据方面的出色应用。因此，现在是时候将最先进的机器学习技术应用于来自经历不同生理、自发和病理过程的组织的大量扩展和分层的DNA断点数据，以揭示它们精细的短期和远程序列依赖关系，以及它们在不同过程中的共性和差异，从而导致DNA链断裂。在整个项目中，学生将获得并实施计算生物学、高级数据分析、机器学习和统计学方面的技能(“定量技能”)，同时使用基因组数据集，并整合从热力学和核酸化学考虑的许多自下而上的发展(“跨学科技能”)。

Genomic insertion/deletion alterations, which occur through the formation of DNA strand breaks, are the second most significant DNA modifications after point mutations. However, unlike point mutations, the long- and short-range sequence-context dependence of DNA strand breakpoints, as well as the detailed regional variation of breakage propensities in the genome, have not been extensively interrogated through cutting edge computational means. This has prevalently been the case because of the relative sparseness of available data on such DNA alterations, coupled with complex multi-etiologic nature of DNA strand breaks that would further stratify already sparse data. Nevertheless, the importance of understanding DNA breakages led to a number of computational works outlining potential associations of CNV breakpoints vs. non-B DNA conformations, frequent SCNA breakpoints vs. cancer genes, induced DNA breaks vs. chromatin packing and abasic sites. Discrete-valued bivariate hidden Markov model developed on CNV data could reach a predictive resolution of ~300bp for about 400 breakpoints in the human genome, showing the promise for higher resolution sequence- based prediction of DNA breakpoints for, at least, CNVs. The past decade has witnessed the advent of specific sequencing methodologies for experimental genome-wide mapping of DNA breakpoints, and, cutting edge machine learning techniques started to become readily available with recent outstanding applications on biological big data. It is therefore the right time to apply the state-of-the-art machine learning techniques on the substantially expanded and well stratified DNA breakpoint data from tissues undergoing different physiological, spontaneous and pathological processes, in order to reveal their fine short- and long-range sequence dependences, with their commonalities and differences across different processes leading to DNA strand breaks. Throughout the project, the student will gain and implement skills in computational biology, advanced data analysis, machine learning and statistics ("Quantitative skills"), while working with genomics datasets and integrating many bottom-to-top developments from thermodynamic and chemical considerations of nucleic acids ("Interdisciplinary skills").