Machine learning approaches for the analysis of circulating tumour DNA

肿瘤细胞产生于细胞的基因变异（通常是突变），导致细胞不受限制的分裂。识别这些基因变异，并随着时间和治疗对其进行监测，将使我们能够在早期发现癌症，针对特定的肿瘤类型定制治疗，并确定治疗反应或缓解。然而，通常很难获得肿瘤组织，通常需要活检，这可能是侵入性的，需要知道潜在肿瘤的确切位置。这使得在一段时间内获得多个肿瘤样本变得不可能。所有细胞都会释放游离细胞DNA （cfDNA）到血液中，然而肿瘤细胞通常会释放更多的循环肿瘤DNA （ctDNA）。这种ctDNA包含与肿瘤细胞相同的遗传变异，因此与其他cfDNA略有不同。识别ctDNA将提供一个易于获取的样本，使我们能够在任何时间点检测肿瘤的存在，并确定肿瘤的精确遗传组成。下一代测序可用于捕获所有cfDNA。确定其中有多少（如果有的话）是ctDNA是很复杂的，因为ctDNA可以占所有cfDNA的1%到90%，而且可能只有极少数的遗传变异只发生在ctDNA中。在一系列癌症类型中对ctDNA的研究在数量和规模上都在增加，该项目将使用机器学习方法来改进ctDNA的检测并确定其临床效用。首先，机器学习将用于学习一组特征，以区分ctDNA中发现的遗传变异与所有cfDNA中发现的遗传变异。匹配肿瘤和cfDNA数据的数据集将提供理想的训练集，因为肿瘤数据将包含仅在ctDNA中可见的一组遗传变异。其次，一旦肿瘤中存在的基因变异被确定，学生将评估特定突变的重要性，或所有基因变异的集合，例如突变的总数，突变的类型，或cfDNA是ctDNA的比例。为此，我们将测试这些特征是否可以用于预测临床因素，如癌症类型、治疗反应或预后。有越来越多的研究关注ctDNA在一系列癌症类型中的作用，我们将使用统计荟萃分析来结合多个研究的数据。该学生将开发机器学习方法并将其应用于前沿的基因测序数据。这个项目有可能改善癌症的检测和诊断，并为治疗提供见解。该学生将受益于利物浦大学转化医学研究所统计遗传学和药物遗传学小组的多学科团队，并与正在开发测序技术的工业合作者密切合作。

Tumour cells arise from genetic variants (usually mutations) in cells, which lead to unrestrained cell division. Identifying these genetic variants, and monitoring them over time and treatment, would allow us to detect cancer at an early stage, tailor treatment to specific tumour type, and determine treatment response or remission. However, it is often difficult to obtain tumour tissue, usually a biopsy is required which may be invasive and involves knowing the exact location of the potential tumour. This can make it impossible to obtain multiple tumour samples over time. All cells release cell free DNA (cfDNA) into the blood stream, however tumour cells often release a much larger amount, known as circulating tumour DNA (ctDNA). This ctDNA contains the same genetic variants as the tumour cells, so will be slightly different from other cfDNA. Identifying ctDNA would provide an easily accessible sample that would allow us to detect tumour presence at any time point and determine the precise genetic composition of the tumour. Next generation sequencing can be used to capture all cfDNA. Determining how much of it (if any) is ctDNA is complex since ctDNA can comprise between around 1% and 90% of all cfDNA, and there may be only a very small number of genetic variants which occur only in ctDNA. Studies of ctDNA across a range of cancer types are increasing in number and size, and the project will use machine learning approaches to improve detection of ctDNA and to determine its clinical utility. Firstly machine learning will be used to learn a set of features which distinguish genetic variants found in ctDNA from those found in all cfDNA. Datasets which have matched tumour and cfDNA data will provide an ideal training set, since the tumour data will contain the set of genetic variants which will be seen only in ctDNA. Secondly, once the genetic variants present in the tumour have been identified, the student will then assess the importance of specific mutations, or of the set of all genetic variants, for example the total number of mutations, type of mutations, or fraction of cfDNA that is ctDNA. To do this, we will test whether any of these features can be used to predict clinical factors such as cancer type, response to treatment, or prognosis. There are an increasing number of studies looking at ctDNA across a range of cancer types, and we will use statistical meta-analysis to combine data from multiple studies. The student will develop machine learning methods and apply them to cutting edge genetic sequencing data. This project has the potential to improve cancer detection and diagnosis, and to provide insights into treatment. The student will benefit from a multidisciplinary team, based in the Statistical Genetics and Pharmacogenetics group within the Institute of Translational Medicine at the University of Liverpool, and also work closely with industrial collaborators who are developing sequencing technology.