From single cells to populations: generalized pseudotime analysis to identify patient trajectories from cross-sectional data in cancer genomics

癌症通过获得突变在遗传水平上不断进化，随后导致正常细胞活动的重新编程并最终导致功能异常。每个患者的癌症演变都是独特的，即使他们是同一类型，尽管他们会分享一些核心相似之处。为了了解癌症如何随着时间的推移而演变，理想的做法是进行研究，其中随着时间的推移对个体患者进行随访，并不断获得肿瘤样本，以了解正在进行的分子变化。在实践中，这在逻辑上是不可能的，也是不道德的，因为进行多次侵入性手术以获得活检对患者来说既昂贵又痛苦，并且不能为了能够对疾病进行前瞻性监测而拒绝治疗。最实用的临床研究涉及从大量患者中获得单个肿瘤活检（或在同一手术期间收集的多个活检）。这种随机患者人群的横截面特征不会给我们关于个体患者的疾病如何演变的直接信息，但我们可以将患者的所有信息联合收割机结合起来，以识别似乎具有相似疾病轨迹的个体亚组。也就是说，假设我们有两个病人，一个处于疾病的晚期，有许多突变，另一个处于相对较早的疾病阶段，两个人都有一组相似的核心突变。如果不治疗，晚期患者的分子状态可能是早期患者未来分子谱的指标。本项目提出开发新的统计机器学习算法，将这种逻辑和原理应用于整合横断面研究中患者全基因组测序分析获得的分子图谱，以识别和学习未直接观察到但可能留下线索的时间信息。我们将这些算法应用于国家基因组学英格兰100，000个基因组项目，旨在未来几年内对一系列癌症类型的数万个癌症基因组进行测序。该项目将深入了解癌症如何演变，重要的是提供了一种开发基于分子信息的预后指标的方法，以告诉我们患者疾病的严重程度及其可能的轨迹。

Cancer continually evolves at the genetic level through the acquisition of mutations that subsequently lead to the reprogramming of normal cellular activity and ultimately abnormal function. The evolution of cancer in each patient is unique, even when they are of the same type, although they will share some core similarities. In order to understand how cancers evolve over time, it would be ideal to conduct studies where individual patients are followed over time and samples of tumours continually obtained to understand the molecular changes that are ongoing. In practice, this is both logistically impossible and unethical as multiple invasive surgeries to obtain biopsies would be both costly and distressing to patients and treatment cannot be withheld to enable prospective monitoring of the disease. The most practical clinical studies involve obtaining a single tumour biopsy from a patient (or multiple biopsies collected during the same surgery) for a large number of patients. This cross-sectional profile across a random patient population would not give us direct information about how the disease of individual patients evolve but we could combine all the information across the patients to identify sub-groups of individuals who appear to have similar disease trajectories. That is, suppose we have two patients, one at an advanced stage of disease with many mutations and another who presents at a relatively earlier disease stage, and both share a similar set of core mutations. The molecular status of the advanced patient could be an indicator of the future molecular profile of the early stage patient, if left untreated. This project proposes to develop novel statistical machine learning algorithms to apply such logic and rationale to integrate molecular profiles obtained from whole genome sequencing analysis of patients in a cross-sectional study to identify and learn temporal information that is not directly observed but may leave tell-tale clues behind.We will apply these algorithms to the national Genomics England 100,000 Genomes Projects which seeks to sequence tens of thousands of cancer genomes across a range of cancer type over the next few years. The project will give insight into how cancers evolves and importantly provide a means of developing prognostic indicators that are based on molecular information to tell us the severity of a patient's disease and their likely trajectories.