Inferring ancestry and relatedness of human genomes using ancient DNA samples' and falls within the EPSRC Artificial Intelligence and Healthcare Techn

在每个基因组位置，两个个体通过导致共同祖先的系谱关系联系在一起。从个体到这个祖先的时间距离称为到最近共同祖先的时间(TMRCA)。这可以通过用树来表示它们的谱系关系来概括到一组个体。随着基因组的移动，随着基因组在减数分裂过程中被重组分解，树木的拓扑结构可能会发生变化。因此，一组样本的进化历史可以用一个称为祖先重组图(ARG)的图来简洁地表示，该图由跨越基因组不同块的单个树组成。有多种计算密集的方法可以从现代DNA样本的高质量测序数据中重建ARG。由于环境条件和污染，古代样本的质量下降；通常只能在非常低的覆盖率下对它们进行测序，这使得将它们纳入一组现代样本的ARG具有挑战性。在现代和古代DNA样本之间重建准确的联合ARG具有多种潜在的应用前景，我们的目标是探索。它可以用于祖先推断，通过恢复现代样本从各种古代祖先群体继承的祖先比例，使我们能够重建历史事件，如人口迁移。我们可以进一步利用ARG拓扑学来检测自然选择，通过定位基因组中与某些特定个体或古老群体异常共享的区域。寻找正或负选择区域，特别是具有已知生物功能的区域，在与医疗保健相关的应用中可能特别有用，例如，此类区域已被用来确定制药环境中的药物靶标。最后，通过测试某些群体的祖先是否与某些表型更紧密地相关，可以评估变异的表型影响。该项目的第一个目标是建立一个关联度推断算法，可以推断现代和古代样本之间的树状拓扑和TMRCA，并使用它来重建与英国生物库和其他来源的数据联合的ARG。跨样本对共享的远程染色体区域对于这种分析来说是信息丰富的，但在低覆盖率的古代DNA中很难检测到，因此我们的算法将需要隐式或显式地建模单倍型共享，尽管缺乏分相信息，或者在存在噪声的计算分相的情况下。对于这个算法，我们利用深度学习(DL)，它在过去十年中改变了许多科学领域。种群遗传学传统上专注于开发复杂的参数模型，但尚未从DL进展中显著受益。测序数据具有空间结构，因此可以将多个样本的序列堆叠在一起形成图像，并使用计算机视觉方法(例如卷积神经网络)进行分析，以调整样本顺序无关的事实(即需要可交换的网络)。由于我们已经可以访问现代样本的ARG，我们的目标是探索使用基于注意力和图形的方法来提取ARG信息，这些信息将有助于推断现代和古代样本之间的祖先。总的来说，我们希望做出两个主要贡献。第一个是算法开发，它将允许使用DL来重建现代和古代DNA样本的联合谱系树，为后者解决质量问题。第二种是使用真实的英国生物库和古代数据的联合ARG推理。然后，我们的目标是分析这一ARG，以回答与自然选择和具有某些古代群体祖先的表型影响有关的问题。

At every genomic position, two individuals are connected through genealogical relationships that lead to a common ancestor. The chronological distance from the individuals to this ancestor is termed time to the most recent common ancestor (TMRCA). This can be generalized to a set of individuals by representing their genealogical relationships by a tree. Moving along the genome, the topology of the trees can change as the genome is broken up by recombination during meiosis. Hence, the evolutionary history of a set of samples can be compactly represented by a graph, called the ancestral recombination graph (ARG), comprised by the individual trees spanning different chunks of the genome. There are multiple computationally intensive methods to reconstruct the ARG from high-quality sequencing data of modern DNA samples. Ancient samples are of degraded quality due to environmental conditions and contamination; usually they can only be sequenced at very low coverage making the task of incorporating them into the ARG of a set of modern samples challenging. Reconstructing an accurate joint ARG between modern and ancient DNA samples has multiple potential applications that we aim to explore. It can be used for ancestry inference, by recovering the ancestry proportion a modern sample inherits from various ancient ancestral groups, enabling us to reconstruct historical events such as population migrations. We can further exploit ARG topology to detect natural selection, by locating regions of the genome that are unusually shared from certain individuals or ancient groups . Finding regions under positive or negative selection, particularly with known biological functionality, can be especially useful in healthcare-related applications and such regions have, for example, been leveraged to determine drug targets in pharmaceutical settings. Finally, the phenotypic impact of variants can be evaluated by testing whether ancestry from certain groups is more closely related to certain phenotypes.The project's first goal is to build a relatedness inference algorithm that can infer tree topology and TMRCAs between modern and ancient samples and use it to reconstruct a joint ARG with data from the UK BioBank and other sources. Long-range chromosomal regions that are shared across pairs of samples are informative for this analysis but hard to detect in low coverage ancient DNA, so our algorithm will need to to implicitly or explicitly model haplotype sharing despite the lack of phasing information, or in the presence of noisy computational phasing. For this algorithm, we leverage Deep Learning (DL), which has transformed many scientific fields in the past decade. Population genetics has traditionally focused on developing complex parametric models and has not yet significantly benefited from DL advances. Sequencing data has a spatial structure, so sequences from multiple samples can be stacked to form an image and analysed using computer vision approaches (e.g. Convolutional Neural Networks),adjusting for the fact that sample order is irrelevant (i.e. require exchangeable networks). As we already have access to an ARG for modern samples, our aim is to explore the use of attention- and graph-based methods to extract ARG information that will help infer ancestry between modern and ancient samples.Overall, we expect to make two main contributions. The first is algorithmic development that will allow the use of DL for reconstructing joint genealogical trees for modern and ancient DNA samples, tackling quality issues for the latter. The second is joint ARG inference using real UK Biobank and ancient data. We then aim to analyse this ARG to answer questions relating to natural selection and phenotypic impact of having ancestry from certain ancient groups.