Robust, scalable, and accurate discovery of mutational signatures

从肿瘤基因组序列中推断出的突变特征有可能提供环境暴露的记录，并可以为致癌的病因提供线索。然而，为了使推断的特征具有生物学意义，每个特征必须准确地代表每个致突变过程中不同突变类型的贡献。使用非负矩阵分解（NMF）的启发式算法主要用于发现突变特征。但是这些方法不灵活，不健壮，并且需要大量的计算。拟议项目的目标是开发计算效率高的算法，尽管建模假设不完善，但可以发现生物学上有意义的特征。Aim 1通过开发可扩展、易于使用和准确的变分推理（一种广泛使用的近似贝叶斯推理方法）的新框架来支持这一目标，该框架适用于突变发现模型。目标2开发统计方法，从使用提议的变分推理框架获得的推论中提取生物学上有意义的特征。通过理论分析和对合成数据和实际数据的数值实验，确保了目标1和目标2中开发的方法的准确性和统计有效性。最后，目标3通过以下方式改进了当前对突变过程的理解：(1)将目标1和目标2中开发的方法应用于大型泛癌症数据集；(2)开发了一个新的模型，该模型允许结构化地合并单碱基和双碱基替换，以及每个特征中的插入和删除。所提出的工作很好地取代了用于发现有意义的数据表示的启发式方法，因此对如何分析其他基因组数据类型（如单细胞RNA-seq）具有长期影响。这项工作也与NIGMS直接相关，因为它属于“DNA和RNA代谢（修复）”，因为许多突变过程与异常DNA修复或与衰老相关的“时钟样”分子机制有关，这可以在组织学上正常的组织中观察到

The mutational signatures inferred from tumor genome sequences have the potential to provide a record of environmental exposure and can give clues about the etiology of carcinogenesis. However, for inferred signatures to be biologically meaningful, each signature must accurately represent the contribution of different mutation types in each mutagenic process. Heuristic algorithms using non-negative matrix factorization (NMF) have primarily been used to discover mutational signatures. But these approaches are inflexible, non-robust, and require massive amounts of computation. The objective of the proposed project is to develop computationally efficient algorithms that, despite imperfect modeling assumptions, can discover biologically meaningful signatures. Aim 1 supports this objective by developing a new framework for scalable, easy-to-use, and accurate variational inference – a widely used approach to approximate Bayesian inference – that is applicable to mutational discovery models. Aim 2 develops statistical methods to extract biologically meaningful signatures from the inferences obtained using the proposed variational inference framework. The accuracy and statistical validity of the methods developed in Aims 1 and 2 is ensured through theoretical analysis and numerical experiments on synthetic and real data. Finally, Aim 3 improves upon the current understanding of mutational processes by (1) applying the methods developed in Aims 1 and 2 to a large Pan-Cancer dataset and (2) by developing a novel model that allows for the structured incorporation of single-base and double-base substitutions, and insertions and deletions in each signature. The proposed work is well-positioned to replace heuristics used for discovering meaningful representations of data, and so have long-term impact on how other genomic data types such as single-cell RNA-seq are analyzed. This work is also directly relevant to the NIGMS as it falls under “DNA and RNA metabolisms (repair)” since many mutational processes are related to aberrant DNA repair or “clock-like” molecular mechanisms that are associated with aging, which can be observed in histologically normal appearing tissue