Sadeqi Azer Erfan, Haghir Ebrahimabadi Mohammad, Malikić Salem, Khardon Roni, Sahinalp S Cenk
Department of Computer Science, Indiana University, Bloomington, IN 47408, USA.
Cancer Data Science Laboratory, Center for Cancer Research, National Cancer Institute, National Institutes of Health, Bethesda, MD 20892, USA.
iScience. 2020 Oct 7;23(11):101655. doi: 10.1016/j.isci.2020.101655. eCollection 2020 Nov 20.
Principled computational approaches for tumor phylogeny reconstruction via single-cell sequencing typically aim to build the most likely perfect phylogeny tree from the noisy genotype matrix - which represents genotype calls of single cells. This problem is NP-hard, and as a result, existing approaches aim to solve relatively small instances of it through combinatorial optimization techniques or Bayesian inference. As expected, even when the goal is to infer basic topological features of the tumor phylogeny, rather than reconstructing the topology entirely, these approaches could be prohibitively slow. In this paper, we introduce fast deep learning solutions to the problems of inferring whether the most likely tree has a linear (chain) or branching topology and whether a perfect phylogeny is feasible from a given genotype matrix. We also present a reinforcement learning approach for reconstructing the most likely tumor phylogeny. This preliminary work demonstrates that data-driven approaches can reconstruct key features of tumor evolution.
通过单细胞测序进行肿瘤系统发育重建的原则性计算方法通常旨在从噪声基因型矩阵构建最可能的完美系统发育树，该矩阵表示单细胞的基因型调用。这个问题是NP难问题，因此，现有方法旨在通过组合优化技术或贝叶斯推理来解决相对较小的实例。不出所料，即使目标是推断肿瘤系统发育的基本拓扑特征，而不是完全重建拓扑结构，这些方法也可能极其缓慢。在本文中，我们针对从给定基因型矩阵推断最可能的树是否具有线性（链状）或分支拓扑以及完美系统发育是否可行的问题，引入了快速深度学习解决方案。我们还提出了一种用于重建最可能的肿瘤系统发育的强化学习方法。这项初步工作表明，数据驱动的方法可以重建肿瘤进化的关键特征。
