Fast and flexible Bayesian phylogenetics via modern machine learning

项目摘要/摘要 SARS-CoV-2大流行凸显了我们对全球病原体爆发的易感性和死亡人数。 病毒基因组的系统发育分析提供了对疾病病理生理学、传播和感染的关键见解。 潜在控制然而，如果这些方法要用于病毒控制策略，它们必须可靠地说明 不确定性，并能够在可行的时间内对1,000个基因组进行推断。缩放贝叶斯网络- 满足这种需要的ICS是一个巨大的挑战，不太可能通过优化现有算法来满足。 我们将用一种全新的方法来迎接这一挑战：贝叶斯变分推理， ics（VIP）使用系统发育树上的可变分布，这些系统发育树使用基于梯度的方法拟合，类似于 如何有效地训练大规模神经网络。通过采用变分方法，我们还可以 将系统发育分析集成到非常强大的开源建模框架中，如TensorFlow 还有PyTorch。这将开辟新的模型类别，如神经网络模型，以整合数据， 作为采样位置和迁移模式与系统发育推断。这些灵活的模型将告知 病毒控制策略。 在目标1中，我们将开发可扩展和可靠的VIP所需的理论，包括子树边缘， 在线算法、收敛诊断和参数支持所需的局部梯度更新 估算我们将在VIP的C++基础库中实现这些算法。在目标2中， 为遗传学开发一个灵活的基于TensorFlow的建模平台， 基于神经网络的系统发育模型，以最少的程序学习动态异质性- 明的努力。我们将通过我们的C++库为这个实现提供有效的梯度。在目标3中， 利用VIP后验是完整数据后验的持久和可扩展描述这一事实， 变分后验的动态在线计算，包括分治贝叶斯遗传学。 这项工作将使基于云的病毒遗传学解决方案能够快速更新我们目前对 当新数据到达或模型被艾德时的后验分布。 1

Project Abstract/Summary The SARS-CoV-2 pandemic underlines both our susceptibility to and the toll of a global pathogen outbreak. Phylogenetic analysis of viral genomes provides key insight into disease pathophysiology, spread and po- tential control. However, if these methods are to be used in a viral control strategy they must reliably account for uncertainty and be able to perform inference on 1,000s of genomes in actionable time. Scaling Bayesian phylogenet- ics to meet this need is a grand challenge that is unlikely to be met by optimizing existing algorithms. We will meet this challenge with a radically new approach: Bayesian variational inference for phylogenet- ics (VIP) using flexible distributions on phylogenetic trees that are fit using gradient-based methods analogous to how one efficiently trains massive neural networks. By taking a variational approach we will also be able to integrate phylogenetic analysis into very powerful open-source modeling frameworks such as TensorFlow and PyTorch. This will open up new classes of models, such as neural network models, to integrate data such as sampling location and migration patterns with phylogenetic inference. These flexible models will inform strategies for viral control. In Aim 1 we will develop the theory necessary for scalable and reliable VIP, including subtree marginal- ization, local gradient updates needed for online algorithms, convergence diagnostics, and parameter support estimates. We will implement these algorithms in our C++ foundation library for VIP. In Aim 2 we will develop a flexible TensorFlow-based modeling platform for phylogenetics, enabling a whole new realm of phylogenetic models based on neural networks to learn phylodynamic heterogeneity with minimal program- ming effort. We will provide efficient gradients to this implementation via our C++ library. In Aim 3 we will use the fact that VIP posteriors are durable and extensible descriptions of the full data posterior to enable dynamic online computation of variational posteriors, including divide-and-conquer Bayesian phylogenetics. This work will enable a cloud-based viral phylogenetics solution to rapidly update our current estimate of the posterior distribution when new data arrive or the model is modified. 1