Müller Nicola F, Bouckaert Remco R
Department of Biosystems Science and Engineering, ETH Zürich, Basel, Switzerland.
Swiss Institute of Bioinformatics, Lausanne, Switzerland.
PeerJ. 2020 Sep 16;8:e9473. doi: 10.7717/peerj.9473. eCollection 2020.
With ever more complex models used to study evolutionary patterns, approaches that facilitate efficient inference under such models are needed. Metropolis-coupled Markov chain Monte Carlo (MCMC) has long been used to speed up phylogenetic analyses and to make use of multi-core CPUs. Metropolis-coupled MCMC essentially runs multiple MCMC chains in parallel. All chains are heated except for one cold chain that explores the posterior probability space like a regular MCMC chain. This heating allows chains to make bigger jumps in phylogenetic state space. The heated chains can then be used to propose new states for other chains, including the cold chain. One of the practical challenges using this approach, is to find optimal temperatures of the heated chains to efficiently explore state spaces. We here provide an adaptive Metropolis-coupled MCMC scheme to Bayesian phylogenetics, where the temperature difference between heated chains is automatically tuned to achieve a target acceptance probability of states being exchanged between individual chains. We first show the validity of this approach by comparing inferences of adaptive Metropolis-coupled MCMC to MCMC on several datasets. We then explore where Metropolis-coupled MCMC provides benefits over MCMC. We implemented this adaptive Metropolis-coupled MCMC approach as an open source package licenced under GPL 3.0 to the Bayesian phylogenetics software BEAST 2, available from https://github.com/nicfel/CoupledMCMC.
随着用于研究进化模式的模型越来越复杂，需要有助于在此类模型下进行高效推断的方法。长期以来， metropolis耦合马尔可夫链蒙特卡罗（MCMC）一直被用于加速系统发育分析并利用多核CPU。 metropolis耦合MCMC本质上是并行运行多个MCMC链。除了一个像常规MCMC链一样探索后验概率空间的冷链之外，所有链都被加热。这种加热使得链能够在系统发育状态空间中进行更大的跳跃。然后可以使用加热的链为其他链（包括冷链）提出新的状态。使用这种方法的一个实际挑战是找到加热链的最佳温度以有效地探索状态空间。我们在此为贝叶斯系统发育学提供一种自适应metropolis耦合MCMC方案，其中加热链之间的温差会自动调整，以实现各个链之间状态交换的目标接受概率。我们首先通过在几个数据集上比较自适应metropolis耦合MCMC和MCMC的推断来展示这种方法的有效性。然后我们探索metropolis耦合MCMC相对于MCMC的优势所在。我们将这种自适应metropolis耦合MCMC方法作为一个开源软件包实现，根据GPL 3.0许可用于贝叶斯系统发育学软件BEAST 2，可从https://github.com/nicfel/CoupledMCMC获取。
