TreeTime: Maximum-likelihood phylodynamic analysis

1Phylogenetics uses differences between homologous sequences to infer the history of the sample and learn about the evolutionary processes that gave rise to the observed diversity. In absence of recombination, this history is a tree along which sequences descend from ancestors with modification. In general, the reconstruction of phylogenetic trees is a computationally difficult problem but efficient heuristics often produce reliable reconstructions in polynomial time. Such heuristics become indispensable for large datasets of hundreds or thousands of sequences.

2Beyond phylogenetic tree building, many research questions require parameter inference and hypothesis testing. Again, exact inference from large datasets is computationally expensive since it requires high-dimensional optimization of complex likelihood functions or extensive sampling of the posterior distribution. Efficient heuristics are needed to cope with the growing datasets available today.

3One particularly common inference problem is estimating the time of historical events from sequence data. This problem goes back to, who hypothesized that changes in protein sequences accumulate at a constant rate and that the number of differences between homologous sequences can be used as a ‘molecular clock’ to date the divergence between sequences. Molecular clock methods have since been used to date the divergence of ancient proteins billions of years ago as well as the spread of RNA viruses on time scales less than a year. Beyond dating of individual divergence events or a common ancestor algorithms have been developed to infer trees where branch lengths correspond directly to elapsed time and each node is placed such that its position reflects its known or inferred date. Such trees are known as time trees, molecular clock phylogenies, or time stamped phylogenies. These methods have been generalized to allow for variation in substitution rates between different branches of the tree and between sites along a sequence. For a recent review of such methods, see.

4In addition to questions regarding natural history, time trees are useful to study epidemiology and pathogen evolution. Time trees of ‘measurably evolving’ pathogens can be used to date cross-species transmissions, introductions into geographic regions, and the time course of pathogen population sizes. In outbreak scenarios such as the recent Ebola virus (EBOV) or Zika virus outbreaks, rapid near real-time analysis of large numbers of viral genomes has the potential to assist epidemiological analysis and containment efforts –provided sample collection, sequencing, and analysis are sufficiently rapid.

5BEAST is one of most sophisticated tools for time tree estimation. BEAST samples many possible histories to evaluate posterior distributions of divergence times, evolutionary rates, and many other parameters. BEAST implements a large number of different phylogenetic and phylogeographic models. The sampling of trees, however, results in run-times of days to weeks for moderately large datasets of a few hundred sequences. On the other end of the spectrum are much simpler distance based tools that infer time scaled phylogenies orders of magnitudes faster.

6We developed a new tool called TreeTime that combines efficient heuristics with probabilistic sequence evolution models. TreeTime infers maximum likelihood time trees of a few thousand tips within a few minutes. TreeTime was designed for applications in molecular epidemiology and analysis of rapidly evolving heterochronous viral sequences. It is already in use as an integral component of the real-time time outbreak tracking tools nextstrain and nextflu. The main applications of TreeTime are ancestral state inference, evolutionary model inference, and time tree estimation. We discuss the core algorithms briefly below.

7TreeTime’s overarching strategy is to find an approximate maximum-likelihood configuration by iterative optimization of simpler subproblems similar in spirit to ‘sequential quadratic programming’ or ‘expectation maximization’. Iteration is used on multiple levels, for example by iterating optimization of branch lengths, ancestral sequences, parameters of the relaxed clock, or coalescent models. Such an iterative procedure typically converges quickly when the branch lengths of the tree are short such that ancestral sequence inference has little ambiguity.

8Ancestral sequences or node positions can be determined to optimize the joint or marginal likelihood. A joint maximum-likelihood assignment corresponds to the global configuration with highest likelihood. In a marginal maximum-likelihood assignment, individual parameters are assigned to the most likely value after summing or integrating over all other unknown states. On a tree, both of these optimal assignments can be calculated in linear time and TreeTime implements both marginal and joint ancestral reconstructions for ancestral sequences and node dates.

9In general, optimizing the branch lengths of a tree is a complicated computational problem with 2N−3 free parameters and a likelihood function that requires steps to evaluate. However, when branch lengths are short and only a minority of sites change on a given branch, a joint optimization of branch lengths and ancestral sequences can be achieved by iteratively inferring branch length and ancestral sequences since corrections due to recurrent substitutions are neglibile. Given a tree topology and the branch length, the maximum-likelihood ancestral sequences can be inferred in linear time. Likewise maximum-likelihood branch length given the parent and offspring sequences are easy to optimize. We use this iterative optimization scheme to rapidly optimize branch length and ancestral sequences. For more divergent sequences, however, subleading states of internal nodes make a substantial contribution and the iterative optimization will underestimate the branch lengths. In this case, TreeTime can use branch lengths provided in the input tree.

10For a fixed tree topology, TreeTime infers ancestral sequences maximizing the joint sequence likelihood (see above). The branch lengths corresponding to the maximum-likelihood molecular clock phylogeny can be computed in linear time using dynamic programming or message passing techniques. This approach is similar to the approach by, but the dynamic programming technique avoids computationally expensive numerical optimization of the branch lengths.

11In analogy to maximum-likelihood inference of ancestral sequences the algorithm proceeds via a post-order tree traversal propagating the maximum-likelihood assignments of subtrees towards the root, and a pre-order traversal selecting the optimal subtree given the placement of the parent node. Specifically, we calculate in post-order for each node n the likelihood that the node sits at position t given the information and constraints propagated from its children. accounts for external contraints imposed on the date of the node (e.g. fossil dating), while the product runs over all children c of node n and multiplies the integrated messages of all subtending nodes. The time t is measured as time before present. Temporal information is propagated along the branches of the tree via where is the probability distribution of the branch length τ between the focal node n and its parent. This distribution is conditional on the sequences assigned to node n and its parent. Intuitively, specifies the distribution of the date tp of the parent of node n, given the constraints from the tips descending from node n and the substitutions that accumulated on the branch to the parent node. The different objects are illustrated in Fig. 1.

12During the post-order traversal, the branch lengths maximizing Equation (2) for a given tp are tabulated and saved for the back-trace. Once the post-order traversal arrives at the root, the root is assigned the time.

13The post-order traversal is followed by a pre-order back-trace during which the branch length of each internal node is assigned to the optimal conditional on the parental position tp. To accelerate the optimization, TreeTime tabulates the branch length likelihood function and the subtree origin likelihoods.

14The above algorithm assigns each node to the time that maximizes the joint likelihood of all branch lengths in analogy to the ancestral state reconstruction algorithm by. The marginally optimal time of each internal node, that is, the time after integration over all other unconstrained nodes, can be determined in a similar manner by replacing the max in Equation (2) by a convolution integral over τ where is the analog of in Equation (1) multiplying the of all children and any external date prior.

15Once the post-order transversal arrives at the root, the marginal distribution of time t of the root node r is given by where Zr is a normalization factor. The corresponding marginal distributions of other nodes are then calculated during a pre-order traversal via

16The factor accounts for the date information coming from the leaves of node n, while the integral contributes the date information from clades other than node n and its children. Note that the contribution of node n to Pp is removed by dividing by.

17The result of the marginal reconstruction is a probability distribution of the node date given the tree, the ancestral sequence assignment, and the evolutionary model while the unknown times of other nodes are traced out. From this distribution, confidence intervals of node dates can be computed in a straight-forward manner.

18TreeTime allows one to compute joint or marginal maximum-likelihood dates, but the algorithm described above can be used for any continuous character on the tree. In Equation (2), can be replaced by any transmission function that depends either on the branch or properties of the child and parent node. We will use an analogous algorithm below to estimate parameters of relaxed molecular clock models.

19The fraction of variance in root-to-tip (RTT) distance explained by a linear regression on sampling date is given by where the sums run over all tips of the tree and ti and di are the sampling date and the distance from the root to node i, respectively. The distances di are measured as the sum of lengths of all branches from the root to the tip, that is, the expected number of substitutions since the root divided by the length of the sequence. The angular brackets denote the sample average. The regression and r2 depend on the choice of root since the di depend on the root. In absence of an outgroup, the root is often chosen to maximize r2 or minimize the squared residuals of a linear fit to the RTT distance. Programs such as TempEst and LSD allow to search for the root that maximizes this correlation and TreeTime has implemented similar functionality.

20This search for the optimal root can be achieved in linear time in the number of sequences N by first calculating for each internal node n. Here, the sum runs over all tips of node n while ti and are the sampling date and the distance of tip i from node n, respectively. The quantities θn, γn, and δn can be calculated recursively from θc, γc, and δc of the child nodes in one post-order traversal. Once those quantities are calculated, the corresponding quantities Θn, Γn, and Δn that sum contributions from all tips—not just the subtending ones—can be calculated in one pre-order traversal.

21With these quantities at hand, r2 can be calculated for any choice of root on the tree as detailed in the Appendix. Hence two tree traversals are sufficient to determine the optimal root. The root position that minimizes the mean squared residual can be calculated analogously.

22In general, the optimal position of the root will not be an internal node, but a position between two nodes on a branch of the tree. Such optimal position on internal branches of the tree can be determined from the quantities calculated above by solving a quadratic equation without any numerical optimization. The required algebra is described in the Appendix.

23Phylogenetic trees of many very similar sequences are often poorly resolved and contain multifurcating nodes also known as polytomies. Tree building software often randomly resolves these polytomies into a series of bifurcations. However, the order of bifurcations will often be inconsistent with the temporal structure of the tree resulting in poor approximations. To overcome this problem, TreeTime can prune all branches of length zero and resolve the resulting polytomies in a manner consistent with the sampling dates. For each pair of nodes, TreeTime calculates by how much the likelihood would increase when grouping this pair of nodes into a clade of size two. The polytomy is then resolved iteratively by always grouping pairs corresponding to the highest gain.

24The likelihood of observing a particular genealogical tree depends on the size of the population, its geographic structure, and fitness variation in the population. Hence parameters of models describing the ensemble of genealogies can be estimated from the data.

25In the simplest case of a panmictic population without fitness variation, the ensemble of genealogies is described by a ) coalescent, possibly with a population size that changes over time. Within the Kingman coalescent, merger events occur at random with a rate that depends on the population size N(t) and the current number of lineages k(t).

26Here, the population size N(t) defines a time scale measured in units of generation time and we will more generally refer to this time scale by and measure it in units of the inverse clock rate.

27The contribution of a branch between time points t0 (child) and t1 (parent) in the tree to the likelihood is then given by where is the rate at which a given lineage merges with any of the other. A merger at time t contributes a factor to the coalescent likelihood.

28TreeTime can estimate population sizes or coalescent time scales by maximizing the likelihood contribution of the coalescent likelihood for a fixed tree. The latter can be evaluated in one tree traversal by summing contributions from branches and merger events. In addition to a constant Tc, TreeTime can model Tc as a piecewise linear function and optimize the parameters of that function. Such piecewise functions are known as ‘skyline’.

29As part of the iterative optimization by TreeTime, the next round of optimization of branch lengths and dates of ancestral nodes will account for the coalescent likelihood. The newly inferred dates will in turn be used to update the parameters of the coalescent model as described earlier.

30Large phylogenies typically contain 100s of substitutions and thus provide enough information to infer substitution models from the data. General time reversible (GTR) substitution models are parameterized by equilibrium state frequencies πi and a symmetric substitution matrix Wij. The substitution rate from state is then.

31TreeTime infers parameters of GTR models via an iterative procedure similar to Expectation–Maximization algorithms. TreeTime first reconstructs ancestral sequences using a standard substitution model specified by the user (Jukes-Cantor by default). From this reconstruction, TreeTime calculates the time Ti spent in different states i across the tree, and the number of substitutions nij between any pair of states (i, j). Then, π and W are determined by iterating the two equations where pc is a small pseudo-count driving the estimate towards a flat Jukes-Cantor model in absence of data, and the mi are the number times state i is observed in the sequence of the root. Wij are evaluated at fixed π, followed by calculating π with the current Wij. After each iteration, π is normalized to one, the diagonal of Wij is set to, and Wij is rescaled such that the total expected substitution rate equals one. The rescaling of π and Wij can be absorbed into an overall rate μ. This algorithm typically converges in a few iterations.

32Substitution rates can vary across the tree and models that assume constant clock rates may give inaccurate inferences. Models that allow for clock rate variation have been proposed. These models typically regularize clock rate variation through a prior and penalize rapid changes of the rate by coupling the rate along branches—known as autocorrelated or local molecular clock.

33TreeTime implements an autocorrelated molecular with a normal prior on variation in clock rates. The choice of the normal prior allows for an exact and linear time solution for the maximum-likelihood substitution rates via the same forward/backward trace algorithm used for the inference of dates of internal nodes. Other priors could be implemented, but would require numerical optimization or approximations.

34TreeTime is implemented in Python (version 2.7) and uses the packages numpy and scipy for optimization, linear algebra, and interpolation ) and. Computationally costly operations are cast into array operations executed by numpy whenever possible.

35TreeTime is organized as a hierarchy of classes. TreeAnc performs maximum-likelihood inference of ancestral sequences, ClockTree infers a time scaled phylogeny given a tree topology, and TreeTime adds an additional layer of functionality including rerooting, polytomy resolution, coalescent models, and relaxed clocks. The substitution model is implemented in the class GTR.

36This structure allows TreeTime to be used in a modular fashion in Python based phylogenetic analysis pipelines. In addition, scripts can be called from the command line to perform standard tasks such as ancestral sequence inference, rerooting of trees, and time tree estimation.

37TreeTime is published under an MIT license and available at github.com/neherlab/treetime. Data and scripts necessary used to validate TreeTime are available at github.com/neherlab/treetime_validation. TreeTime can be used via a web interface at treetime.ch.

38To assess the accuracy of date reconstructions of TreeTime and to compare its performance to existing tools such as Bayesian Evolutionary Analysis Sampling Trees (BEAST) and Least-square dating (LSD), we generated toy data using the FFPopSim forward simulation library. We simulated populations of size N = 100 and used a range evolutionary rates resulting in expected genetic diversity from 0.001 to 0.2. Sequences were sampled every 10, 20, or 50 generations. The length of the simulated sequences was L = 1000.

39Fig. 2 shows the error in the estimates of the clock rate for TreeTime, LSD, and BEAST as a function of genetic diversity. TreeTime and LSD estimated the clock rate accurately at low diversity but tended to underestimate the rates at when diversity exceeds a few percent. This is expected in the case of TreeTime since maximum-likelihood sequence assignment can result in underestimated branch lengths. BEAST produced accurate estimates across the entire range of diversities. By sampling trees, BEAST does not suffer from the atypical maximum-likelihood assignments.

40In a similar manner, TreeTime and LSD estimated the time of the most recent common ancestor to within 10% accuracy at low diversity (relative to the coalescence time) with larger deviations at diversity above 10%, see Fig. 3. BEAST returned accurate estimates across the entire range of diversities.

41We also ran TreeTime on simulated data provided by and compared it to the results reported by for LSD, BEAST, and a number of other methods. Figure 4 compares the accuracy of TMRCA and clock rate estimates, showing that TreeTime achieves similar or better accuracy than other methods.

42Population bottlenecks, selective sweeps, or population structure affect the rate of coalescence in a time-dependent manner. BEAST can infer a history of effective population size (inverse coalescence rate) from a tree—often known as skyline. TreeTime can perform a similar inference by maximizing the coalescence likelihood with respect to the pivots of a piecewise linear approximation of the coalescence rate history Tc(t) (aka effective population size). To test the power and accuracy of this inference, we simulated sinusoidal population size histories of different amplitude and period, uniformly sampled sequences through time, and used these data to estimate the coalescent rate history. True and estimated population size histories agree well with each other as shown in Fig. 5.

43The dense sampling of influenza A virus sequences over many decades makes this virus an ideal test case to evaluate the sensitivity of time tree estimation to sampling depth. We estimated the clock rate and the time of the most recent common ancestor of influenza A/H3N2 HA sequences sampled from 2011 to 2013 for sets of sequences varying from 30 to 3,000, see Fig. 6. TreeTime estimates are stable across this range, while estimates by LSD tend to drift with lower rates and older MRCAs for larger samples. Estimates by BEAST are generally consistent with TreeTime.

44Next, we tested how accurately TreeTime infered dates of tips when only a fraction of tips have dates assigned. Every tip in TreeTime can either be assigned a precise date, an interval within which the date is assumed to be uniformly distributed, or no constraint at all. TreeTime will then determine the probability distribution of the date of the node based on the distribution of the ancestor and the substitutions that occurred since the ancestor.

45Figure 7 shows the distribution of error in leaf date reconstruction as the fraction of missing dates increased from 5 to 95% of all nodes. TreeTime estimated the date of influenza sequences to an average accuracy of ≈0.5 years if >50% of dates are known.

46In 2014, West Africa experienced the largest known outbreak of EBOV in humans. The genomic epidemiology has been studied intensively by multiple groups. Here, we reanalyzed a subset of 350 EBOV sequences sampled throughout the outbreak from 2014–16. Due to the dense sampling, the maximum-likelihood phylogeny has many unresolved nodes and TreeTime was used to resolve polytomies using temporal information. After automatic rooting and GTR model inference, TreeTime produced the time tree shown in Fig. 8. The GTR model inferred from the tree was

47This analysis took 4 min to complete on a 2016 laptop (Dell XPS13) with an i7 processor using a single CPU. In addition to inferring a time tree, TreeTime estimated the time course of the coalescent population size shown in the lower panel of Fig. 8. The estimated population size closely mirrors the case counts reported by the WHO throughout this period.

48TreeTime was developed to analyse large heterochronous viral sequence alignments and we have used TreeTime as the core component of the real-time phylogenetics pipelines nextstrain and nextflu. TreeTime tries to strike a useful compromise between inflexible but fast heuristics and computationally expensive Bayesian approaches that require extensive sampling of treespace. The overarching algorithmic strategy is iterative optimization of efficiently solvable subproblems to arrive at a consistent approximation of the global optimum. Although this strategy is approximate and often assumes short branch length, it converges fast for many applications and trees with thousands of tips can be analyzed in a few minutes. In this paper, we presented analyses of human seasonal influenza A/H3N2 virus sequences and sequences of the recent EBOV outbreak. In both cases, average pairwise distances between strains are 10% and individual branches in the trees are much shorter still. TreeTime assumption of short branches is therefore met.

49Rapid, efficient analysis phylodynamic algorithms are of increasing importance as datasets are increasing in size. For example during the recent outbreaks of EBOV and Zika virus, hundreds of sequences were generated and needed to be analyzed in near real time to inform containment efforts. Similarly, the GISRS network for surveillance of seasonal influenza virus sequences hundreds of viral genomes per month. Timely analysis of these data with Bayesian methods that require extensive tree sampling such as BEAST is difficult. Sequencing from EBOV, Zika virus outbreaks, or seasonal influenza viruses are typically very similar to each other (>90% identity) such that TreeTime assumptions and approximations are justified.

50When compared with other methods recently developed for rapid estimation of time trees, TreeTime uses probabilistic models of evolution, allows inference of ancestral characters, and coalescent models. In TreeTime, every node of the tree can be given a strict or probabilistic date constraint. This higher model complexity results in longer run times, but the scaling of run times remains linear in the size of the dataset and alignments with thousands of sequences can be analyzed routinely. The time tree inference and dating are typically faster than the estimation of the tree topology.

51TreeTime was tested predominantly on sequences from viruses with a pairwise identity above 90%. The iterative optimization procedures are not expected to be accurate for trees were many sites are saturated. In scenarios with extensive uncertainty of ancestral states and tree topology, convergence of the iterative steps cannot be guaranteed. While in many cases TreeTime might still give approximate branch lengths, ancestral assignments and time tree estimates, these need to be checked for plausibility. In general global optimization and sampling of the posterior can not be avoided.

52TreeTime can be used in a number of different ways. The core TreeTime algorithms and classes can be used in larger phylogenetic analysis pipelines as Python scripts. This is the most flexible way to use TreeTime and all the different analysis steps can be combined in custom ways with user specified parameters. In addition, we provide command-line scripts for typical recurring tasks such as ancestral state reconstruction, rerooting to maximize temporal order, and time tree inference. We also implemented a web-server that allows exploration and analysis of heterochronous alignments in the browser without the need to use the command-line.

53Conflict of interest: None declared.