How Do Bayesian and ML Methods Estimate Branch Support?
In computational evolutionary biology, estimating phylogenetic branch support is critical for assessing the statistical confidence of inferred evolutionary trees. Maximum likelihood and Bayesian inference approach this challenge through fundamentally distinct statistical paradigms. Maximum likelihood typically employs non-parametric bootstrapping or fast likelihood ratio tests to evaluate how consistently alignment columns support a given clade across resampled data. In contrast, Bayesian inference calculates posterior probabilities by sampling tree topologies and model parameters using Markov Chain Monte Carlo algorithms. Understanding how these frameworks operate, how their support values are computed, and how their biological interpretations differ is essential for accurately reconstructing phylogenetic histories.
Maximum Likelihood and Non-Parametric Bootstrapping
Maximum likelihood (ML) evaluates a candidate phylogenetic tree topology and its branch lengths by maximizing the likelihood function:
\[L(\tau, \boldsymbol{\theta} \mid D) = P(D \mid \tau, \boldsymbol{\theta})\]
where \(\tau\) represents the tree topology, \(\boldsymbol{\theta}\) denotes substitution model parameters and branch lengths, and \(D\) is the multiple sequence alignment consisting of \(N\) homologous sites.
Because standard ML searches produce a single point estimate for the optimal tree, assessing branch confidence requires empirical resampling strategies. The most common approach is Felsenstein's non-parametric bootstrap:
- Pseudo-replicate datasets \(D^*\) of size \(N\) are generated by sampling alignment columns randomly with replacement from \(D\).
- For each pseudo-replicate, a full heuristic ML tree search is performed to find the optimal topology \(\hat{\tau}^*\).
- The bootstrap support for a specific branch (or clade bipartite split) is calculated as the proportion of bootstrap trees that contain that split:
\[\text{BS}(e) = \frac{1}{B} \sum_{b=1}^{B} I(e \in \hat{\tau}_b^*)\]
where \(B\) is the total number of replicates (typically 100 to 1,000) and \(I(\cdot)\) is an indicator function. High bootstrap percentages indicate that branch recovery is robust against site sampling variance.
Likelihood-Based Fast Branch Tests
Because standard bootstrapping can be computationally prohibitive on large datasets, several direct likelihood-based approximations are frequently used:
The SH-Like Approximate Likelihood Ratio Test
The Shimodaira-Hasegawa-like approximate likelihood ratio test (SH-aLRT) evaluates the likelihood difference between the best tree containing the target branch and alternative nearest-neighbor interchange (NNI) configurations:
\[2\Delta \ln L = 2 \left[ \ln L(\hat{\tau}_0) - \max(\ln L(\hat{\tau}_1), \ln L(\hat{\tau}_2)) \right]\]
The test statistic is compared against a null distribution generated by non-parametric resampling of site likelihoods, yielding a fast, conservative measure of branch support.
Ultrafast Bootstrap Approximation (UFBoot)
UFBoot speeds up traditional bootstrapping by evaluating candidate trees on bootstrap pseudo-replicates using fixed model parameters optimized on the original alignment. It explores tree space via random walks and computes branch support while avoiding full re-optimization for every replicate, providing support values that correlate closely with standard bootstrapping without the severe computational overhead.
Bayesian Inference and Posterior Probabilities
Bayesian phylogenetics shifts the objective from point estimation to estimating the posterior probability distribution of trees and evolutionary parameters using Bayes' theorem:
\[P(\tau, \boldsymbol{\theta} \mid D) = \frac{P(D \mid \tau, \boldsymbol{\theta}) P(\tau) P(\boldsymbol{\theta})}{P(D)}\]
Here, \(P(\tau)\) is the prior probability over tree topologies, \(P(\boldsymbol{\theta})\) is the prior over model parameters, and \(P(D)\) is the marginal likelihood integrated over all possible trees and continuous parameters:
\[P(D) = \sum_{\tau} \int P(D \mid \tau, \boldsymbol{\theta}) P(\tau) P(\boldsymbol{\theta}) \, d\boldsymbol{\theta}\]
Due to the intractable sum across all topologies, Markov Chain Monte Carlo (MCMC) methods, such as the Metropolis-Hastings algorithm, sample states proportional to their posterior probability:
- A proposed tree state \((\tau', \boldsymbol{\theta}')\) is generated from the current state \((\tau, \boldsymbol{\theta})\) via proposal distribution \(q(\cdot \mid \cdot)\).
- The proposal is accepted with probability:
\[\alpha = \min\left(1, \frac{P(D \mid \tau', \boldsymbol{\theta}') P(\tau') P(\boldsymbol{\theta}') q(\tau, \boldsymbol{\theta} \mid \tau', \boldsymbol{\theta}')}{P(D \mid \tau, \boldsymbol{\theta}) P(\tau) P(\boldsymbol{\theta}) q(\tau', \boldsymbol{\theta}' \mid \tau, \boldsymbol{\theta})}\right)\]
After discarding initial burn-in iterations and thinning the chain, the Bayesian posterior probability (BPP) of a clade or branch \(e\) is calculated directly as its marginal frequency among the sampled trees:
\[\text{BPP}(e) = \frac{1}{M} \sum_{m=1}^{M} I(e \in \tau_m)\]
where \(M\) is the number of post-burn-in sampled trees.
Comparing Bootstrap Proportions and Posterior Probabilities
Although both bootstrap percentages and Bayesian posterior probabilities range from 0 to 1 (or 0% to 100%), they measure distinct statistical properties:
Statistical Meaning
ML bootstrap proportions quantify repeatable signal under simulated empirical data resampling, reflecting the variance of the estimator due to finite sequence length. Bayesian posterior probabilities represent the conditional probability that a clade is true given the observed alignment, substitution model, and prior distributions.
Sensitivity and Conservatism
Empirical and simulation studies demonstrate that Bayesian posterior probabilities are generally higher and more sensitive than standard bootstrap support. A branch with a bootstrap support of 70% may exhibit a posterior probability exceeding 95%. Bootstrap support is considered a conservative measure of clade robustness, while posterior probabilities can be sensitive to model misspecification and short internode branches.
Computational Performance
Standard maximum likelihood bootstrapping scales linearly with the number of replicates \(B\), multiplying runtime accordingly. Bayesian MCMC requires long chains to ensure adequate parameter convergence and effective sample size (ESS), but posterior probabilities for all branches are obtained simultaneously from a single converged posterior distribution sample.