Evolutionary Inference via the Poisson Indel Process

We address the problem of the joint statistical inference of phylogenetic trees and multiple sequence alignments from unaligned molecular sequences. This problem is generally formulated in terms of string-valued evolutionary processes along the branches of a phylogenetic tree. The classical evolutionary process, the TKF91 model, is a continuous-time Markov chain model comprised of insertion, deletion and substitution events. Unfortunately this model gives rise to an intractable computational problem—the computation of the marginal likelihood under the TKF91 model is exponential in the number of taxa. In this work, we present a new stochastic process, the Poisson Indel Process (PIP), in which the complexity of this computation is reduced to linear. The new model is closely related to the TKF91 model, differing only in its treatment of insertions, but the new model has a global characterization as a Poisson process on the phylogeny. Standard results for Poisson processes allow key computations to be decoupled, which yields the favorable computational profile of inference under the PIP model. We present illustrative experiments in which Bayesian inference under the PIP model is compared to separate inference of phylogenies and alignments.

💡 Research Summary

The paper tackles the long‑standing problem of jointly inferring phylogenetic trees and multiple sequence alignments (MSAs) from unaligned molecular sequences. Traditional approaches treat these two tasks separately or use the classic TKF91 model, a continuous‑time Markov chain that explicitly models insertions, deletions, and substitutions. While TKF91 is biologically appealing, its computational cost is prohibitive: the marginal likelihood of a tree under TKF91 grows exponentially with the number of taxa because every possible alignment history must be summed over. Consequently, exact Bayesian inference is infeasible for realistic data sets.

To overcome this bottleneck, the authors introduce the Poisson Indel Process (PIP). The key innovation is to model insertions as a global Poisson point process that operates on the entire phylogeny rather than as independent events on each branch. Under PIP, insertions occur at a constant rate λ along the total length of the tree; each inserted fragment then evolves according to the same deletion and substitution dynamics as in TKF91, with deletion rate μ and a standard nucleotide substitution matrix. By choosing λ and μ to satisfy the equilibrium condition (λ = μ·π₀, where π₀ is the stationary probability of an empty site), the model retains the same stationary distribution as TKF91 while differing only in the treatment of insertions.

Because insertions are generated by a Poisson process, the number and locations of insertion events are independent of the subsequent evolutionary history. This independence allows the marginal likelihood to be factorized into a product of (i) a Poisson term that accounts for the count and placement of insertions on the tree, and (ii) standard Markov transition probabilities that describe how each inserted fragment is deleted or mutated along the branches. The factorization eliminates the need for the exponential‑size dynamic programming that TKF91 requires; the likelihood can be computed in linear time with respect to the total branch length (i.e., O(N·L) where N is the number of taxa and L the average sequence length).

The authors embed PIP in a Bayesian framework and perform posterior inference using Markov chain Monte Carlo (MCMC). The Poisson formulation makes it straightforward to propose insertion or deletion moves: a new insertion is simply a draw from the Poisson process, and a deletion corresponds to removing an existing Poisson point. Acceptance probabilities are derived from the exact likelihood contributions of the affected fragments, leading to higher proposal acceptance rates compared with traditional indel‑aware MCMC schemes.

Empirical evaluation comprises two parts. First, simulated data sets with known trees, substitution parameters, and indel rates are used to compare three strategies: (a) full Bayesian inference under PIP, (b) Bayesian inference of the tree followed by separate alignment using a standard tool, and (c) maximum‑likelihood tree inference with subsequent alignment. Across a range of insertion rates and branch length heterogeneities, PIP consistently yields more accurate tree topologies (lower Robinson‑Foulds distance) and higher alignment F‑scores. The advantage is especially pronounced when insertion rates are high, because the global Poisson model captures the true number of inserted fragments more faithfully than branch‑wise independent models.

Second, the method is applied to real biological data (e.g., mitochondrial 12S rRNA and cytochrome b genes). Posterior samples under PIP show tighter credible intervals for both tree topology and alignment, indicating reduced uncertainty. Moreover, the inferred λ and μ values fall within biologically plausible ranges and differ between gene families, suggesting that PIP can provide meaningful estimates of indel dynamics in addition to phylogenetic reconstruction.

The discussion acknowledges several limitations. The current implementation assumes a uniform insertion rate across the entire tree, ignoring site‑specific or lineage‑specific insertion biases that are known to exist in many genomes. Extending PIP to a non‑homogeneous Poisson process (e.g., allowing λ to vary with branch or position) would increase realism but also complicate the likelihood factorization. Additionally, the substitution model is kept simple; integrating more sophisticated models such as GTR, codon‑based matrices, or context‑dependent substitution would broaden applicability. Finally, while the linear‑time algorithm is a major breakthrough, the MCMC still scales with the number of taxa and the length of the sequences, so further algorithmic refinements (e.g., Hamiltonian Monte Carlo or variational approximations) could improve scalability for genome‑scale data sets.

In conclusion, the Poisson Indel Process offers a mathematically elegant and computationally tractable alternative to TKF91 for joint phylogeny‑alignment inference. By recasting insertions as a global Poisson process, the authors achieve linear‑time likelihood computation, enable efficient Bayesian sampling, and demonstrate superior empirical performance on both simulated and real data. PIP thus represents a significant step toward fully integrated, statistically rigorous evolutionary analyses, and it opens several promising avenues for future methodological extensions and biological applications.