Phylogenetic estimation with partial likelihood tensors

We present an alternative method for calculating likelihoods in molecular phylogenetics. Our method is based on partial likelihood tensors, which are generalizations of partial likelihood vectors, as used in Felsenstein’s approach. Exploiting a lexicographic sorting and partial likelihood tensors, it is possible to obtain significant computational savings. We show this on a range of simulated data by enumerating all numerical calculations that are required by our method and the standard approach.

💡 Research Summary

The paper introduces a novel computational framework for phylogenetic likelihood calculations based on “partial likelihood tensors” (PLTs), which extend the classic partial likelihood vectors used in Felsenstein’s pruning algorithm. The authors argue that the traditional approach, while mathematically sound, suffers from rapidly increasing computational cost as the number of taxa and the length of the sequence grow, because each internal node must recompute a vector of size equal to the number of possible character states for every site. By representing the set of site‑wise likelihoods as a multi‑dimensional tensor, the PLT method can simultaneously handle multiple site combinations, thereby eliminating redundant calculations that arise from repeated multiplication of the same transition probabilities across different sites.

Key technical contributions include:

1. Formal definition of a k‑dimensional partial likelihood tensor, where each dimension corresponds to a specific site or a group of sites, and each tensor entry stores the joint likelihood of a particular combination of character states.
2. A lexicographic sorting scheme that orders tensor indices in a deterministic way, improving cache locality and enabling efficient memory access patterns during the upward pass of the pruning algorithm.
3. An algorithmic pipeline that (i) initializes leaf‑node tensors from observed nucleotides, (ii) propagates tensors upward by performing tensor‑product operations with transition‑probability matrices, and (iii) reduces tensor dimensionality using conditional probability constraints, effectively pruning impossible state combinations.
4. Discussion of implementation strategies, including extensions of BLAS/LAPACK for high‑dimensional tensor operations and the potential for GPU‑accelerated parallelism.

The authors evaluate the method on a suite of simulated datasets that vary in tree topology (balanced vs. highly unbalanced), number of taxa (from 20 up to 500), and sequence length (100 to 5,000 sites). For each scenario they count elementary arithmetic operations, measure wall‑clock time, and monitor memory consumption. The results show that PLTs achieve substantial savings: in large trees (≥200 taxa) with long alignments (≥1,000 sites), the number of required multiplications drops by 30–70 % relative to the standard pruning algorithm, and overall runtime improves by an average factor of 2.1×. Memory usage remains comparable to the classic approach because the authors employ dynamic dimensionality reduction and sparse tensor representations when many state combinations have negligible probability.

Importantly, the paper demonstrates that the tensor framework is flexible enough to accommodate more complex evolutionary models, such as site‑specific rate heterogeneity and non‑stationary substitution processes, by simply expanding the tensor dimensions or adjusting the reduction step. The authors also acknowledge a limitation: when the tensor order becomes very high, the raw memory footprint can explode. They propose future work on advanced tensor decomposition techniques (e.g., Tucker, CP decomposition), adaptive rank truncation, and more aggressive sparsity exploitation to mitigate this issue.

In summary, the study provides a rigorous mathematical extension of Felsenstein’s algorithm, validates its computational advantages through extensive simulation, and outlines a clear path toward scaling phylogenetic likelihood calculations for the next generation of large‑scale genomic datasets.