Is multiple-sequence alignment required for accurate inference of phylogeny?

Abstract

The process of inferring phylogenetic trees from molecular sequences almost always starts with a multiple alignment of these sequences but can also be based on methods that do not involve multiple sequence alignment. Very little is known about the accuracy with which such alignment-free methods recover the correct phylogeny or about the potential for increasing their accuracy. We conducted a large-scale comparison of ten alignment-free methods, among them one new approach that does not calculate distances and a faster variant of our pattern-based approach; all distance-based alignment-free methods are freely available from http://www.bioinformatics.org.au (as Python package decaf+py). We show that most methods exhibit a higher overall reconstruction accuracy in the presence of high among-site rate variation. Under all conditions that we considered, variants of the pattern-based approach were significantly better than the other alignment-free methods. The new pattern-based variant achieved a speed-up of an order of magnitude in the distance calculation step, accompanied by a small loss of tree reconstruction accuracy. A method of Bayesian inference from k-mers did not improve on classical alignment-free (and distance-based) methods but may still offer other advantages due to its Bayesian nature. We found the optimal word length k of word-based methods to be stable across various data sets, and we provide parameter ranges for two different alphabets. The influence of these alphabets was analyzed to reveal a trade-off in reconstruction accuracy between long and short branches. We have mapped the phylogenetic accuracy for many alignment-free methods, among them several recently introduced ones, and increased our understanding of their behavior in response to biologically important parameters. In all experiments, the pattern-based approach emerged as superior, at the expense of higher resource consumption. Nonetheless, no alignment-free method that we examined recovers the correct phylogeny as accurately as does an approach based on maximum-likelihood distance estimates of multiply aligned sequences.

Figure 1

RF distance landscape for method B-bin. Average RF distance (y-axis) of method B-bin on three reference sets (top to bottom: set 2, set 4, and set 6) of two synthetic data sets (a, c, e: control; b, d, f: ASRV). Each subfigure shows the behavior as a function of word length k (x-axis) for two alphabets (AA: original amino acids, CE: chemical equivalence classes). Points are joined for ease of visual inspection only.

Figure 2

Average RF distance for six methods. Average RF distance (y-axis) for six selected methods on all seven reference sets (x-axis) of two synthetic data sets (a: control; b: ASRV). For each data set, we show (1) the ML distance estimate based on correct alignments, (2) the best pattern-based variant, (3 and 4) the best word-based method and the best method not based on words, 5) the best composition distance; and (6) the W-metric; the numbers in the inserted legends refer to the far left-hand column of Tables 1 (Figure 2a) and 2 (Figure 2b) respectively.