Adaptive memory programming: local search parallel algorithms for phylogenetic tree construction

Abstract

One of the most important aspect of molecular and computational biology is the reconstruction of evolutionary relationships. The area is well explored after decades of intensive research. Despite this fact there remains a need for good and efficient algorithms that are capable of reconstructing the evolutionary relationship in reasonable time.Since the problem is computationally intractable, exact algorithms are used only for small groups of species. In the Maximum Parsimony approach the time of computation grows so fast when number of sequences increases, that in practice it is possible to find the optimal solution for instances containing about 20 sequences only.It is this reason that in practical applications, heuristic methods are used. In this paper, parallel adaptive memory programming algorithms based on Maximum Parsimony and some known neighborhood search methods for phylogenetic tree construction are proposed, and the results of computational experiments are presented. The proposed algorithms achieve a superlinear speedup and find solutions of good quality.

Computer science and parsimony: a reappraisal, with discussion of methods for poorly structured datasets

“…Blazewicz et al. () compared their results with those of Ribeiro and Vianna (), Viana et al. () and Goëffon et al.…”

“…Blazewicz et al. () made a very similar point:

In order to eliminate the influence of a style of implementation of algorithms on efficiency of applications, a general framework is designed. This work provides an architectural basis for comparison.

…”

“…The only paper testing variable neighbourhoods that has examined TBR is that of Blazewicz et al. (); they used a full down‐pass to evaluate scores and explicitly pointed out (p. 87–89) that TBR was so slow as to be simply prohibitive. Andreatta and Ribeiro (), Blazewicz et al.…”

“…Andreatta and Ribeiro (), Blazewicz et al. () and Ribeiro and Vianna () used neighbourhoods that switch from NNI to SPR as the search proceeds and the preceding swapping method fails to improve results. The latter authors also used what they called “2‐SPR” as the final neighbourhood, where two branches of the tree are cut and then reinserted (as done in the “mswap” command of Nona).…”

“…(). Their runs using NNI did not take much time, but the resulting scores were too far from optimal to be seriously considered; their TBR implementation also found very poor scores, by virtue of being too slow to reasonably explore tree space (Blazewicz et al., ; p. 87–89). For the methods based on VNS (variable neighbourhood search with some NNIs and then SPR) the actual times are as long as for the papers just discussed (see their table 11, which summarizes results of 10 runs or more).…”