On the design of shared memory approaches to parallelize a multiobjective bee-inspired proposal for phylogenetic reconstruction

Abstract

Current efforts in solving computationally demanding optimization problems in bioinformatics rely on the combination of bioinspired computing and parallelism. The multiobjective nature shown by a wide variety of these problems represents an additional challenge, as the optimization of multiple objective functions involves growing computational requirements. In this work, a multiobjective metaheuristic inspired by honey bees is applied to tackle the phylogenetic inference problem. For this purpose, two parallel implementations for shared memory architectures are proposed: a synchronous generational model and a novel non-generational design inspired by the asynchronous behaviour of bees in nature. Experiments on six real biological data sets and comparisons with other parallel biological methods point out the relevance of applying nature-inspired parallelization strategies, addressing the performance pitfalls shown by traditional parallelization schemes.

Introduction

In the last decade, the growing availability of parallel architectures has opened the door to novel programming models where multiple processing units collaborate to address computationally demanding problems. In this sense, several research lines in bioinformatics try to overcome challenging biological problems by taking advantage of shared memory Multiple Instruction Multiple Data (MIMD) multiprocessor systems, commonly represented by symmetric multiprocessors (SMP). A wide variety of biological processes can be modelled as optimization problems which seek to explore a search space Φ with the aim of obtaining an optimal solution x ∈ Φ in accordance with an objective function $f:\Phi \to \mathbb{R}$. As pointed out in [49], the NP-hard complexities shown by such problems (including multiple sequence alignment, DNA fragment assembling, biological sequence analysis, etc. [56]) motivate that the current needs of biological processing cannot be satisfied by using traditional serial approaches. This issue has led to an increasing interest in developing parallel algorithms for computational biology.

In order to model the behaviour of real natural systems, biological optimization procedures are often conducted by considering a number of different principles, defined as optimality criteria. This multiobjective nature has motivated the formulation of a wide range of biological problems as Multiobjective Optimization Problems (MOPs) [14]. When addressing MOPs, we aim to find a Pareto set containing solutions which optimize two or more objective functions simultaneously. This idea represents an additional challenge, as the process of finding a good approximation to the Pareto-optimal set implies a high computational cost. Recent advances in parallel architectures represent an opportunity to deal with these problems, giving support to proposals which combine bioinspired computing and parallelism.

The development of parallel metaheuristics follows two main directions [51]. Firstly, the self-contained parallel cooperation approach aims to improve the quality of results by running independent algorithms which cooperate through the interchange of promising solutions. Secondly, the idea of intra-algorithm parallelization focuses on reducing the execution times required to run an algorithm by distributing tasks among processes or threads. In bioinformatics, this second work line represents the most popular approach to make feasible the analysis of growing amounts of biological sequences. By developing parallel metaheuristics, major computational challenges in bioinformatics can be addressed. In this research, we address one of the most outstanding problems in this field, phylogenetic inference [32].

By reconstructing the evolutionary history of living species, phylogeneticists provide useful knowledge in a number of scientific fields, including biomedicine, chemistry, and paleontology. The need to infer satisfying phylogenies over exponentially increasing search spaces explains the NP-hard complexity of the problem and, therefore, the close relationship between phylogenetics, parallelism, and bioinspired computing. Throughout the years, phylogenetic searches have been carried out following optimality criteria which measure the quality of solutions. However, multiple researches [7], [35] have given account of the major sources of incongruence introduced by different optimality criteria, reporting conflicting evolutionary histories. This key issue can be addressed by formulating a multiobjective view of phylogenetics. By applying multiobjective approaches, we can describe evolutionary trees supported by multiple objective functions simultaneously. In this way, biologists will be able to infer a set of alternative trade-off solutions, along with high-quality phylogenetic topologies according to each objective separately.

In this paper, we tackle the phylogeny reconstruction problem under two criteria: parsimony and likelihood. For this purpose, we study shared memory schemes based on OpenMP [10] to parallelize a Multiobjective Artificial Bee Colony (MOABC) algorithm for phylogenetics [45]. In order to exploit pure shared memory environments, two parallel approaches are proposed. Firstly, a synchronous generational design which parallelizes the main sections that compose each generation of the algorithm, in such a way that the next generation does not take place until all the tasks in the previous one have been completed. Secondly, an asynchronous nature-inspired scheme designed to model the inherently asynchronous behaviour of honey bees, dismissing the traditional concept of generation. The main goal is to decide which parallel design leads to improved parallel performance while keeping the multiobjective and biological quality of the results. Each version will be assessed by experimentation on six real nucleotide data sets, making comparisons with other biological and parallel methods for inferring phylogenies.

This paper is organized as follows. The next section highlights the relationship between parallelism and bioinspired computing and reviews their applications to the field of phylogenetics. Section 3 introduces the problem, discussing its computational complexity and the biological relevance of multiobjective optimization. In Section 4, we detail the algorithmic design of MOABC, defining the proposed parallelization schemes in Section 5. The experimental analysis of parallel, multiobjective, and biological performance is found in Section 6. Finally, Section 7 includes conclusions and future work.