ClustalXeed: a GUI-based grid computation version for high performance and terabyte size multiple sequence alignment

Genetic and protein sequences are being discovered rapidly, and as a result, the number of sequences entered into biological databases is growing exponentially over time. Most of the work currently being done in computational biology involves searching for inter- and intra-sequence homology in massive volumes of genetic and protein sequence data, which are commonly based on a multiple sequence alignments (MSAs) [1]. However, increasing the computational efficiency to solve a variety of real MSA problems is still a challenging task because of the high demand for greater capacity and speed [2, 3].

The oldest and most widely used MSA program that estimates trees as it aligns multiple sequences is ClustalW [4, 5]. ClustalX is an integrated graphical-user-interface (GUI) version of the ClustalW multiple sequence alignment program [6]. It provides an easy-to-use work environment for performing MSA and pattern analyses. The latest version of ClustalX (version 2.0) added two new features [7]. The main advantage of ClustalX 2.0 is that it provides an easier way to maintain code for other applications. The new guided-tree implementation, compared with the older version, enables larger, faster computations. ClustalX 2.0 is now available for a number of platforms, including SUN Solaris, IRIX5.3 on Silicon Graphics, Digital UNIX on DECStations, Microsoft Windows for PCs, Linux ELF for x86 PCs, and Macintosh PowerMac.

Unfortunately, most of the currently available MSA programs are not suitable for large-capacity data storage and massive computation. These programs, including ClustalX (or X 2.0), are still single-PC based, and the storage and computation is entirely dependent on physical random-access memory (RAM). Past MSA performance evaluations focused simply on how compute-intense and sensitive the program was with respect to the longest-common-sequence (LCS)-based exact-string matching algorithm (e.g., the Smith-Waterman or Needleman-Wunsch method) [8]. Depending on both the volume of data to be aligned and the accuracy of the comparisons, computation using dynamic programming is extremely time-consuming when large sequence volumes and high accuracy are required simultaneously. Numerous parallel-computation programs, such as parallelized Praline [9], DiAlign P [10], ClustalW-MPI [11], and a commercial SGI parallel Clustal on a shared memory SGI multiprocessor [12], have been developed, primarily to increase computational speed, rather than for larger capacity data handling. Multiple sequence alignments with the Clustal series programs and the required features have been reviewed [13].

In this study, we report ClustalXeed, which was designed for a wide range of MSA purposes (Figure 1). This is a new ultra-high grid version of ClustalX supports both a large-scale MSA and efficient load-balancing. We also evaluated its overall performance using a typical grid-computation system. ClustalXeed is currently available for Linux 64-bit/32-bit platforms. A major point we emphasize is that ClustalXeed provides dual-mode sequence loading and computation platforms, where if the size of the biological sequence exceeds the available RAM, the user can readily shift to a distributed file-allocation (-disk memory) mode. The disk-storage system uses an idle node-seeking task algorithm (INSTA), which forces sleeping nodes into active modes during the massive file-swapping and pair-align computation to overcome slow computation speed. The GUI-based parallel-computation system is implemented with Linux and can be classified into two subsystems: the GUI engine (a highly modified ClustalX with an efficient load-balancing algorithm) for interactions between users and the parallel-computation system, and the pair-alignment engine for the calculation of pair-align matrices.

ClustalXeed allows the alignment of large numbers of nucleic acid or protein sequences. Generally, the memory space requirement for the dynamic programming algorithm follows O (m × n). Based on this, for over 400000 sequences, MSA requires at least 1.28 terabytes of memory (i.e., 400000 × 400000 × 8/2) just to construct a distance-tree matrix, regardless of sequence length. We evaluated our software using large volumes of real nucleic acid and protein data sets to assess its overall and detailed performance. We set up a typical 100 × CPU AMD Opteron 244 (1.8 GHz) 64-bit cluster system with Linux (Fedora 4 Core). The master node consists of a 10-terabyte HDD (Samsung Electronics): 3ware RAID adapter (RAID 0, striping) and 8 GB DRAM, for a huge-tree matrix construction. Each node is a dual-core Opteron system with 4 GB DRAM and a 1-terabyte HDD.

To confirm the full sequence-loading and speed-up computation, almost six million sequences (23.08 GB) in an nt.gz file were first downloaded from ftp://ftp.ncbi.nih.gov/ blast/FASTA. These DNA sequences were divided into smaller data sets containing 150000 to 280000 sequences. With 200000 sequences as the first test set, which requires a nearly 8-terabyte file-managing environment for the MSA job, the elapsed computation time was 1.5- to 6-times shorter than that without INSTA. The total execution time for the complete alignment of 200,000 nucleic acid sequences by the disk-storage system was about 4700-6100 s with INSTA, and 7500-24000 s without INSTA. Once the INSTA was built, the overall speed-up rate was directly increased with the added computation nodes. The total computation efficiency depends on the pair-align scoring stage (i.e., the "bottleneck" step). For a given number of sequences, the average speed of ClustalXeed performance increased with the number of computation nodes (up to 50 nodes were tested).

As a second performance test, we extracted from NCBI a data set comprised of 52,750 protein sequences from the cytochrome P450 superfamily. The average length of the protein sequences was 440 amino acids (aa), with minimal length around 10 aa and maximal length of 11600 aa. In total, 1.39×10⁹ (i.e., 52750 × 52749/2) pair-wise sequence alignments are required to complete each SP score calculation. We randomly mixed all of these sequences and aligned them using the disk-storage system. Figure 4 summarizes the obtained execution times of different job-balancing conditions as a function of node numbers. When the load-balancing algorithm was applied, the execution time of ClustalXeed was more than four times faster than the usual file-swap mode. The average speed-up for 20 nodes was 11.7 with INSTA and 7.5 without INSTA. The observed speed-up difference was due to the efficient load-balancing of sparsely distributed sequence pairs, and it clearly demonstrated the success of the implemented load-balancing algorithm. We showed that there exists an optimal node number to be assigned to jobs for the effective usage of computation resources. It is important to determine what system size will result in the maximum speed-up. The speed-up rate, as a function of participating node numbers, was not linear, because there was implicit waiting time due to communication latency (i.e., disk access time and network delay). All pairs of sequences in the ClustalXeed main disk-storage are transferred to compute nodes, and the calculated pair scores are then returned to the master and are saved in the disk-storage distance matrix (see Figure 3).

ClustalW is the most widely used multiple sequence alignment program. The program uses a profile-based progressive alignment, with the alignment process divided into three steps: (1) distance matrix/pair-wise alignment, (2) guided tree generation, and (3) progressive alignment. However, ClustalW is not capable of estimating phylogenies and alignments for large data sets, because of slow computation times and limited memory usage. Alternatives to ClustalW include ClustalX (or X2.0), a graphical user interface (GUI)-based version of ClustalW, and ClustalW-MPI, a recently developed, parallel implementation version of ClustalW. In the latter tool, all three required computation steps are performed in parallel to reduce the total computation time. For parallel job scheduling, ClustalW-MPI uses a fixed-size chunk scheduler, which divides each job into uniform, fixed sizes. This method of batch job scheduling sometimes reduces the communication overhead required for a small and moderate number of sequences having very short lengths, such as ESTs (expressed sequence tags). However, it is not always effective at balancing the workload required for a large number of sequences having unequal lengths. Under these circumstances, the iteration for the next job scheduling is often delayed, which may result in high processor idle time [18]. Moreover, because ClustalW-MPI uses a RAM memory-storage mode, it has the inherent problem of a shortage of memory available for large-scale data sets.

ClustalXeed was designed to simultaneously address the dual needs of fast computation and large sequence data handling. Core changes with significant strides in supporting disk-memory access capabilities and enhanced job scheduling can tackle this bottleneck by enhancing both storage capacity and speed. As mentioned previously, ClustalW-MPI delivers fixed-size chunks to every compute node before the calculation. The batch scheduling mode needs high memory on each node, which may result in processor idling when the sequence lengths are not uniform. In contrast, the ClustalXeed disk storage algorithm does not send all pair sequences to the computation nodes and instead master allocates (distributes) the pair sequences individually to an appropriate node using a load-balancing idle node-seeking task algorithm (INSTA) to reduce the initial file transfer cost and slave node memory requirements. ClustalXeed also provides a fixed size chunk algorithm in the memory-storage mode for alignment of small-to-medium-sized sequences.

Because ClustalXeed was designed to work at the capacity of disk memory space, it has no inherent limitation on input sequence amounts. With a 64-bit compiler version, ClustalXeed works with sequence arrays containing up to 18 terabytes in disk-storage mode. In the disk memory operation, alignment speed was also improved by incorporating INSTA, which allocates or re-allocates tasks to idle nodes during pair alignment without the requirement for prior task information. In this way, INSTA determines from moment to moment which imbalanced nodes can be resolved immediately by allocating a new job. The main advantage of INSTA is its prompt scanning and decision-making without requiring decision complexity of rescheduling or reallocating tasks to resources at runtime, as occurs with static or other general dynamic load-balancing algorithms.

While setting up a new sequence batch job, the user has the option to choose an appropriate alternate computing environment using the graphical menu option. Other precise values of multiple sequence alignment tuning parameters can be set for each ClustalXeed menu link individually (similar to the old versions of ClustalX or X 2.0). Users can follow the progress of their alignment execution from the real-time in-process dialogue box. As indicated previously, the new parameter inputs, the easy-to-use realign option, the fully integrated graphical user interface, and the improved speed-up performance for very large volumes of sequence data may provide optimized computation of biological sequence alignments. ClustalXeed demonstrates that the embedded INSTA is a simple and effective means to meet the high computational demands (speed and memory) for massive biological-sequence alignments.

Moreover, large phylogenetic tree construction is becoming increasingly important, as vast amounts of biological sequences are used in the evolutionary and functional analysis of grouping organisms. ClustalXeed provides a great opportunity for the discovery of large and reliable phylogenetic trees between interesting biological sequence data sets. Based on the combination of 38 new protein-weight matrices, encouraging results show that ClustalXeed is one of most powerful MSA tools available for the rapid construction of huge phylogenies.

Project name: ClustalXeed.

Project home page: http://pchm.inje.ac.kr/clustalxeed.

Operating System(s): runs both on single and parallel Linux-based clustal systems.

Programming Language: C-language, Message Passing Interface (MPI).

Other requirements: described in the project home page. More detailed installation instructions and a tutorial manual will also be available for download.

License: ClustalXeed is available in source code, without fee, for academic, non-profit uses. The program is distributed in the hope that it will become further improved and put to practical use.