Parallelization of MAFFT for large-scale multiple sequence alignments

Abstract

Summary

We report an update for the MAFFT multiple sequence alignment program to enable parallel calculation of large numbers of sequences. The G-INS-1 option of MAFFT was recently reported to have higher accuracy than other methods for large data, but this method has been impractical for most large-scale analyses, due to the requirement of large computational resources. We introduce a scalable variant, G-large-INS-1, which has equivalent accuracy to G-INS-1 and is applicable to 50 000 or more sequences.

Availability and implementation

This feature is available in MAFFT versions 7.355 or later at https://mafft.cbrc.jp/alignment/software/mpi.html.

Supplementary information

Supplementary data are available at Bioinformatics online.

A large number of biological sequences from widely divergent organisms are becoming available. Accordingly, the need for multiple alignments of large numbers of sequences is increasing for various kinds of sequence analysis. The G-INS-1 option of MAFFT was recently reported to have higher accuracy than other methods for large multiple sequence alignments (MSAs) in independent benchmarks (Le et al., 2017; Yamada et al., 2016). However, this method was impractical for actual analyses, requiring large computational resources in both space and time to perform all-to-all pairwise alignments by dynamic programming (DP) (Needleman and Wunsch, 1970), which are used for a guide tree and a scoring function similar to COFFEE (Notredame et al., 1998). Here, we introduce a scalable variant, G-large-INS-1, which has equivalent accuracy to G-INS-1 and is applicable to 50 000 or more sequences. Our strategies to reduce computational costs are (i) parallelization across multiple machines and/or processor cores using MPI and Pthreads to increase speed and (ii) the use of a high-speed shared filesystem, which is becoming common for processing big data. An MPI-based parallelization of another high-accuracy MSA method, MSAProbs, was recently released (González-Domínguez et al., 2016), but it cannot be applied to thousands of sequences. The present update of MAFFT is designed to satisfy the need for accurately aligning large numbers of sequences but is not applicable to long genomic sequences since the length dependence of the computational cost is unchanged. The G-large-INS-1 option is available in MAFFT versions 7.355 or later and the online service (Katoh et al., 2017).

Accuracy of G-large-INS-1 was compared with that of conventional G-INS-1 using different benchmarks, QuanTest (Le et al., 2017) (Fig. 1a), HomFam (Sievers et al., 2011), OXFam (Raghava et al., 2003; Yamada et al., 2016) and ContTest (Fox et al., 2016) (Supplementary Table S1). Both methods ran with different input orders and/or minor variations in pairwise alignment and guide tree (see Supplementary Data) in order to assess instability of accuracy scores (Boyce et al., 2015). In all cases, the difference between G-INS-1 (blue lines in Fig. 1a) and G-large-INS-1 (red lines) was small.

Fig. 1.

(a) QuanTest. Accuracy of protein secondary structure prediction based on various sizes of MSAs by G-large-INS-1 (red bold lines), G-INS-1 (version 7.245; blue bold lines) and other popular methods. We used 1940 (out of 2265) entries so that JPred (Drozdetskiy et al., 2015) can be consistently applied to the MSAs by all methods. (b)–(g), Parallelization efficiency of all-to-all alignment stage (b, d and f) and progressive stage (c, e and g) when applying G-large-INS-1 to LSU rRNA (b, c) sdr (d, e) and zf-CCHH (f, g). Green squares and magenta triangles are the computational time on NFS and Lustre filesystem, respectively. Lines are the expected time based on the cases using seven cores [NFS; green solid lines in (b), (d) and (f)], 35 cores [Lustre; magenta dotted lines in (b), (d) and (f)] and single core (c, e and g), assuming a perfect efficiency. The calculations with NFS (green) were performed on a heterogeneous cluster system (each node has 16–20 cores of Intel Xeon E5-2660 v3 2.6 GHz, E5-2680 2.7 GHz and E5-2670 v2 2.50 GHz with 64–128GB RAM). The calculations with the Lustre filesystem (magenta) were performed on Intel Xeon E5-2695 v4 2.10 GHz 36 cores with 256GB RAM per node using Lustre version 2.5.42

Large amounts of RAM are required if conventional tools for high-quality MSAs are applied to a large number of sequences. For example, MAFFT-L-INS-i and MSAProbs-MPI used at most 9.23GB and 74.8GB for a subset of 1000 sequences in QuanTest. For a larger subset (4000 sequences), MAFFT-G-INS-1 and QuickProbs2 (Gudyś and Deorowicz, 2017) used at most 26.0 GB and 411 GB RAM, respectively. In contrast, G-large-INS-1 used only 5.72GB at most, for the subset of 4000 sequences. Memory usage for larger problems (up to ∼90 000 sequences) is shown in Supplementary Table S1, which suggests that this advantage increases with the number of sequences. Note that G-large-INS-1 uses files to save temporary data and thus requires a high-speed filesystem when the input sequences are very short, as discussed below.

Parallelization efficiency in three examples is shown in Figure 1(b–g), separately for two stages: (i) the all-to-all alignment stage (b, d and f) and (ii) the progressive alignment stage (c, e and g).

For LSU rRNA sequences (b, 1521–4102 bases, 1000 sequences randomly selected from the SEED alignment in Silva (Glöckner et al., 2017) and protein sequences with usual lengths (d, 21–297 amino acids, 50 157 sequences, the ‘sdr’ family taken from HomFam), the wall-clock time for the all-to-all alignment stage decreased almost linearly with the number of cores used for the calculation. However, for a dataset with very short sequences (f, 12–35 amino acids, 88 345 sequences, the ‘zf-CCHH’ family taken from HomFam), the efficiency differs depending on filesystem: high in Lustre (shown with magenta triangles) but low in NFS (shown with green squares). This difference is due to the balance between calculation and disk operations. As noted earlier, a considerable amount of temporary data is written in parallel into the filesystem: approximately 218 MB, 100 GB and 142 GB for LSU rRNA, ‘sdr’ and ‘zf-CCHH’, respectively, in the examples shown here. Overhead due to these disk operations is almost negligible in the former two cases but not in the latter case, where alignment of ∼23 amino acids takes only a short time in comparison with the time to write the temporary data to disk using NFS.

Figure 1c, e and g suggest that the wall-clock time of the progressive stage varies for each run and does not linearly decrease, but usually this is not a speed-limiting step. CPU time and wall-clock time for various problems are shown in Supplementary Table S1.

Until now, it was necessary to use highly approximate methods, such as the FFT-NS-2 option of MAFFT or the progressive option of Clustal Omega, in order to construct large MSAs. In terms of the MSA itself, the accuracy of these methods tends to decrease along with the increase in the number of sequences. This was first pointed out by Sievers et al. (2013) and confirmed by Le et al. (2017). The increase in accuracy observed in Figure 1a for more than 200 sequences is due to the prediction phase not due to the alignment phase (see the last section in Supplementary Data and black dashed lines in Supplementary Fig. S1). As a result, it was difficult to know how many sequences should be included in an MSA. With more sequences, the MSA has richer comparative information, but the alignment quality is expected to decrease. The optimal balance between these two factors may differ by case. In contrast, the accuracy of G-large-INS-1 and G-INS-1 (red and blue dashed lines in Supplementary Fig. S1) was robust to data size in this test. The number of sequences to include in the MSA can now be determined simply based on the computational resources available and the requirements for the downstream analysis.