Sensitive protein alignments at tree-of-life scale using DIAMOND

Abstract

We are at the beginning of a genomic revolution in which all known species are planned to be sequenced. Accessing such data for comparative analyses is crucial in this new age of data-driven biology. Here, we introduce an improved version of DIAMOND that greatly exceeds previous search performances and harnesses supercomputing to perform tree-of-life scale protein alignments in hours, while matching the sensitivity of the gold standard BLASTP.

Fig. 1. Benchmark of DIAMOND, MMSeqs2 and BLASTP using various sensitivity modes.

Computational speedup and alignment sensitivity comparisons were carried out between the new version of DIAMOND, v2.0.7 (using default, --sensitive, --very-sensitive and --ultra-sensitive modes), the old version of DIAMOND, v0.7.12 (using default and --sensitive modes), MMSeqs2 release 11 (using modes s = 1.0, s = 2.5, s = 6.0, s = 7.5, s = 7.5** with --max-seqs 100000), BLASTP v2.10.0 and QuickBLAST v0.0.0. a, Alignment sensitivity (AUC1) measured as the fraction of the query’s protein family covered until the first false positive, averaged over all queries in the benchmark dataset. Dashed vertical line, alignment sensitivity level of BLASTP v2.10.0 (AUC1 = 0.622). b, ROC curves of the same benchmark showing the true average error rate per query versus the average coverage of the protein family, depending on the e-value threshold. Source data

Fig. 2. Strong scaling of DIAMOND on up to 520 nodes (20,800 cores) of the Cobra supercomputer of the Max Planck Society.

The parallel scalability of DIAMOND in a distributed computing environment (supercomputer) is shown. Runs are performed in --ultra-sensitive mode, using the NCBI nr database (~281 million sequences) as the query database and the UniRef50 database (~39 million sequences) as the reference database. The full alignment was performed on 520 nodes (20,800 cores total), finishing in less than 18 hours of wall-clock time. On 1, 14, 28 and 56 nodes, only a subset of the query blocks could be processed, and the time for the full alignment was linearly extrapolated for each node count. An almost ideal scaling is observed, with a small super-linear effect caused by the caches of the parallel input–output system. To illustrate the massive parallelism, the inset shows a zoomed-in view of the sequence of tasks that DIAMOND (v2.0.0) has performed over time on 20 of the 520 nodes, indicated by blue rectangles (alignment process) and orange rectangles (join operations). White spaces encode the input–output activity on the supercomputer’s shared parallel file system (Extended Data Fig. 1). Source data

Extended Data Fig. 1. Detailed visual trace of a massively parallel DIAMOND run on 520 nodes (41,600 threads) of the Cobra supercomputer of the Max Planck Society.

To illustrate the scalability of DIAMOND (v2.0.0) in a distributed computing environment (supercomputer), line-by-line worker tasks are shown individually for each worker node (the detailed version of Fig. 2). The sequence of tasks DIAMOND (v2.0.0) has performed over time is indicated by blue rectangles and orange rectangles, in which blue rectangles denote the alignment process and orange rectangles represent join operations. The numbers within the rectangle indicate the indices of the individual query chunks and of their respective reference chunks. White spaces encode the IO activity on the supercomputer’s shared parallel file system. The run shown here was performed in–ultra-sensitive mode and used the full NCBI non-redundant database as the query database, and the UniRef50 database as the reference database, finishing in below 18 hours of wallclock time. The result demonstrates DIAMOND’s efficient distribution of massively parallel work packages at scale, showing that all workers finish around the same time without creating a significant load imbalance (Experimental Study).

Extended Data Fig. 2. Assessment of true positive alignments.

AUC1 sensitivity as reported for our main benchmark, resolved by sequence identity of the query-subject association under our SCOPe annotation (middle = median, hinges=25%/75% quantiles, lower/upper whisker = smallest/largest observation greater/less than or equal to lower/upper hinge -/+ 1.5 * IQR). Source data

Extended Data Fig. 3. Assessment of protein family associations.

The distribution (shown in 5% bins) of a query’s protein family member associations with respect to the sequence identity of the corresponding Needleman Wunsch alignments between the annotated ranges (middle = median, hinges=25%/75% quantiles, lower/upper whisker = smallest/largest observation greater/less than or equal to lower/upper hinge −/+ 1.5 * IQR). Source data

Extended Data Fig. 4. Metagenomic benchmark results using Illumina HiSeq 4000 paired end sequencing (2x150 bp) reads from Stewart et al., 2019.

Computational speedup and alignment sensitivity comparisons for translated searches of 150bp Illumina short reads from rumen metagenome samples (Supplementary Benchmark 1). Alignment sensitivity (AUC1) is measured as the fraction of the query’s protein family covered until the first false positive, averaged over all queries in the benchmark dataset. The y-axis denotes the x-fold computational speedup achieved over BLASTX v2.10.0. Source data

Extended Data Fig. 5. Metagenomic benchmark results using Illumina HiSeq 2500 paired end sequencing (2x250 bp) reads from Bahram et al., 2018.

Computational speedup and alignment sensitivity comparisons for translated searches of 250bp Illumina short reads from topsoil metagenome samples (Supplementary Benchmark 2). Alignment sensitivity (AUC1) is measured as the fraction of the query’s protein family covered until the first false positive, averaged over all queries in the benchmark dataset. The y-axis denotes the x-fold computational speedup achieved over BLASTX v2.10.0. Source data

Extended Data Fig. 6. ROC curves for metagenomic benchmark using Illumina HiSeq 4000 paired end sequencing (2x150bp) reads from Stewart et al., 2019.

We show the true average error rate per query (x-axis) against the average coverage of the protein family (y-axis) depending on the e-value threshold for Supplementary Benchmark 1. Source data

Extended Data Fig. 7. ROC curves for metagenomic benchmark using Illumina HiSeq 2500 paired end sequencing (2x250bp) reads from Bahram et al., 2018.

We show the true average error rate per query (x-axis) against the average coverage of the protein family (y-axis) depending on the e-value threshold for Supplementary Benchmark 2. Source data