We have in below three sections investigated accuracy and runtime metrics for v0.7 on SIM3 and REPEATS datasets included in the preprint, as well as performance of SNV and small indel calling for additional simulated and biological (GIAB) datasets.
For the biological SNV and indel experiments, we used GIAB datasets (HG004; Mother) with 2x150bp reads (subsampled to ~26x coverage) and 2x250bp reads (~17x coverage).
Below shows the accuracy (panel A) runtime (panel B) and %-aligned reads (panel C) for the SIM3 (Fig 1) and REPEATS (Fig 2) datasets in the preprint using strobealign v0.7. On all but the 2x100 datasets, strobealign has comparable or higher accuracy than BWA-MEM while being substantially faster. On the 2x100 datasets, strobealign has the second highest accuracy after BWA-MEM on SIM3 while being substantially faster, and comparable accuracy to minimap2 and BWA-MEM on the REPEATS dataset while being twice as fast.
Figure 1. Accuracy (panel A) runtime (panel B) and %-aligned reads (panel C) for the SIM3 dataset
Figure 2. Accuracy (panel A) runtime (panel B) and %-aligned reads (panel C) for the REPEATS dataset
A small SNV and INDEL calling benchmark with strobealign v0.7 is provided below. We used bcftools to call SNPs and indels on a simulated repetitive genome based on alignments from strobealign, BWA-MEM, and minimap2 (using one core). The genome is a 16.8Mbp sequence consisting of 500 concatenated copies of a 40kbp sequence which is mutated through substitutions (5%) and removing segments of size 1bp-1kbp (0.5%) along the oringinal 20Mbp string.
Then, 2 million paired-end reads (lengths 100, 150, 200, 250, 300) from a related genome with high variation rate: 0.5% SNVs and 0.5% INDELs. The challange is to find the right location of reads in the repetitive genome to predict the SNVs and INDELs in the related genome. In the genome where the reads are simulated from there is about 78k SNVs and INDELS, respectively. Locations of true SNVs and INDELs and provided by the read simulator. The precision (P), recall (R), and F-score are computed based on the true variants (for details see section Variant calling benchmark method). Results in table below.
In the experiments strobealign is in general the fastest tool, has the highest SNV precision, and highest precision, recall, and F-score for indels.
There are frequent indels in this dataset (every 200th bases on average) requiring calls to base level alignments for most reads. Between 65-85% of strobealign's runtime is spent on base level alignments with third-party SSW alignment module. The longer the reads the higher % of time is spent on base level alignment. Speed improvements to base-level alignment libraries will greatly reduce runtime on this dataset.
| Read length | Tool | SNVs (P) | SNVs (R) | SNVs (F-score) | Indels (P) | Indels (R) | Indels (F-score) | Alignment time (s) |
|---|---|---|---|---|---|---|---|---|
| 100 | strobealign | 97.9 | 93.5 | 95.6 | 55.6 | 41.1 | 47.2 | 424 |
| minimap2 | 91.4 | 94.3 | 92.8 | 55.2 | 39.1 | 45.8 | 605 | |
| bwa_mem | 93.7 | 95.9 | 94.8 | 55.3 | 30.0 | 38.9 | 1020 | |
| 150 | strobealign | 96.6 | 92.7 | 94.6 | 55.2 | 46.2 | 50.3 | 350 |
| minimap2 | 89.8 | 94.6 | 92.1 | 54.9 | 44.8 | 49.3 | 902 | |
| bwa_mem | 96.0 | 96.0 | 96.0 | 55.0 | 39.6 | 46.1 | 1010 | |
| 200 | strobealign | 97.4 | 94.1 | 95.7 | 55.3 | 45.8 | 50.1 | 487 |
| minimap2 | 88.1 | 96.7 | 92.2 | 55.0 | 44.7 | 49.3 | 1290 | |
| bwa_mem | 95.2 | 96.5 | 95.8 | 55.1 | 42.3 | 47.8 | 1263 | |
| 250 | strobealign | 96.4 | 93.3 | 94.8 | 55.1 | 45.0 | 49.6 | 697 |
| minimap2 | 87.7 | 94.8 | 91.1 | 54.9 | 43.8 | 48.7 | 998 | |
| bwa_mem | 94.3 | 96.2 | 95.2 | 55.1 | 42.3 | 47.8 | 1593 | |
| 300 | strobealign | 95.7 | 92.7 | 94.1 | 55.1 | 44.5 | 49.2 | 1005 |
| minimap2 | 88.2 | 94.3 | 91.2 | 54.8 | 43.4 | 48.4 | 1046 | |
| bwa_mem | 93.7 | 96.4 | 95.0 | 54.9 | 42.0 | 47.6 | 1988 |
We simulated 2x150 and 2x250 reads at 30x coverage from a human genome with SNV and indel rate according to the SIM3 genome (described in the preprint). We aligned the reads to hg38 without alternative haplotypes as proposed here. We used 16 cores for all aligners.
Results are shown for SNVs and indels separately in Figure 3. For SNVs, predictions with strobealign as the aligner have an F-score on par with most other aligners. BWA has the best performance on this dataset. However, indel predictions have both the highest recall and precision using strobealign. Minimap2 is the close second best aligner for calling indels on this dataset, having only 0.1% lower recall and precision to strobealign.
Figure 3. Recall precision and F-score for the aligners on 2x150 and 2x250 datasets from SIM3.
We used Illumina paired-end reads from the GIAB datasets HG004 (Mother) with the 2x150bp reads (subsampled to ~26x coverage; using only the reads in 140818_D00360_0047_BHA66FADXX/Project_RM8392) and 2x250bp reads (~17x coverage). We aligned the reads to hg38 without alternative haplotypes as proposed here. We used 16 cores for all aligners. We obtain the "true" SNVs and INDELs from the GIAB gold standard predictions formed from several sequencing technologies. They are provided here.
Results are shown for SNVs and indels separately in Figure 4. For SNVs, predictions with strobealign as the aligner have the highest F-score of all benchmarked aligners on both datasets. Strobealign's alignments yield the highest precision at the cost of a slightly lower recall. As for indels, predictions have a low recall, precision, and F-score with all aligners. This may be because we benchmarked against all gold standard SVs for HG004 that were not SNVs (see below method for the evaluation). Overall, predictions using Bowtie2 are the most desirable on these datasets.
Figure 4. Recall precision and F-score for the aligners on 2x150 and 2x250 datasets from HG004.
For the four larger datasets above we show the runtime of aligners using 16 threads in Figure 5. The two SIM3 datasets are denoted SIM150 and SIM250, and the two GIAB datasets are denoted BIO150 and BIO250. Urmap was excluded from the timing benchmark because we can only get it to run with 1 core on our server as reported here. Strobealign is the fastest aligner across datasets. While urmap could be faster (based on the singlethreaded benchmarks), strobealign has substaintially better accuracy and downstream SV calling statistics (as seen in previous sections).
Figure 5. Runtime of aligners using 16 threads on two simulated and two biological datasets of about 20-30x coverage of a human genome.
For the results, we ran
bcftools mpileup -O z --fasta-ref ref aligned.bam > aligned.vcf.gz
bcftools call -v -c -O v aligned.vcf.gz > aligned.variants.vcf.gz
# Split into SNP and INDELS
grep -v -E -e "INDEL;" aligned.variants.vcf.gz > aligned.variants.SNV.vcf
grep "#" aligned.variants.vcf.gz > aligned.variants.INDEL.vcf
grep -E -e "INDEL;" aligned.variants.vcf.gz >> aligned.variants.INDEL.vcf
# Separate GIAB SNVs and INDELS
shell('zgrep "#" true.variants.vcf > true.variants.SNV.vcf')
shell('zgrep -P "\t[ACGT]\t[ACGT]\t" true.variants.vcf >> true.variants.SNV.vcf')
shell('zgrep -v -P "\t[ACGT]\t[ACGT]\t" true.variants.vcf > true.variants.INDEL.vcf')
for type in SNV INDEL
do
bcftools sort -Oz aligned.variants.$type.vcf.gz -o aligned.variants.sorted.$type.vcf.gz
bcftools index aligned.variants.sorted.$type.vcf.gz
bcftools isec --nfiles 2 -O u true_variants.sorted.$type.vcf.gz aligned.variants.sorted.$type.vcf -p out_$type
done