PEAR: a fast and accurate Illumina Paired-End reAd mergeR

When fragment lengths vary, and especially when the fragment size is shorter than a single-end read or longer than twice the single-end read, most state-of-the-art paired-end mergers fail to generate reliable results.

“However, when fragment lengths vary and, in particular, when either the fragment size is shorter than a single-end read, or longer than twice the size of a single-end read, most state-of-the-art mergers fail to generate reliable results.”

When DNA fragment size is shorter than single-end read length, current mergers generate erroneous results.

“When this does not hold, for example when sequencing the V6 region of 16S rRNA genes of bacterial samples [fragment sizes range between 110 and 130-bp (Gloor et al., 2010)] with read lengths of 150-bp (see case C in Fig. 1), current mergers will generate erroneous results.”

• cites Gloor et al., 2010 sourceNot checked yet

  Not checked in the free run.

PEAR can merge ∼40% of the reads when the mean overlap is only 10-bp.

“PEAR is robust with respect to short overlaps because it can still merge ∼40% of the reads when the mean overlap is only 10-bp.”

When running with four cores, PEAR and FLASH are equally fast.

“Surprisingly, when running the experiment with four cores, which corresponds to a standard desktop PC, PANDAseq is slower than for the sequential run, whereas PEAR and FLASH are equally fast and twice as fast as PANDAseq.”

PEAR implements a statistical test for minimizing false-positive results.

“It also implements a statistical test for minimizing false-positive results.”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

On simulated paired-end reads with a mean overlap of 20-bp, PEAR correctly merges 70.06% of the fragments with a FPR of only 0.48% when the significance level of the test is set to 1%.

“It correctly merges 70.06% of the fragments with a FPR of only 0.48% when the significance level of the test is set to 1%.”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

PANDAseq incorrectly merges over one-third of the reads in scenario C.

“PANDAseq incorrectly merges over one-third of the reads in this scenario.”

A quarter of the reads merged by PANDAseq were not mapped to the reference genome using Bowtie2.

“Nonetheless, a quarter of the reads merged by PANDAseq were not mapped to the reference genome using Bowtie2.”

If the overlap is known to be short (35-bp in simulations), PEAR's statistical test will reject up to 4% of correctly merged sequences because of low quality scores.

“If, however, the overlap is known to be short (35-bp in our simulations), our statistical test will reject up to 4% (based on our simulations) correctly merged sequences because of low quality scores.”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

For the single template dataset, PEAR merges the highest number of reads when the statistical test is disabled.

“For this dataset, PEAR merges the highest number of reads when the statistical test is disabled (Table 3).”

The optimal RAM setting on a 48 core Magny-Cours system is 200 MB.

“We observed that the optimal RAM setting on a 48 core Magny-Cours system (see Section 3.4 for details) is 200 MB.”

PEAR merges fewer correct fragments when the statistical test is enabled.

“PEAR merges fewer correct fragments when the statistical test is enabled.”

PEAR is accurate on datasets with short overlaps.

“The program is accurate on datasets with (i) short overlaps”

COPE, PANDAseq and FLASH were unable to merge reads under application scenario C where the DNA fragment size is smaller than a single-end read.

“COPE, PANDAseq and FLASH were unable to merge reads under application scenario C (see Fig. 1) where the DNA fragment size is smaller than a single-end read (Table 1).”

PEAR does not require the target fragment size as input.

“does not require the target fragment size as input”

PEAR is accurate on datasets with DNA target fragment sizes that are smaller than single-end read lengths.

“and (ii) DNA target fragment sizes that are smaller than single-end read lengths.”

COPE shows a lower FPR than PEAR when the statistical test in PEAR is disabled.

“COPE merges less reads than PEAR and shows a lower FPR when the statistical test in PEAR is disabled.”

PANDAseq correctly merges the highest amount of reads in the S.aureus dataset.

“PANDAseq correctly merges the highest amount of reads; PEAR ∼2% less (statistical test disabled).”

PANDAseq incorrectly merges 25.11% of the reads when the mean overlap is set to 10-bp.

“Furthermore, PANDAseq incorrectly merges 55.4% of the reads that do not overlap and 25.11% of the reads when the mean overlap is set to 10-bp.”

PEAR correctly merges ∼2% less reads than PANDAseq when the statistical test is disabled.

“PANDAseq correctly merges the highest amount of reads; PEAR ∼2% less (statistical test disabled).”

The FPR of 0.64% (MAP = 0.01) under the 10-bp mean overlap setting is 10 times lower than for FLASH and PANDAseq.

“The FPR of 0.64% (MAP = 0.01) under this setting is 10 times lower than for FLASH and PANDAseq.”

When reads do not overlap, PEAR classifies them as unmerged with an FPR of 0.03%.

“When reads do not overlap, PEAR classifies them as unmerged with an FPR of 0.03%.”

When running with four cores, PANDAseq is slower than for the sequential run.

“Surprisingly, when running the experiment with four cores, which corresponds to a standard desktop PC, PANDAseq is slower than for the sequential run, whereas PEAR and FLASH are equally fast and twice as fast as PANDAseq.”

COPE did not merge any reads on the single template dataset.

“COPE did not merge any reads, however.”

Only 4.9% of the merged reads from PEAR could not be mapped.

“In contrast only 4.9% of the merged reads from PEAR could not be mapped.”

When MAP is used to evaluate merged reads, PEAR produces more merged reads with an FPR analogous to the FPR generated by PEAR with the statistical test.

“When we use MAP to evaluate the merged reads, PEAR produces more merged reads with an FPR that is analogous to the FPR generated by PEAR with the statistical test.”

PANDAseq's FPR increases with decreasing overlap size, regardless of the minimal overlap size setting.

“However, its FPR increases with decreasing overlap size, regardless of the minimal overlap size setting.”

Neither PANDAseq nor FLASH yields substantial speedups when executed on several cores.

“However, the main observation is that neither PANDAseq nor FLASH yields substantial speedups when executed on several cores (see Fig. 2).”

With the exception of the no-overlap test case, PEAR consistently generates a larger number of correctly merged sequences when the statistical test is disabled.

“With the exception of the first test case (no overlaps), PEAR consistently generates a larger number of correctly merged sequences when the statistical test is disabled.”

PEAR takes quality scores and sequence matches into account.

“Our approach takes quality scores and sequence matches into account.”

Tests on simulated and empirical data show that PEAR consistently generates highly accurate merged paired-end reads.

“Tests on simulated and empirical data show that PEAR consistently generates highly accurate merged paired-end reads.”

On simulated paired-end reads with a mean overlap of 20-bp, PEAR correctly merges 90.44% of the fragments with a FPR of 2.78% when the statistical test is disabled.

“On simulated paired-end reads with a mean overlap of 20-bp (Section 3.1), PEAR correctly merges 90.44% of the fragments with a FPR of 2.78% when our statistical test is disabled.”

Only 0.03% of the merged reads are false positives when using the test.

“When setting and using the test, less reads are merged, but only 0.03% of the merged reads are false positives.”

PEAR does not require any prior knowledge regarding overlap lengths.

“In addition, it does not require any prior knowledge regarding overlap lengths.”

On simulated paired-end reads with a mean overlap of 20-bp, the best competing merger PANDAseq correctly merges 83.51% of the fragments with a FPR of 6.65%.

“The best competing merger (PANDAseq) correctly merges 83.51% of the fragments, but with a FPR of 6.65%.”

PANDAseq is the fastest tool on a single core.

“PANDAseq is the fastest tool on a single core.”

PEAR can reliably identify reads that can either be merged or need to be discarded.

“It can reliably identify reads that can either be merged or need to be discarded.”

PEAR's statistical test is based on observed expected alignment scores (OESs).

“To identify false-positive merged reads, we propose a statistical test that is based on the observed expected alignment scores (OESs).”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

PANDAseq performs equally well as PEAR for the majority of cases where the overlaps exceed 20-bp.

“PANDAseq performs equally well as PEAR for the majority of cases where the overlaps exceed 20-bp.”

FLASH failed to merge the majority of reads for small overlap sizes.

“FLASH failed to merge the majority of reads for small overlap sizes, but exhibits low FPRs for merged sequences.”

Although the overlap size is only 18-bp, all mergers decrease the ER.

“Although the overlap size is only 18-bp all mergers decrease the ER.”

When running with four cores, PEAR and FLASH are twice as fast as PANDAseq.

“Surprisingly, when running the experiment with four cores, which corresponds to a standard desktop PC, PANDAseq is slower than for the sequential run, whereas PEAR and FLASH are equally fast and twice as fast as PANDAseq.”

PEAR can produce 4% more merged sequences using MAP at the cost of a slight (~0.1%) increase in FPRs.

“Our tests show that PEAR can produce 4% more merged sequences using MAP at the cost of a slight (∼0.1%) increase in FPRs.”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

On multi-core architectures, the parallel version of PEAR shows linear speedups compared with the sequential version of PEAR.

“On multi-core architectures, the parallel version of PEAR shows linear speedups compared with the sequential version of PEAR.”

When setting the P-value or MAP to 0.01, PEAR shows lower FPRs than all three competing mergers.

“When setting the P-value or MAP to 0.01, PEAR shows lower FPRs than all three competing mergers.”

A highly optimized implementation allows PEAR to merge millions of paired-end reads within a few minutes on a standard desktop computer.

“A highly optimized implementation allows for merging millions of paired-end reads within a few minutes on a standard desktop computer.”

Merged reads produced by FLASH and PANDAseq show ERs that are slightly lower than PEAR (statistical test disabled).

“Merged reads produced by FLASH and PANDAseq show ERs that are slightly lower than PEAR (statistical test disabled).”

All mergers work well in the S.aureus setting.

“All mergers work well in this setting.”

COPE merges less reads than PEAR in the S.aureus dataset.

“COPE merges less reads than PEAR and shows a lower FPR when the statistical test in PEAR is disabled.”

PANDAseq incorrectly merges 55.4% of the reads that do not overlap.

“Furthermore, PANDAseq incorrectly merges 55.4% of the reads that do not overlap and 25.11% of the reads when the mean overlap is set to 10-bp.”

Overall, PEAR shows low FPRs across all test scenarios (overlap lengths).

“Overall, PEAR shows low FPRs across all test scenarios (overlap lengths).”

PEAR yields linear speedups.

“In fact, although PEAR yields linear speedups, FLASH and PANDAseq yield no speedup at all.”

PEAR yields three times lower ERs when the statistical test is enabled.

“However, PEAR yields three times lower ERs when the statistical test is enabled.”

When setting and using the test, less reads are merged.

“When setting and using the test, less reads are merged, but only 0.03% of the merged reads are false positives.”

FLASH exhibits low FPRs for merged sequences.

“FLASH failed to merge the majority of reads for small overlap sizes, but exhibits low FPRs for merged sequences.”

If the sample contains DNA fragments that are at least twice as long as the single-end reads, PANDAseq exhibits a high false-positive rate (FPR).

“Thus, if the sample contains DNA fragments that are at least twice as long as the single-end reads, PANDAseq exhibits a high false-positive rate (FPR).”

PEAR becomes faster when using less memory.

“In Section 2.4 we outline why PEAR becomes faster when using less memory.”

FLASH performs poorly when the overlaps between reads are short.

“Furthermore, our tests show that FLASH performs poorly when the overlaps between reads are short (Section 3).”

Using 48 cores, PEAR requires 62 s to finish on the human chromosome 14 dataset.

“Using 48 cores, PEAR only requires 62 s to finish, whereas FLASH and PANDAseq need 7 and 21.5 min, respectively.”

Using 48 cores, PANDAseq requires 21.5 min to finish on the human chromosome 14 dataset.

“Using 48 cores, PEAR only requires 62 s to finish, whereas FLASH and PANDAseq need 7 and 21.5 min, respectively.”

Using 48 cores, FLASH requires 7 min to finish on the human chromosome 14 dataset.

“Using 48 cores, PEAR only requires 62 s to finish, whereas FLASH and PANDAseq need 7 and 21.5 min, respectively.”

Most current paired-end mergers assume that DNA fragments are longer than individual single-end reads.

“Finally, most current paired-end mergers assume that the DNA fragments are longer than the individual single-end reads.”

If the target DNA fragment size is smaller than twice the length of the single-end reads, the corresponding paired-end reads can be merged into a fragment.

“If the target DNA fragment size is smaller than twice the length of the single-end reads, that is, if there exists an overlap, the corresponding paired-end reads can be merged into a fragment.”

The results are conservative because some correctly merged reads might be missed by Bowtie2 due to sequencing errors.

“Note that, the results are conservative because some of the correctly merged reads might be missed by Bowtie2 due to sequencing errors.”

Overlaps that exclusively maximize f might not yield correctly merged sequences.

“However, overlaps that exclusively maximize f might not yield correctly merged sequences.”

There is an exponential increase in error rates along Illumina reads.

“Furthermore, there is an exponential increase in error rates (ERs) along the reads (Cox et al., 2010).”

• cites Cox et al., 2010 sourceNot checked yet

  Not checked in the free run.

When using the MAP work-around, the P-value of the statistical test is not valid anymore, as ω depends on the output of the algorithm.

“However, when using this work-around, the P-value of the statistical test is not valid anymore, as ω depends on the output of our algorithm.”

• AssumptionThe authors state this as an assumption: In the null model for the statistical test, the two corresponding reads are independent random sequences and there are no sequencing errors.
• AssumptionThe authors state this as an assumption: The null hypothesis is that the two corresponding reads are independent from each other, meaning any overlap between the two reads is purely by chance.

PEAR is robust with respect to short overlaps.

“PEAR is robust with respect to short overlaps because it can still merge ∼40% of the reads when the mean overlap is only 10-bp.”

PEAR can be used for merging sequences with varying fragment sizes.

“Therefore, PEAR can be used for merging sequences with varying fragment sizes.”

FLASH's default setting can be shown to have a P-value of 0.00156 for merged reads by using the statistical test introduced in Section 2.2 and replacing with f.

“This setting can be shown to have a P-value of 0.00156 for merged reads by using the statistical test introduced in Section 2.2 and replacing with f.”

Using Bowtie2 in end-to-end mode without allowing gaps guarantees that all merged reads that can be mapped to the reference genome are correctly merged.

“This guarantees that all merged reads that can be mapped to the reference genome are correctly merged.”

• cites Langmead and Salzberg, 2012 sourceNot checked yet

  Not checked in the free run.

FLASH can only merge paired-end reads into fragments of almost identical size.

“Therefore, it can only merge paired-end reads into fragments of ‘almost’ identical size.”

FLASH requires the mean DNA fragment size and standard deviation of the fragment size as input parameters.

“FLASH requires the mean DNA fragment size and standard deviation of the fragment size as input parameters.”

• cites Magoč and Salzberg, 2011 sourceNot checked yet

  Not checked in the free run.

COPE is substantially slower than the other programs tested.

“Hence, COPE is substantially slower than the other programs we tested.”

When the bases have an average error probability >0.039, PANDAseq will favor merging randomly generated sequences.

“In other words, when the bases have an average error probability that is >0.039, PANDAseq will favor merging randomly generated sequences.”

• AssumptionThe authors state this as an assumption: The merged sequences are generated randomly with all base frequencies equally likely with probability 0.25 and all bases have an equal error probability e.

PEAR can merge paired-end read datasets under settings where most competing mergers fail.

“It can merge paired-end read datasets under settings where most competing mergers fail.”

Because the quality of Illumina reads decreases toward the end of the reads, PANDAseq will incorrectly merge reads that do not overlap.

“Because the quality of Illumina reads decreases toward the end of the reads, PANDAseq will therefore incorrectly merge reads that do not overlap.”

• AssumptionRests on an assumption: The quality of Illumina reads decreases toward the end of the reads.

PEAR scales well on most server and desktop computers.

“Finally, PEAR scales well on most server and desktop computers.”