Why paired end sequencing

Therefore, the models produced by Edgar and Flyvbjerg [ 20 ] are invalid. One reason for the lack of independence of errors in paired-end sequencing stems from the beginning of a sequencing run, during first-strand synthesis Fig.

Since the original DNA fragment is denatured after it is copied, any errors made during this step will be propagated throughout the cluster that is formed during bridge amplification. Thus, both paired reads will contain the errors, but the erroneous bases will not have reduced quality scores.

First-strand synthesis on the flow cell. A single-stranded DNA fragment to be sequenced anneals to an oligonucleotide that is covalently attached to the flow cell surface. The primer is extended to copy the DNA fragment, which is then removed by denaturation [ 1 ]. Thus, they may not work as well with other Illumina platforms, such as the NextSeq. NGmerge provides the option to forgo its default quality score profiles and instead to utilize calculations similar to those of fastq-join and FLASH , which, though simplistic, are conservative over most of the score ranges.

A third option is for the user to supply custom matrices of quality score profiles to NGmerge. Inaccuracies in reference sequences are a persistent problem that adversely affect error rate calculations [ 22 ], and in fact that proved to be the case here. Once the PhiX reference genome was corrected to account for the five sequence variants, the calculated error rates closely followed the expected relationship shown in Eq.

It is important to note that, in general, errors occurring during the library preparation process e. PCR amplification can be misconstrued as sequencing errors, leading to specious conclusions [ 23 ]. This is another reason why unamplified PhiX remains an enduring control in Illumina sequencing applications. We have examined errors produced by Illumina sequencing technology via reads derived from PhiX.

We have found that variants from the canonical PhiX reference genome account for most of the discrepancy between the actual and theoretical relationships between quality scores and error rates. Furthermore, in the course of developing empirical models for error rates of paired-end sequence reads, we have demonstrated the fallacy of the assumption that has been repeatedly made, both implicitly and explicitly, that errors in such reads are independent.

Finally, we have described a free and open-source program, NGmerge, that merges paired-end sequence reads, thus correcting errors and ambiguous bases, and assigning quality scores that are consistent with the measured error rates. The program can also be run in an alternative mode simply to remove contaminating sequencing adapters. The program is written in C and is parallelized with OpenMP 4. In either mode, NGmerge tests all possible gapless alignments of a pair of reads in attempting to find an optimal alignment.

If multiple valid alignments are found, the one with the lowest fraction mismatch is selected as the optimum. In all of these calculations, ambiguous bases Ns are considered neither matches nor mismatches. The bases and quality scores of any non-overlapping regions are copied into the new read. NGmerge v0. Fastq-join v1. Because fastq-join does not allow for dovetailed alignments, adapters were removed from the reads with NGmerge prior to analysis with fastq-join. PEAR v0.

The cap on output quality scores was increased from the default value of 40 -c With the other merging programs, a custom Python script findDiffs.

The reads of each of the datasets were aligned to this genome using Bowtie2 [ 24 ], as described below. Pileup files were created from the alignment files using SAMtools v1. The downloaded PhiX genome was modified to incorporate the five variants observed in the datasets Table 1.

Reads were aligned to the modified PhiX reference genome using Bowtie2 v2. Error rates were calculated by quality score for the alignments in the SAM alignment files by a custom Python script countErrors. When analyzing SAM files of merged reads, the script was provided the original read length s and a list of merging mismatches and Ns, in order to further categorize the errors based on the nucleotides in the R1 and R2 original reads, into matches, mismatches, and Ns Additional file 1 : Figure S3.

The list of mismatches and Ns was produced by NGmerge or findDiffs. In order to create the error profiles of NGmerge, we used the script countErrors3D. The paired-end reads were aligned to the modified PhiX genome after adapter-trimming with NGmerge, as described above. A LOESS regression function relating the quality scores to the logarithm base 10 of the error rates was calculated in R v3.

This formed the baseline error profile for subsequent analyses. To create the quality score profiles of NGmerge, the same reads were processed with NGmerge in stitch mode, allowing dovetailed alignments -d. The merged reads were aligned to PhiX, and error rates were calculated for each combination of the quality scores of the R1 and R2 reads with countErrors3D.

Then, for each table, a two-dimensional LOESS regression function relating both quality scores to the log base 10 of the error rates and predicted error rates were calculated in R.

These error rates were then transformed back into quality scores using the baseline error profile calculated for the original paired-end reads. The sequencing runs of over SRA studies were examined, though some were eliminated immediately for various reasons mislabeled as paired-end; actual read lengths shorter than stated; reads already trimmed. Those with at least 10, read pairs aligning to PhiX in a properly-paired configuration were further analyzed. The details of these 33 datasets, which contained sequencing runs and a total of 2.

The reads of the 33 SRA datasets were analyzed in a similar fashion to the Harvard datasets. The baseline error rates were calculated from the original reads, and error rates were also determined after processing the reads with each of the merging programs. For each set of error rates, LOESS regression functions were computed, relating the quality scores to the log base 10 of the error rates.

Accurate whole human genome sequencing using reversible terminator chemistry. High-throughput sequencing technologies. Mol Cell. Nucleic Acids Res. Ewing B, Green P. Base-calling of automated sequencer traces using phred. Error probabilities. Learn More. Sequencing Read Length Choosing the right sequencing read length depends on your sample type, application, and coverage requirements.

RNA-Seq Overview This method offers a high-resolution view of coding and noncoding regions of the transcriptome for a deeper understanding of biology. NGS is Revealing the Mysterious World of Microbes Researchers are using 16sRNA to investigate the genomes of microbes and improve our understanding of human health, disease, and microbial evolution. Library Preparation Innovative, comprehensive library prep solutions are a key part of the Illumina sequencing workflow. Explore Library Prep.

Interested in receiving newsletters, case studies, and information from Illumina based on your area of interest? Sign up now. View Video x Sequencing Technology Video. It can also improve the assembly of repetitive regions. This degree of accuracy may not be required for all experiments, however, and paired-end reads are more expensive and time-consuming to perform than single-end reads.

The depth of coverage is a measure of the number of times that a specific genomic site is sequenced during a sequencing run. In exome sequencing, for example, the target might be 60X coverage, meaning that — on average — each targeted base is sequenced 60 times.

This does not mean that every targeted base is sequenced every time; some segments may be read or more times, while others might only be read once or twice, or not at all. The higher the number of times that a base is sequenced, the better the quality of the data.

For RNA-seq, we generally recommend a minimum of 20 million reads per sample. The quality score assigned to the assembled sequence is the geometric mean of the quality scores calculated above, which compensates for the variable lengths of the final sequences. PANDAseq enables users to reject sequences based on low quality score, lengths that are too short or too long, or the presence of uncalled bases. A module system is also available within PANDAseq to allow more sophisticated validation of user sequences, such as verification of known secondary structure or conserved regions.

Note that there is a detailed manual included with the software that describes example usage scenarios. Simulated data was useful in determining how real quality scores affect sequence assembly.

We used a previously published Illumina sequencing run of V3 hypervariable regions from a defined library described below [ 1 ] and replaced the sequence with the corresponding region from Sinorhizobium meliloti bases, region amplified by f and r excluding primers [ 9 ] , up to the length of the original reads. Although this V3 sequence was taken from the published genome, it corresponds to the region being sequenced in the experimental data such that any sequencing quality problems due to secondary structure are preserved.

This provides simulated error-free reads with experimental quality scores. Though the assembly was then performed without a quality filter, all 1 synthesized paired-end sequences assembled with quality scores greater than 0. This value establishes an upper limit on the quality score independent of sequencing errors; that is, setting the quality threshold higher than 0.

Quality scores of assembled masked data. A histogram of quality scores for the assembled sequences is shown. Of these, sequences were assembled with an assembly quality score greater than or equal to 0. We assembled the same single-template data with a quality threshold of 0. The errors in the original, individual reads and the reconstructed sequences were counted and error information is shown in Table 1.

Only two reads contained uncalled bases and were excluded. PANDAseq improved the correctness of the reconstructed sequence relative to the original reads or preserved the correctness of good reads. Depending on the quality threshold, only about 0.

Given an assembly threshold of 0. We determined the geometric mean of the read qualities of the sequences which assembled to be no lower than 0. Only 0. Therefore, if a sequence assembles, it is probably correct, given the quality of the underlying read, regardless of quality score.

For this M. If assembling sequences where the overlap region is large, it is possible that the end of one read would overlap the primer region of the other see the highlyoverlapping scenario shown in Figure 1.

Shown in Figure 3 is a scatter plot of accuracy versus coverage for the four different methods we considered. PANDAseq assembles the fewest reads in the dataset, but was, by far, the most accurate. SHERA assembled all sequences in the dataset, but it is worth noting that, upon inspection, many of the products assembled exclusively by SHERA were incorrect as an erroneous overlap region had been selected data not shown.

The number of error-free sequences in the overlap region is shown in Table 2. Comparison of output of various assemblers. A scatter plot of the percentage of paired-end sequence assemblies from sequenced V3-region amplicons of Methylococcus capsulatus strain Bath against the average number of mismatching nucleotides between the assembled sequence and the reference sequence.

In this composite library, the most abundant sequences are from the added pure cultures, but there are other contaminant sequences, likely from the growth media used [ 1 ]. At a threshold of 0. Relaxing the quality threshold increased sequence recoveries substantially. When the quality threshold was reduced to 0.