View markdown source on GitHub

Quality Control


Bérénice Batut



last_modification Last modification: Jul 26, 2021

Why Quality Control?

Speaker Notes

Potential audience poll ideas:


Where is my data coming from?

Cartoon of different types of sequencing and where they appear in the genome. Bisulfite and ChIP-Seq have arrows pointing to nucleosomes. DNaseq-seq points to the region between nucleosomes. Hi-C and ChIA-PET point to the long range chromatin interactions. RNA-Seq points to a subset of the genome showing a promoter and transcribed region.

Ecker et al, Nature, 2012

Speaker Notes

Segue: Might be concerned about different processing for each

From experiments to data

RNA Seq, Exome Seq, ChIP-Seq, and DNA-Seq all point to a large sequencing box and produce files. Then come bioinformatic analysis, namely quality control, on all of the different types.

Quality control = First step of the bioinformatics analyses

Speaker Notes

Segue: So let’s look at how that data is stored

Sequences: FASTA

>Identifier1 (comment)
>Identifier2 (comment)

Speaker Notes

Segue: But this is just sequence, and we have data from a sequencer, which includes quality

Sequences: FASTQ

@Identifier1 (comment)
@Identifier2 (comment)

Speaker Notes

Segue: so what do the quality chars mean?

Quality score

Measure of the quality of the identification of the nucleobases
generated by automated DNA sequencing

Phred Quality Score Probability of incorrect base call Base call accuracy
10 1 in 10 90%
20 1 in 100 99%
30 1 in 1000 99.9%
40 1 in 10,000 99.99%
50 1 in 100,000 99.999%
60 1 in 1,000,000 99.9999%


Speaker Notes

Quality score

Graph of quality score vs probability of misidentification. There are two lines, red shows -10log(p) while solexa has a different formula

Speaker Notes

Quality score encoding

Encoding of the quality score with ASCII characters for different Phred encoding. The ascii code sequence is shown at the top with symbols for 33 to 64, upper case letters, more symbols, and then lowercase letters. Sanger maps from 33 to 73 while solexa is shifted, starting at 59 and going to 104. Illumina 1.3 starts at 54 and goes to 104, Illumina 1.5 is shifted three scores to the right but still ends at 104. Illumina 1.8+ goes back to the Sanger except one single score wider. Illumina

Speaker Notes

Identifying Potential Quality Issues


Screenshot of FastQC report, showing the table of contents with green checks on nearly every result, and the base statistics and per-base sequence quality graphs shown.

Speaker Notes

Quality score: Per-base

Fastqc quality score plot, most results are in the green region but the box portion of the box and whisker plot start to dip into the yellow, medium quality (less than 30) region near 34+ base position in read. The whiskers begin extending to the red region (less than 20) by base 31 and get progressively worse.

Good quality score

Per-base Quality

Fastqc quality score plot, most results are in the green region up until 30. The whiskers extend to the yellow region from the start, and after base 30 get progressively worse, goign to the worst possible score by the end. The boxes cover the yellow region by base 40.

Bad quality score

Per-base Quality

A graph in between the previous two, mostly green, but gts to yellow  by base 30 and red by base 33.

Intermediate quality score

Per-sequence Quality

A per-sequence quality showing quality score distribution histogram with average quality per read plotted. Most reads pile up around quality 30, but another peak appears at 17. very few reads have quality less than 10.

Per-tile Quality

A heatmap with several red and green squares, but overall largely blue.

Speaker Notes In Illumina libraries, the original sequence identifier is retained. Encoded in these is the flowcell tile from which each read came.

There might be transient problems such as bubbles going through the flowcell, or more permanent problems such as smudges on the flowcell, or debris inside the flowcell lane.

This graph will only appear with Illumina libraries which retain their original sequence identifiers. The graph allows to check the quality scores from each tile across all bases, to see if there was a loss in quality associated with only one portion of the flowcell. The plot shows the deviation from the average quality for each tile. The colours are on a cold to hot scale, with cold colours being positions where the quality was at or below the average for that base in the run, and hot colours to indicate that a tile had worse quality reads than other tiles for that base. In the example below you can see that certain tiles show consistently poor quality. A good plot should be blue all over.

Per-base Sequence Content

A line chart mapping sequence content across all bases with % of the four nucleotides plotted. The graph starts off very jagged and stabilises by base 12.

Speaker Notes The per-base sequence content highlights the proportion of each base in each position of a sequence for which each of the four DNA bases have been called. In a random library there would be little to no difference between the different bases of a sequence run. The relative amount of each base should reflect the overall amount of these bases, but in any case they should not be hugely imbalanced from one another. It is worth noting that some types of libraries will always produce biased sequence composition, normally at the start of the read. Libraries produced by priming with random hexamers (including nearly all RNA-Seq libraries) and those which were fragmented using transposases, inherit an intrinsic bias in the positions at which the reads start. This bias does not concern an absolute sequence, but instead provides an enrichment of a number of different K-mers at the 5’ end of the reads. Whilst this is a true technical bias, it isn’t something which can be corrected by trimming and in most cases doesn’t seem to adversely affect the downstream analysis. It will however produce a warning or error in this module.

There are a number of common scenarios for these issues:

Per-sequence GC content

A line chart showing mean GC content and threoretical distribution as largely overlapping peaks.

Speaker Notes The GC content distribution of most samples should follow a normal distribution. In some cases, a bi-modal distribution can be observed, especially for metagenomic data sets. An unusually shaped distribution could indicate a contaminated library or some other kinds of biased subset. A normal distribution which is shifted indicates some systematic bias which is independent of base position. Such a systematic bias creating a shifted normal distribution won’t be flagged as an error, since the tool cannot guess what the provided genome’s GC content should be.

Issues in the GC content distribution usually indicate a problem with the library. Sharp peaks on an otherwise smooth distribution are normally the result of a specific contaminant (adapter dimers for example), which may well be picked up by the over-represented sequences module. Broader peaks may represent contamination with a different species.

Per-base N content

A line graph of N content across all bases. It shows several peaks to 65 at specific positions and goes to 100 near the end.

Speaker Notes Sequences can contain the ambiguous base N for positions that could not be identified as a particular base. A high number of Ns can be a sign for a low quality sequence or even dataset. If no quality scores are available, the sequence quality can be inferred from the percent of Ns found in a sequence or dataset.

If a sequencer is unable to make a base call with sufficient confidence then it will normally substitute an N rather than a conventional base call. It’s not unusual to see a very low proportion of Ns appearing in a sequence, especially nearer the end of a sequence. However, if this proportion rises above a few percent it suggests that the analysis pipeline was unable to interpret the data well enough to make valid base calls.

Sequence length distribution

A graph with a single peak at 75, and 0 outside of this region.

Speaker Notes Some high throughput sequencers generate sequence fragments of uniform length, while others can output reads of wildly varying lengths. The length distribution can be then used as quality measure. You would expect a normal distribution for the best result. However, most sequencing results show a slowly increasing and then a steep falling distribution.

FastQC generates a graph showing the distribution of fragment sizes in the file which was analysed. In many cases this will produce a simple graph showing a peak only at one size, but for variable length FASTQ files this will show the relative amounts of each different size of sequence fragment.

This module will raise a warning if all sequences are not the same length. This module will raise an error if any of the sequences have zero length.

Duplicated sequences

Two line graphs, deduplicated sequences in red, and total sequences in blue. They start off near 100 and go rapidly to zero by sequence duplication level 2 and 3.

Speaker Notes This quality check module counts the degree of duplication for every sequence in the library, and creates a plot showing the relative number of sequences with different degrees of duplication:

In genomic projects, sequence duplication is investigated. Duplicated sequences can arise when there are too few fragments present at any stage prior to sequencing.

This module issues a warning if non-unique sequences make up for more than 20% of the total sequences. An error is raised instead if non-unique sequences make up for more than 50% of the total.

Tag sequences: Adapter contamination

The graph shows a line at zero for the five possible datasets.

Speaker Notes Tag sequences are artifacts at the ends of sequence reads such as multiplex identifiers, adapters, and primer sequences that were introduced during pre-amplification with primer-based methods. The base frequencies across the reads present an easy way to check for tag sequences. If the distribution seems uneven (high frequencies for certain bases over several positions), it could indicate some residual tag sequences. This doesn’t indicate a problem as such - just that the sequences will need to be adapter trimmed before proceeding with any downstream analysis.

To investigate tag or adapter content, FastQC generates a plot showing a cumulative percentage count of the proportion of the library which has seen each of the adapter sequences at each position. Once a sequence has been seen in a read it is counted as being present right through to the end of the read so the percentages you see will only increase as the read length goes on.

Tag sequences: K-mer content

several peaks are visible on the log2 obs/exp graph of different k-mers at different positions in the read.

Speaker Notes Another way to find tag sequences is to look at the K-mer content, and find those which do not have even coverage through the length of your reads and could correspond to tag sequences.

K-mers with positionally biased enrichment are reported. The top 6 most biased K-mer are additionally plotted to show their distribution.

Over-represented K-mers will appear as sharp spikes at a single point in the sequence, deviating from what should be a progressive or broad enrichment.

Improving the quality of sequences

Key Points

Thank you!

This material is the result of a collaborative work. Thanks to the Galaxy Training Network and all the contributors! Galaxy Training Network This material is licensed under the Creative Commons Attribution 4.0 International License.