What the FastQC Per Base Sequence Quality Plot and Phred Scores Mean
A Phred quality score is the sequencer’s estimate of the probability that a single base call is wrong, expressed on a logarithmic scale. The definition is Q = −10 × log₁₀(Perror). Q20 means a 1 in 100 chance the base is wrong, Q30 means 1 in 1,000, and Q40 means 1 in 10,000. FastQC’s “Per base sequence quality” module takes every read in your FASTQ file, groups the quality values by position along the read, and draws a box-and-whisker summary for each position. You are looking at the distribution of per-base confidence as a function of cycle number, and the usual question it answers is whether the run degraded toward the 3’ end.
How the score is stored in a FASTQ record
Each read in a FASTQ file has four lines: an identifier, the sequence, a separator, and a quality string of exactly the same length as the sequence. The quality string encodes each integer Q as a single ASCII character using the Phred+33 convention, so the character code is Q + 33. ! is Q0 and 5 is Q20. : is Q25 and ? is Q30. F is Q37 and I is Q40.
Older Illumina pipelines used Phred+64, which shifts everything by 31 characters. Anything sequenced in the last decade is Phred+33 and FastQC will tell you which encoding it inferred in the Basic Statistics module. If you ever see quality characters in the range a–h, you have a legacy file and every downstream tool needs an explicit encoding flag.
Quality strings are expensive. They compress far worse than the bases themselves and often dominate the size of a compressed FASTQ, which is why lossy quality compression and binning schemes have been studied so heavily 1. Two-level and binned schemes can cut quality storage by large factors with little measured effect on variant calls 2. That matters here because the same binning changes what your FastQC plot looks like.
Reading the plot, and where the thresholds come from
The x-axis is the base position within the read, grouped into bins for longer reads. For each position FastQC draws a yellow box spanning the 25th to 75th percentile, with whiskers at the 10th and 90th percentiles. A red line marks the median, and a blue line traces the mean across positions. The background is shaded green above Q28, orange between Q20 and Q28, and red below Q20.
The module raises a warning when the lower quartile at any position falls below 10 or the median falls below 25. It fails when the lower quartile falls below 5 or the median falls below 20. These are fixed heuristics rather than anything tuned to your application. A whole-genome library at 30× can tolerate a fail on the last few cycles with no practical consequence. A low-input amplicon panel where you are trying to call variants at 1% allele fraction cannot.
The companion modules matter as much as this one. “Per sequence quality scores” plots the distribution of mean quality per read and warns when the most frequently observed mean is below 27, failing below 20. A bimodal shape there tells you a subset of reads is bad rather than a subset of cycles. “Per tile sequence quality” splits the same data by flowcell tile and flags tiles whose mean falls several Phred units below the average at that position. Tile-level and flowcell-level error structure is real and measurable, and tools built specifically to quantify it can identify error-prone regions that a global quality summary hides 3.
Binned quality scores change the shape of the plot
On NovaSeq and other RTA3 instruments, Illumina no longer reports 40 distinct quality values. The real-time analysis software bins every call into four values: 2, 12, 23 and 37. Your per-base boxplots therefore collapse onto a small number of horizontal levels, quartiles sit exactly on bin boundaries, and the median is Q37 across almost the entire read until quality drops off a cliff.
The practical consequence is that the plot loses resolution as a diagnostic. A gentle decline that you would have seen on a HiSeq 2500 becomes invisible until it crosses a bin edge. We treat binned data as a pass/fail signal at the tile and read level and lean on alignment metrics for anything finer. Platform differences of this kind are well documented in cross-platform reference datasets, where quality value distributions and error profiles differ noticeably between instruments sequencing the same sample 4.
What we would run
FastQC is fine, but it is single-threaded per file and slow on a 100 GB WGS FASTQ. We run falco, a drop-in reimplementation that produces the same modules and HTML from the same input, and reserve FastQC for cases where a collaborator expects its exact output.
falco -t 8 -o qc/ sample_R1.fastq.gz sample_R2.fastq.gz
multiqc qc/ -o qc/report
For a per-cycle view you can compute yourself, seqtk fqchk gives a compact table of mean quality and base composition by position:
seqtk fqchk -q20 sample_R1.fastq.gz | head -20
On trimming, our view is to do very little for short-read DNA. Adapter removal is worth it, quality trimming usually is not. Modern aligners soft-clip low-quality ends, and aggressive 3’ trimming shortens reads and costs you mappability without improving calls. fastp with --detect_adapter_for_pe and no quality cutting is a reasonable default. If you do trim, tell the recalibration step about it, because post-alignment processing choices measurably shift variant discovery results 5. Published whole-genome QC pipelines for Illumina data lay out this sequence of steps in detail 6.
For long reads the per-cycle framing does not apply. Read length varies by orders of magnitude and error is distributed along the read rather than accumulating with cycle number. Use a tool built for it, such as LongReadSum, which summarizes per-read quality and length jointly and can also read basecaller signal files 7.
Failure modes worth recognizing
A steady decline across all cycles with a widening lower whisker is normal phasing and signal decay. It is expected and usually benign above Q25 at the median.
A sharp collapse at a specific cycle range across all tiles points to a run-level event, such as a bubble or a reagent problem. A collapse confined to a few tiles shows up in the per-tile module and can often be resolved by filtering reads from those tiles rather than discarding the run.
High median quality with an unexpected variant call profile is the harder case. Reported Phred values are estimates from the basecaller and they are not always calibrated, which is the reason base quality score recalibration exists in most variant pipelines 5. If you want to know whether your pipeline behaves as expected under a known truth, simulate FASTQ files with controlled error and quality profiles and run them end to end 8. Coverage and callability also vary by genomic context, and regulatory regions are systematically harder to cover and call than coding regions 9.
None of these metrics say anything about your health. They tell you whether the data are fit to analyze. Any interpretation of a variant or biomarker for a medical decision belongs with a qualified clinician.
Questions people also ask
What does the Phred quality score represent in a FASTQ file? It is the sequencer’s estimated probability that the base call at that position is incorrect, on a −10 log₁₀ scale. Each base gets one score, stored as one ASCII character offset by 33.
What is a good Phred quality score? For short-read Illumina data, a median at or above Q30 across the read is good and Q37 is what binned NovaSeq data reports for most positions. FastQC warns below a median of 25 and fails below 20. Long-read platforms use different scales and a mean read quality near Q20 is strong for nanopore data.
How do you calculate a Phred quality score? From an error probability, Q = −10 × log₁₀(P). Going the other way, P = 10^(−Q/10). To recover Q from a FASTQ quality character, subtract 33 from its ASCII code.
Why do my quality scores only take four values? Your data came off an instrument using RTA3 binning. That binning collapses quality into the values 2, 12, 23 and 37 to reduce file size. This is expected and has little measured impact on variant calling 2.
Footnotes
-
Jan Voges. Compression of DNA Sequencing Data. 2022. https://doi.org/10.51202/9783186878106 ↩
-
Jan Voges, Ali Fotouhi, Jörn Ostermann, et al. A Two-Level Scheme for Quality Score Compression. Journal of Computational Biology, 2018. https://doi.org/10.1089/cmb.2018.0065 ↩ ↩2
-
Eric M. Davis, Yu Sun, Yanling Liu, et al. SequencErr: measuring and suppressing sequencer errors in next-generation sequencing data. Genome Biology, 2021. https://doi.org/10.1186/s13059-020-02254-2 ↩
-
Jie Huang, Xinming Liang, Yuankai Xuan, et al. A reference human genome dataset of the BGISEQ-500 sequencer. GigaScience, 2017. https://doi.org/10.1093/gigascience/gix024 ↩
-
Shulan Tian, Huihuang Yan, Michael Kalmbach, et al. Impact of post-alignment processing in variant discovery from whole exome data. BMC Bioinformatics, 2016. https://doi.org/10.1186/s12859-016-1279-z ↩ ↩2
-
Marvin N. Wright, Damian Gola, Andreas Ziegler. Preprocessing and Quality Control for Whole-Genome Sequences from the Illumina HiSeq X Platform. Methods in Molecular Biology, 2017. https://doi.org/10.1007/978-1-4939-7274-6_30 ↩
-
Jonathan Elliot Perdomo, Mian Umair Ahsan, Qian Liu, et al. LongReadSum: A fast and flexible quality control and signal summarization tool for long-read sequencing data. Computational and Structural Biotechnology Journal, 2025. https://doi.org/10.1016/j.csbj.2025.01.019 ↩
-
Matthew Frampton, Richard Houlston. Generation of Artificial FASTQ Files to Evaluate the Performance of Next-Generation Sequencing Pipelines. PLoS ONE, 2012. https://doi.org/10.1371/journal.pone.0049110 ↩
-
Weixin Wang, Zhi Wei, Tak-Wah Lam, et al. Next generation sequencing has lower sequence coverage and poorer SNP-detection capability in the regulatory regions. Scientific Reports, 2011. https://doi.org/10.1038/srep00055 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.