Read quality
What a Phred score means, the five ways the tools here trim a read, and why they do not all cut in the same place.
This page exists because every tool that reads sequencing data reports quality on the same scale but trims by different rules. Each tool page names its own rule; this page puts them side by side.
The scale
A base's quality is a Phred score, Q = −10 log₁₀ P, where P is the probability that the base was called wrong.
| Q | Chance the base is wrong | Read it as |
|---|---|---|
| 10 | 1 in 10 | One call in ten wrong: the floor the Chromatogram offers for dimming |
| 20 | 1 in 100 | The threshold papers quote for "readable" |
| 30 | 1 in 1,000 | The threshold quoted for a confident call |
| 40 | 1 in 10,000 | The top of a good Sanger or Illumina read |
That is why every quality plot here draws its rules at Q20 and Q30 rather than at round fractions of the axis, and why the ≥ Q20 and ≥ Q30 fractions are the summary a read is judged by. Mean quality over a whole read is a weak summary: a read can have a good mean and a useless first and last hundred bases, which is why every tool also shows where quality falls.
Where the number comes from matters. In a trace file, the Phred values are the instrument's own basecaller's opinion under its own calibration. The Chromatogram can also recompute quality from the peak shapes, against a calibration table for a named instrument and polymer; expect the two scales to disagree on most bases, and quote the instrument's values when you report what your facility produced.
In FASTQ the score is one ASCII character per base, and the offset varies. Phred+33 is written by every tool since Illumina 1.8; Phred+64 by Illumina 1.3 to 1.7; Solexa+64 by earlier machines, on a different scale. Read a Phred+64 file as Phred+33 and every score is inflated by 31, so the Read quality viewer infers the encoding from the characters present and says which it chose.
Five trimming rules, and what each answers
Sanger and short reads are both unreliable at their ends, so every tool that grades a read trims it first. They do not use one rule, because they are not answering one question:
| Rule | How it decides | Used by | Reach for it when |
|---|---|---|---|
| Mott | Converts each base's quality to an error probability, subtracts a cutoff, and keeps the contiguous stretch whose cumulative score is highest. The sequencer computes it and stores it in ABI files | Chromatogram (Auto uses it when present), Sequencing Verification | You want the answer your facility's software gave. SCF and ZTR files carry no Mott record |
| Sliding mean-quality window | Moves a window of fixed length in from each end and starts where the window's mean first reaches a floor | Chromatogram, Sequencing Verification, the workflow Load trace step | The file has no Mott record, or you want one rule applied to a mixed set of files |
| Consecutive run | Scans in from each end until three consecutive bases clear the threshold | Compare Traces to Reference | You want a noisy tail kept from voting in a consensus |
| Running sum from the 3′ end | Subtracts the cutoff from each quality, accumulates from the end inward, and cuts where the total peaks; the rule BWA uses | Read Mapping read QC | Short reads before mapping, where a lone dip should not truncate a good read |
| Fixed window | Skips a stated number of bases at the start: TIDE's published default is the first 100 | CRISPR Edit Inference | Decomposing an edited trace. Fit to this read replaces it with the read's quality-trimmed window |
Mott and the running sum find the best interval; a sliding window does not. A window stops where its mean crosses the floor, so roughly half a window of the poor stretch survives at each end. A per-base threshold, which none of these tools uses, would stop at the first bad base and truncate a good read on one dip. Shorten the window if you want a tighter cut.
The workflow step trims the way the Chromatogram trims, by the same window rule, so the read a pipeline verifies is the read the panel showed you.
Masking is not trimming
Compare Traces to Reference can also mask single bases below a minimum Phred. A masked base becomes an N, which never votes and never counts as coverage. Raising the threshold makes disagreements disappear, and it does so by discarding evidence: watch coverage fall at the same time.
A heterozygous call is two calls
When a trace is read as a mix of two alleles, the call is right only if both alleles are, so the two error probabilities combine: a pair of Q30 alleles is about a Q27 call. The Chromatogram shows each allele's own Phred value beside the combined one.
Related
Chromatogram · Read Quality Profile · Sequencing Verification · Compare Traces to Reference · Read Mapping · Read quality viewer · CRISPR Edit Inference
Melting temperature
Which melting-temperature model each tool computes with, why two panels can print different numbers for the same primer, and which number to take to the bench.
Restriction enzymes
The dataset four tools share, where it comes from, and the two properties that decide whether a digest, a ligation or an assembly is right.