One Wrong Base in a Thousand - Choosing the Q30 cutoff
One Wrong Base in a Thousand
Every base a sequencer reads comes with a confidence score. Learn to read that one number — the Phred quality score — and most of what people mean by "good data" stops being mysterious.
A sequencing run does not hand you DNA. It hands you millions of guesses — one letter at a time, A, C, G, or T — and, quietly attached to each letter, a number that says how sure the machine is. That number is the most useful and most ignored value in all of next-generation sequencing. Once you can read it, "the data looks fine" becomes something you can actually check rather than hope.
"A base without a quality score is an opinion. A base with one is a measurement."
A confidence score attached to every letter
When an instrument reads a fragment of DNA, each position is a small act of pattern recognition — interpreting a signal and deciding which of the four bases produced it. Some calls are easy and some are borderline, and the instrument knows the difference. So alongside the sequence itself, it records a per-base quality score: a compact estimate of how likely that particular call is to be wrong.
This idea is older than modern sequencing. It comes from a program called phred, written in the 1990s to base-call the automated sequencers of the Human Genome Project, which introduced a standardized, log-scaled quality value that turned out to predict real error rates remarkably well (Ewing & Green, Genome Research, 1998). Every FASTQ file produced today still carries a direct descendant of that score, one character per base, riding along beside the sequence.
The Phred scale, in one simple rule
The whole system rests on a single formula, and it is worth seeing once:
Q = −10 × log₁₀(P)
Here P is the probability that the base is wrong, and Q is the quality score. The only thing you need to remember is what the logarithm does: every jump of 10 points makes the error ten times smaller. A score of 20 means one error in a hundred. A score of 30 means one in a thousand. A score of 40 means one in ten thousand. The scale is steep on purpose — a handful of points is a large change in trust.
| Quality score (Q) | Chance the base is wrong | Base-call accuracy | Plain reading |
|---|---|---|---|
| Q10 | 1 in 10 | 90% | Barely a guess — usually trimmed away |
| Q20 | 1 in 100 | 99% | Acceptable for many uses |
| Q30 | 1 in 1,000 | 99.9% | The industry benchmark for "good" |
| Q40 | 1 in 10,000 | 99.99% | Excellent; common on healthy short-read runs |
Why the industry settled on Q30
Of all these numbers, one became shorthand for quality: Q30. When a sequencing vendor advertises that a run delivered "85% of bases above Q30," they are saying that 85% of the letters have no worse than a one-in-a-thousand chance of being wrong (Illumina, Quality Scores for Next-Generation Sequencing). Q30 is not a law of nature — it is a sensible, widely adopted line in the sand. Below it, errors start to accumulate fast enough to matter; at or above it, most downstream analyses behave.
Crucially, quality is not spread evenly across a read. Early cycles tend to be clean and later cycles drift, so the same read can begin at Q38 and end near Q20. This is why the first thing an analyst opens is not the sequence but the quality profile, produced by a standard tool such as FastQC, which draws exactly this shape (Andrews, Babraham Bioinformatics).
What to do when the quality drops
Seeing the drop is not the same as fixing it, and the fix is where judgment enters. The common response is quality trimming: cut the low-scoring tail off each read before alignment, using a tool like Trimmomatic or fastp. Done well, it removes noise that would otherwise turn into false variant calls. Done reflexively, it does harm — trim too aggressively and you throw away real sequence, shorten your reads until they no longer map uniquely, and quietly bias the very regions you were trying to measure.
There is also a subtler point that keeps quality scores honest. A high score is a claim about a random sequencing error; it says nothing about a systematic one. A base can be called with serene Q40 confidence and still be wrong because the read sits over a PCR artifact, a mismapping, or a repetitive region the instrument cannot resolve. Q30 tells you the letter is probably right. It does not tell you the letter is in the right place. That distinction is exactly where an experienced analyst earns their keep.
The Strategic Takeaway
Quality scores are the closest thing sequencing has to a built-in honesty check, and they are cheap to read once the scale stops being intimidating. The single formula, the one-in-a-thousand meaning of Q30, and the drooping shape of a per-base plot are enough to let anyone on a team ask sharper questions: How much of this run cleared Q30? Where does quality fall off, and did trimming account for it? Are we trusting a confident base that happens to be in the wrong location?
None of that requires a specialist to understand — but knowing which of those questions actually threatens a result, and which are noise, still does. The score tells you how much to trust a letter. Deciding what that means for your experiment is the part that never automates.
"Q30 says the base is probably right. It never says the base is in the right place — and that gap is the whole job."
Key References
- Ewing B, Green P. Base-calling of automated sequencer traces using phred. II. Error probabilities. Genome Research. 1998;8(3):186–194. doi:10.1101/gr.8.3.186
- Ewing B, Hillier L, Wendl MC, Green P. Base-calling of automated sequencer traces using phred. I. Accuracy assessment. Genome Research. 1998;8(3):175–185. doi:10.1101/gr.8.3.175
- Illumina. Quality Scores for Next-Generation Sequencing (technical note / knowledge article). illumina.com — technote_Q-Scores.pdf
- Andrews S. FastQC: A Quality Control Tool for High Throughput Sequence Data. Babraham Bioinformatics. bioinformatics.babraham.ac.uk/projects/fastqc