A Phred quality score is a compact estimate of base-call uncertainty, not a verdict on an entire sequence. Its value is greatest when retained at base resolution and interpreted alongside the primary signal, assay design, and downstream question. A single mean score discards exactly the spatial information needed to judge whether a key variant is supported.
The logarithmic scale
For an estimated probability $P_e$ that a base call is wrong, the Phred score is
$$Q = -10\log_{10}(P_e)$$and therefore
$$P_e = 10^{-Q/10}.$$| Phred score | Estimated error probability | Nominal call accuracy |
|---|---|---|
| Q10 | 1 in 10 | 90% |
| Q20 | 1 in 100 | 99% |
| Q30 | 1 in 1,000 | 99.9% |
| Q40 | 1 in 10,000 | 99.99% |
A ten-point increase is a tenfold decrease in estimated error probability. A three-point increase is approximately a twofold decrease, not a three-percentage-point improvement.
The probability is model-derived. In the original Phred system, features of Sanger electropherogram peaks were calibrated against observed errors. Modern instruments and processing pipelines may emit Phred-scaled values from different models. The equation defines the scale, but it does not guarantee that Q30 calls from every chemistry, instrument, basecaller, and sequence context have an empirical error rate of exactly 0.1%. Calibration should be evaluated on representative data when thresholds carry consequential decisions.
Per-base quality and expected errors
For a read with qualities $Q_1,\ldots,Q_n$, a useful summary is the expected number of base-call errors:
$$E = \sum_{i=1}^{n}10^{-Q_i/10}.$$This is not a prediction of exactly $E$ errors in that read. It is the sum of modeled error probabilities and is useful across many comparable reads when scores are calibrated. Importantly, converting the arithmetic mean of Q scores back to one error probability does not equal the mean error probability, because the transformation is nonlinear.
Two 700-base reads can have the same mean Q while differing materially. One may decline smoothly at the end; another may contain a short, severe quality collapse around the only phylogenetically informative position. Mean, median, percentiles, minimum, and expected errors summarize different features, but none replaces the quality vector.
What quality scores do not measure
A high Q score says that the basecaller considers its chosen symbol reliable under its model. It does not establish that:
- the tube or sample identifier is correct;
- the sequenced amplicon is mitochondrial rather than a NUMT or off-target product;
- contamination is absent;
- primer sequences and low-quality ends were removed correctly;
- strand orientation or alignment is correct;
- a consensus accurately represents mixed molecules;
- a secondary Sanger peak is or is not heteroplasmy;
- the biological interpretation attached to the base is valid.
These are separate quality domains. Sequence QC should not collapse laboratory controls, identity checks, alignment QC, and interpretive review into one Phred threshold.
Reading a Sanger trace locally
An AB1 file can contain base calls, per-base qualities, and four fluorescence channels. Review the trace around consequential positions rather than relying on the called text alone.
A strong interval generally shows distinct, regularly spaced peaks with low baseline signal and limited overlap. Warning patterns include broad or compressed peaks, rapidly declining intensity, dye blobs, high baseline noise, loss of spacing, and overlapping signals extending across many positions. Mixed templates may produce persistent secondary peaks, but isolated shoulders and noisy read ends are not sufficient evidence of a biological mixture.
Bidirectional sequencing is especially useful. Concordant forward and reverse calls can resolve strand-specific artifacts, although agreement is not independent confirmation if both reactions share the same amplification product or contamination source. A critical discrepancy should trigger trace review and, according to the study’s validation plan, repeat amplification or sequencing.
Trimming is a decision, not a cosmetic step
Read ends frequently have lower quality, but “trim to Q20” is underspecified. A trimming algorithm might remove every terminal base below Q20, find the longest window above a threshold, maximize a quality function, or trim after a sustained decline. These methods return different boundaries.
Window-based trimming is less sensitive to one low base but can retain that base inside an otherwise strong window. Single-base trimming can fragment reads or overreact to a solitary dip. Internal low-quality regions require masking or review; end trimming does not fix them. Primer-derived sequence should be removed according to primer coordinates, not merely because it happens to have high or low quality.
Every trim should preserve the original read and record:
- the algorithm and version;
- score threshold and window length;
- whether boundaries are zero- or one-based and whether the end is inclusive;
- the pre-trim and post-trim lengths;
- any internal masking rule;
- whether a reviewer accepted or changed the suggestion.
Thresholds must follow the use case
A Q20 base may be adequate for exploratory consensus generation when supported by an overlapping read, yet inadequate as the sole observation underlying a reported difference. Conversely, a rigid Q30 rule can erase usable data in difficult regions without improving the validity of the remaining calls.
Thresholds should be established against known controls and representative failure modes. For Sanger mtDNA work, the policy should address read ends, strand agreement, ambiguous symbols, homopolymer-adjacent positions, variant confirmation, and minimum overlap. For quality-bearing FASTQ consensus records, confirm that the qualities refer to the consensus bases; they should not be treated as raw-read depth or variant allele evidence.
Common interpretation mistakes
Reporting only mean Q. The mean hides location and distribution. Include at least the analyzed range and flagged intervals, then inspect decisive positions individually.
Treating Q20 as a universal pass line. Q20 has a mathematical meaning, not universal fitness for purpose.
Averaging Q scores across replicates. Replicates provide separate evidence. Pooling their scores can conceal disagreement.
Calling a read “99.9% accurate” because its mean is Q30. Per-base probabilities do not translate directly into the probability that an entire read is error-free.
Using high Q to validate a mixed call. Standard base quality generally evaluates the selected base, not the quantitative reliability of a minority peak or allele fraction.
Review and reporting checklist
| Review item | Record | Why it matters |
|---|---|---|
| Source | Instrument, chemistry, basecaller, file type | Phred-scale calibration is pipeline-dependent |
| Distribution | Mean, median, percentiles, min/max | Reveals skew hidden by one average |
| Location | Per-base scores and weak intervals | Connects quality to interpreted positions |
| Signal | AB1 channels or underlying reads | Distinguishes model output from primary evidence |
| Trimming | Rule, threshold, window, coordinates | Makes the analyzed span reproducible |
| Concordance | Forward/reverse or replicate status | Exposes unresolved disagreement |
| Sequence policy | Ambiguity, masking, consensus rules | Defines how uncertainty enters the result |
| Decision | Reviewer, disposition, rationale | Separates automated suggestion from acceptance |
Where GeneFlow fits
GeneFlow extracts and stores AB1 base calls, Phred scores, and A/C/G/T trace channels. Its current QC implementation reports mean, median, 10th, 25th, 75th, and 90th percentiles plus minimum and maximum. It flags contiguous regions where a 20-base rolling mean falls below Q20 and suggests the span between the first and last 20-base windows whose means exceed Q25. Overall labels are based on mean Q: above 40 is high, above 30 medium, above 20 low, and 20 or below failed.
Those are transparent screening rules, not assay validation. A trim coordinate is a suggestion and does not alter the preserved primary evidence. Reviewers should inspect the trace around important calls and apply their study’s acceptance criteria. GeneFlow can also store quality-bearing consensus FASTQ records, but it does not accept raw high-throughput read sets as an mtDNA variant-calling workflow, and its Phred QC does not detect heteroplasmy.