Mapping Quality Estimation

Phred-scaled confidence in a read placement

Bioinformatics

Algorithms

Read Mapping

Statistical Inference

How aligners summarize uncertainty that a reported read alignment is at the correct reference location.

Author

Jiahong Sun

Published

August 19, 2026

Modified

August 19, 2026

1 Topic at a Glance

One-sentence summary: Mapping quality (MAPQ) is a Phred-scaled estimate of the probability that a reported alignment is wrong, combining evidence for the chosen locus with evidence for competing loci.

Item	Description
Biological question	How much should a downstream analysis trust this read’s genomic origin?
Computational problem	Convert competing alignment evidence into a calibrated placement-confidence score
Method type	Probabilistic or empirical confidence estimation
Input	Primary and alternative alignment scores, seeds/chains, read properties, and reference context
Output	The `MAPQ` field in a SAM/BAM alignment record
Prerequisites	read mapping and seed chaining, probability, Phred scores

2 What MAPQ means

For an alignment reported at locus $L$, the intended interpretation is

\[ \mathrm{MAPQ} = -10\log_{10} P(\text{reported placement is incorrect}). \]

Thus MAPQ 10 corresponds to an estimated error probability of 0.1, MAPQ 20 to 0.01, and MAPQ 30 to 0.001. It concerns the locus, not whether every aligned base is correct: base qualities and the QUAL values in a variant call answer different questions.

MAPQ is an estimate, not a universal measurement. The SAM specification permits any integer 0–254 and reserves 255 for unavailable information. Different aligners use different models, caps, and conventions, so equal MAPQ values from different tools or modes need not be identically calibrated.

3 From alignment competition to probability

If candidate loci $L_1,\ldots,L_k$ have likelihoods $P(R\mid L_i)$ for read $R$ and priors $P(L_i)$, a model-based posterior is

\[ P(L_1\mid R)=\frac{P(R\mid L_1)P(L_1)}{\sum_{i=1}^{k}P(R\mid L_i)P(L_i)}. \]

The ideal MAPQ for the selected $L_1$ would use $1-P(L_1\mid R)$. In real mappers, enumerating every locus and modeling sequencing error, polymorphism, reference incompleteness, and paired-read evidence exactly is expensive. They therefore approximate this quantity from features such as the best-versus-second-best alignment or chain score, aligned length, seed uniqueness, base qualities, and whether candidates were filtered.

3.1 Score separation is the central signal

Consider two otherwise comparable candidates. If the primary alignment score is 100 and the next is 99, either may be the true locus; MAPQ should be low. If the next best score is 40, the primary is much more distinctive. A short perfect read in a repeat can still have low MAPQ because many unreported candidates may tie it. Conversely, a long read may contain enough unique anchors to earn high MAPQ despite base errors.

The exact mapping from score differences to MAPQ is aligner-specific. For example, minimap2 estimates primary-chain MAPQ from the best chain score, the strongest competing secondary chain, and the number of anchors, then applies additional heuristics. Treat its score as its documented empirical confidence signal, not as a generic posterior recomputed from the displayed CIGAR alone.

4 Interpretation in SAM/BAM workflows

MAPQ 0       no distinction from an alternative placement (or an explicitly ambiguous placement)
MAPQ 20      approximately 1% estimated placement error under the aligner's model
MAPQ 30      approximately 0.1% estimated placement error under the aligner's model
MAPQ 255     unavailable, not exceptionally high confidence

An alignment flag identifies whether a record is primary, secondary, or supplementary; MAPQ does not replace those flags. A supplementary record may be meaningful evidence for a split read but is not an independent observation. Secondary-alignment reporting is configurable, so absence of a secondary record does not prove unique mapping.

4.1 Minimal inspection

samtools view -q 30 -F 0x904 alignments.bam

This selects records with MAPQ at least 30 while excluding unmapped (0x4), secondary (0x100), and supplementary (0x800) records. It is a common starting point, not a universally valid quality-control rule: RNA-seq, ancient DNA, metagenomics, duplicated genes, or structural-variant evidence may require a different treatment. Always retain an unfiltered file for auditability.

5 Pitfalls and calibration

Pitfall	Why it matters	Better practice
Equating MAPQ with base quality	One describes locus uncertainty; the other describes base-call uncertainty	Filter and model each at its appropriate stage
Treating 255 as high confidence	It means unavailable in SAM	Handle it explicitly according to tool documentation
Comparing MAPQ across aligners as a universal probability	Heuristics and caps differ	Validate within the mapper, version, reference, and read type used
Removing every low-MAPQ read	Discards real signal from duplicated or polymorphic loci	Match the filtering rule to the biological question

MAPQ can be overconfident when the true source is absent from the reference, when alternate haplotypes are missing, or when a mapper fails to generate a close competitor. It can be underconfident in biologically duplicated regions where multi-mapping is expected but aggregate evidence is still useful. Benchmark calibration with simulated reads or a truth set whenever the downstream decision is sensitive to placement uncertainty.

6 Connections

MAPQ follows the candidate generation and alignment described in read mapping and seed chaining. Its ambiguity arises especially from repeated seeds in k-mers and sequence composition and sparse-seed collisions with minimizers and sequence sketching. Variant callers later combine mapping quality with base quality, allele balance, and local realignment evidence; a high MAPQ alone does not establish a variant.

7 Take-home Messages

MAPQ quantifies estimated error in the read’s placement, on a Phred scale.
Near-tied candidate loci, especially repeats, should yield low MAPQ even for a clean alignment.
MAPQ is mapper- and context-dependent; 255 means unavailable.
Use MAPQ with alignment flags and other quality evidence, and validate a threshold for the analysis at hand.

8 References

Li, H. et al. 2008. Mapping short DNA sequencing reads and calling variants using mapping quality scores. Genome Research 18:1851–1858.
Li, H. 2018. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34:3094–3100.
Sequence Alignment/Map Format Specification, §1.4 (MAPQ field).
minimap2 documentation.

--- title: "Mapping Quality Estimation" subtitle: "Phred-scaled confidence in a read placement" description: "How aligners summarize uncertainty that a reported read alignment is at the correct reference location." author: "Jiahong Sun" date: "2026-08-19" date-modified: "2026-08-19" categories: [Bioinformatics, Algorithms, Read Mapping, Statistical Inference] toc: true number-sections: true code-fold: show draft: false --- # Topic at a Glance > **One-sentence summary:** Mapping quality (MAPQ) is a Phred-scaled estimate of the probability that a reported alignment is wrong, combining evidence for the chosen locus with evidence for competing loci. | Item | Description | |---|---| | **Biological question** | How much should a downstream analysis trust this read's genomic origin? | | **Computational problem** | Convert competing alignment evidence into a calibrated placement-confidence score | | **Method type** | Probabilistic or empirical confidence estimation | | **Input** | Primary and alternative alignment scores, seeds/chains, read properties, and reference context | | **Output** | The `MAPQ` field in a SAM/BAM alignment record | | **Prerequisites** | [read mapping and seed chaining](read-mapping-and-seed-chaining.qmd), probability, Phred scores | # What MAPQ means For an alignment reported at locus $L$, the intended interpretation is $$ \mathrm{MAPQ} = -10\log_{10} P(\text{reported placement is incorrect}). $$ Thus MAPQ 10 corresponds to an estimated error probability of 0.1, MAPQ 20 to 0.01, and MAPQ 30 to 0.001. It concerns the **locus**, not whether every aligned base is correct: base qualities and the `QUAL` values in a variant call answer different questions. MAPQ is an estimate, not a universal measurement. The SAM specification permits any integer 0--254 and reserves 255 for unavailable information. Different aligners use different models, caps, and conventions, so equal MAPQ values from different tools or modes need not be identically calibrated. # From alignment competition to probability If candidate loci $L_1,\ldots,L_k$ have likelihoods $P(R\mid L_i)$ for read $R$ and priors $P(L_i)$, a model-based posterior is $$ P(L_1\mid R)=\frac{P(R\mid L_1)P(L_1)}{\sum_{i=1}^{k}P(R\mid L_i)P(L_i)}. $$ The ideal MAPQ for the selected $L_1$ would use $1-P(L_1\mid R)$. In real mappers, enumerating every locus and modeling sequencing error, polymorphism, reference incompleteness, and paired-read evidence exactly is expensive. They therefore approximate this quantity from features such as the best-versus-second-best alignment or chain score, aligned length, seed uniqueness, base qualities, and whether candidates were filtered. ## Score separation is the central signal Consider two otherwise comparable candidates. If the primary alignment score is 100 and the next is 99, either may be the true locus; MAPQ should be low. If the next best score is 40, the primary is much more distinctive. A short perfect read in a repeat can still have low MAPQ because many unreported candidates may tie it. Conversely, a long read may contain enough unique anchors to earn high MAPQ despite base errors. The exact mapping from score differences to MAPQ is aligner-specific. For example, minimap2 estimates primary-chain MAPQ from the best chain score, the strongest competing secondary chain, and the number of anchors, then applies additional heuristics. Treat its score as its documented empirical confidence signal, not as a generic posterior recomputed from the displayed CIGAR alone. # Interpretation in SAM/BAM workflows ```text MAPQ 0 no distinction from an alternative placement (or an explicitly ambiguous placement) MAPQ 20 approximately 1% estimated placement error under the aligner's model MAPQ 30 approximately 0.1% estimated placement error under the aligner's model MAPQ 255 unavailable, not exceptionally high confidence ``` An alignment flag identifies whether a record is primary, secondary, or supplementary; MAPQ does not replace those flags. A supplementary record may be meaningful evidence for a split read but is not an independent observation. Secondary-alignment reporting is configurable, so absence of a secondary record does not prove unique mapping. ## Minimal inspection ```bash samtools view -q 30 -F 0x904 alignments.bam ``` This selects records with MAPQ at least 30 while excluding unmapped (`0x4`), secondary (`0x100`), and supplementary (`0x800`) records. It is a common starting point, not a universally valid quality-control rule: RNA-seq, ancient DNA, metagenomics, duplicated genes, or structural-variant evidence may require a different treatment. Always retain an unfiltered file for auditability. # Pitfalls and calibration | Pitfall | Why it matters | Better practice | |---|---|---| | Equating MAPQ with base quality | One describes locus uncertainty; the other describes base-call uncertainty | Filter and model each at its appropriate stage | | Treating 255 as high confidence | It means unavailable in SAM | Handle it explicitly according to tool documentation | | Comparing MAPQ across aligners as a universal probability | Heuristics and caps differ | Validate within the mapper, version, reference, and read type used | | Removing every low-MAPQ read | Discards real signal from duplicated or polymorphic loci | Match the filtering rule to the biological question | MAPQ can be overconfident when the true source is absent from the reference, when alternate haplotypes are missing, or when a mapper fails to generate a close competitor. It can be underconfident in biologically duplicated regions where multi-mapping is expected but aggregate evidence is still useful. Benchmark calibration with simulated reads or a truth set whenever the downstream decision is sensitive to placement uncertainty. # Connections MAPQ follows the candidate generation and alignment described in [read mapping and seed chaining](read-mapping-and-seed-chaining.qmd). Its ambiguity arises especially from repeated seeds in [k-mers and sequence composition](k-mers-and-sequence-composition.qmd) and sparse-seed collisions with [minimizers and sequence sketching](minimizers-and-sequence-sketching.qmd). Variant callers later combine mapping quality with base quality, allele balance, and local realignment evidence; a high MAPQ alone does not establish a variant. # Take-home Messages 1. MAPQ quantifies estimated error in the read's *placement*, on a Phred scale. 2. Near-tied candidate loci, especially repeats, should yield low MAPQ even for a clean alignment. 3. MAPQ is mapper- and context-dependent; 255 means unavailable. 4. Use MAPQ with alignment flags and other quality evidence, and validate a threshold for the analysis at hand. # References - Li, H. et al. 2008. [Mapping short DNA sequencing reads and calling variants using mapping quality scores](https://doi.org/10.1101/gr.078212.108). *Genome Research* 18:1851--1858. - Li, H. 2018. [Minimap2: pairwise alignment for nucleotide sequences](https://doi.org/10.1093/bioinformatics/bty191). *Bioinformatics* 34:3094--3100. - [Sequence Alignment/Map Format Specification](https://samtools.github.io/hts-specs/SAMv1.pdf), §1.4 (MAPQ field). - [minimap2 documentation](https://lh3.github.io/minimap2/minimap2.html).