Why Read Placement Can Be Uncertain
A sequenced read is useful only if it can be related to a genomic location. An aligner compares the read with a reference genome and selects one or more placements according to sequence similarity and other rules. When a read comes from a repetitive region, its sequence may match several locations equally well. Even in a region with a unique best placement, a read carrying an alternate allele can match the reference less closely than a read carrying the reference allele. Both situations can affect which reads are assigned to a locus and therefore what the resulting data appear to show.
These are mapping problems: they concern where a read is placed, or whether it is placed at all. They are different from a sequencing error, which concerns whether a read's recorded bases differ from the molecule that was sequenced. A read can have accurately recorded bases yet be mapped ambiguously or in a way that favors the reference allele.
Repeated Sequences and Multiple Placements
A genome contains repeated elements and duplicated regions. A short read wholly inside a sequence that occurs at two locations may match both locations perfectly. If the read has no distinguishing bases, its sequence provides no evidence for choosing one location over the other. An aligner may report both placements, choose one according to an implementation-specific rule, assign a low mapping-quality score, or exclude the read under a filtering rule. The precise behavior depends on the aligner and analysis settings.
Mapping quality is commonly expressed on a Phred scale for the estimated probability that a reported placement is wrong. Under that interpretation, a mapping quality of 20 corresponds to an estimated error probability of \(10^{-2}\), or 1%. It describes uncertainty about location, not the probability that an individual base was sequenced incorrectly. Its calibration and meaning in practice depend on the aligner, so it should not be treated as a universal guarantee.
A repeated sequence can cause difficulty even when some reads from the region map uniquely. Reads that extend into unique flanking sequence may identify their source, while shorter reads contained entirely within the repeat may not. Consequently, the set of reads retained at a locus can be a selected subset rather than a random sample of all molecules originating there. Apparent depth may be low, or reads from a related genomic location may be assigned to the locus.
Worked Example: A Read That Fits Two Repeated Locations
Suppose a hypothetical reference has two locations, A and B, containing the same 30-base sequence. In a small set of 40 reads originating from A, 24 include a unique flanking base that identifies A. The other 16 consist only of the shared 30-base sequence. Assume the sequence calls are accurate and that no additional information, such as a paired read or a longer read, distinguishes the two locations.
The 24 reads with the unique flanking base have a clear placement at A. For each of the remaining 16 reads, the sequence matches A and B equally well. Therefore, those 16 reads do not provide sequence-based evidence for choosing A rather than B.
If an analysis retains only reads that map uniquely, it keeps 24 of the 40 reads from A. The retained fraction is \(24/40=0.60\), or 60%, and the apparent depth at A is reduced by 16 reads in this example. This reduction does not show that only 24 molecules came from A; it shows that 16 reads could not be assigned uniquely under the stated information and rule.
If an aligner instead assigns every ambiguous read to one location, the resulting count depends on its assignment rule, not on a distinction contained in these reads. A random allocation would have an expected \(16/2=8\) reads assigned to each location, but that expectation is not evidence that exactly eight reads came from each. The underlying placement information remains unresolved.
How Alignment Can Favor the Reference Allele
At a variant site, reads carrying the reference allele often match the reference sequence at that position, while reads carrying an alternate allele differ from it. An aligner may penalize mismatches or gaps when evaluating candidate placements. If the alternate-bearing read receives a worse score, it may be mapped with less confidence, placed elsewhere, or omitted by a mapping-quality filter. This is especially relevant when the read also overlaps a repetitive or otherwise similar sequence.
This mechanism does not mean that all aligners always discard alternate reads. The effect depends on read length, the sequence around the variant, the competing placements, alignment scoring, and filtering choices. The key condition is that reference and alternate reads have different probabilities of being retained at the locus.
Worked Examples of Allele-Count Distortion
Worked Example: Unequal Retention in a Heterozygous Sample
Consider a hypothetical heterozygous sample with 100 reference-bearing and 100 alternate-bearing molecules sampled before alignment. Assume reads from the reference allele are retained with probability \(r_r=0.95\), while reads from the alternate allele are retained with probability \(r_a=0.70\). These values are illustrative, and the calculation concerns expected retained counts.
The expected retained counts are \(100(0.95)=95\) reference reads and \(100(0.70)=70\) alternate reads. The expected alternate fraction among retained reads is $$ f_a^{\mathrm{obs}}=\frac{70}{70+95}=\frac{70}{165}=0.4242. $$
Before alignment the alternate fraction was \(100/(100+100)=0.5000\). Under these assumptions, unequal retention shifts the expected fraction to about 42.4%, below one half. This is a mapping-related distortion of allele evidence; it does not change the sample's assumed starting proportions.
Worked Example: Checking the Direction of the Bias
Suppose a different hypothetical sample has starting proportions \(p_r=0.60\) and \(p_a=0.40\). Reads are retained with probabilities \(r_r=0.90\) and \(r_a=0.60\). The expected retained alternate fraction is $$ f_a^{\mathrm{obs}}=\frac{(0.40)(0.60)}{(0.40)(0.60)+(0.60)(0.90)} =\frac{0.24}{0.24+0.54} =\frac{0.24}{0.78}=0.3077. $$
The alternate fraction falls from 40.0% before alignment to about 30.8% among retained reads. The direction follows from the smaller retention probability for alternate reads. In general, if both allele proportions are positive and \(r_a<r_r\), then $$ \frac{p_a r_a}{p_a r_a+p_r r_r}<p_a. $$ To verify this, multiply both sides by the positive denominator and cancel \(p_a^2r_a\); the inequality reduces to \(p_a p_r r_a < p_a p_r r_r\), which holds because \(p_a p_r>0\) and \(r_a<r_r\).
Justifying the Allele-Fraction Formula
The formula for the observed fraction follows by counting the expected retained reads from each allele. For a large set of sampled molecules, the expected retained contribution from the reference allele is proportional to \(p_r r_r\), and the expected retained contribution from the alternate allele is proportional to \(p_a r_a\). The total retained contribution is their sum. Dividing the alternate contribution by that total gives the formula.
This is a ratio of expected contributions used to explain systematic distortion; for finite read counts, realized fractions also vary from sample to sample. It does not establish that retention probabilities are known or constant. In a real analysis, the formula clarifies what must be checked: whether the two allele types have comparable opportunities to align and pass the chosen filters.
Worked Example: A Sequence Difference Near a Competing Placement
Suppose a hypothetical 25-base read from a unique genomic location carries two alternate bases compared with the reference. At that location, a simple illustrative scoring rule assigns 0 penalty for a match and 2 penalty units for each mismatch. The read therefore receives a penalty of \(2(2)=4\) at its true location. Assume a similar repeated location has a candidate alignment with one mismatch, for a penalty of \(2(1)=2\).
Under this simplified score alone, the repeated location has the better alignment by 2 penalty units, even though the read originated from the unique location. An aligner might therefore prefer the competing placement, report uncertainty, or use further information to resolve the choice. This arithmetic illustrates a possible scoring conflict; actual aligners use more detailed algorithms and may also use paired-read information, sequence context, and other criteria.
The example also shows why an alternate allele can contribute to mapping ambiguity: the bases that distinguish it from the reference may make its alignment less favorable, while a reference-bearing read at the same locus would not incur those mismatch penalties. The example does not require any sequencing error; it assumes the read bases accurately represent the molecule.
Diagnostics, Alternatives, and Limitations
Mapping problems are more plausible when evidence has a recognizable pattern: many reads have low mapping quality; multiple placements are reported; coverage drops in duplicated or repetitive regions; allele fractions shift toward the reference; or the alternate-supporting reads are preferentially removed by a mapping filter. Comparing allele counts before and after filtering can reveal a change, but a change alone does not prove reference bias. The region's sequence and competing placements must also be examined.
Useful comparisons include inspecting read alignments around the locus, checking whether reads extend into unique flanks, and comparing results under clearly documented mapping and filtering settings. Paired-end reads can help when the mate reaches distinctive sequence, because the pair provides more location information than either short sequence alone. Longer reads may span the repeat and its unique flanks. Neither approach guarantees resolution if the relevant sequence remains identical or the available reads do not bridge the ambiguous region.
| Approach | Potential benefit | Important limitation |
|---|---|---|
| Keep only uniquely mapped reads | Reduces assignments based on unresolved alternatives | Can discard many genuine reads from repetitive regions and reduce usable depth |
| Retain or distribute ambiguous reads | Preserves information about reads with multiple plausible placements | Counts at an individual locus may depend on an allocation rule rather than a unique placement |
| Use paired-end or longer reads | May connect a repeated sequence to unique flanking sequence | Cannot resolve a region when the additional sequence is also non-distinguishing |
| Compare reference-based and allele-aware references | Can expose whether reference choice changes alignment of alternate-bearing reads | Results depend on how candidate sequences are represented and compared; this is a diagnostic, not automatic proof of the correct placement |
No single strategy removes every ambiguity. Excluding low-confidence placements may improve the specificity of locus-level evidence while reducing the number of usable reads. Retaining ambiguous reads may preserve overall information but does not justify assigning each one to a particular locus. Alternative reference representations can help assess sensitivity to the reference sequence, but still require appropriate sequences and transparent interpretation.
A common mistake is to interpret a lower alternate-read count after mapping as evidence that the alternate allele is absent. That conclusion confuses the reads retained at a locus with all molecules originating there. The correction is to ask whether alternate-bearing reads had the same chance of aligning and passing filters as reference-bearing reads, and whether the locus has competing placements. A second mistake is to call every low mapping-quality read a sequencing error. Mapping quality concerns the confidence in placement; base accuracy is a separate issue.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A read matches two genomic locations exactly and contains no distinguishing sequence. What can be concluded about its placement from that read alone?
- In the retention model, let \(p_r=p_a=0.50\), \(r_r=0.80\), and \(r_a=0.80\). Calculate the expected observed alternate fraction.
- A sample has equal starting proportions of reference and alternate molecules, but alternate reads are retained less often. Will the observed alternate fraction be below, equal to, or above one half? Explain.
- Why is a low mapping-quality score not itself evidence that the read contains a sequencing error?
- Name one advantage and one limitation of retaining only uniquely mapped reads in a repetitive region.
Solutions. (1) The read alone cannot distinguish the two locations; it supports both placements equally. (2) \(f_a^{\mathrm{obs}}=(0.50)(0.80)/[(0.50)(0.80)+(0.50)(0.80)]=0.40/0.80=0.5000\). Equal retention preserves the equal fraction in this model. (3) It will be below one half, because \(p_a=p_r\) but \(r_a<r_r\), so fewer alternate reads are retained. (4) Mapping quality describes uncertainty about where a read belongs, not whether its bases were recorded correctly. (5) Unique-only retention reduces uncertain locus assignments, but it can discard genuine reads from the repeat and lower apparent depth.