The Allele That Never Aligned

Written By Kanna Nandakumar
The Allele That Never Aligned — Zetobit Bioinformatics Insight Series No. 17

Zetobit · Bioinformatics Insight Series · No. 17

The Allele That Never Aligned

Every RNA-seq experiment measures expression against a reference that is, by construction, somebody else's genome. The reads that disagree with it are the ones most likely to disappear — and they disappear in a direction.

By Kanna Nandakumar, PhD · Zetobit, LLC

A read either aligns or it does not, and the decision looks like a technicality. It is made by a scoring function, in microseconds, millions of times per sample, and it produces a number that nobody inspects. But the scoring function has an opinion. It rewards agreement with the reference sequence, and it penalizes departure from it. A read carrying a non-reference allele arrives at the aligner already carrying a mismatch, already a little worse than its counterpart, already slightly less likely to survive. Multiply that small asymmetry by every heterozygous site in a transcriptome and you no longer have a technicality. You have a systematic distortion with a known sign.

This is reference bias, and it is the quietest failure mode in RNA-seq. It does not crash the pipeline. It does not raise a QC flag. It produces clean BAM files, plausible expression matrices, and allelic ratios that look almost right — tilted, consistently, toward the allele that happens to be in the reference.

The mechanism is embarrassingly simple

The linear reference genome is flat. Every position holds exactly one allele — generally the more common one — and no position holds two. But your sample is diploid. At every heterozygous site, half the transcripts carry a base that the reference does not contain.

The aligner does not know this. It sees a read with a mismatch and treats that mismatch the way it treats any other: as evidence, however slight, against this alignment. If the read carries one non-reference base, it usually survives — the penalty is small relative to the mismatch budget. If it carries two or three, clustered within the read's span, it may fall below threshold and go unmapped, or map somewhere else entirely. Its reference-allele counterpart, carrying no penalty at all, maps cleanly every time.

The result is not noise. Noise would be tolerable; noise averages out. This is bias, and bias accumulates. Clusters of differentiating sites block alternate-allele reads from aligning at all, and the effect scales with divergence between the two alleles — the more the haplotypes differ, the harder the aligner pushes them toward the reference.2

Indels are worse than SNVs. An ungapped alignment supporting the reference will out-score a gapped alignment supporting the insertion, essentially by design.

Indels deserve their own sentence. An aligner deciding between an ungapped alignment that matches the reference and a gapped alignment that matches the alternate allele is not weighing two equal hypotheses — gap penalties are, by construction, larger than mismatch penalties. The reference wins on points before the biology is consulted.

How large is it, actually

Large enough to have been discovered immediately and never fully solved. Degner and colleagues sequenced mRNA from two HapMap Yoruba individuals in 2009 and found a significant excess of reference-allele reads at heterozygous SNPs.1 They tried the obvious fix — masking the SNP positions in the reference so neither allele has an advantage — and it helped, but it did not close the gap: roughly 5–10% of SNPs retained an inherent mapping bias even after masking, and removing the biased sites eliminated 40% of the strongest apparent ASE signals.1

Read that last figure slowly. Four out of ten of the top allele-specific expression hits were artifacts of the aligner.

Modern tooling has not made the underlying problem disappear. Benchmarking STAR against GRCh38, the difference between reference- and alternate-allele read proportions, averaged genome-wide, ranged from 4% to 10% across samples, with a median around 6.4% — where the expected value, averaged over all variants, is zero.3 The same analysis found that the reads that bias-aware filtering discards carry a REF–ALT skew of 44–80%, which tells you these are not marginal cases.3

Reference bias increases with allelic divergence A schematic showing that as the number of differentiating sites within a read increases, the proportion of alternate-allele reads that successfully align falls sharply, while reference-allele reads align at a constant high rate. 0% 25% 50% 75% 100% 0 1 2 3 4+ Differentiating sites within a single read Reads aligning correctly the bias Reference-allele reads Alternate-allele reads alleles identical — no bias possible
Figure 1. The bias is a function of divergence. When a read spans no differentiating sites, both haplotypes align identically and there is no bias. As differentiating sites accumulate within a single read's span, alternate-allele reads accrue mismatch penalties and progressively fail to align, while reference-allele reads are unaffected. The widening gap is the reference bias — which is why it is worst exactly where diversity is highest, and near-absent in the conserved regions where nobody worries about it. Schematic; shape after Degner et al. (2009) and Stevenson et al. (2013).

Where it concentrates

Because bias tracks divergence, it is not spread evenly across the transcriptome. It pools in the regions you are most likely to care about.

The HLA locus is the extreme case and the clearest warning. Benchmarking 1000 Genomes Phase I calls against Sanger sequencing of 930 samples, 18.6% of SNP genotype calls in HLA genes were wrong, and roughly a quarter of HLA SNPs had allele frequency estimates off by more than ±0.1 — with the error running toward overestimating the reference allele, exactly as mapping bias predicts.4 That is a fifth of the genotypes in the most immunologically consequential region of the genome, wrong in a consistent direction.

The pattern generalizes. Reference bias is strongest at HLA and KIR, at immunoglobulin loci, across highly polymorphic gene families, in segmental duplications, and — the part that should concern anyone doing translational work — in samples from individuals whose ancestry is underrepresented in the reference. The reference is not neutral ground. It is a specific set of haplotypes, and the further your sample sits from those haplotypes, the more of its signal the aligner quietly declines to see.

Where this actually bites

Allele-specific expression. The canonical victim. Reference bias inflates the reference allele's apparent expression, manufacturing allelic imbalance where none exists and masking real imbalance running the other way.

eQTL mapping. Genotype-correlated mapping efficiency produces genotype-correlated expression estimates. That is precisely the signature an eQTL scan is designed to detect.

RNA-seq variant calling. Alternate-allele reads that never align cannot support a call. Heterozygotes get called homozygous-reference; VAFs skew low.

Differential expression. Usually the safest of the four — bias is roughly consistent across samples and partially cancels in the contrast. It stops cancelling when your groups differ in ancestry, or when the gene is polymorphic enough that mapping rate itself varies by genotype.

What actually works

The fixes form a ladder, and they trade cost against completeness.

  • SNP masking replaces known variant positions with N so neither allele scores better. Cheap, and it helps — but Degner's 5–10% residual is the ceiling. It cannot fix indels, and it cannot fix bias arising from homology elsewhere in the genome.1
  • Filtering biased sites — simulating reads from both alleles and discarding sites that fail to map symmetrically — is honest and effective. The cost is coverage: you are buying accuracy by deleting the hardest and often most interesting positions.
  • WASP-style re-mapping takes each read overlapping a heterozygous site, swaps in the other allele, re-aligns, and keeps the read only if it lands in the same place. It directly tests the thing you actually care about. STAR+WASP folds this into the aligner rather than bolting it on afterward.3
  • Personalized references — building a genome from the sample's own phased genotypes — eliminate bias at known sites almost completely. They require that genotype data, plus phasing, plus a coordinate-lift back to standard space. When you have matched DNA, this is the strongest option.
  • Graph and pangenome references encode alternate alleles as paths rather than penalties, so a non-reference read is no longer a worse read. This is now the direction of travel for RNA-seq specifically, and it is the structural fix rather than a patch.5 The caveat is worth stating plainly: pangenome pipelines that retain a linear reference backbone can reintroduce the very bias they were built to remove.6

Notice what is missing from that list: increasing depth. Sequencing deeper does not help. The alternate-allele reads are not being lost to sampling — they are being lost to a decision rule, and the decision rule does not change when you hand it more reads. You will simply measure the same biased ratio with tighter confidence intervals, which is worse than measuring it badly, because now you believe it.

Depth buys precision. It does not buy accuracy. A systematic bias measured at 100× is the same systematic bias, reported with more confidence.

The pragmatic position

Not every experiment needs a pangenome. The proportionate response depends on what you are measuring.

If you are running standard differential expression on a reasonably homogeneous cohort across ordinary genes, reference bias is real but largely cancels in the contrast, and the effort is better spent elsewhere. If you are quantifying allele-specific expression, mapping eQTLs, calling variants from RNA, or working anywhere near HLA — or if your cohort's ancestry is not well represented in GRCh38 — then bias correction is not an enhancement. It is a correctness requirement, and skipping it means your effect sizes have a sign you did not choose.

The uncomfortable part is that none of this shows up in QC. Alignment rate looks fine. Duplication rate looks fine. The distribution of the biased quantity looks like a distribution. The only way to know how much reference bias your pipeline carries is to measure it deliberately: simulate reads from both alleles, align them, and count. If the two numbers differ, you have quantified your bias. If you never look, you have not avoided it — you have only agreed not to know its size.

The reference genome was never a description of your sample. It was a scaffold, useful precisely because it is fixed. The error is not in using it. The error is in forgetting that every read it rejects was rejected for a reason, and that the reason has a direction.

References

  1. Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK. Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009;25(24):3207–3212. doi:10.1093/bioinformatics/btp579
  2. Stevenson KR, Coolon JD, Wittkopp PJ. Sources of bias in measures of allele-specific expression derived from RNA-seq data aligned to a single reference genome. BMC Genomics. 2013;14:536. doi:10.1186/1471-2164-14-536
  3. Kaminow B, Ballouz S, Gillis J, Dobin A. STAR+WASP reduces reference bias in the allele-specific mapping of RNA-seq reads. bioRxiv. 2024. doi:10.1101/2024.01.21.576391
  4. Brandt DYC, Aguiar VRC, Bitarello BD, Nunes K, Goudet J, Meyer D. Mapping bias overestimates reference allele frequencies at the HLA genes in the 1000 Genomes Project Phase I data. G3 (Bethesda). 2015;5(5):931–941. doi:10.1534/g3.114.015784
  5. Bu et al. PanGraphRNA: an efficient and flexible bioinformatics platform for graph pangenome-based RNA-seq data analysis. Journal of Integrative Plant Biology. 2026. doi:10.1111/jipb.70231
  6. Frontiers in Genetics. Beyond single references: pangenome graphs and the future of genomic medicine. 2025. doi:10.3389/fgene.2025.1679660
  7. Salavati M, Bush SJ, Palma-Vera S, McCulloch ME, Hume DA, Clark EL. Elimination of reference mapping bias reveals robust immune related allele-specific expression in crossbred sheep. Frontiers in Genetics. 2019;10:863. doi:10.3389/fgene.2019.00863

Zetobit Bioinformatics Insight Series No. 17. Zetobit, LLC is a boutique bioinformatics consultancy in Lexington, Kentucky, building CAP/CLIA-compliant NGS pipelines and multi-omics analysis for biotech, pharma, and academic partners. If reference bias is load-bearing in your analysis and you are not sure how much of it you are carrying, we can help you measure it — zetobit.com.

Kanna Nandakumar https://www.zetobit.com
Previous

The Variant That Only Showed Its Hand in RNA
Next

The Normal You Never Sequenced