Allele-Specific Expression from RNA-Seq: Why the Reference Genome Biases the Very Imbalance You're Trying to Measure

Zetobit · Bioinformatics Insight Series

Allele-Specific Expression from RNA-Seq: Why the Reference Genome Biases the Very Imbalance You're Trying to Measure

Allele-specific expression asks a precise question: at a site where a person is heterozygous, are both copies of the gene expressed equally? The trouble is that the standard analysis has a built-in thumb on the scale — it favors the reference allele before any biology is measured.

By Kanna Nandakumar, PhD Zetobit LLC

Most expression analysis compares a gene's total output between samples or conditions. Allele-specific expression (ASE) asks something more surgical and internally controlled: within a single sample, at a heterozygous site, is the maternal copy expressed at the same level as the paternal one? Because both alleles sit in the same cell, exposed to the same environment and the same trans-acting factors, any consistent imbalance points to a cis-acting cause — a regulatory variant, imprinting, nonsense-mediated decay, X-inactivation.¹ That internal control is exactly what makes ASE powerful for interpreting non-coding and rare regulatory variants, where standard differential expression is blind.²

The measurement itself sounds trivial: align RNA-seq reads, find the ones covering a heterozygous SNP, count how many carry each allele, and test whether the split departs from 50/50.³ But that final ratio is only as trustworthy as the read counting, and the read counting is systematically distorted by the one thing the whole method depends on — the reference genome.

Reference bias: the thumb on the scale

Reads are aligned to a reference genome in which every heterozygous position contains only the reference allele. A read carrying the reference base matches perfectly; a read carrying the alternate base has at least one mismatch, which lowers its alignment score, can push it below a threshold, or can cause it to map elsewhere or not at all.⁴ The result is a preferential alignment of reference-allele reads — reference mapping bias — that inflates the reference count at exactly the sites ASE interrogates.⁵ The method's core measurement is biased in a fixed direction before any real biology enters.

This is the central confounder of ASE, and its magnitude is not trivial. In a simulation mapping RNA-seq reads over 9.5 million common variants, roughly 15.6% of variants showed a biased mapping rate between reference and non-reference reads.⁶ Left uncorrected, that bias masquerades as allelic imbalance — a signal that looks biological but is an artifact of the aligner preferring the sequence it was given.

Even when both alleles are transcribed equally, alternate-allele reads carry a mismatch against the reference and are disproportionately lost during alignment. The counted ratio drifts toward the reference allele, producing apparent allelic imbalance that reflects the aligner, not the cell.

Correcting it is necessary but not sufficient

The field has built several defenses against reference bias, none of them free. N-masking replaces the reference base at each heterozygous site with an ambiguous N so neither allele is favored. Personalized or diploid genomes insert the individual's known variants so both alleles have a correct template. Variant-aware and graph aligners encode alternate alleles in the alignment itself. And WASP-style filtering re-maps each read after swapping its alleles and discards any read whose mapping changes — removing precisely the reads that would have introduced bias.⁴ WASP has been shown to map multi-allele reads more consistently and yield lower false-discovery rates than N-masking or personalized-genome mapping alone.⁴

But even with bias correction, the statistics fight back. ASE read counts are overdispersed — noisier than a simple binomial 50/50 model assumes — so a naïve binomial test produces overly optimistic p-values and false positives; a beta-binomial model is the more honest choice.⁷ And residual issues remain: even after accounting for mapping bias, ASE p-values have been observed to stay inflated, meaning correction narrows the problem without fully eliminating it.⁷ Getting ASE right is a chain of defenses, and the weakest link sets the reliability of the result.

Bias correction has to be validated, not assumed. A recent quality-control method applied to over 15,000 GTEx transcriptomes flagged 563 samples with inflated ASE and splicing-outlier rates that had passed both DNA-seq and RNA-seq QC — transcriptional noise invisible to standard pipelines, removable only by looking at the ASE signal itself.⁸ Passing generic QC does not mean the allelic ratios are trustworthy.

Phasing: one SNP is a weak instrument

A further limitation is structural. Conventional ASE relies on single heterozygous SNPs, and single-SNP data is sparse and low-count: each variant is informative in at most half of individuals, and the reads covering one SNP capture only a fraction of a gene's output.⁹ Which SNP you pick, and where it sits relative to the regulatory variant driving the imbalance, changes the answer.⁹

The fix is phasing: assign the heterozygous SNPs across a gene to their parental haplotypes, then aggregate reads by haplotype rather than by individual SNP. Haplotype-level aggregation consistently improves ASE signal quality over single-SNP approaches, pooling sparse counts into a stronger, less noisy estimate and expanding the set of testable genes.⁸ The corollary is that an ASE pipeline is only as good as its phasing — and unphased, single-SNP ASE is a weaker instrument than a haplotype-aware one, on the same data.

What a real ASE pipeline requires

Turning allelic read counts into a trustworthy imbalance call is a sequence of deliberate corrections:

• Confident heterozygous genotypes. ASE is only defined at true het sites; genotyping errors create phantom imbalance, so start from high-confidence variant calls, ideally from matched DNA.³
• Explicit reference-bias correction. Apply N-masking, a personalized/diploid genome, a variant-aware aligner, or WASP filtering — and treat this as mandatory, not optional.⁴
• Overdispersion-aware statistics. Test allelic ratios with a beta-binomial (or equivalent) model rather than a plain binomial, so p-values aren't inflated.⁷
• Haplotype-level aggregation. Phase SNPs and pool reads per haplotype to reduce sparsity and stabilize estimates.⁸
• ASE-aware quality control. Screen samples for globally inflated allelic imbalance that signals transcriptional noise, beyond standard DNA/RNA QC.⁸
• Interpret the cause. Real imbalance still has many origins — cis-regulatory variation, imprinting, NMD, X-inactivation, somatic copy-number change — so a positive ASE call is a starting hypothesis about mechanism, not a mechanism.¹

As with the other blind spots in this series, the trustworthy output states its own construction: which bias correction was applied, whether counts were phased, what statistical model was used, and whether the sample passed ASE-specific QC. A raw allelic ratio reported without that provenance looks like a measurement but is closer to a rumor.

The takeaway

Allele-specific expression is one of the most elegant ideas in transcriptomics — an internally controlled readout of cis-regulatory biology that standard differential expression cannot see. But its central measurement is compromised by the reference genome itself, which favors one allele before any biology is counted. Reference-bias correction, overdispersion-aware testing, phasing, and ASE-specific QC are not optional refinements; they are what separate a real regulatory signal from an artifact of the aligner. The imbalance is easy to measure and easy to measure wrong — and knowing the difference is the whole job.

References

1. Allele-specific expression reveals genes with recurrent cis-regulatory alterations in neuroblastoma (causes: cis-regulatory, NMD, methylation, CNV; ASE insensitive to trans/environmental factors). Genome Biol. 2022. s13059-022-02640-y
2. ASE for complex genetic phenotypes (cis eQTL/sQTL, NMD, X-inactivation, imprinting; per-sample detection of low-MAF variants). Sci Rep. 2023. nature.com
3. STAR+WASP — ASE workflow (map, overlap het variants, assign to allele, compute count imbalance). bioRxiv. 2024. biorxiv.org
4. van de Geijn B, et al. WASP: allele-specific software for robust molecular QTL discovery (mapping-bias removal; lower FDR than N-masking/personal genome). Nat Methods. 2015. stanford.edu
5. Elimination of reference mapping bias reveals robust immune-related ASE (reference bias definition; N-masking vs WASP). Front Genet / bioRxiv. 2019. PMC6761296
6. Allelic mapping bias in RNA-seq is not a major confounder in eQTL studies (9.5M variants simulated; 15.6% show biased mapping rate). Genome Biol. 2014. PMC4212091
7. Overdispersion in ASE — binomial gives optimistic p-values, beta-binomial preferred; p-values remain inflated even after bias correction (Castel et al.). WASP suppl. / RG. researchgate.net
8. ASE quality control fills a critical gap in rare-variant interpretation (563 GTEx samples flagged past standard QC; haplotype-level aggregation improves quality). bioRxiv. 2025. biorxiv.org
9. ANEVA-h — haplotype-aggregated ASE (single-SNP sparsity; SNP choice and LD affect signal; >2-fold more testable genes with phasing). medRxiv. 2025. medrxiv.org

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com