phASER

Summary

phASER phases genetic variants using the reads that overlap them, and its key insight is that RNA-seq reads span introns — so a single spliced read can phase two variants hundreds of kilobases apart in the genome, a distance no DNA-seq read can bridge.[1] It then uses that phase to quantify haplotypic expression: the number of unique reads assigned to each haplotype of a gene, rather than allelic counts at individual heterozygous sites. This is a different kind of tool from the splicing and sequence models elsewhere in this vault — it is a phenotype-construction and phasing method, not a predictor.

Method

Two-step read-backed haplotype assembly:[1]

1. Edge definition — for each pair of variants observed on the same sequencing fragment, choose the phase with most read support and apply a binomial test against a uniform sequencing-error model; edges failing at nominal P < 0.01 are dropped.
2. Block resolution — count edges supporting each of the 2ⁿ possible haplotype configurations and take the best-supported. To avoid exponential blow-up, large haplotypes are split into sub-blocks (default max 15 variants) at the points spanned by the fewest edges, then the sub-blocks are phased relative to each other.

Phasing runs chromosome-wide with no distance restriction between variants. When population-phased data are available, phase anchoring assigns a genome-wide phase to each block by weighting each variant's population phase by its allele frequency — common variants are more likely to be correctly population-phased, so they anchor the block.[1]

DNA-seq and RNA-seq libraries can be combined, giving both close-proximity phasing from WES/WGS and long-range phasing from spliced RNA reads.[1]

Performance

• Accuracy: against Illumina's NA12878 Platinum Genome (~200×, phased by transmission), phASER with default settings phased 98% of variants correctly. Unlike population-based phasing, read-backed phasing performed well at low allele frequency — the regime that matters for rare-variant work.[1]
• Benchmarking: comparable to HapCUT in accuracy, runtime and haplotype length; both dramatically more accurate than the GATK Read Backed Phasing tool.[1]
• Distance: with 250 bp paired-end RNA-seq, an average of 4,300 variants (5.8%) were phased at ≥5 kb. At the same distance WES phased 0 and WGS phased 7.[1]
• Coverage of rare variants: RNA-seq from accessible tissues genome-wide-phased up to 15.4% of rare coding variants (MAF < 1%), rising to 21.3% when tissues were combined, versus 19% for WES and 11.1% for 5× WGS. Adding combined RNA-seq gave a 1.5× increase over WES alone and 2.4× over WGS alone.[1] Jointly phasing 14 GTEx tissues phased nearly 50% of all heterozygous coding variants with at least one other variant.[1]

Two applications

Compound heterozygote resolution (medical genetics). Population phasing is systematically wrong more often for exactly the variants that matter: stop-gain variants were ~2.9× more likely to be mis-phased than others.[1] Adding LCL RNA-seq to WES increased identification of compound heterozygotes in the most severe class (LoF × damaging) by 1.3×.[1] In a congenital diaphragmatic hernia cohort (20 patients, paired WES and fibroblast RNA-seq), phASER prioritised a median of 25 alleles per individual in trans while down-prioritising 44 in cis; including RNA-seq boosted the trans count 2.6×.[1]

Haplotypic expression (functional genomics). Allelic expression is usually measured at single heterozygous sites, but the unit of biological interest is the gene. Naively summing allele counts across sites double-counts reads covered by the same fragment and propagates phasing errors. phASER instead counts unique reads per haplotype. Benchmarked on 345 GEUVADIS samples at genes with known eQTLs, this removed false-positive allelic imbalance calls at 56.2% of genes tested and uncovered false negatives at 7.3%.[1]

Relationship to sequence-based variant effect models

Synthesis: phASER does not predict anything — it constructs a measurement. Its relevance to AlphaGenome, Borzoi and SpliceAI is that haplotypic expression is an individual-level, allele-resolved expression phenotype, whereas the eQTL effect sizes those models are benchmarked against are population-level regression coefficients. The AlphaGenome authors name personal-genome prediction as an unbenchmarked weakness of the sequence-model class [2]; haplotype-resolved expression from phASER is precisely the kind of phenotype such a benchmark would need. No cited study has run that evaluation.

Two further connections, both inferential rather than reported:

• Compound heterozygosity is invisible to current variant scorers. AlphaGenome and Borzoi score variants one at a time against a reference background; whether two damaging alleles sit in cis or trans changes the functional consequence completely and is a phasing question, not a sequence-prediction question.[1]
• Allelic imbalance is an orthogonal validation channel for a predicted regulatory effect in the individual carrying the variant, complementing population QTL evidence — relevant to the direction-of-effect problem discussed in Direction-of-Effect Assignment at GWAS Loci.

See Also

• RNA-seq-Derived Molecular Phenotypes — where haplotypic expression sits among the phenotypes derivable from RNA-seq.
• LeafCutter — the splicing-side counterpart: a different phenotype from the same raw data type.
• AlphaGenome — sequence model whose personal-genome performance is unbenchmarked.

Citations

[1] Castel, S. E., Mohammadi, P., Chung, W. K., Shen, Y. & Lappalainen, T. (2016). Rare variant phasing and haplotypic expression from RNA sequencing with phASER. Nature Communications 7:12817. Supports: the two-step haplotype assembly and phase anchoring; the 98% NA12878 concordance and HapCUT/GATK comparison; all phasing-distance and rare-variant coverage figures; the stop-gain mis-phasing rate; compound heterozygote and CDH cohort results; the 56.2%/7.3% allelic-imbalance false-positive/false-negative figures. Location: Results ("Haplotype assembly and phasing accuracy"; "Contribution of RNA-seq to variant phasing"; "Application of phASER to genetic studies"; "Application of phASER to allelic expression studies"); Figs. 1–2; Methods. Source paper: ncomms12817.pdf

[2] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature 649:1206–1217. Supports: the statement that personal-genome prediction is a known, unbenchmarked weakness of this model class. Location: Discussion (limitations). Source paper: s41586-025-10014-0.pdf