LeafCutter

Summary

LeafCutter quantifies alternative splicing by measuring what is removed rather than what is retained: it identifies intron excision events directly from split RNA-seq reads and expresses splicing as the proportion of reads supporting each intron within a cluster of overlapping introns. This sidesteps transcript assembly and isoform deconvolution entirely, which is why it scales, and why it detects splicing events absent from any annotation.[1] It is the method that produces the intron-excision phenotypes underlying most published sQTL catalogues — including the GTEx sQTLs that AlphaGenome and Borzoi are benchmarked against.

Method

1. Pool mapped reads across all samples in a study and find overlapping introns demarcated by split reads.
2. Cluster: build a graph connecting every pair of overlapping introns that share a donor or acceptor site; connected components become clusters representing alternative intron-excision events.
3. Filter and re-cluster iteratively, removing rarely used introns based on their read proportion within the cluster. Without this step, clusters grow arbitrarily large at high read depth because noisy splicing events accumulate multiple supporting reads.
4. Quantify as intron excision proportions within each cluster; test differential splicing with a Dirichlet-multinomial GLM over the cluster, which is more sensitive than a beta-binomial GLM testing each intron independently, and whose multinomial normalisation removes the need to estimate library-size parameters.[1]

The design choice that matters: because reads spanning exon–exon junctions unambiguously identify the excised intron, quantification of usage ratios is accurate without the statistically hard step of assigning ambiguous reads to isoforms.[1] The trade-off is that alternative transcription start sites and alternative polyadenylation are not measured, since neither is captured by intron excision.[1]

Key results

Unannotated splicing is common. Across 2,192 GTEx samples in 14 tissues, 10.8%–19.3% of alternatively excised introns were absent from GENCODE v19, Ensembl and UCSC — rising to 48.5% in testis, consistent with the "out-of-testis" hypothesis of permissive transcription. Overall 31.5% of detected alternatively excised introns were unannotated, and 86% of novel junctions were independently recoverable in 21,504 SRA samples via Intropolis.[1] 15–25% of novel splice sites show vertebrate conservation (mean phastCons ≥ 0.6), indicating a substantial functional subset.[1]

Scalability. On YRI-versus-CEU differential splicing comparisons, LeafCutter completed all comparisons within an hour using under 400 MB RAM, while Cufflinks2, rMATS and MAJIQ took up to 7.8, 55.7 and 66.2 hours respectively, and MAJIQ required over 50 GB for the larger comparisons.[1]

sQTL yield — the headline result. On 372 European GEUVADIS LCLs, LeafCutter found 5,774 sQTLs at 5% FDR versus 620 transcript-ratio QTLs in the original study (a ninefold difference). In a controlled comparison on 85 YRI samples with identical downstream processing, LeafCutter identified 1.36–1.46× more sQTLs than Cufflinks2 and 1.83–2.06× more than Altrans.[1]

Tissue sharing. Across heart, lung, thyroid and whole blood, 75–93% of sQTLs replicate across tissue pairs — substantially higher than the 9–48% previously reported for these data, which the authors attribute to increased detection power. The remaining 7–25% show genuine tissue specificity, though many are trivial cases where the intron is only variable in one tissue.[1]

Disease interpretation. Applying S-PrediXcan to LeafCutter intron quantifications across 40 complex traits increased the number of associated disease genes at 5% FDR by an average of 2.1-fold relative to gene expression alone. For rheumatoid arthritis, 18 putative disease genes were identified through splicing, 13 of which were not detectable via expression — including CD40.[1] Multiple sclerosis variants are more enriched among sQTLs than eQTLs.[1]

Why this matters for sequence models

Synthesis: LeafCutter defines the molecular phenotype that splicing variant-effect models are scored against. The sQTLs used to benchmark Borzoi and AlphaGenome are intron-excision-ratio QTLs, so "sQTL classification auPRC" is a statement about agreement with LeafCutter-style intron-cluster quantification, not with splicing in the abstract. Two consequences follow, neither stated in the cited papers: a model that predicts splice-site probability (such as SpliceAI) is being evaluated on a phenotype it does not directly parameterise; and the 31.5% of intron-excision events that are unannotated [1] are, by construction, invisible to any predictor whose output space is keyed to an annotated transcript model. See RNA-seq-Derived Molecular Phenotypes.

See Also

• LeafCutter2 — the successor, which additionally classifies each junction as productive or NMD-inducing.
• Pangolin — the splicing predictor from the same group; LeafCutter sQTL calls were used to test its causal-variant nominations.
• RNA-seq-Derived Molecular Phenotypes — how phenotype definition determines what a sequence model can be benchmarked on.
• Borzoi, AlphaGenome — sequence models benchmarked on LeafCutter-derived sQTLs.
• S-PrediXcan — the TWAS method applied to LeafCutter intron quantifications.
• phASER — a complementary RNA-seq phenotype: haplotypic expression rather than intron usage.

Citations

[1] Li, Y. I., Knowles, D. A., Humphrey, J., Barbeira, A. N., Dickinson, S. P., Im, H. K. & Pritchard, J. K. (2018). Annotation-free quantification of RNA splicing using LeafCutter. Nature Genetics 50:151–158. Supports: the clustering and Dirichlet-multinomial method; the ATSS/APA scope limitation; unannotated-intron and conservation figures; runtime and memory comparisons; all sQTL yield, tissue-sharing and S-PrediXcan figures. Location: Results ("Overview of LeafCutter"; "De novo identification of RNA splicing in mammalian organs"; "Fast and robust identification of differential splicing"; "Mapping splicing QTLs using LeafCutter"; "LeafCutter sQTLs link disease variants to mechanism"); Figs. 2–5; Table 1. Full text retrieved via PubMed Central (PMC5742080); source copy: s41588-017-0004-9.md